- Stateful TCP offload drivers for Terminator 3 and 4 (T3 and T4) ASICs.
These are available as t3_tom and t4_tom modules that augment cxgb(4)
and cxgbe(4) respectively. The cxgb/cxgbe drivers continue to work as
usual with or without these extra features.
- iWARP driver for Terminator 3 ASIC (kernel verbs). T4 iWARP in the
works and will follow soon.
Build-tested with make universe.
30s overview
============
What interfaces support TCP offload? Look for TOE4 and/or TOE6 in the
capabilities of an interface:
# ifconfig -m | grep TOE
Enable/disable TCP offload on an interface (just like any other ifnet
capability):
# ifconfig cxgbe0 toe
# ifconfig cxgbe0 -toe
Which connections are offloaded? Look for toe4 and/or toe6 in the
output of netstat and sockstat:
# netstat -np tcp | grep toe
# sockstat -46c | grep toe
Reviewed by: bz, gnn
Sponsored by: Chelsio communications.
MFC after: ~3 months (after 9.1, and after ensuring MFC is feasible)
TCP_KEEPCNT, that allow to control initial timeout, idle time, idle
re-send interval and idle send count on a per-socket basis.
Reviewed by: andre, bz, lstewart
- The existing ipi_lock continues to protect the global inpcb list and
inpcb counter. This lock is now relegated to a small number of
allocation and free operations, and occasional operations that walk
all connections (including, awkwardly, certain UDP multicast receive
operations -- something to revisit).
- A new ipi_hash_lock protects the two inpcbinfo hash tables for
looking up connections and bound sockets, manipulated using new
INP_HASH_*() macros. This lock, combined with inpcb locks, protects
the 4-tuple address space.
Unlike the current ipi_lock, ipi_hash_lock follows the individual inpcb
connection locks, so may be acquired while manipulating a connection on
which a lock is already held, avoiding the need to acquire the inpcbinfo
lock preemptively when a binding change might later be required. As a
result, however, lookup operations necessarily go through a reference
acquire while holding the lookup lock, later acquiring an inpcb lock --
if required.
A new function in_pcblookup() looks up connections, and accepts flags
indicating how to return the inpcb. Due to lock order changes, callers
no longer need acquire locks before performing a lookup: the lookup
routine will acquire the ipi_hash_lock as needed. In the future, it will
also be able to use alternative lookup and locking strategies
transparently to callers, such as pcbgroup lookup. New lookup flags are,
supplementing the existing INPLOOKUP_WILDCARD flag:
INPLOOKUP_RLOCKPCB - Acquire a read lock on the returned inpcb
INPLOOKUP_WLOCKPCB - Acquire a write lock on the returned inpcb
Callers must pass exactly one of these flags (for the time being).
Some notes:
- All protocols are updated to work within the new regime; especially,
TCP, UDPv4, and UDPv6. pcbinfo ipi_lock acquisitions are largely
eliminated, and global hash lock hold times are dramatically reduced
compared to previous locking.
- The TCP syncache still relies on the pcbinfo lock, something that we
may want to revisit.
- Support for reverting to the FreeBSD 7.x locking strategy in TCP input
is no longer available -- hash lookup locks are now held only very
briefly during inpcb lookup, rather than for potentially extended
periods. However, the pcbinfo ipi_lock will still be acquired if a
connection state might change such that a connection is added or
removed.
- Raw IP sockets continue to use the pcbinfo ipi_lock for protection,
due to maintaining their own hash tables.
- The interface in6_pcblookup_hash_locked() is maintained, which allows
callers to acquire hash locks and perform one or more lookups atomically
with 4-tuple allocation: this is required only for TCPv6, as there is no
in6_pcbconnect_setup(), which there should be.
- UDPv6 locking remains significantly more conservative than UDPv4
locking, which relates to source address selection. This needs
attention, as it likely significantly reduces parallelism in this code
for multithreaded socket use (such as in BIND).
- In the UDPv4 and UDPv6 multicast cases, we need to revisit locking
somewhat, as they relied on ipi_lock to stablise 4-tuple matches, which
is no longer sufficient. A second check once the inpcb lock is held
should do the trick, keeping the general case from requiring the inpcb
lock for every inpcb visited.
- This work reminds us that we need to revisit locking of the v4/v6 flags,
which may be accessed lock-free both before and after this change.
- Right now, a single lock name is used for the pcbhash lock -- this is
undesirable, and probably another argument is required to take care of
this (or a char array name field in the pcbinfo?).
