to do a window update to the peer (thru an ACK) from soreceive()
itself. TCP will do that upon return from the socket callback.
Sending a window update from soreceive() results in a lock reversal.
Submitted by: Mohan Srinivasan mohans at yahoo-inc dot com
Reviewed by: rwatson
soreceive(), then pass in M_DONTWAIT to m_copym(). Also fix up error
handling for the case where m_copym() returns failure.
Submitted by: Mohan Srinivasan mohans at yahoo-inc dot com
Reviewed by: rwatson
or pru_attach is NULL. With loadable protocols the SPACER dummy protocols
have valid function pointers for all methods to functions returning just
EOPNOTSUPP. Thus the early abort check would not detect immediately that
attach is not supported for this protocol. Instead it would correctly
get the EOPNOTSUPP error later on when it calls the protocol specific
attach function.
Add testing against the pru_attach_notsupp() function pointer to the
early abort check as well.
(sorele()/sotryfree()):
- This permits the caller to acquire the accept mutex before the socket
mutex, avoiding sofree() having to drop the socket mutex and re-order,
which could lead to races permitting more than one thread to enter
sofree() after a socket is ready to be free'd.
- This also covers clearing of the so_pcb weak socket reference from
the protocol to the socket, preventing races in clearing and
evaluation of the reference such that sofree() might be called more
than once on the same socket.
This appears to close a race I was able to easily trigger by repeatedly
opening and resetting TCP connections to a host, in which the
tcp_close() code called as a result of the RST raced with the close()
of the accepted socket in the user process resulting in simultaneous
attempts to de-allocate the same socket. The new locking increases
the overhead for operations that may potentially free the socket, so we
will want to revise the synchronization strategy here as we normalize
the reference counting model for sockets. The use of the accept mutex
in freeing of sockets that are not listen sockets is primarily
motivated by the potential need to remove the socket from the
incomplete connection queue on its parent (listen) socket, so cleaning
up the reference model here may allow us to substantially weaken the
synchronization requirements.
RELENG_5_3 candidate.
MFC after: 3 days
Reviewed by: dwhite
Discussed with: gnn, dwhite, green
Reported by: Marc UBM Bocklet <ubm at u-boot-man dot de>
Reported by: Vlad <marchenko at gmail dot com>
Sockets in the listen queues have reference counts of 0, so if the
protocol decides to disconnect the pcb and try to free the socket, this
triggered a race with accept() wherein accept() would bump the reference
count before sofree() had removed the socket from the listen queues,
resulting in a panic in sofree() when it discovered it was freeing a
referenced socket. This might happen if a RST came in prior to accept()
on a TCP connection.
The fix is two-fold: to expand the coverage of the accept mutex earlier
in sofree() to prevent accept() from grabbing the socket after the "is it
really safe to free" tests, and to expand the logic of the "is it really
safe to free" tests to check that the refcount is still 0 (i.e., we
didn't race).
RELENG_5 candidate.
Much discussion with and work by: green
Reported by: Marc UBM Bocklet <ubm at u-boot-man dot de>
Reported by: Vlad <marchenko at gmail dot com>
state test as well as set, or we risk a race between a socket wakeup
and registering for select() or poll() on the socket. This does
increase the cost of the poll operation, but can probably be optimized
some in the future.
This appears to correct poll() "wedges" experienced with X11 on SMP
systems with highly interactive applications, and might affect a plethora
of other select() driven applications.
RELENG_5 candidate.
Problem reported by: Maxim Maximov <mcsi at mcsi dot pp dot ru>
Debugged with help of: dwhite
buffers with kqueue filters is no longer required: the kqueue framework
will guarantee that the mutex is held on entering the filter, either
due to a call from the socket code already holding the mutex, or by
explicitly acquiring it. This removes the last of the conditional
socket locking.
sockets are connection-oriented for the purposes of kqueue
registration. Since UDP sockets aren't connection-oriented, this
appeared to break a great many things, such as RPC-based
applications and services (i.e., NFS). Since jmg isn't around I'm
backing this out before too many more feet are shot, but intend to
investigate the right solution with him once he's available.
Apologies to: jmg
Discussed with: imp, scottl
a more complete subsystem, and removes the knowlege of how things are
implemented from the drivers. Include locking around filter ops, so a
module like aio will know when not to be unloaded if there are outstanding
knotes using it's filter ops.
Currently, it uses the MTX_DUPOK even though it is not always safe to
aquire duplicate locks. Witness currently doesn't support the ability
to discover if a dup lock is ok (in some cases).
Reviewed by: green, rwatson (both earlier versions)
that get certain types of control messages (ping6 and rtsol are
examples). This gets the new code closer to working:
1) Collect control mbufs for processing in the controlp ==
NULL case, so that they can be freed by externalize.
2) Loop over the list of control mbufs, as the externalize
function may not know how to deal with chains.
3) In the case where there is no externalize function,
remember to add the control mbuf to the controlp list so
that it will be returned.
