PCI flush to get correct status block update. Add an optimized
interrupt handler that is activated for MSI case. Actual interrupt
handling is done by taskqueue such that the handler does not
require driver lock for Rx path. The MSI capable bge(4) controllers
automatically disables further interrupt once it enters interrupt
state so we don't need PIO access to disable interrupt in interrupt
handler.
update and then clear status block. Previously it used to access
these index without synchronization which may cause problems when
bounce buffers are used. Also add missing bus_dmamap_sync(9) in
polling handler. Since we now update status block in driver, adjust
bus_dmamap_sync(9) for status block.
checking IFF_DRV_RUNNING and IFF_DRV_OACTIVE flags. Also if we
have less than 16 free send BDs set IFF_DRV_OACTIVE and try it
later. Previously bge(4) used to reserve 16 free send BDs after
loading dma maps but hardware just need one reserved send BD. If
prouder index has the same value of consumer index it means the Tx
queue is empty.
While I'm here check IFQ_DRV_IS_EMPTY first to save one lock
operation.
directly access them at fixed address. While I'm here don't touch
other bits of PCIe device control register except max payload size.
Reviewed by: marius
Revision 1.158 says only lower ten bits of
BGE_RXLP_LOCSTAT_IFIN_DROPS register is valid. For BCM5761 case it
seems the controller maintains 16bits value for the register.
However 16bits are still too small to count all dropped packets
happened in a second. To get a correct counter we have to read the
register in bge_rxeof() which would be too expensive.
Pointed out by: bde
MAC in bge_tick. Previously it used to show more number of input
errors. I noticed actual input errors were less than 8% even for
64 bytes UDP frames generated by netperf.
Since we always access BGE_RXLP_LOCSTAT_IFIN_DROPS register in
bge_tick, remove useless code protected by #ifdef notyet.
initializes it to ETHER_HDR_LEN so we have to override it after
calling ether_ifattch().
While I'm here remove setting if_mtu value, it's initialized in
ether_ifattach().
Introduce two spare dma maps for standard buffer and jumbo buffer
respectively. If loading a dma map failed reuse previously loaded
dma map. This should fix unloaded dma map is used in case of dma
map load failure. Also don't blindly unload dma map and defer
dma map sync and unloading operation until we know dma map for new
buffer is successfully loaded. This change saves unnecessary dma
load/unload operation. Previously bge(4) tried to reuse mbuf
with unloaded dma map which is really bad thing in bus_dma(9)
perspective.
While I'm here update if_iqdrops if we can't allocate Rx buffers.
standard buffer size. If controller is not capable of handling
jumbo frame, interface MTU couldn't be larger than standard MTU
which in turn the received should be fit in standard buffer. This
fixes bus_dmamap_sync call for jumbo ring is called even if
interface is configured to use standard MTU.
Also if total frame size could be fit into standard buffer don't
use jumbo buffers.
for buffer allocation. If driver know we are out of Rx buffers let
controller stop. This should fix panic when interface is run even
if it had no configured Rx buffers.
and Rx DMA tag separately. Previously it used a common mbuf DMA tag
for both Tx and Rx path but Rx buffer(standard ring case) should
have a single DMA segment and maximum buffer size of the segment
should be less than or equal to MCLBYTES. This change also make it
possible to add TSO with minor changes.
bge_newbuf_std still has a bug for handling dma map load failure
under high network load. Just reusing mbuf is not enough as driver
already unloaded the dma map of the mbuf. Graceful recovery needs
more work.
Ideally we can just update dma address part of a Rx descriptor
because the controller never overwrite the Rx descriptor. This
requires some Rx initialization code changes and it would be done
later after fixing other incorrect bus_dma(9) usages.
instead of POSTREAD: the hardware do not touch this memory (CPU
updates it). It is already synchronized as PREWRITE after the
processing is done.
- Synchronize RX return ring memory in rx_eof. This is needed
as the deviced updates this memory when receives packets.
- Decouple the synchronization of BGE status block in the interrupt
service routine: perfrom PREREAD synchronization only all accesses
to this block are finished. This seems to be more natural.