This is not an MFC candidate for 8.x due to its impact on lookup and
locking semantics. It's possible some of these issues could be worked
around with compatibility wrappers, if necessary.
Reviewed by: bz
Sponsored by: Juniper Networks, Inc.
persist state and the retransmit timer. However, the code that implements
"bad retransmit recovery" only checks t_rxtshift to see if an ACK has been
received in during the first retransmit timeout window. As a result, if
ticks has wrapped over to a negative value and a socket is in the persist
state, it can incorrectly treat an ACK from the remote peer as a
"bad retransmit recovery" and restore saved values such as snd_ssthresh and
snd_cwnd. However, if the socket has never had a retransmit timeout, then
these saved values will be zero, so snd_ssthresh and snd_cwnd will be set
to 0.
If the socket is in fast recovery (this can be caused by excessive
duplicate ACKs such as those fixed by 220794), then each ACK that arrives
triggers either NewReno or SACK partial ACK handling which clamps snd_cwnd
to be no larger than snd_ssthresh. In effect, the socket's send window
is permamently stuck at 0 even though the remote peer is advertising a
much larger window and pending data is only sent via TCP window probes
(so one byte every few seconds).
Fix this by adding a new TCP pcb flag (TF_PREVVALID) that indicates that
the various snd_*_prev fields in the pcb are valid and only perform
"bad retransmit recovery" if this flag is set in the pcb. The flag is set
on the first retransmit timeout that occurs and is cleared on subsequent
retransmit timeouts or when entering the persist state.
Reviewed by: bz
MFC after: 2 weeks
Control Algorithms for FreeBSD" FreeBSD Foundation funded project. More details
about the project are available at: http://caia.swin.edu.au/freebsd/5cc/
- Add a KPI and supporting infrastructure to allow modular congestion control
algorithms to be used in the net stack. Algorithms can maintain per-connection
state if required, and connections maintain their own algorithm pointer, which
allows different connections to concurrently use different algorithms. The
TCP_CONGESTION socket option can be used with getsockopt()/setsockopt() to
programmatically query or change the congestion control algorithm respectively
from within an application at runtime.
- Integrate the framework with the TCP stack in as least intrusive a manner as
possible. Care was also taken to develop the framework in a way that should
allow integration with other congestion aware transport protocols (e.g. SCTP)
in the future. The hope is that we will one day be able to share a single set
of congestion control algorithm modules between all congestion aware transport
protocols.
- Introduce a new congestion recovery (TF_CONGRECOVERY) state into the TCP stack
and use it to decouple the meaning of recovery from a congestion event and
recovery from packet loss (TF_FASTRECOVERY) a la RFC2581. ECN and delay based
congestion control protocols don't generally need to recover from packet loss
and need a different way to note a congestion recovery episode within the
stack.
- Remove the net.inet.tcp.newreno sysctl, which simplifies some portions of code
and ensures the stack always uses the appropriate mechanisms for recovering
from packet loss during a congestion recovery episode.
- Extract the NewReno congestion control algorithm from the TCP stack and
massage it into module form. NewReno is always built into the kernel and will
remain the default algorithm for the forseeable future. Implementations of
additional different algorithms will become available in the near future.
- Bump __FreeBSD_version to 900025 and note in UPDATING that rebuilding code
that relies on the size of "struct tcpcb" is required.
Many thanks go to the Cisco University Research Program Fund at Community
Foundation Silicon Valley and the FreeBSD Foundation. Their support of our work
at the Centre for Advanced Internet Architectures, Swinburne University of
Technology is greatly appreciated.
In collaboration with: David Hayes <dahayes at swin edu au> and
Grenville Armitage <garmitage at swin edu au>
Sponsored by: Cisco URP, FreeBSD Foundation
Reviewed by: rpaulo
Tested by: David Hayes (and many others over the years)
MFC after: 3 months
feature when you have a seemingly stuck socket and want to figure
out why it has not been closed yet.
No plans to MFC this, as it changes the netstat sysctl ABI.
Reviewed by: andre, rwatson, Eric Van Gyzen
vnet.h, we now use jails (rather than vimages) as the abstraction
for virtualization management, and what remained was specific to
virtual network stacks. Minor cleanups are done in the process,
and comments updated to reflect these changes.
Reviewed by: bz
Approved by: re (vimage blanket)
unused custom mutex/condvar-based sleep locks with two locks: an
rwlock (for non-sleeping use) and sxlock (for sleeping use). Either
acquired for read is sufficient to stabilize the vnet list, but both
must be acquired for write to modify the list.