4) After adding stuff to the controlp list, walk to the
end of the list of stuff that was added, incase we added
a chain.
This code can be further improved, but this is enough to get most
things working again.
Reviewed by: rwatson
a better name. I have a kern_[sg]etsockopt which I plan to commit
shortly, but the arguments to these function will be quite different
from so_setsockopt.
Approved by: alfred
in soreceive() after removing an MT_SONAME mbuf from the head of the
socket buffer.
When processing MT_CONTROL mbufs in soreceive(), first remove all of
the MT_CONTROL mbufs from the head of the socket buffer to a local
mbuf chain, then feed them into dom_externalize() as a set, which
both avoids thrashing the socket buffer lock when handling multiple
control mbufs, and also avoids races with other threads acting on
the socket buffer when the socket buffer mutex is released to enter
the externalize code. Existing races that might occur if the protocol
externalize method blocked during processing have also been closed.
Now that we synchronize socket buffer and stack state following
modifications to the socket buffer, turn the manual synchronization
that previously followed control mbuf processing with a set of
assertions. This can eventually be removed.
The soreceive() code is now substantially more MPSAFE.
the head of the mbuf chains in a socket buffer, re-synchronizes the
cache pointers used to optimize socket buffer appends. This will be
used by soreceive() before dropping socket buffer mutexes to make sure
a consistent version of the socket buffer is visible to other threads.
While here, update copyright to account for substantial rewrite of much
socket code required for fine-grained locking.
locking on 'nextrecord' and concerns regarding potentially inconsistent
or stale use of socket buffer or stack fields if they aren't carefully
synchronized whenever the socket buffer mutex is released. Document
that the high-level sblock() prevents races against other readers on
the socket.
Also document the 'type' logic as to how soreceive() guarantees that
it will only return one of normal data or inline out-of-band data.
associated with a PR_ADDR protocol, make sure to update the m_nextpkt
pointer of the new head mbuf on the chain to point to the next record.
Otherwise, when we release the socket buffer mutex, the socket buffer
mbuf chain may be in an inconsistent state.
so_gencnt, numopensockets, and the per-socket field so_gencnt. Annotate
this this might be better done with atomic operations.
Annotate what accept_mtx protects.
- sowakeup() now asserts the socket buffer lock on entry. Move
the call to KNOTE higher in sowakeup() so that it is made with
the socket buffer lock held for consistency with other calls.
Release the socket buffer lock prior to calling into pgsigio(),
so_upcall(), or aio_swake(). Locking for this event management
will need revisiting in the future, but this model avoids lock
order reversals when upcalls into other subsystems result in
socket/socket buffer operations. Assert that the socket buffer
lock is not held at the end of the function.
- Wrapper macros for sowakeup(), sorwakeup() and sowwakeup(), now
have _locked versions which assert the socket buffer lock on
entry. If a wakeup is required by sb_notify(), invoke
sowakeup(); otherwise, unconditionally release the socket buffer
lock. This results in the socket buffer lock being released
whether a wakeup is required or not.
- Break out socantsendmore() into socantsendmore_locked() that
asserts the socket buffer lock. socantsendmore()
unconditionally locks the socket buffer before calling
socantsendmore_locked(). Note that both functions return with
the socket buffer unlocked as socantsendmore_locked() calls
sowwakeup_locked() which has the same properties. Assert that
the socket buffer is unlocked on return.
- Break out socantrcvmore() into socantrcvmore_locked() that
asserts the socket buffer lock. socantrcvmore() unconditionally
locks the socket buffer before calling socantrcvmore_locked().
Note that both functions return with the socket buffer unlocked
as socantrcvmore_locked() calls sorwakeup_locked() which has
similar properties. Assert that the socket buffer is unlocked
on return.
- Break out sbrelease() into a sbrelease_locked() that asserts the
socket buffer lock. sbrelease() unconditionally locks the
socket buffer before calling sbrelease_locked().
sbrelease_locked() now invokes sbflush_locked() instead of
sbflush().
- Assert the socket buffer lock in socket buffer sanity check
functions sblastrecordchk(), sblastmbufchk().
- Assert the socket buffer lock in SBLINKRECORD().
- Break out various sbappend() functions into sbappend_locked()
(and variations on that name) that assert the socket buffer
lock. The !_locked() variations unconditionally lock the socket
buffer before calling their _locked counterparts. Internally,
make sure to call _locked() support routines, etc, if already
holding the socket buffer lock.
- Break out sbinsertoob() into sbinsertoob_locked() that asserts
the socket buffer lock. sbinsertoob() unconditionally locks the
socket buffer before calling sbinsertoob_locked().
- Break out sbflush() into sbflush_locked() that asserts the
socket buffer lock. sbflush() unconditionally locks the socket
buffer before calling sbflush_locked(). Update panic strings
for new function names.
- Break out sbdrop() into sbdrop_locked() that asserts the socket
buffer lock. sbdrop() unconditionally locks the socket buffer
before calling sbdrop_locked().