Reviewed by: yongari, marius
MFC after: 2 weeks
to the lock we hold, disable interrupts, and announce to the firmware
that we are shutting down. Especially do this before disabling blocks.
This makes some types of machines with asf enabled no longer hang upon
boot, when we start configuring the interface.
PR: i386/96382, kern/100410, kern/122252, kern/116328
Reported by: erwin
Hardware provided by: TDC A/S
Reviewed by: stas
Tested by: stas
BGE_PCI_PRODID_ASICREV register to store the chip identifier and its revision.
- Add new grouping macro for 7575+ chips (BGE_IS_5755_PLUS).
- Add IDs for Fujitsu-branded Broadcom adapters.
PR: kern/127587
Tested by: Thomas Quinot <thomas@quinot.org> (BCM7561 A0)
MFC after: 2 weeks
Obtained from: OpenBSD
loop iteration as it can be updated by the card while we
process the RX ring forcing us to process RX descriptors
for which DMA synchronisation operation has not been
performed. This fixes the bug when bge(4) drops packets
under high load.
Discussed with: yongari, marius
Approved by: re (kib)
MFC after: 1 week
IF_ADDR_UNLOCK() across network device drivers when accessing the
per-interface multicast address list, if_multiaddrs. This will
allow us to change the locking strategy without affecting our driver
programming interface or binary interface.
For two wireless drivers, remove unnecessary locking, since they
don't actually access the multicast address list.
Approved by: re (kib)
MFC after: 6 weeks
CPU for too long period than necessary. Additively, interfaces are kept
polled (in the tick) even if no more packets are available.
In order to avoid such situations a new generic mechanism can be
implemented in proactive way, keeping track of the time spent on any
packet and fragmenting the time for any tick, stopping the processing
as soon as possible.
In order to implement such mechanism, the polling handler needs to
change, returning the number of packets processed.
While the intended logic is not part of this patch, the polling KPI is
broken by this commit, adding an int return value and the new flag
IFCAP_POLLING_NOCOUNT (which will signal that the return value is
meaningless for the installed handler and checking should be skipped).
Bump __FreeBSD_version in order to signal such situation.
Reviewed by: emaste
Sponsored by: Sandvine Incorporated
drops and re-grabs the softc mutex in the middle, resulting in kernel
trap 12. This may happen when a lot of traffic is being hammered on
one bge(4) interface while the system is shutting down.
Reported by: Alexander Sack <pisymbol gmail com>
PR: kern/134548
MFC After: 2 weeks
quirk requiring it to be enabled even when using MSI. This makes
the latter work again after r189285.
- Remove a comment which no longer applies since r190194.
- If boot verbose, print asicrev, chiprev and bus type on attach.
- For PCI Express devices:
1) Adjust max read request size to 4Kbytes
2) Turn on FIFO_LONG_BURST in RDMA during bge_blockinit()
Though 1) does not seem to have much to do with the poor TX performance
observed on PCI Express bge(4), 2) does fix the problem. [1]
- Nuke the RX CPU self-diag, which prevents working cards from working
(Linux tg3 does not have this diag neither does OpenBSD's bge(4)).
The increasing of the firmware handshaking timeout to 20000 retries
done as part of the original commit isn't merged as way already have a
way higher BGE_TIMEOUT of 100000.
PR: 119361 [1]
Obtained from: tg3 via DragonflyBSD [1], DragonflyBSD
causes data corruption in combination with certain bridges.
Information about this problem was kindly provided by davidch. [1]
- As BGE_FLAG_PCIX is meant to indicate that the controller is in
PCI-X mode, revert to the pre __FreeBSD_version 602101 method of
reading the bus mode register rather than checking the mere
existence of a PCI-X capability, which is also there when the
NIC f.e. is put into a 32-bit slot causing it not to be in PCI-X
mode. Setting BGE_FLAG_PCIX inappropriately could cause the NIC
to be tuned incorrectly.
PR: 128833 [1]
Reviewed by: jhb
MFC after: 3 days
for the BCM5714 revision A0 when in a multi-port configuration
and unconditionally for the remainder of the class of BCM575X
and beyond chips.
This was prodded by mav and is based on a suggestion and a
patch submitted by jhb.
