> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
/*-
|
|
|
|
* Copyright (c) 1992, 1993, University of Vermont and State
|
|
|
|
* Agricultural College.
|
|
|
|
* Copyright (c) 1992, 1993, Garrett A. Wollman.
|
|
|
|
*
|
|
|
|
* Portions:
|
|
|
|
* Copyright (c) 1990, 1991, William F. Jolitz
|
|
|
|
* Copyright (c) 1990, The Regents of the University of California
|
|
|
|
*
|
1994-08-25 20:16:59 +00:00
|
|
|
* 3Com 3C507 support:
|
|
|
|
* Copyright (c) 1993, 1994, Charles M. Hannum
|
|
|
|
*
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
* All rights reserved.
|
|
|
|
*
|
|
|
|
* Redistribution and use in source and binary forms, with or without
|
|
|
|
* modification, are permitted provided that the following conditions
|
|
|
|
* are met:
|
|
|
|
* 1. Redistributions of source code must retain the above copyright
|
|
|
|
* notice, this list of conditions and the following disclaimer.
|
|
|
|
* 2. Redistributions in binary form must reproduce the above copyright
|
|
|
|
* notice, this list of conditions and the following disclaimer in the
|
|
|
|
* documentation and/or other materials provided with the distribution.
|
|
|
|
* 3. All advertising materials mentioning features or use of this software
|
|
|
|
* must display the following acknowledgement:
|
|
|
|
* This product includes software developed by the University of
|
|
|
|
* Vermont and State Agricultural College and Garrett A. Wollman,
|
|
|
|
* by William F. Jolitz, by the University of California,
|
1994-08-25 20:16:59 +00:00
|
|
|
* Berkeley, by Larwence Berkeley Laboratory, by Charles M. Hannum,
|
|
|
|
* and their contributors.
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
* 4. Neither the names of the Universities nor the names of the authors
|
|
|
|
* may be used to endorse or promote products derived from this software
|
|
|
|
* without specific prior written permission.
|
|
|
|
*
|
|
|
|
* THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
|
|
|
|
* ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
|
|
|
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
|
|
|
|
* ARE DISCLAIMED. IN NO EVENT SHALL THE UNIVERSITY OR AUTHORS BE LIABLE
|
|
|
|
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
|
|
|
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
|
|
|
|
* OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
|
|
|
|
* HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
|
|
|
|
* LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
|
|
|
|
* OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
|
|
|
|
* SUCH DAMAGE.
|
|
|
|
*
|
1996-02-06 18:51:28 +00:00
|
|
|
* $Id: if_ie.c,v 1.32 1996/01/26 09:27:26 phk Exp $
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
*/
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Intel 82586 Ethernet chip
|
|
|
|
* Register, bit, and structure definitions.
|
|
|
|
*
|
|
|
|
* Written by GAW with reference to the Clarkson Packet Driver code for this
|
|
|
|
* chip written by Russ Nelson and others.
|
|
|
|
*
|
|
|
|
* BPF support code stolen directly from hpdev/if_le.c, supplied with
|
|
|
|
* tcpdump.
|
|
|
|
*/
|
|
|
|
|
|
|
|
/*
|
|
|
|
* The i82586 is a very versatile chip, found in many implementations.
|
|
|
|
* Programming this chip is mostly the same, but certain details differ
|
1995-05-30 08:16:23 +00:00
|
|
|
* from card to card. This driver is written so that different cards
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
* can be automatically detected at run-time. Currently, only the
|
|
|
|
* AT&T EN100/StarLAN 10 series are supported.
|
|
|
|
*/
|
|
|
|
|
|
|
|
/*
|
|
|
|
Mode of operation:
|
|
|
|
|
|
|
|
We run the 82586 in a standard Ethernet mode. We keep NFRAMES received
|
|
|
|
frame descriptors around for the receiver to use, and NBUFFS associated
|
|
|
|
receive buffer descriptors, both in a circular list. Whenever a frame is
|
|
|
|
received, we rotate both lists as necessary. (The 586 treats both lists
|
|
|
|
as a simple queue.) We also keep a transmit command around so that packets
|
|
|
|
can be sent off quickly.
|
|
|
|
|
|
|
|
We configure the adapter in AL-LOC = 1 mode, which means that the
|
|
|
|
Ethernet/802.3 MAC header is placed at the beginning of the receive buffer
|
|
|
|
rather than being split off into various fields in the RFD. This also
|
|
|
|
means that we must include this header in the transmit buffer as well.
|
|
|
|
|
|
|
|
By convention, all transmit commands, and only transmit commands, shall
|
|
|
|
have the I (IE_CMD_INTR) bit set in the command. This way, when an
|
|
|
|
interrupt arrives at ieintr(), it is immediately possible to tell
|
|
|
|
what precisely caused it. ANY OTHER command-sending routines should
|
|
|
|
run at splimp(), and should post an acknowledgement to every interrupt
|
|
|
|
they generate.
|
|
|
|
|
|
|
|
The 82586 has a 24-bit address space internally, and the adaptor's
|
|
|
|
memory is located at the top of this region. However, the value we are
|
|
|
|
given in configuration is normally the *bottom* of the adaptor RAM. So,
|
|
|
|
we must go through a few gyrations to come up with a kernel virtual address
|
|
|
|
which represents the actual beginning of the 586 address space. First,
|
|
|
|
we autosize the RAM by running through several possible sizes and trying
|
|
|
|
to initialize the adapter under the assumption that the selected size
|
|
|
|
is correct. Then, knowing the correct RAM size, we set up our pointers
|
|
|
|
in ie_softc[unit]. `iomem' represents the computed base of the 586
|
|
|
|
address space. `iomembot' represents the actual configured base
|
|
|
|
of adapter RAM. Finally, `iosize' represents the calculated size
|
|
|
|
of 586 RAM. Then, when laying out commands, we use the interval
|
|
|
|
[iomembot, iomembot + iosize); to make 24-pointers, we subtract
|
|
|
|
iomem, and to make 16-pointers, we subtract iomem and and with 0xffff.
|
|
|
|
|
|
|
|
*/
|
|
|
|
|
|
|
|
#include "ie.h"
|
|
|
|
#if NIE > 0
|
|
|
|
|
1994-08-13 03:50:34 +00:00
|
|
|
#include <sys/param.h>
|
|
|
|
#include <sys/systm.h>
|
1995-08-16 16:14:28 +00:00
|
|
|
#include <sys/kernel.h>
|
1995-11-18 08:35:49 +00:00
|
|
|
#include <sys/conf.h>
|
1994-08-13 03:50:34 +00:00
|
|
|
#include <sys/mbuf.h>
|
|
|
|
#include <sys/protosw.h>
|
|
|
|
#include <sys/socket.h>
|
|
|
|
#include <sys/ioctl.h>
|
|
|
|
#include <sys/errno.h>
|
|
|
|
#include <sys/syslog.h>
|
1994-10-19 01:59:11 +00:00
|
|
|
#include <sys/devconf.h>
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
|
1994-08-13 03:50:34 +00:00
|
|
|
#include <net/if.h>
|
|
|
|
#include <net/if_types.h>
|
|
|
|
#include <net/if_dl.h>
|
|
|
|
#include <net/route.h>
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
|
|
|
|
#include "bpfilter.h"
|
|
|
|
|
|
|
|
#ifdef INET
|
1994-08-13 03:50:34 +00:00
|
|
|
#include <netinet/in.h>
|
|
|
|
#include <netinet/in_systm.h>
|
|
|
|
#include <netinet/in_var.h>
|
|
|
|
#include <netinet/ip.h>
|
|
|
|
#include <netinet/if_ether.h>
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
#endif
|
|
|
|
|
1995-10-26 20:31:59 +00:00
|
|
|
#ifdef IPX
|
|
|
|
#include <netipx/ipx.h>
|
|
|
|
#include <netipx/ipx_if.h>
|
|
|
|
#endif
|
|
|
|
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
#ifdef NS
|
1994-08-13 03:50:34 +00:00
|
|
|
#include <netns/ns.h>
|
|
|
|
#include <netns/ns_if.h>
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
#endif
|
|
|
|
|
1995-03-28 07:58:53 +00:00
|
|
|
#include <machine/clock.h>
|
|
|
|
|
1994-08-13 03:50:34 +00:00
|
|
|
#include <i386/isa/isa.h>
|
|
|
|
#include <i386/isa/isa_device.h>
|
|
|
|
#include <i386/isa/ic/i82586.h>
|
|
|
|
#include <i386/isa/if_iereg.h>
|
|
|
|
#include <i386/isa/icu.h>
|
1994-08-24 22:32:44 +00:00
|
|
|
#include <i386/isa/if_ie507.h>
|
|
|
|
#include <i386/isa/elink.h>
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
|
1994-08-13 03:50:34 +00:00
|
|
|
#include <vm/vm.h>
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
|
|
|
|
#if NBPFILTER > 0
|
1994-08-13 03:50:34 +00:00
|
|
|
#include <net/bpf.h>
|
|
|
|
#include <net/bpfdesc.h>
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
#endif
|
|
|
|
|
1995-12-10 13:40:44 +00:00
|
|
|
static int check_ie_present __P((int unit, caddr_t where, unsigned size));
|
1995-11-18 08:35:49 +00:00
|
|
|
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
static struct mbuf *last_not_for_us;
|
|
|
|
|
|
|
|
#ifdef DEBUG
|
|
|
|
#define IED_RINT 1
|
|
|
|
#define IED_TINT 2
|
|
|
|
#define IED_RNR 4
|
|
|
|
#define IED_CNA 8
|
|
|
|
#define IED_READFRAME 16
|
|
|
|
int ie_debug = IED_RNR;
|
|
|
|
#endif
|
|
|
|
|
|
|
|
#ifndef ETHERMINLEN
|
|
|
|
#define ETHERMINLEN 60
|
|
|
|
#endif
|
|
|
|
|
|
|
|
#define IE_BUF_LEN 1512 /* length of transmit buffer */
|
|
|
|
|
|
|
|
/* Forward declaration */
|
|
|
|
struct ie_softc;
|
|
|
|
|
1993-11-25 01:38:01 +00:00
|
|
|
static int ieprobe(struct isa_device *dvp);
|
|
|
|
static int ieattach(struct isa_device *dvp);
|
|
|
|
static void ieinit(int unit);
|
1995-11-18 08:35:49 +00:00
|
|
|
static void ie_stop __P((int unit));
|
1993-11-25 01:38:01 +00:00
|
|
|
static int ieioctl(struct ifnet *ifp, int command, caddr_t data);
|
|
|
|
static void iestart(struct ifnet *ifp);
|
1994-08-24 22:32:44 +00:00
|
|
|
static void el_reset_586(int unit);
|
|
|
|
static void el_chan_attn(int unit);
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
static void sl_reset_586(int unit);
|
|
|
|
static void sl_chan_attn(int unit);
|
1994-05-27 04:02:10 +00:00
|
|
|
static void iereset(int unit);
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
static void ie_readframe(int unit, struct ie_softc *ie, int bufno);
|
|
|
|
static void ie_drop_packet_buffer(int unit, struct ie_softc *ie);
|
|
|
|
static void sl_read_ether(int unit, unsigned char addr[6]);
|
|
|
|
static void find_ie_mem_size(int unit);
|
1995-11-18 08:35:49 +00:00
|
|
|
static void chan_attn_timeout __P((void * rock));
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
static int command_and_wait(int unit, int command, void volatile *pcmd, int);
|
1995-11-18 08:35:49 +00:00
|
|
|
static void run_tdr __P((int unit, struct ie_tdr_cmd *cmd));
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
static int ierint(int unit, struct ie_softc *ie);
|
|
|
|
static int ietint(int unit, struct ie_softc *ie);
|
|
|
|
static int iernr(int unit, struct ie_softc *ie);
|
|
|
|
static void start_receiver(int unit);
|
|
|
|
static int ieget(int, struct ie_softc *, struct mbuf **,
|
|
|
|
struct ether_header *, int *);
|
|
|
|
static caddr_t setup_rfa(caddr_t ptr, struct ie_softc *ie);
|
|
|
|
static int mc_setup(int, caddr_t, volatile struct ie_sys_ctl_block *);
|
|
|
|
static void ie_mc_reset(int unit);
|
|
|
|
|
|
|
|
#ifdef DEBUG
|
|
|
|
void print_rbd(volatile struct ie_recv_buf_desc *rbd);
|
|
|
|
|
|
|
|
int in_ierint = 0;
|
|
|
|
int in_ietint = 0;
|
|
|
|
#endif
|
|
|
|
|
|
|
|
/*
|
|
|
|
* This tells the autoconf code how to set us up.
|
|
|
|
*/
|
|
|
|
struct isa_driver iedriver = {
|
|
|
|
ieprobe, ieattach, "ie",
|
|
|
|
};
|
|
|
|
|
|
|
|
enum ie_hardware {
|
|
|
|
IE_STARLAN10,
|
|
|
|
IE_EN100,
|
|
|
|
IE_SLFIBER,
|
1994-08-24 22:32:44 +00:00
|
|
|
IE_3C507,
|
1994-09-05 22:28:31 +00:00
|
|
|
IE_NI5210,
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
IE_UNKNOWN
|
|
|
|
};
|
|
|
|
|
1995-12-10 13:40:44 +00:00
|
|
|
static const char *ie_hardware_names[] = {
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
"StarLAN 10",
|
|
|
|
"EN100",
|
|
|
|
"StarLAN Fiber",
|
1994-08-24 22:32:44 +00:00
|
|
|
"3C507",
|
1994-09-05 22:28:31 +00:00
|
|
|
"NI5210",
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
"Unknown"
|
|
|
|
};
|
|
|
|
|
1995-05-30 08:16:23 +00:00
|
|
|
/*
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
sizeof(iscp) == 1+1+2+4 == 8
|
|
|
|
sizeof(scb) == 2+2+2+2+2+2+2+2 == 16
|
|
|
|
NFRAMES * sizeof(rfd) == NFRAMES*(2+2+2+2+6+6+2+2) == NFRAMES*24 == 384
|
|
|
|
sizeof(xmit_cmd) == 2+2+2+2+6+2 == 18
|
|
|
|
sizeof(transmit buffer) == 1512
|
|
|
|
sizeof(transmit buffer desc) == 8
|
|
|
|
-----
|
|
|
|
1946
|
|
|
|
|
|
|
|
NBUFFS * sizeof(rbd) == NBUFFS*(2+2+4+2+2) == NBUFFS*12
|
|
|
|
NBUFFS * IE_RBUF_SIZE == NBUFFS*256
|
|
|
|
|
|
|
|
NBUFFS should be (16384 - 1946) / (256 + 12) == 14438 / 268 == 53
|
|
|
|
|
|
|
|
With NBUFFS == 48, this leaves us 1574 bytes for another command or
|
|
|
|
more buffers. Another transmit command would be 18+8+1512 == 1538
|
|
|
|
---just barely fits!