Replace previous no-op read locking macros, used in various places
in the stack, with actual locking to prevent race conditions. Callers
must declare when they may perform unbounded sleeps or not when
selecting how to lock.
Refactor vnet sysinits so that the vnet list and locks are initialized
before kernel modules are linked, as the kernel linker will use them
for modules loaded by the boot loader.
Update various consumers of these KPIs based on whether they may sleep
or not.
Reviewed by: bz
Approved by: re (kib)
(DPCPU), as suggested by Peter Wemm, and implement a new per-virtual
network stack memory allocator. Modify vnet to use the allocator
instead of monolithic global container structures (vinet, ...). This
change solves many binary compatibility problems associated with
VIMAGE, and restores ELF symbols for virtualized global variables.
Each virtualized global variable exists as a "reference copy", and also
once per virtual network stack. Virtualized global variables are
tagged at compile-time, placing the in a special linker set, which is
loaded into a contiguous region of kernel memory. Virtualized global
variables in the base kernel are linked as normal, but those in modules
are copied and relocated to a reserved portion of the kernel's vnet
region with the help of a the kernel linker.
Virtualized global variables exist in per-vnet memory set up when the
network stack instance is created, and are initialized statically from
the reference copy. Run-time access occurs via an accessor macro, which
converts from the current vnet and requested symbol to a per-vnet
address. When "options VIMAGE" is not compiled into the kernel, normal
global ELF symbols will be used instead and indirection is avoided.
This change restores static initialization for network stack global
variables, restores support for non-global symbols and types, eliminates
the need for many subsystem constructors, eliminates large per-subsystem
structures that caused many binary compatibility issues both for
monitoring applications (netstat) and kernel modules, removes the
per-function INIT_VNET_*() macros throughout the stack, eliminates the
need for vnet_symmap ksym(2) munging, and eliminates duplicate
definitions of virtualized globals under VIMAGE_GLOBALS.
Bump __FreeBSD_version and update UPDATING.
Portions submitted by: bz
Reviewed by: bz, zec
Discussed with: gnn, jamie, jeff, jhb, julian, sam
Suggested by: peter
Approved by: re (kensmith)
TCPSTAT_INC(), rather than directly manipulating the fields across the
kernel. This will make it easier to change the implementation of
these statistics, such as using per-CPU versions of the data structures.
MFC after: 3 days
certain flags that should have been in inp_flags ended up in inp_vflag,
meaning that they were inconsistently locked, and in one case,
interpreted. Move the following flags from inp_vflag to gaps in the
inp_flags space (and clean up the inp_flags constants to make gaps
more obvious to future takers):
INP_TIMEWAIT
INP_SOCKREF
INP_ONESBCAST
INP_DROPPED
Some aspects of this change have no effect on kernel ABI at all, as these
are UDP/TCP/IP-internal uses; however, netstat and sockstat detect
INP_TIMEWAIT when listing TCP sockets, so any MFC will need to take this
into account.
MFC after: 1 week (or after dependencies are MFC'd)
Reviewed by: bz
The new behaviour is on by default, and can be disabled by setting the
net.inet.tcp.rfc3465 sysctl to 0 to obtain previous behaviour.
The patch changes struct tcpcb in sys/netinet/tcp_var.h which breaks
the ABI. Bump __FreeBSD_version to 800061 accordingly. User space tools
that rely on the size of struct tcpcb (e.g. sockstat) need to be recompiled.
Reviewed by: rpaulo, gnn
Approved by: gnn, kmacy (mentors)
Sponsored by: FreeBSD Foundation
directly include only the header files needed. This reduces the
unneeded spamming of various headers into lots of files.
For now, this leaves us with very few modules including vnet.h
and thus needing to depend on opt_route.h.
Reviewed by: brooks, gnn, des, zec, imp
Sponsored by: The FreeBSD Foundation
from the vimage project, as per plan established at devsummit 08/08:
http://wiki.freebsd.org/Image/Notes200808DevSummit
Introduce INIT_VNET_*() initializer macros, VNET_FOREACH() iterator
macros, and CURVNET_SET() context setting macros, all currently
resolving to NOPs.
Prepare for virtualization of selected SYSCTL objects by introducing a
family of SYSCTL_V_*() macros, currently resolving to their global
counterparts, i.e. SYSCTL_V_INT() == SYSCTL_INT().
Move selected #defines from sys/sys/vimage.h to newly introduced header
files specific to virtualized subsystems (sys/net/vnet.h,
sys/netinet/vinet.h etc.).