Reviewed by: jhb
MFC after: 2 months
containing an Ethernet address fitted as this is yet another thing
that fails in that case in order to avoid the one second delay
until pci_read_vpd_reg() times out.
- Const'ify the bge_devs array.
- Rename BGE_FLAG_EEPROM to BGE_FLAG_EADDR to underline it's absence means
"there's no chip containing an Ethernet address fitted to the BGE chip
so we have to get it from the firmware instead" rather than "there's no
EEPROM, but maybe NVRAM or something else".
- Don't treat BCM5906[M] generally like chips w/o BGE_FLAG_EADDR set, just
in the two cases really necessary. This gets us line with the original
patch for DragonFlyBSD.
- For sparc64 restore the intended behavior of obtaining the Ethernet
address from the firmware in case BGE_FLAG_EADDR is not set, even for
BCM5906[M].
- Fix some style(9) bugs introduced with rev. 1.208 of if_bge.c
Approved by: jhb
Additional testing by: Thomas Nystroem (BCM5906)
all cards/modes.
In addition to the intr forcing added with rev. 1.205 adopt the other
places to use the same logic.
We need to exclude a few chips/revisions (5700, 5788) from using the
enhanced version and fall back to the old way as that is the only
method they support.
Tested by: phk
Suggested by: davidch, Broadcom (thanks a lot for the help!)
MFC after: 16 days
10/100 operation and place the mailbox registers at a different offset.
They also do not have an EEPROM, so the MAC address must be read from
NVRAM instead.
MFC after: 1 month
PR: kern/118975
Submitted by: benjsc, Thomas Nyström thn at saeab dot se
Submitted by: sephe (original patch for DragonflyBSD)
when creating the parent bus DMA tag. While at it correct the style
and a nearby comment.
- Take advantage of m_collapse(9) for performance reasons.
MFC after: 2 weeks
Because of this we were not getting further interrupts for link state
changes, thus never went into iface UP state and thus could not transmit.
The only way out of this was an incoming packet generating an rx interrupt
and making us call into bge_link_upd.
Up to rev. 1.101, in bge_start_locked, we only returned instantly
if there was 'no link AND nothing queued for tx'. So with a packet queued
for tx, we hit the register scrubbing at the end of bge_start_locked
and were out fine. We simply lost a packet or two but got the interrupts
need to get into UP state.
With rev. 1.102 this was turned into 'if there is no link OR there is
nothing to send' (correct behaviour) and as long as there is no link
we never hit the register scrubbing and consequently never got the link UP.
What we do now is force an interrupt at the end of bge_ifmedia_upd_locked
so we will call bge_link_upd, clear the link state attention and get
further interrupts.
This helps to get the iface UP on an idle network or at least to get
it UP faster not depending on an rx intr anymore.
In case you could not get a DHCP lease or it took very long,
it was because of this.
It is unknown which chips are affected by this. ASIC rev. 0x2003 was the
most popular trouble candidate.
At least the fiber cards should have been working fine.
Which register to scrub is currently under discussion. The comitted
solution was tested and found to work for a lot of setups. It might
not help with MSI.
The reason why we end up in such a situation is entirely unknown.
PR: kern/111804
Tested by: phk, scottl at Y!
MFC after: 14 days
The register layout is little different from memory-mapped stats
in the previous generation chips. In fact, it is bad because
registers in this range are cleared after reading them.
Reviewed by: scottl
MFC after: 3 days
support machines having multiple independently numbered PCI domains
and don't support reenumeration without ambiguity amongst the
devices as seen by the OS and represented by PCI location strings.
This includes introducing a function pci_find_dbsf(9) which works
like pci_find_bsf(9) but additionally takes a domain number argument
and limiting pci_find_bsf(9) to only search devices in domain 0 (the
only domain in single-domain systems). Bge(4) and ofw_pcibus(4) are
changed to use pci_find_dbsf(9) instead of pci_find_bsf(9) in order
to no longer report false positives when searching for siblings and
dupe devices in the same domain respectively.
Along with this change the sole host-PCI bridge driver converted to
actually make use of PCI domain support is uninorth(4), the others
continue to use domain 0 only for now and need to be converted as
appropriate later on.