|
|
|
|
|
|
|
|
Obviously all these would have to be reduced for smaller memory sizes.
|
|
|
|
With a larger memory, it would be possible to roughly double the number of
|
|
|
|
both transmit and receive buffers.
|
|
|
|
*/
|
|
|
|
|
|
|
|
#define NFRAMES 16 /* number of frames to allow for receive */
|
|
|
|
#define NBUFFS 48 /* number of buffers to allocate */
|
|
|
|
#define IE_RBUF_SIZE 256 /* size of each buffer, MUST BE POWER OF TWO */
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Ethernet status, per interface.
|
|
|
|
*/
|
1995-12-10 13:40:44 +00:00
|
|
|
static struct ie_softc {
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
struct arpcom arpcom;
|
|
|
|
void (*ie_reset_586)(int);
|
|
|
|
void (*ie_chan_attn)(int);
|
|
|
|
enum ie_hardware hard_type;
|
|
|
|
int hard_vers;
|
|
|
|
|
|
|
|
u_short port;
|
|
|
|
caddr_t iomem;
|
|
|
|
caddr_t iomembot;
|
|
|
|
unsigned iosize;
|
1994-09-05 22:28:31 +00:00
|
|
|
int bus_use; /* 0 means 16bit, 1 means 8 bit adapter */
|
1995-05-30 08:16:23 +00:00
|
|
|
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
int want_mcsetup;
|
|
|
|
int promisc;
|
|
|
|
volatile struct ie_int_sys_conf_ptr *iscp;
|
|
|
|
volatile struct ie_sys_ctl_block *scb;
|
|
|
|
volatile struct ie_recv_frame_desc *rframes[NFRAMES];
|
|
|
|
volatile struct ie_recv_buf_desc *rbuffs[NBUFFS];
|
|
|
|
volatile char *cbuffs[NBUFFS];
|
|
|
|
int rfhead, rftail, rbhead, rbtail;
|
|
|
|
|
|
|
|
volatile struct ie_xmit_cmd *xmit_cmds[2];
|
|
|
|
volatile struct ie_xmit_buf *xmit_buffs[2];
|
|
|
|
int xmit_count;
|
|
|
|
u_char *xmit_cbuffs[2];
|
|
|
|
|
|
|
|
struct ie_en_addr mcast_addrs[MAXMCAST + 1];
|
|
|
|
int mcast_count;
|
|
|
|
} ie_softc[NIE];
|
|
|
|
|
|
|
|
#define MK_24(base, ptr) ((caddr_t)((u_long)ptr - (u_long)base))
|
|
|
|
#define MK_16(base, ptr) ((u_short)(u_long)MK_24(base, ptr))
|
|
|
|
|
|
|
|
#define PORT ie_softc[unit].port
|
|
|
|
#define MEM ie_softc[unit].iomem
|
|
|
|
|
1994-08-24 22:32:44 +00:00
|
|
|
static int sl_probe(struct isa_device *);
|
|
|
|
static int el_probe(struct isa_device *);
|
1994-09-05 22:28:31 +00:00
|
|
|
static int ni_probe(struct isa_device *);
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
|
1994-10-19 01:59:11 +00:00
|
|
|
static struct kern_devconf kdc_ie[NIE] = { {
|
|
|
|
0, 0, 0, /* filled in by dev_attach */
|
1994-10-23 21:28:03 +00:00
|
|
|
"ie", 0, { MDDT_ISA, 0, "net" },
|
|
|
|
isa_generic_externalize, 0, 0, ISA_EXTERNALLEN,
|
|
|
|
&kdc_isa0, /* parent */
|
|
|
|
0, /* parentdata */
|
1995-04-12 20:48:13 +00:00
|
|
|
DC_UNCONFIGURED, /* state */
|
|
|
|
"Ethernet adapter", /* description */
|
|
|
|
DC_CLS_NETIF /* class */
|
1994-10-19 01:59:11 +00:00
|
|
|
} };
|
|
|
|
|
|
|
|
static inline void
|
1995-04-12 20:48:13 +00:00
|
|
|
ie_registerdev(struct isa_device *id)
|
1994-10-19 01:59:11 +00:00
|
|
|
{
|
|
|
|
if(id->id_unit)
|
|
|
|
kdc_ie[id->id_unit] = kdc_ie[0];
|
|
|
|
kdc_ie[id->id_unit].kdc_unit = id->id_unit;
|
|
|
|
kdc_ie[id->id_unit].kdc_isa = id;
|
|
|
|
dev_attach(&kdc_ie[id->id_unit]);
|
|
|
|
}
|
|
|
|
|
1994-08-25 20:16:59 +00:00
|
|
|
/* This routine written by Charles Martin Hannum. */
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
int ieprobe(dvp)
|
|
|
|
struct isa_device *dvp;
|
1994-08-24 22:32:44 +00:00
|
|
|
{
|
1994-08-25 20:16:59 +00:00
|
|
|
int ret;
|
1994-08-24 22:32:44 +00:00
|
|
|
|
1995-04-12 20:48:13 +00:00
|
|
|
ie_registerdev(dvp);
|
|
|
|
|
1994-08-25 20:16:59 +00:00
|
|
|
ret = sl_probe(dvp);
|
|
|
|
if(!ret) ret = el_probe(dvp);
|
1994-09-05 22:28:31 +00:00
|
|
|
if(!ret) ret = ni_probe(dvp);
|
1994-08-25 20:16:59 +00:00
|
|
|
return(ret);
|
1994-08-24 22:32:44 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static int sl_probe(dvp)
|
|
|
|
struct isa_device *dvp;
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
{
|
|
|
|
int unit = dvp->id_unit;
|
|
|
|
u_char c;
|
|
|
|
|
|
|
|
ie_softc[unit].port = dvp->id_iobase;
|
|
|
|
ie_softc[unit].iomembot = dvp->id_maddr;
|
|
|
|
ie_softc[unit].iomem = 0;
|
1994-09-05 22:28:31 +00:00
|
|
|
ie_softc[unit].bus_use = 0;
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
|
|
|
|
c = inb(PORT + IEATT_REVISION);
|
|
|
|
switch(SL_BOARD(c)) {
|
|
|
|
case SL10_BOARD:
|
|
|
|
ie_softc[unit].hard_type = IE_STARLAN10;
|
|
|
|
ie_softc[unit].ie_reset_586 = sl_reset_586;
|
|
|
|
ie_softc[unit].ie_chan_attn = sl_chan_attn;
|
|
|
|
break;
|
|
|
|
case EN100_BOARD:
|
|
|
|
ie_softc[unit].hard_type = IE_EN100;
|
|
|
|
ie_softc[unit].ie_reset_586 = sl_reset_586;
|
|
|
|
ie_softc[unit].ie_chan_attn = sl_chan_attn;
|
|
|
|
break;
|
|
|
|
case SLFIBER_BOARD:
|
|
|
|
ie_softc[unit].hard_type = IE_SLFIBER;
|
|
|
|
ie_softc[unit].ie_reset_586 = sl_reset_586;
|
|
|
|
ie_softc[unit].ie_chan_attn = sl_chan_attn;
|
|
|
|
break;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Anything else is not recognized or cannot be used.
|
|
|
|
*/
|
|
|
|
default:
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
ie_softc[unit].hard_vers = SL_REV(c);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Divine memory size on-board the card. Ususally 16k.
|
|
|
|
*/
|
|
|
|
find_ie_mem_size(unit);
|
|
|
|
|
|
|
|
if(!ie_softc[unit].iosize) {
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
dvp->id_msize = ie_softc[unit].iosize;
|
|
|
|
|
|
|
|
switch(ie_softc[unit].hard_type) {
|
|
|
|
case IE_EN100:
|
|
|
|
case IE_STARLAN10:
|
|
|
|
case IE_SLFIBER:
|
|
|
|
sl_read_ether(unit, ie_softc[unit].arpcom.ac_enaddr);
|
|
|
|
break;
|
|
|
|
|
|
|
|
default:
|
1995-05-30 08:16:23 +00:00
|
|
|
printf("ie%d: unknown AT&T board type code %d\n", unit,
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
ie_softc[unit].hard_type);
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
return 1;
|
|
|
|
}
|
|
|
|
|
1994-08-25 20:16:59 +00:00
|
|
|
/* This routine written by Charles Martin Hannum. */
|
1994-08-24 22:32:44 +00:00
|
|
|
static int el_probe(dvp)
|
|
|
|
struct isa_device *dvp;
|
|
|
|
{
|
1994-08-25 20:16:59 +00:00
|
|
|
struct ie_softc *sc = &ie_softc[dvp->id_unit];
|
|
|
|
u_char c;
|
|
|
|
int i;
|
|
|
|
u_char signature[] = "*3COM*";
|
|
|
|
int unit = dvp->id_unit;
|
1995-05-30 08:16:23 +00:00
|
|
|
|
1994-08-25 20:16:59 +00:00
|
|
|
sc->port = dvp->id_iobase;
|
|
|
|
sc->iomembot = dvp->id_maddr;
|
1994-09-05 22:28:31 +00:00
|
|
|
sc->bus_use = 0;
|
1995-05-30 08:16:23 +00:00
|
|
|
|
1994-08-25 20:16:59 +00:00
|
|
|
/* Need this for part of the probe. */
|
|
|
|
sc->ie_reset_586 = el_reset_586;
|
|
|
|
sc->ie_chan_attn = el_chan_attn;
|
1995-05-30 08:16:23 +00:00
|
|
|
|
1994-08-25 20:16:59 +00:00
|
|
|
/* Reset and put card in CONFIG state without changing address. */
|
|
|
|
elink_reset();
|
|
|
|
outb(ELINK_ID_PORT, 0x00);
|
|
|
|
elink_idseq(ELINK_507_POLY);
|
|
|
|
elink_idseq(ELINK_507_POLY);
|
|
|
|
outb(ELINK_ID_PORT, 0xff);
|
1995-05-30 08:16:23 +00:00
|
|
|
|
1994-08-25 20:16:59 +00:00
|
|
|
c = inb(PORT + IE507_MADDR);
|
|
|
|
if(c & 0x20) {
|
1994-09-02 23:23:57 +00:00
|
|
|
#ifdef DEBUG
|
1994-08-25 20:16:59 +00:00
|
|
|
printf("ie%d: can't map 3C507 RAM in high memory\n", unit);
|
1994-09-02 23:23:57 +00:00
|
|
|
#endif
|
1994-08-25 20:16:59 +00:00
|
|
|
return 0;
|
|
|
|
}
|
1995-05-30 08:16:23 +00:00
|
|
|
|
1994-08-25 20:16:59 +00:00
|
|
|
/* go to RUN state */
|
|
|
|
outb(ELINK_ID_PORT, 0x00);
|
|
|
|
elink_idseq(ELINK_507_POLY);
|
|
|
|
outb(ELINK_ID_PORT, 0x00);
|
1995-05-30 08:16:23 +00:00
|
|
|
|
1994-08-25 20:16:59 +00:00
|
|
|
outb(PORT + IE507_CTRL, EL_CTRL_NRST);
|
1995-05-30 08:16:23 +00:00
|
|
|
|
1994-08-25 20:16:59 +00:00
|
|
|
for (i = 0; i < 6; i++)
|
|
|
|
if (inb(PORT + i) != signature[i])
|
|
|
|
return 0;
|
1995-05-30 08:16:23 +00:00
|
|
|
|
1994-08-25 20:16:59 +00:00
|
|
|
c = inb(PORT + IE507_IRQ) & 0x0f;
|
1995-05-30 08:16:23 +00:00
|
|
|
|
1994-08-25 20:16:59 +00:00
|
|
|
if (dvp->id_irq != (1 << c)) {
|
|
|
|
printf("ie%d: kernel configured irq %d doesn't match board configured irq %d\n",
|
|
|
|
unit, ffs(dvp->id_irq) - 1, c);
|
|
|
|
return 0;
|
|
|
|
}
|
1995-05-30 08:16:23 +00:00
|
|
|
|
1994-08-25 20:16:59 +00:00
|
|
|
c = (inb(PORT + IE507_MADDR) & 0x1c) + 0xc0;
|
1995-05-30 08:16:23 +00:00
|
|
|
|
1994-08-25 20:16:59 +00:00
|
|
|
if (kvtop(dvp->id_maddr) != ((int)c << 12)) {
|
1994-10-26 00:16:20 +00:00
|
|
|
printf("ie%d: kernel configured maddr %lx doesn't match board configured maddr %x\n",
|
1994-08-25 20:16:59 +00:00
|
|
|
unit, kvtop(dvp->id_maddr),(int)c << 12);
|
|
|
|
return 0;
|
|
|
|
}
|
1995-05-30 08:16:23 +00:00
|
|
|
|
1994-08-25 20:16:59 +00:00
|
|
|
outb(PORT + IE507_CTRL, EL_CTRL_NORMAL);
|
1995-05-30 08:16:23 +00:00
|
|
|
|
1994-08-25 20:16:59 +00:00
|
|
|
sc->hard_type = IE_3C507;
|
|
|
|
sc->hard_vers = 0; /* 3C507 has no version number. */
|
1995-05-30 08:16:23 +00:00
|
|
|
|
1994-08-25 20:16:59 +00:00
|
|
|
/*
|
|
|
|
* Divine memory size on-board the card.