All the changes are verified to have zero functional impact at this
point in time by doing MD5 comparision between pre- and post-change
object files(*).
(*) netipsec/keysock.c did not validate depending on compile time options.
Implemented by: julian, bz, brooks, zec
Reviewed by: julian, bz, brooks, kris, rwatson, ...
Approved by: julian (mentor)
Obtained from: //depot/projects/vimage-commit2/...
X-MFC after: never
Sponsored by: NLnet Foundation, The FreeBSD Foundation
virtualization work done by Marko Zec (zec@).
This is the first in a series of commits over the course
of the next few weeks.
Mark all uses of global variables to be virtualized
with a V_ prefix.
Use macros to map them back to their global names for
now, so this is a NOP change only.
We hope to have caught at least 85-90% of what is needed
so we do not invalidate a lot of outstanding patches again.
Obtained from: //depot/projects/vimage-commit2/...
Reviewed by: brooks, des, ed, mav, julian,
jamie, kris, rwatson, zec, ...
(various people I forgot, different versions)
md5 (with a bit of help)
Sponsored by: NLnet Foundation, The FreeBSD Foundation
X-MFC after: never
V_Commit_Message_Reviewed_By: more people than the patch
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
explicitly select write locking for all use of the inpcb mutex.
Update some pcbinfo lock assertions to assert locked rather than
write-locked, although in practice almost all uses of the pcbinfo
rwlock main exclusive, and all instances of inpcb lock acquisition
are exclusive.
This change should introduce (ideally) little functional change.
However, it lays the groundwork for significantly increased
parallelism in the TCP/IP code.
MFC after: 3 months
Tested by: kris (superset of committered patch)
- Reintegrate the ANSI C function declaration change
from tcp_timer.c rev 1.92
- Reorganize the tcpcb structure so that it has a single
pointer to the "tcp_timer" structure which contains all
of the tcp timer callouts. This change means that when
the single tcp timer change is reintegrated, tcpcb will
not change in size, and therefore the ABI between
netstat and the kernel will not change.
Neither of these changes should have any functional
impact.
Reviewed by: bmah, rrs
Approved by: re (bmah)
TCP timers as a single timer, but retain the API changes necessary to
reintroduce this change. This will back out the source of at least two
reported problems: lock leaks in certain timer edge cases, and TCP timers
continuing to fire after a connection has closed (a bug previously fixed and
then reintroduced with the timer rewrite).
In a follow-up commit, some minor restylings and comment changes performed
after the TCP timer rewrite will be reapplied, and a further change to allow
the TCP timer rewrite to be added back without disturbing the ABI. The new
design is believed to be a good thing, but the outstanding issues are
leading to significant stability/correctness problems that are holding
up 7.0.
This patch was generated by silby, but is being committed by proxy due to
poor network connectivity for silby this week.
Approved by: re (kensmith)
Submitted by: silby
Tested by: rwatson, kris
Problems reported by: peter, kris, others
a timer issues a shutdown and a simultaneous close on the socket
happens. This race condition is inherent in the current socket/
inpcb life cycle system but can be handled well.
Reported by: kris
Tested by: kris (on 8-core machine)
other than repo copied tcp_subr.c into tcp_timewait.c#1.284:
tcp_input.c#1.350 tcp_timewait() -> tcp_twcheck()
tcp_timer.c#1.92 tcp_timer_2msl_reset() -> tcp_tw_2msl_reset()
tcp_timer.c#1.92 tcp_timer_2msl_stop() -> tcp_tw_2msl_stop()
tcp_timer.c#1.92 tcp_timer_2msl_tw() -> tcp_tw_2msl_scan()
This is a mechanical move with appropriate renames and making
them static if used only locally.
The tcp_tw_2msl_scan() cleanup function is still run from the
tcp_slowtimo() in tcp_timer.c.
directly to a merged model where only one callout, the next to fire,
is registered.
Instead of callout_reset(9) and callout_stop(9) the new function
tcp_timer_activate() is used which then internally manages the callout.
The single new callout is a mutex callout on inpcb simplifying the
locking a bit.
tcp_timer() is the called function which handles all race conditions
in one place and then dispatches the individual timer functions.
Reviewed by: rwatson (earlier version)
potential issues where the peer does not close, potentially leaving
thousands of connections in FIN_WAIT_2. This is controlled by a new sysctl
fast_finwait2_recycle, which is disabled by default.
Reviewed by: gnn, silby.
timeouts for TCP and T/TCP connections in the TIME_WAIT
state, and we had two separate timed wait queues for them.
