This gives significant performance improvements when many raw sockets used.
Benchmarks of mpd handeling 1000 simultaneous PPTP connections show up to 50%
performance boost. With higher number of connections benefit becomes even
bigger. PopTop snd others should also get some benefits.
- removing 'const' qualifier from an input parameter to conform to the type
required by rw_assert();
- using in_addr->s_addr to retrive 32 bits address value.
Observed by: tinderbox
information from rip_input() to rip_append(). Instead, pass the source
address for an IP datagram to rip_append() using a stack-allocated
sockaddr_in, similar to udp_input() and udp_append().
Prior to the move to rwlocks for inpcbinfo, this was not a problem, as
use of the global was synchronized using the ripcbinfo mutex, but with
read-locking there is the potential for a race during concurrent
receive.
This problem is not present in the IPv6 raw IP socket code, which
already used a stack variable for the address.
Spotted by: mav
MFC after: 1 week (before inpcbinfo rwlock changes)
completes the move to a fully parallel UDP transmit path by using
global read, rather than write, locking of inpcbinfo in further
semi-connected cases:
- Add macros to allow try-locking of inpcb and inpcbinfo.
- Always acquire an incpcb read lock in udp_output(), which stablizes the
local inpcb address and port bindings in order to determine what further
locking is required:
- If the inpcb is currently not bound (at all) and are implicitly
connecting, we require inpcbinfo and inpcb write locks, so drop the
read lock and re-acquire.
- If the inpcb is bound for at least one of the port or address, but an
explicit source or destination is requested, trylock the inpcbinfo
lock, and if that fails, drop the inpcb lock, lock the global lock,
and relock the inpcb lock.
- Otherwise, no further locking is required (common case).
- Update comments.
In practice, this means that the vast majority of consumers of UDP sockets
will not acquire any exclusive locks at the socket or UDP levels of the
network stack. This leads to a marked performance improvement in several
important workloads, including BIND, nsd, and memcached over UDP, as well
as significant improvements in pps microbenchmarks.
The plan is to MFC all of the rwlock changes to RELENG_7 once they have
settled for a weeks in the tree.
Tested by: ps, kris (older revision), bde
MFC after: 3 weeks
udp_output() so that argument validation occurs before jail processing.
Add additional comments explaining what's going on when we process
addresses and binding during udp_output().
MFC after: 3 weeks
2) Adds some __UserSpace__ on some of the common defines that
the user space code needs
3) Fixes a bug when we send up data to a user that failed. We
need to a) trim off the data chunk headers, if present, and
b) make sure the frag bit is communicated properly for the
msgs coming off the stream queues... i.e. we see if some
of the msg has been taken.
Obtained from: jeli contributed the VIMAGE changes on this pass Thanks Julain!
inpcb. When directly invoking udp_notify() from udp_ctlinput(), acquire
only a read lock; we may still see write locks in udp_notify() as the
in_pcbnotifyall() routine is shared with TCP and always uses a write lock
on the inpcb being notified.
MFC after: 1 month
some code paths, global or inpcb write locks are required, but for other
code paths, read locks or no locking at all are sufficient for the data
structures.
MFC after: 1 month
source or a specific destination address is requested as part of a send
on a UDP socket, read lock the inpcb rather than write lock it. This
will allow fully parallel transmit down to the IP layer when sending
simultaneously from multiple threads on a connected UDP socket.
Parallel transmit for more complex cases, such as when sendto(2) is
invoked with an address and there's already a local binding, will
follow.
MFC after: 1 month
possible to exhaust and garble stack with a packet that contains a couple
of hundreds nested encapsulation levels.
Submitted by: Ming Fu <fming@borderware.com>
Reviewed by: rwatson
PR: kern/85320
dispatched without Giant, and add NETISR_FORCEQUEUE, which allows specific
netisr handlers to always be dispatched via a queue (deferred). Mark the
usb and if_ppp netisr handlers as NETISR_FORCEQUEUE, and explicitly
acquire Giant in those handlers.
Previously, any netisr handler not marked NETISR_MPSAFE would necessarily
run deferred and with Giant acquired. This change removes Giant
scaffolding from the netisr infrastructure, but NETISR_FORCEQUEUE allows
non-MPSAFE handlers to continue to force deferred dispatch so as to avoid
lock order reversals between their acqusition of Giant and any calling
context.
It is likely we will be able to remove NETISR_FORCEQUEUE once
IFF_NEEDSGIANT is removed, as non-MPSAFE usb and if_ppp drivers will no
longer be supported.
Reviewed by: bz
MFC after: 1 month
X-MFC note: We can't remove NETISR_MPSAFE from stable/7 for KPI reasons,
but the rest can go back.
generating an RTM_MISS for every IP packet forwarded making user space
routing daemons unhappy.
PR: kern/123621, kern/124540, kern/122338
Reported by: Paul <paul gtcomm.net>, Mike Tancsa <mike sentex.net> on net@
Tested by: Paul and Mike
Reviewed by: andre
MFC after: 3 days
datagram-only protocols, such as UDP. This version removes use of
sblock(), which is not required due to an inability to interlace data
improperly with datagrams, as well as avoiding some of the larger loops
and state management that don't apply on datagram sockets.
This is experimental code, so hook it up only for UDPv4 for testing; if
there are problems we may need to revise it or turn it off by default,
but it offers *significant* performance improvements for threaded UDP
applications such as BIND9, nsd, and memcached using UDP.
Tested by: kris, ps
rather than write locking: while we need to maintain a valid reference
to the inpcb and fix its state, no protocol layer state is modified
during an IPv4 UDP receive -- there are only changes at the socket
layer, which is separately protected by socket locking.
While parallel concurrent receive on a single UDP socket is currently
relatively unusual, introducing read locking in the transmit path,
allowing concurrent receive and transmit, will significantly improve
performance for loads such as BIND, memcached, etc.
MFC after: 2 months
Tested by: gnn, kris, ps
in_ifaddrhashtbl in in_ifinit because error handler in in_control removes
entries only for AF_INET addresses. If in_ifinit is called for the cloned
inteface that has just been created its address family is not AF_INET and
therefor LIST_REMOVE is not called for respective LIST_INSERT_HEAD and
freed entries remain in in_ifaddrhashtbl and lead to memory corruption.
PR: kern/124384
This is needed for correct behavior when packets are lost or reordered.
PR: kern/123950
Reviewed by: andre@, silby@
Reported by: Yahoo!, Wang Jin
MFC after: 1 week
- only one functino to destroy an SCTP stack sctp_finish()
- Make it so this function also arranges for any threads
created by the image to do a kthread_exit()
- Vimage prep - these are major restructures to move
all global variables to be accessed via a macro or two.
The variables all go into a single structure.
- Asconf address addition tweaks (add_or_del Interfaces)
- Fix rwnd calcualtion to be more conservative.
- Support SACK_IMMEDIATE flag to skip delayed sack
by demand of peer.
- Comment updates in the sack mapping calculations
- Invarients panic added.
- Pre-support for UDP tunneling (we can do this on
MAC but will need added support from UDP to
get a "pipe" of UDP packets in.
- clear trace buffer sysctl added when local tracing on.
Note the majority of this huge patch is all the vimage prep stuff :-)
template, use an M_TEMP malloc(9) allocation rather than an mbuf
with mtod(9) and dtom(9). This eliminates the last use of
dtom(9) in TCP.
MFC after: 3 weeks