|
|
|
|
*/
|
1994-09-05 22:28:31 +00:00
|
|
|
find_ie_mem_size(unit);
|
1995-05-30 08:16:23 +00:00
|
|
|
|
1994-08-25 20:16:59 +00:00
|
|
|
if (!sc->iosize) {
|
|
|
|
printf("ie%d: can't find shared memory\n", unit);
|
|
|
|
outb(PORT + IE507_CTRL, EL_CTRL_NRST);
|
|
|
|
return 0;
|
|
|
|
}
|
1995-05-30 08:16:23 +00:00
|
|
|
|
1994-08-25 20:16:59 +00:00
|
|
|
if(!dvp->id_msize)
|
|
|
|
dvp->id_msize = sc->iosize;
|
|
|
|
else if (dvp->id_msize != sc->iosize) {
|
|
|
|
printf("ie%d: kernel configured msize %d doesn't match board configured msize %d\n",
|
|
|
|
unit, dvp->id_msize, sc->iosize);
|
|
|
|
outb(PORT + IE507_CTRL, EL_CTRL_NRST);
|
|
|
|
return 0;
|
|
|
|
}
|
1995-05-30 08:16:23 +00:00
|
|
|
|
1994-08-25 20:16:59 +00:00
|
|
|
sl_read_ether(unit, ie_softc[unit].arpcom.ac_enaddr);
|
1995-05-30 08:16:23 +00:00
|
|
|
|
1994-08-25 20:16:59 +00:00
|
|
|
/* Clear the interrupt latch just in case. */
|
|
|
|
outb(PORT + IE507_ICTRL, 1);
|
1995-05-30 08:16:23 +00:00
|
|
|
|
1994-08-25 20:16:59 +00:00
|
|
|
return 16;
|
1994-08-24 22:32:44 +00:00
|
|
|
}
|
|
|
|
|
1994-09-05 22:28:31 +00:00
|
|
|
|
|
|
|
static int ni_probe(dvp)
|
|
|
|
struct isa_device *dvp;
|
|
|
|
{
|
|
|
|
int unit = dvp->id_unit;
|
|
|
|
int boardtype, c;
|
|
|
|
|
|
|
|
ie_softc[unit].port = dvp->id_iobase;
|
|
|
|
ie_softc[unit].iomembot = dvp->id_maddr;
|
|
|
|
ie_softc[unit].iomem = 0;
|
|
|
|
ie_softc[unit].bus_use = 1;
|
|
|
|
|
|
|
|
boardtype = inb(PORT + IEATT_REVISION);
|
|
|
|
c = inb(PORT + IEATT_REVISION + 1);
|
|
|
|
boardtype = boardtype + (c << 8);
|
|
|
|
switch(boardtype) {
|
|
|
|
case 0x5500: /* This is the magic cookie for the NI5210 */
|
|
|
|
ie_softc[unit].hard_type = IE_NI5210;
|
|
|
|
ie_softc[unit].ie_reset_586 = sl_reset_586;
|
|
|
|
ie_softc[unit].ie_chan_attn = sl_chan_attn;
|
|
|
|
break;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Anything else is not recognized or cannot be used.
|
|
|
|
*/
|
|
|
|
default:
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
ie_softc[unit].hard_vers = 0;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Divine memory size on-board the card. Either 8 or 16k.
|
|
|
|
*/
|
|
|
|
find_ie_mem_size(unit);
|
|
|
|
|
|
|
|
if(!ie_softc[unit].iosize) {
|
1995-05-30 08:16:23 +00:00
|
|
|
return 0;
|
1994-09-05 22:28:31 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
if(!dvp->id_msize)
|
|
|
|
dvp->id_msize = ie_softc[unit].iosize;
|
|
|
|
else if (dvp->id_msize != ie_softc[unit].iosize) {
|
|
|
|
printf("ie%d: kernel configured msize %d doesn't match board configured msize %d\n",
|
|
|
|
unit, dvp->id_msize, ie_softc[unit].iosize);
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
sl_read_ether(unit, ie_softc[unit].arpcom.ac_enaddr);
|
|
|
|
|
|
|
|
return 8;
|
|
|
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
/*
|
|
|
|
* Taken almost exactly from Bill's if_is.c, then modified beyond recognition.
|
|
|
|
*/
|
|
|
|
int
|
|
|
|
ieattach(dvp)
|
|
|
|
struct isa_device *dvp;
|
|
|
|
{
|
|
|
|
int unit = dvp->id_unit;
|
|
|
|
struct ie_softc *ie = &ie_softc[unit];
|
|
|
|
struct ifnet *ifp = &ie->arpcom.ac_if;
|
|
|
|
|
1996-02-06 18:51:28 +00:00
|
|
|
ifp->if_softc = ie;
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
ifp->if_unit = unit;
|
|
|
|
ifp->if_name = iedriver.name;
|
|
|
|
ifp->if_mtu = ETHERMTU;
|
1996-01-26 09:29:29 +00:00
|
|
|
printf(" <%s R%d> ethernet address %6D\n",
|
1996-02-06 18:51:28 +00:00
|
|
|
ie_hardware_names[ie->hard_type],
|
|
|
|
ie->hard_vers + 1,
|
1996-01-26 09:29:29 +00:00
|
|
|
ie->arpcom.ac_enaddr, ":");
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
|
1995-10-13 19:48:06 +00:00
|
|
|
ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
ifp->if_output = ether_output;
|
|
|
|
ifp->if_start = iestart;
|
|
|
|
ifp->if_ioctl = ieioctl;
|
|
|
|
ifp->if_type = IFT_ETHER;
|
|
|
|
ifp->if_addrlen = 6;
|
|
|
|
ifp->if_hdrlen = 14;
|
1995-05-30 08:16:23 +00:00
|
|
|
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
#if NBPFILTER > 0
|
1996-02-06 18:51:28 +00:00
|
|
|
bpfattach(ifp, DLT_EN10MB, sizeof(struct ether_header));
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
#endif
|
|
|
|
|
|
|
|
if_attach(ifp);
|
1996-02-06 18:51:28 +00:00
|
|
|
ether_ifattach(ifp);
|
1995-04-12 20:48:13 +00:00
|
|
|
kdc_ie[unit].kdc_description = ie_hardware_names[ie_softc[unit].hard_type];
|
1996-02-06 18:51:28 +00:00
|
|
|
return 1;
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* What to do upon receipt of an interrupt.
|
|
|
|
*/
|
1995-09-19 18:55:37 +00:00
|
|
|
void ieintr(unit)
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
int unit;
|
|
|
|
{
|
|
|
|
register struct ie_softc *ie = &ie_softc[unit];
|
|
|
|
register u_short status;
|
|
|
|
|
|
|
|
status = ie->scb->ie_status;
|
|
|
|
|
1994-08-25 20:16:59 +00:00
|
|
|
/* This if statement written by Charles Martin Hannum. */
|
1994-08-24 22:32:44 +00:00
|
|
|
if ((status & IE_ST_WHENCE) == 0) {
|
1994-08-25 20:16:59 +00:00
|
|
|
/* Clear the interrupt latch on the 3C507. */
|
|
|
|
if (ie->hard_type == IE_3C507 &&
|
|
|
|
(inb(PORT + IE507_CTRL) & EL_CTRL_INTL))
|
|
|
|
outb(PORT + IE507_ICTRL, 1);
|
1994-08-24 22:32:44 +00:00
|
|
|
}
|
|
|
|
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
loop:
|
|
|
|
if(status & (IE_ST_RECV | IE_ST_RNR)) {
|
|
|
|
#ifdef DEBUG
|
|
|
|
in_ierint++;
|
|
|
|
if(ie_debug & IED_RINT)
|
|
|
|
printf("ie%d: rint\n", unit);
|
|
|
|
#endif
|
|
|
|
ierint(unit, ie);
|
|
|
|
#ifdef DEBUG
|
|
|
|
in_ierint--;
|
|
|
|
#endif
|
|
|
|
}
|
|
|
|
|
|
|
|
if(status & IE_ST_DONE) {
|
|
|
|
#ifdef DEBUG
|
|
|
|
in_ietint++;
|
|
|
|
if(ie_debug & IED_TINT)
|
|
|
|
printf("ie%d: tint\n", unit);
|
|
|
|
#endif
|
|
|
|
ietint(unit, ie);
|
|
|
|
#ifdef DEBUG
|
|
|
|
in_ietint--;
|
|
|
|
#endif
|
|
|
|
}
|
|
|
|
|
|
|
|
if(status & IE_ST_RNR) {
|
|
|
|
#ifdef DEBUG
|
|
|
|
if(ie_debug & IED_RNR)
|
|
|
|
printf("ie%d: rnr\n", unit);
|
|
|
|
#endif
|
|
|
|
iernr(unit, ie);
|
|
|
|
}
|
|
|
|
|
|
|
|
#ifdef DEBUG
|
|
|
|
if((status & IE_ST_ALLDONE)
|
|
|
|
&& (ie_debug & IED_CNA))
|
|
|
|
printf("ie%d: cna\n", unit);
|
|
|
|
#endif
|
|
|
|
|
|
|
|
/* Don't ack interrupts which we didn't receive */
|
|
|
|
ie_ack(ie->scb, IE_ST_WHENCE & status, unit, ie->ie_chan_attn);
|
|
|
|
|
|
|
|
if((status = ie->scb->ie_status) & IE_ST_WHENCE)
|
|
|
|
goto loop;
|
|
|
|
|
1994-08-25 20:16:59 +00:00
|
|
|
/* This comment and if statement written by Charles Martin Hannum. */
|
1994-08-24 22:32:44 +00:00
|
|
|
/* Clear the interrupt latch on the 3C507. */
|
|
|
|
if (ie->hard_type == IE_3C507)
|
1994-08-25 20:16:59 +00:00
|
|
|
outb(PORT + IE507_ICTRL, 1);
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Process a received-frame interrupt.
|
|
|
|
*/
|
|
|
|
static int ierint(unit, ie)
|
|
|
|
int unit;
|
|
|
|
struct ie_softc *ie;
|
|
|
|
{
|
|
|
|
int i, status;
|
|
|
|
static int timesthru = 1024;
|
|
|
|
|
|
|
|
i = ie->rfhead;
|
|
|
|
while(1) {
|
|
|
|
status = ie->rframes[i]->ie_fd_status;
|
|
|
|
|
|
|
|
if((status & IE_FD_COMPLETE) && (status & IE_FD_OK)) {
|
|
|
|
ie->arpcom.ac_if.if_ipackets++;
|
|
|
|
if(!--timesthru) {
|
|
|
|
ie->arpcom.ac_if.if_ierrors += ie->scb->ie_err_crc + ie->scb->ie_err_align +
|
|
|
|
ie->scb->ie_err_resource + ie->scb->ie_err_overrun;
|
|
|
|
ie->scb->ie_err_crc = 0;
|
|
|
|
ie->scb->ie_err_align = 0;
|
|
|
|
ie->scb->ie_err_resource = 0;
|
|
|
|
ie->scb->ie_err_overrun = 0;
|
|
|
|
timesthru = 1024;
|
|
|
|
}
|
|
|
|
ie_readframe(unit, ie, i);
|
|
|
|
} else {
|
|
|
|
if(status & IE_FD_RNR) {
|
|
|
|
if(!(ie->scb->ie_status & IE_RU_READY)) {
|
|
|
|
ie->rframes[0]->ie_fd_next = MK_16(MEM, ie->rbuffs[0]);
|
|
|
|
ie->scb->ie_recv_list = MK_16(MEM, ie->rframes[0]);
|
|
|
|
command_and_wait(unit, IE_RU_START, 0, 0);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
i = (i + 1) % NFRAMES;
|
|
|
|
}
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Process a command-complete interrupt. These are only generated by
|
|
|
|
* the transmission of frames. This routine is deceptively simple, since
|
|
|
|
* most of the real work is done by iestart().
|
|
|
|
*/
|
|
|
|
static int ietint(unit, ie)
|
|
|
|
int unit;
|
|
|
|
struct ie_softc *ie;
|
|
|
|
{
|
|
|
|
int status;
|
|
|
|
int i;
|
|
|
|
|
|
|
|
ie->arpcom.ac_if.if_timer = 0;
|
|
|
|
ie->arpcom.ac_if.if_flags &= ~IFF_OACTIVE;
|
|
|
|
|
|
|
|
for(i = 0; i < ie->xmit_count; i++) {
|
|
|
|
status = ie->xmit_cmds[i]->ie_xmit_status;
|
|
|
|
|
|
|
|
if(status & IE_XS_LATECOLL) {
|
|
|
|
printf("ie%d: late collision\n", unit);
|
|
|
|
ie->arpcom.ac_if.if_collisions++;
|
|
|
|
ie->arpcom.ac_if.if_oerrors++;
|
|
|
|
} else if(status & IE_XS_NOCARRIER) {
|
|
|
|
printf("ie%d: no carrier\n", unit);
|
|
|
|
ie->arpcom.ac_if.if_oerrors++;
|
|
|
|
} else if(status & IE_XS_LOSTCTS) {
|
|
|
|
printf("ie%d: lost CTS\n", unit);
|
|
|
|
ie->arpcom.ac_if.if_oerrors++;
|
|
|
|
} else if(status & IE_XS_UNDERRUN) {
|
|
|
|
printf("ie%d: DMA underrun\n", unit);
|
|
|
|
ie->arpcom.ac_if.if_oerrors++;
|
|
|
|
} else if(status & IE_XS_EXCMAX) {
|
|
|
|
printf("ie%d: too many collisions\n", unit);
|
|
|
|
ie->arpcom.ac_if.if_collisions += 16;
|
|
|
|
ie->arpcom.ac_if.if_oerrors++;
|
|
|
|
} else {
|
|
|
|
ie->arpcom.ac_if.if_opackets++;
|
|
|
|
ie->arpcom.ac_if.if_collisions += status & IE_XS_MAXCOLL;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
ie->xmit_count = 0;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* If multicast addresses were added or deleted while we were transmitting,
|
|
|
|
* ie_mc_reset() set the want_mcsetup flag indicating that we should do it.