Now that is has gone, the timeout is always 2*MSL again,
and there is no reason to keep two queues (the first was
unused anyway!).
Also, reimplement the remaining queue using a TAILQ (it
was technically impossible before, with two queues).
bad under high load. For example with 40k sockets and 25k tcptw
entries, connect() syscall can run for seconds. Debugging showed
that it iterates the cycle millions times and purges thousands of
tcptw entries at a time.
Besides practical unusability this change is architecturally
wrong. First, in_pcblookup_local() is used in connect() and bind()
syscalls. No stale entries purging shouldn't be done here. Second,
it is a layering violation.
o Return back the tcptw purging cycle to tcp_timer_2msl_tw(),
that was removed in rev. 1.78 by rwatson. The commit log of this
revision tells nothing about the reason cycle was removed. Now
we need this cycle, since major cleaner of stale tcptw structures
is removed.
o Disable probably necessary, but now unused
tcp_twrecycleable() function.
Reviewed by: ru
dropped. This prevents a bug introduced during the socket/pcb refcounting
work from occuring, in which occasionally the retransmit timer may fire
after a connection has been reset, resulting in the resulting R|A TCP
packet having a source port of 0, as the port reservation has been
released.
While here, fixing up some RUNLOCK->WUNLOCK bugs.
MFC after: 1 month
return NULL. In principle this shouldn't change the behavior, but
avoids returning a potentially invalid/inappropriate pointer to
the caller.
Found with: Coverity Prevent (tm)
Submitted by: pjd
MFC after: 3 months
pru_abort(), pru_detach(), and in_pcbdetach():
- Universally support and enforce the invariant that so_pcb is
never NULL, converting dozens of unnecessary NULL checks into
assertions, and eliminating dozens of unnecessary error handling
cases in protocol code.
- In some cases, eliminate unnecessary pcbinfo locking, as it is no
longer required to ensure so_pcb != NULL. For example, the receive
code no longer requires the pcbinfo lock, and the send code only
requires it if building a new connection on an otherwise unconnected
socket triggered via sendto() with an address. This should
significnatly reduce tcbinfo lock contention in the receive and send
cases.
- In order to support the invariant that so_pcb != NULL, it is now
necessary for the TCP code to not discard the tcpcb any time a
connection is dropped, but instead leave the tcpcb until the socket
is shutdown. This case is handled by setting INP_DROPPED, to
substitute for using a NULL so_pcb to indicate that the connection
has been dropped. This requires the inpcb lock, but not the pcbinfo
lock.
- Unlike all other protocols in the tree, TCP may need to retain access
to the socket after the file descriptor has been closed. Set
SS_PROTOREF in tcp_detach() in order to prevent the socket from being
freed, and add a flag, INP_SOCKREF, so that the TCP code knows whether
or not it needs to free the socket when the connection finally does
close. The typical case where this occurs is if close() is called on
a TCP socket before all sent data in the send socket buffer has been
transmitted or acknowledged. If INP_SOCKREF is found when the
connection is dropped, we release the inpcb, tcpcb, and socket instead
of flagging INP_DROPPED.
- Abort and detach protocol switch methods no longer return failures,
nor attempt to free sockets, as the socket layer does this.
- Annotate the existence of a long-standing race in the TCP timer code,
in which timers are stopped but not drained when the socket is freed,
as waiting for drain may lead to deadlocks, or have to occur in a
context where waiting is not permitted. This race has been handled
by testing to see if the tcpcb pointer in the inpcb is NULL (and vice
versa), which is not normally permitted, but may be true of a inpcb
and tcpcb have been freed. Add a counter to test how often this race
has actually occurred, and a large comment for each instance where
we compare potentially freed memory with NULL. This will have to be
fixed in the near future, but requires is to further address how to
handle the timer shutdown shutdown issue.
- Several TCP calls no longer potentially free the passed inpcb/tcpcb,
so no longer need to return a pointer to indicate whether the argument
passed in is still valid.
- Un-macroize debugging and locking setup for various protocol switch
methods for TCP, as it lead to more obscurity, and as locking becomes
more customized to the methods, offers less benefit.
- Assert copyright on tcp_usrreq.c due to significant modifications that
have been made as part of this work.
These changes significantly modify the memory management and connection
logic of our TCP implementation, and are (as such) High Risk Changes,
and likely to contain serious bugs. Please report problems to the
current@ mailing list ASAP, ideally with simple test cases, and
optionally, packet traces.
MFC after: 3 months