Note that this means that the format of the location strings as used
by pciconf(8) has been changed and that consumers of <sys/pciio.h>
potentially need to be recompiled.
Suggested by: jhb
Reviewed by: grehan, jhb, marcel
Approved by: re (kensmith), jhb (PCI maintainer hat)
sysctl_handle_int is not sizeof the int type you want to export.
The type must always be an int or an unsigned int.
Remove the instances where a sizeof(variable) is passed to stop
people accidently cut and pasting these examples.
In a few places this was sysctl_handle_int was being used on 64 bit
types, which would truncate the value to be exported. In these
cases use sysctl_handle_quad to export them and change the format
to Q so that sysctl(1) can still print them.
Blade 2500, Fire V210 and probably some other sparc64 machines.
These chips are typically not fitted with an EEPROM which means
that we have to obtain the MAC address via OFW and that some chip
tests will just always fail.
These changes are based on the respective code found in OpenBSD
with some additional info obtained from OpenSolaris and some style
suggestions by jkim@. They also have the desired side-effect of
respecting the 'local-mac-address?' system configuration variable
for the affected BGEs.
- In bge_attach() factor out calling bge_release_resources() before
going to the fail label into the fail label as well as replace a
magic 6 with ETHER_ADDR_LEN.
Reviewed by: yongari (before style changes), jkim
- Remove some excessive parentheses around shift operators.
- Use macro instead of magic number where it is applicable.
- Change lower-case hexdecimals to upper cases to match wpaul's style.
- Revert some unnecessary line wraps and changes from the previous commit.
Pointed out by: bde
- Move some PHY bug detections from brgphy.c to if_bge.c.
- Do not penalize working PHYs.
- Re-arrange bge_flags roughly by their categories.
- Fix minor style(9) nits.
PR: kern/107257
Obtained from: OpenBSD
Tested by: Mike Hibler <mike at flux dot utah dot edu>
bge_intr(). Some of them are used in bge_poll(). Simplify by only
initializing these for polling mode and not toggling them when switching
modes. This also fixes missing synchronization with the coalescing
engine in the toggling.
been handled instead of when at least one descriptor was just handled.
For bge, it is normal to get a txeof when only a small fraction of the
queued tx descriptors have been handled, so the bug broke the watchdog
in a usual case.
- moved the synchronizing bus read to after the bus write for the first
interrupt ack so that it actually synchronizes everything necessary.
We were acking not only the status update that triggered the interrupt
together with any status updates that occurred before we got around
to the bus write for the ack, but also any status updates that occur
after we do the bus write but before the write reaches the device.
The corresponding race for the second interrupt ack resulted in
sometimes returning from the interrupt handler with acked but
unserviced interrupt events. Such events then remain unserviced
until further events cause another interrupt or the watchdog times
out.
The race was often lost on my 5705, apparently since my 5705 has broken
event coalescing which causes a status update for almost every packet,
so another status update is quite likely to occur while the interrupt
handler is running. Watchdog timeouts weren't very noticeable,
apparently because bge_txeof() has one of the usual bugs resetting the
watchdog.
- don't disable device interrupts while bge_intr() is running. Doing this
just had the side effects of:
- entering a device mode in which different coalescing parameters apply.
Different coalescing parameters can be used to either inhibit or
enhance the chance of getting another status update while in the
interrupt handler. This feature is useless with the current
organization of the interrupt handler but might be useful with a
taskqueue handler.
- giving a race for ack+reenable/return. This cannot be handled
by simply rearranging the order of bus accesses like the race for
ack+keepenable/entry. It is necessary to sync the ack and then
check for new events.
- taking longer, especially with the extra code to avoid the race on
ack+reenable/return.
Reviewed by: ru, gleb, scottl
- Do not repeatedly read vendor/device IDs while probing.
- Remove redundant bzero(3) for softc. device_get_softc(9) does it for free[1].
Reviewed by: glebius
Suggested by: glebius[1]
have been added erroneously, and it causes problems on some chips. A larger
change is needed to do this write at a more appropriate place, but that
change requires reworking the ASF logic. That will be worked on in the
future.
Submitted by: Bruce Evans