|
|
|
|
*/
|
|
|
|
if(ie->want_mcsetup) {
|
|
|
|
mc_setup(unit, (caddr_t)ie->xmit_cbuffs[0], ie->scb);
|
|
|
|
ie->want_mcsetup = 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Wish I knew why this seems to be necessary... */
|
|
|
|
ie->xmit_cmds[0]->ie_xmit_status |= IE_STAT_COMPL;
|
|
|
|
|
|
|
|
iestart(&ie->arpcom.ac_if);
|
|
|
|
return 0; /* shouldn't be necessary */
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Process a receiver-not-ready interrupt. I believe that we get these
|
|
|
|
* when there aren't enough buffers to go around. For now (FIXME), we
|
|
|
|
* just restart the receiver, and hope everything's ok.
|
|
|
|
*/
|
|
|
|
static int iernr(unit, ie)
|
|
|
|
int unit;
|
|
|
|
struct ie_softc *ie;
|
|
|
|
{
|
|
|
|
#ifdef doesnt_work
|
|
|
|
setup_rfa((caddr_t)ie->rframes[0], ie);
|
|
|
|
|
|
|
|
ie->scb->ie_recv_list = MK_16(MEM, ie_softc[unit].rframes[0]);
|
|
|
|
command_and_wait(unit, IE_RU_START, 0, 0);
|
|
|
|
#else
|
|
|
|
/* This doesn't work either, but it doesn't hang either. */
|
|
|
|
command_and_wait(unit, IE_RU_DISABLE, 0, 0); /* just in case */
|
|
|
|
setup_rfa((caddr_t)ie->rframes[0], ie); /* ignore cast-qual */
|
|
|
|
|
|
|
|
ie->scb->ie_recv_list = MK_16(MEM, ie_softc[unit].rframes[0]);
|
|
|
|
command_and_wait(unit, IE_RU_START, 0, 0); /* was ENABLE */
|
|
|
|
|
|
|
|
#endif
|
|
|
|
ie_ack(ie->scb, IE_ST_WHENCE, unit, ie->ie_chan_attn);
|
|
|
|
|
|
|
|
ie->arpcom.ac_if.if_ierrors++;
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Compare two Ether/802 addresses for equality, inlined and
|
|
|
|
* unrolled for speed. I'd love to have an inline assembler
|
|
|
|
* version of this...
|
|
|
|
*/
|
|
|
|
static inline int ether_equal(u_char *one, u_char *two) {
|
|
|
|
if(one[0] != two[0]) return 0;
|
|
|
|
if(one[1] != two[1]) return 0;
|
|
|
|
if(one[2] != two[2]) return 0;
|
|
|
|
if(one[3] != two[3]) return 0;
|
|
|
|
if(one[4] != two[4]) return 0;
|
|
|
|
if(one[5] != two[5]) return 0;
|
|
|
|
return 1;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Check for a valid address. to_bpf is filled in with one of the following:
|
|
|
|
* 0 -> BPF doesn't get this packet
|
|
|
|
* 1 -> BPF does get this packet
|
|
|
|
* 2 -> BPF does get this packet, but we don't
|
|
|
|
* Return value is true if the packet is for us, and false otherwise.
|
|
|
|
*
|
|
|
|
* This routine is a mess, but it's also critical that it be as fast
|
|
|
|
* as possible. It could be made cleaner if we can assume that the
|
|
|
|
* only client which will fiddle with IFF_PROMISC is BPF. This is
|
|
|
|
* probably a good assumption, but we do not make it here. (Yet.)
|
|
|
|
*/
|
1995-05-30 08:16:23 +00:00
|
|
|
static inline int check_eh(struct ie_softc *ie,
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
struct ether_header *eh,
|
|
|
|
int *to_bpf) {
|
|
|
|
int i;
|
|
|
|
|
|
|
|
switch(ie->promisc) {
|
|
|
|
case IFF_ALLMULTI:
|
|
|
|
/*
|
|
|
|
* Receiving all multicasts, but no unicasts except those destined for us.
|
|
|
|
*/
|
|
|
|
#if NBPFILTER > 0
|
1996-02-06 18:51:28 +00:00
|
|
|
*to_bpf = (ie->arpcom.ac_if.if_bpf != 0); /* BPF gets this packet if anybody cares */
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
#endif
|
|
|
|
if(eh->ether_dhost[0] & 1) {
|
|
|
|
return 1;
|
|
|
|
}
|
|
|
|
if(ether_equal(eh->ether_dhost, ie->arpcom.ac_enaddr)) return 1;
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
case IFF_PROMISC:
|
|
|
|
/*
|
|
|
|
* Receiving all packets. These need to be passed on to BPF.
|
|
|
|
*/
|
|
|
|
#if NBPFILTER > 0
|
1996-02-06 18:51:28 +00:00
|
|
|
*to_bpf = (ie->arpcom.ac_if.if_bpf != 0);
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
#endif
|
|
|
|
/* If for us, accept and hand up to BPF */
|
|
|
|
if(ether_equal(eh->ether_dhost, ie->arpcom.ac_enaddr)) return 1;
|
|
|
|
|
|
|
|
#if NBPFILTER > 0
|
|
|
|
if(*to_bpf) *to_bpf = 2; /* we don't need to see it */
|
|
|
|
#endif
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Not a multicast, so BPF wants to see it but we don't.
|
|
|
|
*/
|
|
|
|
if(!(eh->ether_dhost[0] & 1)) return 1;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* If it's one of our multicast groups, accept it and pass it
|
|
|
|
* up.
|
|
|
|
*/
|
|
|
|
for(i = 0; i < ie->mcast_count; i++) {
|
|
|
|
if(ether_equal(eh->ether_dhost, (u_char *)&ie->mcast_addrs[i])) {
|
|
|
|
#if NBPFILTER > 0
|
|
|
|
if(*to_bpf) *to_bpf = 1;
|
|
|
|
#endif
|
|
|
|
return 1;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return 1;
|
|
|
|
|
|
|
|
case IFF_ALLMULTI | IFF_PROMISC:
|
|
|
|
/*
|
|
|
|
* Acting as a multicast router, and BPF running at the same time.
|
|
|
|
* Whew! (Hope this is a fast machine...)
|
|
|
|
*/
|
|
|
|
#if NBPFILTER > 0
|
1996-02-06 18:51:28 +00:00
|
|
|
*to_bpf = (ie->arpcom.ac_if.if_bpf != 0);
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
#endif
|
|
|
|
/* We want to see multicasts. */
|
|
|
|
if(eh->ether_dhost[0] & 1) return 1;
|
1995-05-30 08:16:23 +00:00
|
|
|
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
/* We want to see our own packets */
|
|
|
|
if(ether_equal(eh->ether_dhost, ie->arpcom.ac_enaddr)) return 1;
|
|
|
|
|
|
|
|
/* Anything else goes to BPF but nothing else. */
|
|
|
|
#if NBPFILTER > 0
|
|
|
|
if(*to_bpf) *to_bpf = 2;
|
|
|
|
#endif
|
|
|
|
return 1;
|
|
|
|
|
|
|
|
default:
|
|
|
|
/*
|
|
|
|
* Only accept unicast packets destined for us, or multicasts
|
|
|
|
* for groups that we belong to. For now, we assume that the
|
|
|
|
* '586 will only return packets that we asked it for. This
|
|
|
|
* isn't strictly true (it uses hashing for the multicast filter),
|
|
|
|
* but it will do in this case, and we want to get out of here
|
|
|
|
* as quickly as possible.
|
|
|
|
*/
|
|
|
|
#if NBPFILTER > 0
|
1996-02-06 18:51:28 +00:00
|
|
|
*to_bpf = (ie->arpcom.ac_if.if_bpf != 0);
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
#endif
|
|
|
|
return 1;
|
|
|
|
}
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* We want to isolate the bits that have meaning... This assumes that
|
|
|
|
* IE_RBUF_SIZE is an even power of two. If somehow the act_len exceeds
|
|
|
|
* the size of the buffer, then we are screwed anyway.
|
|
|
|
*/
|
|
|
|
static inline int ie_buflen(struct ie_softc *ie, int head) {
|
|
|
|
return (ie->rbuffs[head]->ie_rbd_actual
|
|
|
|
& (IE_RBUF_SIZE | (IE_RBUF_SIZE - 1)));
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline int ie_packet_len(int unit, struct ie_softc *ie) {
|
|
|
|
int i;
|
|
|
|
int head = ie->rbhead;
|
|
|
|
int acc = 0;
|
|
|
|
|
|
|
|
do {
|
|
|
|
if(!(ie->rbuffs[ie->rbhead]->ie_rbd_actual & IE_RBD_USED)) {
|
|
|
|
#ifdef DEBUG
|
|
|
|
print_rbd(ie->rbuffs[ie->rbhead]);
|
|
|
|
#endif
|
|
|
|
log(LOG_ERR, "ie%d: receive descriptors out of sync at %d\n",
|
|
|
|
unit, ie->rbhead);
|
1994-05-27 04:02:10 +00:00
|
|
|
iereset(unit);
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
return -1;
|
|
|
|
}
|
|
|
|
|
|
|
|
i = ie->rbuffs[head]->ie_rbd_actual & IE_RBD_LAST;
|
|
|
|
|
|
|
|
acc += ie_buflen(ie, head);
|
|
|
|
head = (head + 1) % NBUFFS;
|
|
|
|
} while(!i);
|
|
|
|
|
|
|
|
return acc;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Read data off the interface, and turn it into an mbuf chain.
|
|
|
|
*
|
|
|
|
* This code is DRAMATICALLY different from the previous version; this
|
|
|
|
* version tries to allocate the entire mbuf chain up front, given the
|
|
|
|
* length of the data available. This enables us to allocate mbuf
|
|
|
|
* clusters in many situations where before we would have had a long
|
|
|
|
* chain of partially-full mbufs. This should help to speed up the
|
|
|
|
* operation considerably. (Provided that it works, of course.)
|
|
|
|
*/
|
|
|
|
static inline int ieget(unit, ie, mp, ehp, to_bpf)
|
|
|
|
int unit;
|
|
|
|
struct ie_softc *ie;
|
|
|
|
struct mbuf **mp;
|
|
|
|
struct ether_header *ehp;
|
|
|
|
int *to_bpf;
|
|
|
|
{
|
|
|
|
struct mbuf *m, *top, **mymp;
|
|
|
|
int i;
|
|
|
|
int offset;
|
|
|
|
int totlen, resid;
|
|
|
|
int thismboff;
|
|
|
|
int head;
|
|
|
|
|
|
|
|
totlen = ie_packet_len(unit, ie);
|
|
|
|
if(totlen <= 0) return -1;
|
|
|
|
|
|
|
|
i = ie->rbhead;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Snarf the Ethernet header.
|
|
|
|
*/
|
|
|
|
bcopy((caddr_t)ie->cbuffs[i], (caddr_t)ehp, sizeof *ehp);
|
|
|
|
/* ignore cast-qual warning here */
|
|
|
|
|
|
|
|
/*
|
|
|
|
* As quickly as possible, check if this packet is for us.
|
|
|
|
* If not, don't waste a single cycle copying the rest of the
|
|
|
|
* packet in.
|
|
|
|
* This is only a consideration when FILTER is defined; i.e., when
|
|
|
|
* we are either running BPF or doing multicasting.
|
|
|
|
*/
|
|
|
|
if(!check_eh(ie, ehp, to_bpf)) {
|
|
|
|
ie_drop_packet_buffer(unit, ie);
|
|
|
|
ie->arpcom.ac_if.if_ierrors--; /* just this case, it's not an error */
|
|
|
|
return -1;
|
|
|
|
}
|
|
|
|
totlen -= (offset = sizeof *ehp);
|
|
|
|
|
|
|
|
MGETHDR(*mp, M_DONTWAIT, MT_DATA);
|
|
|
|
if(!*mp) {
|
|
|
|
ie_drop_packet_buffer(unit, ie);
|
|
|
|
return -1;
|
|
|
|
}
|
|
|
|
|
|
|
|
m = *mp;
|
|
|
|
m->m_pkthdr.rcvif = &ie->arpcom.ac_if;
|
|
|
|
m->m_len = MHLEN;
|
|
|
|
resid = m->m_pkthdr.len = totlen;
|
|
|
|
top = 0;
|
|
|
|
mymp = ⊤
|
|
|
|
|
|
|
|
/*
|
|
|
|
* This loop goes through and allocates mbufs for all the data we will
|
|
|
|
* be copying in. It does not actually do the copying yet.
|
|
|
|
*/
|
|
|
|
do { /* while(resid > 0) */
|
|
|
|
/*
|
1995-05-30 08:16:23 +00:00
|
|
|
* Try to allocate an mbuf to hold the data that we have. If we
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
* already allocated one, just get another one and stick it on the
|
|
|
|
* end (eventually). If we don't already have one, try to allocate
|
|
|
|
* an mbuf cluster big enough to hold the whole packet, if we think it's
|
|
|
|
* reasonable, or a single mbuf which may or may not be big enough.
|
|
|
|
* Got that?
|
|
|
|
*/
|
|
|
|
if(top) {
|
|
|
|
MGET(m, M_DONTWAIT, MT_DATA);
|
|
|
|
if(!m) {
|
|
|
|
m_freem(top);
|
|
|
|
ie_drop_packet_buffer(unit, ie);
|
|
|
|
return -1;
|
|
|
|
}
|
|
|
|
m->m_len = MLEN;
|
|
|
|
}
|
|
|
|
|
|
|
|
if(resid >= MINCLSIZE) {
|
|
|
|
MCLGET(m, M_DONTWAIT);
|
|
|
|
if(m->m_flags & M_EXT)
|
|
|
|
m->m_len = min(resid, MCLBYTES);
|
|
|
|
} else {
|
|
|
|
if(resid < m->m_len) {
|
|
|
|
if(!top && resid + max_linkhdr <= m->m_len)
|
|
|
|
m->m_data += max_linkhdr;
|
|
|
|
m->m_len = resid;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
resid -= m->m_len;
|
|
|
|
*mymp = m;
|
|
|
|
mymp = &m->m_next;
|
|
|
|
} while(resid > 0);
|
|
|
|
|
|
|
|
resid = totlen;
|
|
|
|
m = top;
|
|
|
|
thismboff = 0;
|
|
|
|
head = ie->rbhead;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Now we take the mbuf chain (hopefully only one mbuf most of the
|
|
|
|
* time) and stuff the data into it. There are no possible failures
|
|
|
|
* at or after this point.
|
|
|
|
*/
|
|
|
|
while(resid > 0) { /* while there's stuff left */
|
|
|
|
int thislen = ie_buflen(ie, head) - offset;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* If too much data for the current mbuf, then fill the current one
|
|
|
|
* up, go to the next one, and try again.
|
|
|
|
*/
|
|
|
|
if(thislen > m->m_len - thismboff) {
|
|
|
|
int newlen = m->m_len - thismboff;
|
|
|
|
bcopy((caddr_t)(ie->cbuffs[head] + offset),
|
|
|
|
mtod(m, caddr_t) + thismboff, (unsigned)newlen);
|
|
|
|
/* ignore cast-qual warning */
|
|
|
|
m = m->m_next;
|
|
|
|
thismboff = 0; /* new mbuf, so no offset */
|
|
|
|
offset += newlen; /* we are now this far into the packet */
|
|
|
|
resid -= newlen; /* so there is this much left to get */
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* If there is more than enough space in the mbuf to hold the
|
|
|
|
* contents of this buffer, copy everything in, advance pointers,
|
|
|
|
* and so on.
|
|
|
|
*/
|
|
|
|
if(thislen < m->m_len - thismboff) {
|
|
|
|
bcopy((caddr_t)(ie->cbuffs[head] + offset), /* ignore warning */
|
|
|
|
mtod(m, caddr_t) + thismboff, (unsigned)thislen);
|
|
|
|
thismboff += thislen; /* we are this far into the mbuf */
|
|
|
|
resid -= thislen; /* and this much is left */
|
|
|
|
goto nextbuf;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Otherwise, there is exactly enough space to put this buffer's
|
|
|
|
* contents into the current mbuf. Do the combination of the above
|
|
|
|
* actions.
|
|
|
|
*/
|
|
|
|
bcopy((caddr_t)(ie->cbuffs[head] + offset), /* ignore warning */
|
|
|
|
mtod(m, caddr_t) + thismboff, (unsigned)thislen);
|
|
|
|
m = m->m_next;
|
|
|
|
thismboff = 0; /* new mbuf, start at the beginning */
|
|
|
|
resid -= thislen; /* and we are this far through */
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Advance all the pointers. We can get here from either of the
|
|
|
|
* last two cases, but never the first.
|
|
|
|
*/
|
|
|
|
nextbuf:
|
|
|
|
offset = 0;
|
|
|
|
ie->rbuffs[head]->ie_rbd_actual = 0;
|
|
|
|
ie->rbuffs[head]->ie_rbd_length |= IE_RBD_LAST;
|
|
|
|
ie->rbhead = head = (head + 1) % NBUFFS;
|
|
|
|
ie->rbuffs[ie->rbtail]->ie_rbd_length &= ~IE_RBD_LAST;
|
|
|
|
ie->rbtail = (ie->rbtail + 1) % NBUFFS;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Unless something changed strangely while we were doing the copy,
|
|
|
|
* we have now copied everything in from the shared memory.
|
|
|
|
* This means that we are done.
|
|
|
|
*/
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Read frame NUM from unit UNIT (pre-cached as IE).
|
|
|
|
*
|
|
|
|
* This routine reads the RFD at NUM, and copies in the buffers from
|
|
|
|
* the list of RBD, then rotates the RBD and RFD lists so that the receiver
|
|
|
|
* doesn't start complaining. Trailers are DROPPED---there's no point
|
|
|
|
* in wasting time on confusing code to deal with them. Hopefully,
|
|
|
|
* this machine will never ARP for trailers anyway.
|
|
|
|
*/
|
|
|
|
static void ie_readframe(unit, ie, num)
|
|
|
|
int unit;
|
|
|
|
struct ie_softc *ie;
|
|
|
|
int num; /* frame number to read */
|
|
|
|
{
|
|
|
|
struct ie_recv_frame_desc rfd;
|
|
|
|
struct mbuf *m = 0;
|
|
|
|
struct ether_header eh;
|
|
|
|
#if NBPFILTER > 0
|
|
|
|
int bpf_gets_it = 0;
|
|
|
|
#endif
|
|
|
|
|
|
|
|
bcopy((caddr_t)(ie->rframes[num]), &rfd, sizeof(struct ie_recv_frame_desc));
|
|
|
|
|
|
|
|
/* Immediately advance the RFD list, since we we have copied ours now. */
|
|
|
|
ie->rframes[num]->ie_fd_status = 0;
|
|
|
|
ie->rframes[num]->ie_fd_last |= IE_FD_LAST;
|
|
|
|
ie->rframes[ie->rftail]->ie_fd_last &= ~IE_FD_LAST;
|
|
|
|
ie->rftail = (ie->rftail + 1) % NFRAMES;
|
|
|
|
ie->rfhead = (ie->rfhead + 1) % NFRAMES;
|
|
|
|
|
|
|
|
if(rfd.ie_fd_status & IE_FD_OK) {
|
|
|
|
if(
|
|
|
|
#if NBPFILTER > 0
|
|
|
|
ieget(unit, ie, &m, &eh, &bpf_gets_it)
|
|
|
|
#else
|
|
|
|
ieget(unit, ie, &m, &eh, (int *)0)
|
|
|
|
#endif
|
|
|
|
) {
|
|
|
|
ie->arpcom.ac_if.if_ierrors++; /* this counts as an error */
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
#ifdef DEBUG
|
|
|
|
if(ie_debug & IED_READFRAME) {
|
1996-01-26 09:29:29 +00:00
|
|
|
printf("ie%d: frame from ether %6D type %x\n", unit,
|
|
|
|
eh.ether_shost, ":", (unsigned)eh.ether_type);
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
}
|
|
|
|
if(ntohs(eh.ether_type) > ETHERTYPE_TRAIL
|
|
|
|
&& ntohs(eh.ether_type) < (ETHERTYPE_TRAIL + ETHERTYPE_NTRAILER))
|
|
|
|
printf("received trailer!\n");
|
|
|
|
#endif
|
|
|
|
|
|
|
|
if(!m) return;
|
|
|
|
|
|
|
|
if(last_not_for_us) {
|
|
|
|
m_freem(last_not_for_us);
|
|
|
|
last_not_for_us = 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
#if NBPFILTER > 0
|
|
|
|
/*
|
|
|
|
* Check for a BPF filter; if so, hand it up.
|
|
|
|
* Note that we have to stick an extra mbuf up front, because
|
|
|
|
* bpf_mtap expects to have the ether header at the front.
|
|
|
|
* It doesn't matter that this results in an ill-formatted mbuf chain,
|
|
|
|
* since BPF just looks at the data. (It doesn't try to free the mbuf,
|
|
|
|
* tho' it will make a copy for tcpdump.)
|
|
|
|
*/
|
|
|
|
if(bpf_gets_it) {
|
|
|
|
struct mbuf m0;
|
|
|
|
m0.m_len = sizeof eh;
|
|
|
|
m0.m_data = (caddr_t)&eh;
|
|
|
|
m0.m_next = m;
|
|
|
|
|
|
|
|
/* Pass it up */
|
1996-02-06 18:51:28 +00:00
|
|
|
bpf_mtap(&ie->arpcom.ac_if, &m0);
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
}
|
|
|
|
/*
|
|
|
|
* A signal passed up from the filtering code indicating that the
|
|
|
|
* packet is intended for BPF but not for the protocol machinery.
|
|
|
|
* We can save a few cycles by not handing it off to them.
|
|
|
|
*/
|
|
|
|
if(bpf_gets_it == 2) {
|
|
|
|
last_not_for_us = m;
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
#endif /* NBPFILTER > 0 */
|
|
|
|
/*
|
|
|
|
* In here there used to be code to check destination addresses upon
|
|
|
|
* receipt of a packet. We have deleted that code, and replaced it
|
|
|
|
* with code to check the address much earlier in the cycle, before
|
|
|
|
* copying the data in; this saves us valuable cycles when operating
|
|
|
|
* as a multicast router or when using BPF.
|
|
|
|
*/
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Finally pass this packet up to higher layers.
|
|
|
|
*/
|
|
|
|
ether_input(&ie->arpcom.ac_if, &eh, m);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ie_drop_packet_buffer(int unit, struct ie_softc *ie) {
|
|
|
|
int i;
|
|
|
|
|
|
|
|
do {
|
|
|
|
/*
|
|
|
|
* This means we are somehow out of sync. So, we reset the
|
|
|
|
* adapter.
|
|
|
|
*/
|
|
|
|
if(!(ie->rbuffs[ie->rbhead]->ie_rbd_actual & IE_RBD_USED)) {
|
|
|
|
#ifdef DEBUG
|
|
|
|
print_rbd(ie->rbuffs[ie->rbhead]);
|
|
|
|
#endif
|
|
|
|
log(LOG_ERR, "ie%d: receive descriptors out of sync at %d\n",
|
|
|
|
unit, ie->rbhead);
|
1994-05-27 04:02:10 +00:00
|
|
|
iereset(unit);
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
i = ie->rbuffs[ie->rbhead]->ie_rbd_actual & IE_RBD_LAST;
|
|
|
|
|
|
|
|
ie->rbuffs[ie->rbhead]->ie_rbd_length |= IE_RBD_LAST;
|
|
|
|
ie->rbuffs[ie->rbhead]->ie_rbd_actual = 0;
|
|
|
|
ie->rbhead = (ie->rbhead + 1) % NBUFFS;
|
|
|
|
ie->rbuffs[ie->rbtail]->ie_rbd_length &= ~IE_RBD_LAST;
|
|
|
|
ie->rbtail = (ie->rbtail + 1) % NBUFFS;
|
|
|
|
} while(!i);
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Start transmission on an interface.
|
|
|
|
*/
|
1993-11-25 01:38:01 +00:00
|
|
|
static void
|
|
|
|
iestart(ifp)
|
|
|
|
struct ifnet *ifp;
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
{
|
1996-02-06 18:51:28 +00:00
|
|
|
struct ie_softc *ie = ifp->if_softc;
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
struct mbuf *m0, *m;
|
|
|
|
unsigned char *buffer;
|
|
|
|
u_short len;
|
|
|
|
/* This is not really volatile, in this routine, but it makes gcc happy. */
|
|
|
|
volatile u_short *bptr = &ie->scb->ie_command_list;
|
|
|
|
|
|
|
|
if(!(ifp->if_flags & IFF_RUNNING))
|
1993-11-25 01:38:01 +00:00
|
|
|
return;
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
if(ifp->if_flags & IFF_OACTIVE)
|
1993-11-25 01:38:01 +00:00
|
|
|
return;
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
|
|
|
|
do {
|
|
|
|
IF_DEQUEUE(&ie->arpcom.ac_if.if_snd, m);
|
|
|
|
if(!m)
|
|
|
|
break;
|
|
|
|
|
|
|
|
buffer = ie->xmit_cbuffs[ie->xmit_count];
|
|
|
|
len = 0;
|
|
|
|
|
|
|
|
for(m0 = m; m && len < IE_BUF_LEN; m = m->m_next) {
|
|
|
|
bcopy(mtod(m, caddr_t), buffer, m->m_len);
|
|
|
|
buffer += m->m_len;
|
|
|
|
len += m->m_len;
|
|
|
|
}
|
|
|
|
|
|
|
|
m_freem(m0);
|
1994-05-26 13:30:20 +00:00
|
|
|
len = max(len, ETHERMINLEN);
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
|
|
|
|
#if NBPFILTER > 0
|
|
|
|
/*
|
|
|
|
* See if bpf is listening on this interface, let it see the packet
|
|
|
|
* before we commit it to the wire.
|
|
|
|
*/
|
1996-02-06 18:51:28 +00:00
|
|
|
if(ie->arpcom.ac_if.if_bpf)
|
|
|
|
bpf_tap(&ie->arpcom.ac_if, ie->xmit_cbuffs[ie->xmit_count], len);
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
#endif
|
|
|
|
|
|
|
|
ie->xmit_buffs[ie->xmit_count]->ie_xmit_flags = IE_XMIT_LAST | len;
|
|
|
|
ie->xmit_buffs[ie->xmit_count]->ie_xmit_next = 0xffff;
|
1995-05-30 08:16:23 +00:00
|
|
|
ie->xmit_buffs[ie->xmit_count]->ie_xmit_buf =
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
MK_24(ie->iomem, ie->xmit_cbuffs[ie->xmit_count]);
|
|
|
|
|
|
|
|
ie->xmit_cmds[ie->xmit_count]->com.ie_cmd_cmd = IE_CMD_XMIT;
|
|
|
|
ie->xmit_cmds[ie->xmit_count]->ie_xmit_status = 0;
|
1995-05-30 08:16:23 +00:00
|
|
|
ie->xmit_cmds[ie->xmit_count]->ie_xmit_desc =
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
MK_16(ie->iomem, ie->xmit_buffs[ie->xmit_count]);
|
|
|
|
|
|
|
|
*bptr = MK_16(ie->iomem, ie->xmit_cmds[ie->xmit_count]);
|
|
|
|
bptr = &ie->xmit_cmds[ie->xmit_count]->com.ie_cmd_link;
|
|
|
|
ie->xmit_count++;
|
|
|
|
} while(ie->xmit_count < 2);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* If we queued up anything for transmission, send it.
|
|
|
|
*/
|
|
|
|
if(ie->xmit_count) {
|
1995-05-30 08:16:23 +00:00
|
|
|
ie->xmit_cmds[ie->xmit_count - 1]->com.ie_cmd_cmd |=
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
IE_CMD_LAST | IE_CMD_INTR;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* By passing the command pointer as a null, we tell
|
|
|
|
* command_and_wait() to pretend that this isn't an action
|
|
|
|
* command. I wish I understood what was happening here.
|
|
|
|
*/
|
|
|
|
command_and_wait(ifp->if_unit, IE_CU_START, 0, 0);
|
|
|
|
ifp->if_flags |= IFF_OACTIVE;
|
|
|
|
}
|
|
|
|
|
1993-11-25 01:38:01 +00:00
|
|
|
return;
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Check to see if there's an 82586 out there.
|
|
|
|
*/
|
1995-12-10 13:40:44 +00:00
|
|
|
static int
|
|
|
|
check_ie_present(unit, where, size)
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
int unit;
|
|
|
|
caddr_t where;
|
|
|
|
unsigned size;
|
|
|
|
{
|
|
|
|
volatile struct ie_sys_conf_ptr *scp;
|
|
|
|
volatile struct ie_int_sys_conf_ptr *iscp;
|
|
|
|
volatile struct ie_sys_ctl_block *scb;
|
|
|
|
u_long realbase;
|
|
|
|
int s;
|
|
|
|
|
|
|
|
s = splimp();
|
|
|
|
|
|
|
|
realbase = (u_long)where + size - (1 << 24);
|
|
|
|
|
|
|
|
scp = (volatile struct ie_sys_conf_ptr *)(realbase + IE_SCP_ADDR);
|
|
|
|
bzero((char *)scp, sizeof *scp); /* ignore cast-qual */
|
1995-05-30 08:16:23 +00:00
|
|
|
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
/*
|
|
|
|
* First we put the ISCP at the bottom of memory; this tests to make
|
|
|
|
* sure that our idea of the size of memory is the same as the controller's.
|
|
|
|
* This is NOT where the ISCP will be in normal operation.
|
|
|
|
*/
|
|
|
|
iscp = (volatile struct ie_int_sys_conf_ptr *)where;
|
|
|
|
bzero((char *)iscp, sizeof *iscp); /* ignore cast-qual */
|
|
|
|
|
|
|
|
scb = (volatile struct ie_sys_ctl_block *)where;
|
|
|
|
bzero((char *)scb, sizeof *scb); /* ignore cast-qual */
|
|
|
|
|
1994-09-05 22:28:31 +00:00
|
|
|
scp->ie_bus_use = ie_softc[unit].bus_use; /* 8-bit or 16-bit */
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
scp->ie_iscp_ptr = (caddr_t)((volatile caddr_t)iscp - /* ignore cast-qual */
|
|
|
|
(volatile caddr_t)realbase);
|
|
|
|
|
|
|
|
iscp->ie_busy = 1;
|
|
|
|
iscp->ie_scb_offset = MK_16(realbase, scb) + 256;
|
|
|
|
|
|
|
|
(*ie_softc[unit].ie_reset_586)(unit);
|
|
|
|
(*ie_softc[unit].ie_chan_attn)(unit);
|
|
|
|
|
|
|
|
DELAY(100); /* wait a while... */
|
|
|
|
|
|
|
|
if(iscp->ie_busy) {
|
|
|
|
splx(s);
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Now relocate the ISCP to its real home, and reset the controller
|
|
|
|
* again.
|
|
|
|
*/
|
1995-05-30 08:16:23 +00:00
|
|
|
iscp = (void *)Align((caddr_t)(realbase + IE_SCP_ADDR -
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
sizeof(struct ie_int_sys_conf_ptr)));
|
|
|
|
bzero((char *)iscp, sizeof *iscp); /* ignore cast-qual */
|
|
|
|
|
|
|
|
scp->ie_iscp_ptr = (caddr_t)((caddr_t)iscp - (caddr_t)realbase);
|
|
|
|
/* ignore cast-qual */
|
|
|
|
|
|
|
|
iscp->ie_busy = 1;
|
|
|
|
iscp->ie_scb_offset = MK_16(realbase, scb);
|
|
|
|
|
|
|
|
(*ie_softc[unit].ie_reset_586)(unit);
|
|
|
|
(*ie_softc[unit].ie_chan_attn)(unit);
|
|
|
|
|
|
|
|
DELAY(100);
|
|
|
|
|
|
|
|
if(iscp->ie_busy) {
|
|
|
|
splx(s);
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
ie_softc[unit].iosize = size;
|
|
|
|
ie_softc[unit].iomem = (caddr_t)realbase;
|
|
|
|
|
|
|
|
ie_softc[unit].iscp = iscp;
|
|
|
|
ie_softc[unit].scb = scb;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Acknowledge any interrupts we may have caused...
|
|
|
|
*/
|
|
|
|
ie_ack(scb, IE_ST_WHENCE, unit, ie_softc[unit].ie_chan_attn);
|
|
|
|
splx(s);
|
|
|
|
|
|
|
|
return 1;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Divine the memory size of ie board UNIT.
|
|
|
|
* Better hope there's nothing important hiding just below the ie card...
|
|
|
|
*/
|
|
|
|
static void find_ie_mem_size(unit)
|
|
|
|
int unit;
|
|
|
|
{
|
|
|
|
unsigned size;
|
|
|
|
|
|
|
|
ie_softc[unit].iosize = 0;
|
|
|
|
|
1994-09-05 22:28:31 +00:00
|
|
|
for(size = 65536; size >= 8192; size -= 8192) {
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
if(check_ie_present(unit, ie_softc[unit].iomembot, size)) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
1994-08-24 22:32:44 +00:00
|
|
|
void el_reset_586(unit)
|
|
|
|
int unit;
|
|
|
|
{
|
1994-08-25 20:16:59 +00:00
|
|
|
outb(PORT + IE507_CTRL, EL_CTRL_RESET);
|
|
|
|
DELAY(100);
|
|
|
|
outb(PORT + IE507_CTRL, EL_CTRL_NORMAL);
|
|
|
|
DELAY(100);
|
1994-08-24 22:32:44 +00:00
|
|
|
}
|
|
|
|
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
void sl_reset_586(unit)
|
|
|
|
int unit;
|
|
|
|
{
|
|
|
|
outb(PORT + IEATT_RESET, 0);
|
|
|
|
}
|
|
|
|
|
1994-08-24 22:32:44 +00:00
|
|
|
void el_chan_attn(unit)
|
1994-08-25 20:16:59 +00:00
|
|
|
int unit;
|
1994-08-24 22:32:44 +00:00
|
|
|
{
|
1994-08-25 20:16:59 +00:00
|
|
|
outb(PORT + IE507_ATTN, 1);
|
1994-08-24 22:32:44 +00:00
|
|
|
}
|
|
|
|
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
void sl_chan_attn(unit)
|
|
|
|
int unit;
|
|
|
|
{
|
|
|
|
outb(PORT + IEATT_ATTN, 0);
|
|
|
|
}
|
|
|
|
|
|
|
|
void sl_read_ether(unit, addr)
|
|
|
|
int unit;
|
|
|
|
unsigned char addr[6];
|
|
|
|
{
|
|
|
|
int i;
|
|
|
|
|
|
|
|
for(i = 0; i < 6; i++)
|
|
|
|
addr[i] = inb(PORT + i);
|
|
|
|
}
|
|
|
|
|
|
|
|
|
1993-11-25 01:38:01 +00:00
|
|
|
static void
|
1994-05-27 04:02:10 +00:00
|
|
|
iereset(unit)
|
|
|
|
int unit;
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
{
|
|
|
|
int s = splimp();
|
|
|
|
|
|
|
|
if(unit >= NIE) {
|
|
|
|
splx(s);
|
1993-11-25 01:38:01 +00:00
|
|
|
return;
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
printf("ie%d: reset\n", unit);
|
|
|
|
ie_softc[unit].arpcom.ac_if.if_flags &= ~IFF_UP;
|
|
|
|
ieioctl(&ie_softc[unit].arpcom.ac_if, SIOCSIFFLAGS, 0);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Stop i82586 dead in its tracks.
|
|
|
|
*/
|
|
|
|
if(command_and_wait(unit, IE_RU_ABORT | IE_CU_ABORT, 0, 0))
|
|
|
|
printf("ie%d: abort commands timed out\n", unit);
|
|
|
|
|
|
|
|
if(command_and_wait(unit, IE_RU_DISABLE | IE_CU_STOP, 0, 0))
|
|
|
|
printf("ie%d: disable commands timed out\n", unit);
|
|
|
|
|
|
|
|
#ifdef notdef
|
|
|
|
if(!check_ie_present(unit, ie_softc[unit].iomembot, ie_softc[unit].iosize))
|
1995-03-19 14:29:26 +00:00
|
|
|
panic("ie disappeared!");
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
#endif
|
|
|
|
|
|
|
|
ie_softc[unit].arpcom.ac_if.if_flags |= IFF_UP;
|
|
|
|
ieioctl(&ie_softc[unit].arpcom.ac_if, SIOCSIFFLAGS, 0);
|
1995-05-30 08:16:23 +00:00
|
|
|
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
splx(s);
|
1993-11-25 01:38:01 +00:00
|
|
|
return;
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* This is called if we time out.
|
|
|
|
*/
|
1993-11-25 01:38:01 +00:00
|
|
|
static void
|
1994-05-25 09:21:21 +00:00
|
|
|
chan_attn_timeout(rock)
|
1995-11-18 08:35:49 +00:00
|
|
|
void *rock;
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
{
|
|
|
|
*(int *)rock = 1;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Send a command to the controller and wait for it to either
|
|
|
|
* complete or be accepted, depending on the command. If the
|
|
|
|
* command pointer is null, then pretend that the command is
|
|
|
|
* not an action command. If the command pointer is not null,
|
|
|
|
* and the command is an action command, wait for
|
|
|
|
* ((volatile struct ie_cmd_common *)pcmd)->ie_cmd_status & MASK
|
|
|
|
* to become true.
|
|
|
|
*/
|
|
|
|
static int command_and_wait(unit, cmd, pcmd, mask)
|
|
|
|
int unit;
|
|
|
|
int cmd;
|
|
|
|
volatile void *pcmd;
|
|
|
|
int mask;
|
|
|
|
{
|
|
|
|
volatile struct ie_cmd_common *cc = pcmd;
|
|
|
|
volatile int timedout = 0;
|
|
|
|
|
|
|
|
ie_softc[unit].scb->ie_command = (u_short)cmd;
|
1995-05-30 08:16:23 +00:00
|
|
|
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
if(IE_ACTION_COMMAND(cmd) && pcmd) {
|
|
|
|
(*ie_softc[unit].ie_chan_attn)(unit);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* According to the packet driver, the minimum timeout should be
|
|
|
|
* .369 seconds, which we round up to .37.
|
|
|
|
*/
|
|
|
|
timeout(chan_attn_timeout, (caddr_t)&timedout, 37 * hz / 100);
|
|
|
|
/* ignore cast-qual */
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Now spin-lock waiting for status. This is not a very nice
|
|
|
|
* thing to do, but I haven't figured out how, or indeed if, we
|
|
|
|
* can put the process waiting for action to sleep. (We may
|
|
|
|
* be getting called through some other timeout running in the
|
|
|
|
* kernel.)
|
|
|
|
*/
|
|
|
|
while(1) {
|
|
|
|
if((cc->ie_cmd_status & mask) || timedout)
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
untimeout(chan_attn_timeout, (caddr_t)&timedout);
|
|
|
|
/* ignore cast-qual */
|
|
|
|
|
|
|
|
return timedout;
|
|
|
|
} else {
|
1995-05-30 08:16:23 +00:00
|
|
|
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
/*
|
|
|
|
* Otherwise, just wait for the command to be accepted.
|
|
|
|
*/
|
|
|
|
(*ie_softc[unit].ie_chan_attn)(unit);
|
|
|
|
|
|
|
|
while(ie_softc[unit].scb->ie_command)
|
|
|
|
; /* spin lock */
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Run the time-domain reflectometer...
|
|
|
|
*/
|
|
|
|
static void run_tdr(unit, cmd)
|
|
|
|
int unit;
|
|
|
|
struct ie_tdr_cmd *cmd;
|
|
|
|
{
|
|
|
|
int result;
|
|
|
|
|
|
|
|
cmd->com.ie_cmd_status = 0;
|
|
|
|
cmd->com.ie_cmd_cmd = IE_CMD_TDR | IE_CMD_LAST;
|
|
|
|
cmd->com.ie_cmd_link = 0xffff;
|
|
|
|
cmd->ie_tdr_time = 0;
|
|
|
|
|
|
|
|
ie_softc[unit].scb->ie_command_list = MK_16(MEM, cmd);
|
|
|
|
cmd->ie_tdr_time = 0;
|
|
|
|
|
|
|
|
if(command_and_wait(unit, IE_CU_START, cmd, IE_STAT_COMPL))
|
|
|
|
result = 0x2000;
|
|
|
|
else
|
|
|
|
result = cmd->ie_tdr_time;
|
|
|
|
|
1995-05-30 08:16:23 +00:00
|
|
|
ie_ack(ie_softc[unit].scb, IE_ST_WHENCE, unit,
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
ie_softc[unit].ie_chan_attn);
|
|
|
|
|
|
|
|
if(result & IE_TDR_SUCCESS)
|
|
|
|
return;
|
|
|
|
|
|
|
|
if(result & IE_TDR_XCVR) {
|
|
|
|
printf("ie%d: transceiver problem\n", unit);
|
|
|
|
} else if(result & IE_TDR_OPEN) {
|
1995-05-30 08:16:23 +00:00
|
|
|
printf("ie%d: TDR detected an open %d clocks away\n", unit,
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
result & IE_TDR_TIME);
|
|
|
|
} else if(result & IE_TDR_SHORT) {
|
|
|
|
printf("ie%d: TDR detected a short %d clocks away\n", unit,
|
|
|
|
result & IE_TDR_TIME);
|
|
|
|
} else {
|
1994-10-26 00:16:20 +00:00
|
|
|
printf("ie%d: TDR returned unknown status %x\n", unit, result);
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static void start_receiver(unit)
|
|
|
|
int unit;
|
|
|
|
{
|
|
|
|
int s = splimp();
|
|
|
|
|
|
|
|
ie_softc[unit].scb->ie_recv_list = MK_16(MEM, ie_softc[unit].rframes[0]);
|
|
|
|
command_and_wait(unit, IE_RU_START, 0, 0);
|
|
|
|
|
|
|
|
ie_ack(ie_softc[unit].scb, IE_ST_WHENCE, unit, ie_softc[unit].ie_chan_attn);
|
|
|
|
|
|
|
|
splx(s);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Here is a helper routine for iernr() and ieinit(). This sets up
|
|
|
|
* the RFA.
|
|
|
|
*/
|
|
|
|
static caddr_t setup_rfa(caddr_t ptr, struct ie_softc *ie) {
|
|
|
|
volatile struct ie_recv_frame_desc *rfd = (void *)ptr;
|
|
|
|
volatile struct ie_recv_buf_desc *rbd;
|
|
|
|
int i;
|
|
|
|
int unit = ie - &ie_softc[0];
|
|
|
|
|
|
|
|
/* First lay them out */
|
|
|
|
for(i = 0; i < NFRAMES; i++) {
|
|
|
|
ie->rframes[i] = rfd;
|
|
|
|
bzero((char *)rfd, sizeof *rfd); /* ignore cast-qual */
|
|
|
|
rfd++;
|
|
|
|
}
|
|
|
|
|
|
|
|
ptr = (caddr_t)Align((caddr_t)rfd); /* ignore cast-qual */
|
1995-05-30 08:16:23 +00:00
|
|
|
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
/* Now link them together */
|
|
|
|
for(i = 0; i < NFRAMES; i++) {
|
|
|
|
ie->rframes[i]->ie_fd_next =
|
|
|
|
MK_16(MEM, ie->rframes[(i + 1) % NFRAMES]);
|
|
|
|
}
|
1995-05-30 08:16:23 +00:00
|
|
|
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
/* Finally, set the EOL bit on the last one. */
|
|
|
|
ie->rframes[NFRAMES - 1]->ie_fd_last |= IE_FD_LAST;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Now lay out some buffers for the incoming frames. Note that
|
|
|
|
* we set aside a bit of slop in each buffer, to make sure that
|
|
|
|
* we have enough space to hold a single frame in every buffer.
|
|
|
|
*/
|
|
|
|
rbd = (void *)ptr;
|
|
|
|
|
|
|
|
for(i = 0; i < NBUFFS; i++) {
|
|
|
|
ie->rbuffs[i] = rbd;
|
|
|
|
bzero((char *)rbd, sizeof *rbd); /* ignore cast-qual */
|
|
|
|
ptr = (caddr_t)Align(ptr + sizeof *rbd);
|
|
|
|
rbd->ie_rbd_length = IE_RBUF_SIZE;
|
|
|
|
rbd->ie_rbd_buffer = MK_24(MEM, ptr);
|
|
|
|
ie->cbuffs[i] = (void *)ptr;
|
|
|
|
ptr += IE_RBUF_SIZE;
|
|
|
|
rbd = (void *)ptr;
|
|
|
|
}
|
1995-05-30 08:16:23 +00:00
|
|
|
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
/* Now link them together */
|
|
|
|
for(i = 0; i < NBUFFS; i++) {
|
|
|
|
ie->rbuffs[i]->ie_rbd_next = MK_16(MEM, ie->rbuffs[(i + 1) % NBUFFS]);
|
|
|
|
}
|
1995-05-30 08:16:23 +00:00
|
|
|
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
/* Tag EOF on the last one */
|
|
|
|
ie->rbuffs[NBUFFS - 1]->ie_rbd_length |= IE_RBD_LAST;
|
|
|
|
|
|
|
|
/* We use the head and tail pointers on receive to keep track of
|
|
|
|
* the order in which RFDs and RBDs are used. */
|
|
|
|
ie->rfhead = 0;
|
|
|
|
ie->rftail = NFRAMES - 1;
|
|
|
|
ie->rbhead = 0;
|
|
|
|
ie->rbtail = NBUFFS - 1;
|
|
|
|
|
|
|
|
ie->scb->ie_recv_list = MK_16(MEM, ie->rframes[0]);
|
|
|
|
ie->rframes[0]->ie_fd_buf_desc = MK_16(MEM, ie->rbuffs[0]);
|
|
|
|
|
|
|
|
ptr = Align(ptr);
|
|
|
|
return ptr;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Run the multicast setup command.
|
|
|
|
* Call at splimp().
|
|
|
|
*/
|
1995-05-30 08:16:23 +00:00
|
|
|
static int mc_setup(int unit, caddr_t ptr,
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
volatile struct ie_sys_ctl_block *scb) {
|
|
|
|
struct ie_softc *ie = &ie_softc[unit];
|
|
|
|
volatile struct ie_mcast_cmd *cmd = (void *)ptr;
|
1995-05-30 08:16:23 +00:00
|
|
|
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
cmd->com.ie_cmd_status = 0;
|
|
|
|
cmd->com.ie_cmd_cmd = IE_CMD_MCAST | IE_CMD_LAST;
|
|
|
|
cmd->com.ie_cmd_link = 0xffff;
|
1995-05-30 08:16:23 +00:00
|
|
|
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
/* ignore cast-qual */
|
|
|
|
bcopy((caddr_t)ie->mcast_addrs, (caddr_t)cmd->ie_mcast_addrs,
|
|
|
|
ie->mcast_count * sizeof *ie->mcast_addrs);
|
|
|
|
|
|
|
|
cmd->ie_mcast_bytes = ie->mcast_count * 6; /* grrr... */
|
1995-05-30 08:16:23 +00:00
|
|
|
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
scb->ie_command_list = MK_16(MEM, cmd);
|
|
|
|
if(command_and_wait(unit, IE_CU_START, cmd, IE_STAT_COMPL)
|
|
|
|
|| !(cmd->com.ie_cmd_status & IE_STAT_OK)) {
|
|
|
|
printf("ie%d: multicast address setup command failed\n", unit);
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
return 1;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* This routine takes the environment generated by check_ie_present()
|
|
|
|
* and adds to it all the other structures we need to operate the adapter.
|
|
|
|
* This includes executing the CONFIGURE, IA-SETUP, and MC-SETUP commands,
|
|
|
|
* starting the receiver unit, and clearing interrupts.
|
|
|
|
*
|
|
|
|
* THIS ROUTINE MUST BE CALLED AT splimp() OR HIGHER.
|
|
|
|
*/
|
1993-11-25 01:38:01 +00:00
|
|
|
static void
|
|
|
|
ieinit(unit)
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
int unit;
|
|
|
|
{
|
|
|
|
struct ie_softc *ie = &ie_softc[unit];
|
|
|
|
volatile struct ie_sys_ctl_block *scb = ie->scb;
|
|
|
|
caddr_t ptr;
|
|
|
|
|
|
|
|
ptr = (caddr_t)Align((caddr_t)scb + sizeof *scb); /* ignore cast-qual */
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Send the configure command first.
|
|
|
|
*/
|
|
|
|
{
|
|
|
|
volatile struct ie_config_cmd *cmd = (void *)ptr;
|
|
|
|
|
|
|
|
ie_setup_config(cmd, ie->promisc, ie->hard_type == IE_STARLAN10);
|
|
|
|
cmd->com.ie_cmd_status = 0;
|
|
|
|
cmd->com.ie_cmd_cmd = IE_CMD_CONFIG | IE_CMD_LAST;
|
|
|
|
cmd->com.ie_cmd_link = 0xffff;
|
|
|
|
|
|
|
|
scb->ie_command_list = MK_16(MEM, cmd);
|
|
|
|
|
|
|
|
if(command_and_wait(unit, IE_CU_START, cmd, IE_STAT_COMPL)
|
|
|
|
|| !(cmd->com.ie_cmd_status & IE_STAT_OK)) {
|
|
|
|
printf("ie%d: configure command failed\n", unit);
|
1993-11-25 01:38:01 +00:00
|
|
|
return;
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
/*
|
|
|
|
* Now send the Individual Address Setup command.
|
|
|
|
*/
|
|
|
|
{
|
|
|
|
volatile struct ie_iasetup_cmd *cmd = (void *)ptr;
|
|
|
|
|
|
|
|
cmd->com.ie_cmd_status = 0;
|
|
|
|
cmd->com.ie_cmd_cmd = IE_CMD_IASETUP | IE_CMD_LAST;
|
|
|
|
cmd->com.ie_cmd_link = 0xffff;
|
|
|
|
|
|
|
|
bcopy((char *)ie_softc[unit].arpcom.ac_enaddr, (char *)&cmd->ie_address,
|
|
|
|
sizeof cmd->ie_address); /* ignore cast-qual */
|
|
|
|
|
|
|
|
scb->ie_command_list = MK_16(MEM, cmd);
|
|
|
|
if(command_and_wait(unit, IE_CU_START, cmd, IE_STAT_COMPL)
|
|
|
|
|| !(cmd->com.ie_cmd_status & IE_STAT_OK)) {
|
|
|
|
printf("ie%d: individual address setup command failed\n", unit);
|
1993-11-25 01:38:01 +00:00
|
|
|
return;
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Now run the time-domain reflectometer.
|
|
|
|
*/
|
|
|
|
run_tdr(unit, (void *)ptr);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Acknowledge any interrupts we have generated thus far.
|
|
|
|
*/
|
|
|
|
ie_ack(ie->scb, IE_ST_WHENCE, unit, ie->ie_chan_attn);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Set up the RFA.
|
|
|
|
*/
|
|
|
|
ptr = setup_rfa(ptr, ie);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Finally, the transmit command and buffer are the last little bit of work.
|
|
|
|
*/
|
|
|
|
ie->xmit_cmds[0] = (void *)ptr;
|
|
|
|
ptr += sizeof *ie->xmit_cmds[0];
|
|
|
|
ptr = Align(ptr);
|
|
|
|
ie->xmit_buffs[0] = (void *)ptr;
|
|
|
|
ptr += sizeof *ie->xmit_buffs[0];
|
|
|
|
ptr = Align(ptr);
|
|
|
|
|
|
|
|
/* Second transmit command */
|
|
|
|
ie->xmit_cmds[1] = (void *)ptr;
|
|
|
|
ptr += sizeof *ie->xmit_cmds[1];
|
|
|
|
ptr = Align(ptr);
|
|
|
|
ie->xmit_buffs[1] = (void *)ptr;
|
|
|
|
ptr += sizeof *ie->xmit_buffs[1];
|
|
|
|
ptr = Align(ptr);
|
|
|
|
|
|
|
|
/* Both transmit buffers */
|
|
|
|
ie->xmit_cbuffs[0] = (void *)ptr;
|
|
|
|
ptr += IE_BUF_LEN;
|
|
|
|
ptr = Align(ptr);
|
|
|
|
ie->xmit_cbuffs[1] = (void *)ptr;
|
|
|
|
|
|
|
|
bzero((caddr_t)ie->xmit_cmds[0], sizeof *ie->xmit_cmds[0]); /* ignore */
|
|
|
|
bzero((caddr_t)ie->xmit_buffs[0], sizeof *ie->xmit_buffs[0]); /* cast-qual */
|
|
|
|
bzero((caddr_t)ie->xmit_cmds[1], sizeof *ie->xmit_cmds[0]); /* warnings */
|
|
|
|
bzero((caddr_t)ie->xmit_buffs[1], sizeof *ie->xmit_buffs[0]); /* here */
|
|
|
|
|
|
|
|
/*
|
|
|
|
* This must be coordinated with iestart() and ietint().
|
|
|
|
*/
|
|
|
|
ie->xmit_cmds[0]->ie_xmit_status = IE_STAT_COMPL;
|
|
|
|
|
|
|
|
ie->arpcom.ac_if.if_flags |= IFF_RUNNING; /* tell higher levels that we are here */
|
|
|
|
start_receiver(unit);
|
1993-11-25 01:38:01 +00:00
|
|
|
return;
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static void ie_stop(unit)
|
|
|
|
int unit;
|
|
|
|
{
|
|
|
|
command_and_wait(unit, IE_RU_DISABLE, 0, 0);
|
|
|
|
}
|
|
|
|
|
1993-11-25 01:38:01 +00:00
|
|
|
static int
|
|
|
|
ieioctl(ifp, command, data)
|
|
|
|
struct ifnet *ifp;
|
|
|
|
int command;
|
|
|
|
caddr_t data;
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
{
|
|
|
|
struct ifaddr *ifa = (struct ifaddr *)data;
|
1996-02-06 18:51:28 +00:00
|
|
|
struct ie_softc *ie = ifp->if_softc;
|
1994-08-12 06:06:19 +00:00
|
|
|
struct ifreq *ifr = (struct ifreq *) data;
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
int s, error = 0;
|
|
|
|
|
|
|
|
s = splimp();
|
|
|
|
|
|
|
|
switch(command) {
|
|
|
|
case SIOCSIFADDR:
|
|
|
|
ifp->if_flags |= IFF_UP;
|
|
|
|
|
|
|
|
switch(ifa->ifa_addr->sa_family) {
|
|
|
|
#ifdef INET
|
|
|
|
case AF_INET:
|
|
|
|
ieinit(ifp->if_unit);
|
1994-12-22 21:56:22 +00:00
|
|
|
arp_ifinit((struct arpcom *)ifp, ifa);
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
break;
|
|
|
|
#endif /* INET */
|
|
|
|
|
1995-10-26 20:31:59 +00:00
|
|
|
#ifdef IPX
|
|
|
|
/* This magic copied from if_is.c; I don't use XNS, so I have no
|
|
|
|
* way of telling if this actually works or not.
|
|
|
|
*/
|
|
|
|
case AF_IPX:
|
|
|
|
{
|
|
|
|
struct ipx_addr *ina = &(IA_SIPX(ifa)->sipx_addr);
|
|
|
|
|
|
|
|
if(ipx_nullhost(*ina)) {
|
|
|
|
ina->x_host = *(union ipx_host *)(ie->arpcom.ac_enaddr);
|
|
|
|
} else {
|
|
|
|
ifp->if_flags &= ~IFF_RUNNING;
|
|
|
|
bcopy((caddr_t)ina->x_host.c_host,
|
|
|
|
(caddr_t)ie->arpcom.ac_enaddr,
|
|
|
|
sizeof ie->arpcom.ac_enaddr);
|
|
|
|
}
|
|
|
|
|
|
|
|
ieinit(ifp->if_unit);
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
#endif /* IPX */
|
|
|
|
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
#ifdef NS
|
|
|
|
/* This magic copied from if_is.c; I don't use XNS, so I have no
|
|
|
|
* way of telling if this actually works or not.
|
|
|
|
*/
|
|
|
|
case AF_NS:
|
|
|
|
{
|
|
|
|
struct ns_addr *ina = &(IA_SNS(ifa)->sns_addr);
|
|
|
|
|
|
|
|
if(ns_nullhost(*ina)) {
|
|
|
|
ina->x_host = *(union ns_host *)(ie->arpcom.ac_enaddr);
|
|
|
|
} else {
|
|
|
|
ifp->if_flags &= ~IFF_RUNNING;
|
|
|
|
bcopy((caddr_t)ina->x_host.c_host,
|
|
|
|
(caddr_t)ie->arpcom.ac_enaddr,
|
|
|
|
sizeof ie->arpcom.ac_enaddr);
|
|
|
|
}
|
|
|
|
|
|
|
|
ieinit(ifp->if_unit);
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
#endif /* NS */
|
|
|
|
|
|
|
|
default:
|
|
|
|
ieinit(ifp->if_unit);
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
|
|
|
|
case SIOCSIFFLAGS:
|
|
|
|
/*
|
|
|
|
* Note that this device doesn't have an "all multicast" mode, so we
|
|
|
|
* must turn on promiscuous mode and do the filtering manually.
|
|
|
|
*/
|
|
|
|
if((ifp->if_flags & IFF_UP) == 0 &&
|
|
|
|
(ifp->if_flags & IFF_RUNNING)) {
|
|
|
|
ifp->if_flags &= ~IFF_RUNNING;
|
|
|
|
ie_stop(ifp->if_unit);
|
|
|
|
} else if((ifp->if_flags & IFF_UP) &&
|
|
|
|
(ifp->if_flags & IFF_RUNNING) == 0) {
|
1995-05-30 08:16:23 +00:00
|
|
|
ie_softc[ifp->if_unit].promisc =
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
ifp->if_flags & (IFF_PROMISC | IFF_ALLMULTI);
|
|
|
|
ieinit(ifp->if_unit);
|
|
|
|
} else if(ie_softc[ifp->if_unit].promisc ^
|
|
|
|
(ifp->if_flags & (IFF_PROMISC | IFF_ALLMULTI))) {
|
|
|
|
ie_softc[ifp->if_unit].promisc =
|
|
|
|
ifp->if_flags & (IFF_PROMISC | IFF_ALLMULTI);
|
|
|
|
ieinit(ifp->if_unit);
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
|
|
|
|
case SIOCADDMULTI:
|
|
|
|
case SIOCDELMULTI:
|
|
|
|
/*
|
|
|
|
* Update multicast listeners
|
|
|
|
*/
|
|
|
|
error = ((command == SIOCADDMULTI)
|
1994-08-12 06:06:19 +00:00
|
|
|
? ether_addmulti(ifr, &ie->arpcom)
|
|
|
|
: ether_delmulti(ifr, &ie->arpcom));
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
|
|
|
|
if(error == ENETRESET) {
|
|
|
|
/* reset multicast filtering */
|
|
|
|
ie_mc_reset(ifp->if_unit);
|
|
|
|
error = 0;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
|
1994-08-08 13:33:16 +00:00
|
|
|
case SIOCSIFMTU:
|
|
|
|
/*
|
|
|
|
* Set the interface MTU.
|
|
|
|
*/
|
|
|
|
if (ifr->ifr_mtu > ETHERMTU) {
|
|
|
|
error = EINVAL;
|
|
|
|
} else {
|
|
|
|
ifp->if_mtu = ifr->ifr_mtu;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
|
> Gcc2 generates memcpy calls when doing some structure copying in the
> code. As this memcpy generation is done by the compiler it doesn't
> work to have memcpy a macro. Instead I made it a static function that
> calls bcopy.
Okay, I tracked down the d*mn thing that is causing the memcpy to be
emmitted by gcc2. In the following patch is a fix to eliminate the
one place that was causing gcc to do the memcpy in function ie_readframe
for the ie_recv_frame_desc rfd = *(ie->rframes[num]) structure copy.
I also fixed the code so that the NS (XNS) support would compile with
out error and removed the cruft #define ns_* stuff from the softc
structures (I did this a while back for all the other either drivers)
Removed the $Log$ and all of the old log stuff, that is all kept in
the cvs tree, no need to duplicate it in the file... I'll copy it
into my commit message so that it is all retained, except the
revision numbers are going to get screwed up.
It now compiles cleanly under my LINT kernel...
> I have not yet tested this on the AT&T fiver cards yet. But, Wollman
> has used it on various other AT&T cards. I'm pretty sure Terry also
> uses this driver.
$Log$ stuff from Garretts original file
* $Log$
* Revision 1.32 1993/06/20 15:56:15 wollman
* Fixed warning message.
*
* Revision 1.31 93/06/20 15:52:28 wollman
* More compilation lint.
*
* Revision 1.30 93/06/20 15:48:16 wollman
* Commented all the rest of the cast-qual warnings. Might be
* fixed someday by changing the prototypes of bcopy(), bzero(), etc.
*
* Revision 1.29 93/06/20 15:40:13 wollman
* Started fixing some of the cast-qual warnings (most just commented).
*
* Revision 1.28 93/06/12 20:39:14 wollman
* Changed definition of MK_24 to silence -Wcast-qual.
*
* Revision 1.27 93/06/08 17:05:57 wollman
* Attach routine now returns void.
*
* Revision 1.26 93/06/08 16:31:21 wollman
* Deleted attempt to set watchdog timer, since we don't use it.
*
* Revision 1.25 93/04/16 22:15:31 wollman
* Fixed infinite loop in iestart() which would cause the transmitter
* to constantly step on the Ethernet, sending the same two packets over
* and over again. This crashes the router (really!) and makes other
* network users very angry.
*
* Revision 1.24 93/04/14 23:57:47 wollman
* Increased transmit buffer size from 1504 to 1512, just in case I don't
* understand how the MTU is used.
*
* Revision 1.23 93/04/14 23:16:19 wollman
* Fix iestart() so that it only tries to execute a command if one was
* queued up (oops...).
*
* Revision 1.22 93/04/14 23:02:16 wollman
* Unknown changes.
*
* Revision 1.21 93/04/12 18:29:33 wollman
* Added multiple transmit commands. Note that they are always queued in
* pairs (there are only two to begin with), and the I flag is only set
* on the second one. This should decrease interrupt activity somewhat.
*
* Revision 1.20 93/04/12 15:36:05 wollman
* Changed setup of config block to take into account difference between
* StarLAN (10BASE-T) and Ethernet: StarLAN is Manchester, Ethernet is NRZ.
*
* Revision 1.19 93/03/23 18:58:44 wollman
* Raised count of receive buffers from 32 to 48 and included analysis of
* memory usage in a comment near that definition.
*
* Revision 1.18 93/03/18 03:59:53 wollman
* Last try to get large TCP windows to work...
*
* Revision 1.17 93/03/18 03:54:10 wollman
* Added RNR processing as a part of receive interrupts. Hopefully this
* will make everything Just Work now.
*
* Revision 1.16 93/03/18 03:23:24 wollman
* Still having problems with RNR conditions. Basic problem: say we
* set TCP window to 16k, which is a semsible value. Then smart
* partners will blast 16k to us, of which we can fit 8k minus a bit
* into our buffers. (Need to look at that memory map..) So, we get
* an RNR interrupt. The problem is, what do we need to do in this case?
* The original code doesn't work, but it doesn't hang the system either.
*
* Revision 1.15 93/03/18 02:37:00 wollman
* Disabled watchdog timer. This gets called from a timeout somewhere,
* which is running up at softclock. This is at a higher priority than
* we are, so we can get interrupted in the process of receiving something
* by the timer expiration, which is not a nice thing to have happen.
* Question: how can we keep this functionality, but do it in sync with
* our interrupts? (Ugh!)
*
* Revision 1.14 93/03/18 02:35:46 wollman
* Fixed some more problems with receive code---it works!
*
* Revision 1.13 93/03/17 20:15:59 wollman
* Think we fixed the serious (showstopper) bugs in 1.12.
*
* Revision 1.12 93/03/15 17:45:39 wollman
* Changed filtering logic dramatically to avoid the unnecessary copyin
* for BPF or multicast routing when the packet is not for us. (We
* have to run in promiscuous mode in this case.) We still need to
* deal with the problem of allocating one mbuf per receive buffer;
* we should be able to get away with a single mbuf for all reasonable
* incoming packets, and using an mbuf cluster will help avoid unnecessary
* copying of data.
*
* Revision 1.11 93/01/30 20:23:15 wollman
* Fiddled with reset code to hopefully make it work better when the
* watchdog timer expires. Question: why does it make a difference that
* we do the same thing through ieioctl() as we were doing before. The
* old code, when iereset() was called, the receive unit would completely
* die. But, a cycle of ifconfig ie0 down ; ifconfig ie0 up' would bring
* things back to normal. So, we try to mimic that now. Grrr...
*
* Revision 1.10 93/01/29 17:26:50 wollman
* Finally fixed multicast routing/BPF mb_map full problem!!!!
* It turned out that, when we received a packet not for us, we
* would just return from ie_readframe(), *WITHOUT* freeing the
* mbuf that the packet was in. So, when running in promiscuous
* mode (i.e., multicast routing or BPF without -p), we would have
* (on our extremely busy subnet 4) a memory leak of epic proportions.
*
* We now hold on to a single mbuf chain allocated in this way,
* and the next time a packet is read, we free it. Also, the
* next time a packet is transmitted, we free it. This way,
* we never lose more than 6 mbufs total (and on average more like
* 3).
*
* Revision 1.9 93/01/23 14:22:21 wollman
* Added some StarLAN Fiber modifications from L. Jonas Olsson.
*
* Revision 1.8 93/01/23 14:02:34 wollman
* Added support for IP multicasting.
*
* Revision 1.7 92/11/21 17:51:59 wollman
* Added StarLAN Fiber version point to my driver.
*
* Revision 1.6 92/11/20 18:21:45 wollman
* Fixed documentation, made out-of-mbuf condition an error.
*
* Revision 1.5 92/11/20 18:14:02 wollman
* Move change of ethertype to host byte order so that it's after
* bpf_mtap. Hopefully, this will allow us to listen to incoming
* IP packets.
*
* Revision 1.4 92/11/20 18:11:00 wollman
* Added code to drop packets when we run out of mbufs rather than
* panicking. This should alleviate the problems observed when using tcpdump.
*
* Revision 1.3 92/11/14 14:12:40 wollman
* Added BPF support code.
*
* Revision 1.2 92/11/14 13:30:26 wollman
* Added RCS keywords
*
1993-10-12 06:52:34 +00:00
|
|
|
default:
|
|
|
|
error = EINVAL;
|
|
|
|
}
|
|
|
|
|
|
|
|
splx(s);
|
|
|
|
return error;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ie_mc_reset(int unit) {
|
|
|
|
struct ie_softc *ie = &ie_softc[unit];
|
|
|
|
struct ether_multi *enm;
|
|
|
|
struct ether_multistep step;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Step through the list of addresses.
|
|
|
|
*/
|
|
|
|
ie->mcast_count = 0;
|
|
|
|
ETHER_FIRST_MULTI(step, &ie->arpcom, enm);
|
|
|
|
while(enm) {
|
|
|
|
if(ie->mcast_count >= MAXMCAST
|
|
|
|
|| bcmp(enm->enm_addrlo, enm->enm_addrhi, 6) != 0) {
|
|
|
|
ie->arpcom.ac_if.if_flags |= IFF_ALLMULTI;
|
|
|
|
ieioctl(&ie->arpcom.ac_if, SIOCSIFFLAGS, (void *)0);
|
|
|
|
goto setflag;
|
|
|
|
}
|
|
|
|
|
|
|
|
bcopy(enm->enm_addrlo, &(ie->mcast_addrs[ie->mcast_count]), 6);
|
|
|
|
ie->mcast_count++;
|
|
|
|
ETHER_NEXT_MULTI(step, enm);
|
|
|
|
}
|
|
|
|
|
|
|
|
setflag:
|
|
|
|
ie->want_mcsetup = 1;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
#ifdef DEBUG
|
|
|
|
void print_rbd(volatile struct ie_recv_buf_desc *rbd) {
|
|
|
|
printf("RBD at %08lx:\n"
|
|
|
|
"actual %04x, next %04x, buffer %08x\n"
|
|
|
|
"length %04x, mbz %04x\n",
|
|
|
|
(unsigned long)rbd,
|
|
|
|
rbd->ie_rbd_actual, rbd->ie_rbd_next, rbd->ie_rbd_buffer,
|
|
|
|
rbd->ie_rbd_length, rbd->mbz);
|
|
|
|
}
|
|
|
|
#endif /* DEBUG */
|
|
|
|
#endif /* NIE > 0 */
|
|
|
|
|