- Break out sbdroprecord() into sbdroprecord_locked() that asserts
the socket buffer lock. sbdroprecord() unconditionally locks
the socket buffer before calling sbdroprecord_locked().
- sofree() now calls socantsendmore_locked() and re-acquires the
socket buffer lock on return. It also now calls
sbrelease_locked().
- sorflush() now calls socantrcvmore_locked() and re-acquires the
socket buffer lock on return. Clean up/mess up other behavior
in sorflush() relating to the temporary stack copy of the socket
buffer used with dom_dispose by more properly initializing the
temporary copy, and selectively bzeroing/copying more carefully
to prevent WITNESS from getting confused by improperly
initialized mutexes. Annotate why that's necessary, or at
least, needed.
- soisconnected() now calls sbdrop_locked() before unlocking the
socket buffer to avoid locking overhead.
Some parts of this change were:
Submitted by: sam
Sponsored by: FreeBSD Foundation
Obtained from: BSD/OS
socket lock over pulling so_options and so_linger out of the socket
structure in order to retrieve a consistent snapshot. This may be
overkill if user space doesn't require a consistent snapshot.
resolved by socket locking: in particular, that we test the connection
state at the socket layer without locking, request that the protocol
begin listening, and then set the listen state on the socket
non-atomically, resulting in a non-atomic cross-layer test-and-set.
lock state. Convert tsleep() into msleep() with socket buffer mutex
as argument. Hold socket buffer lock over sbunlock() to protect sleep
lock state.
Assert socket buffer lock in sbwait() to protect the socket buffer
wait state. Convert tsleep() into msleep() with socket buffer mutex
as argument.
Modify sofree(), sosend(), and soreceive() to acquire SOCKBUF_LOCK()
in order to call into these functions with the lock, as well as to
start protecting other socket buffer use in their implementation. Drop
the socket buffer mutexes around calls into the protocol layer, around
potentially blocking operations, for copying to/from user space, and
VM operations relating to zero-copy. Assert the socket buffer mutex
strategically after code sections or at the beginning of loops. In
some cases, modify return code to ensure locks are properly dropped.
Convert the potentially blocking allocation of storage for the remote
address in soreceive() into a non-blocking allocation; we may wish to
move the allocation earlier so that it can block prior to acquisition
of the socket buffer lock.
Drop some spl use.
NOTE: Some races exist in the current structuring of sosend() and
soreceive(). This commit only merges basic socket locking in this
code; follow-up commits will close additional races. As merged,
these changes are not sufficient to run without Giant safely.
Reviewed by: juli, tjr
rwatson_netperf:
Introduce conditional locking of the socket buffer in fifofs kqueue
filters; KNOTE() will be called holding the socket buffer locks in
fifofs, but sometimes the kqueue() system call will poll using the
same entry point without holding the socket buffer lock.
Introduce conditional locking of the socket buffer in the socket
kqueue filters; KNOTE() will be called holding the socket buffer
locks in the socket code, but sometimes the kqueue() system call
will poll using the same entry points without holding the socket
buffer lock.
Simplify the logic in sodisconnect() since we no longer need spls.
NOTE: To remove conditional locking in the kqueue filters, it would
make sense to use a separate kqueue API entry into the socket/fifo
code when calling from the kqueue() system call.
- Lock down low hanging fruit use of sb_flags with socket buffer
lock.
- Lock down low hanging fruit use of so_state with socket lock.
- Lock down low hanging fruit use of so_options.
- Lock down low-hanging fruit use of sb_lowwat and sb_hiwat with
socket buffer lock.
- Annotate situations in which we unlock the socket lock and then
grab the receive socket buffer lock, which are currently actually
the same lock. Depending on how we want to play our cards, we
may want to coallesce these lock uses to reduce overhead.
- Convert a if()->panic() into a KASSERT relating to so_state in
soaccept().
- Remove a number of splnet()/splx() references.
More complex merging of socket and socket buffer locking to
follow.
flags relating to several aspects of socket functionality. This change
breaks out several bits relating to send and receive operation into a
new per-socket buffer field, sb_state, in order to facilitate locking.
This is required because, in order to provide more granular locking of
sockets, different state fields have different locking properties. The
following fields are moved to sb_state:
SS_CANTRCVMORE (so_state)
SS_CANTSENDMORE (so_state)
SS_RCVATMARK (so_state)
Rename respectively to:
SBS_CANTRCVMORE (so_rcv.sb_state)
SBS_CANTSENDMORE (so_snd.sb_state)
SBS_RCVATMARK (so_rcv.sb_state)
This facilitates locking by isolating fields to be located with other
identically locked fields, and permits greater granularity in socket
locking by avoiding storing fields with different locking semantics in
the same short (avoiding locking conflicts). In the future, we may
wish to coallesce sb_state and sb_flags; for the time being I leave
them separate and there is no additional memory overhead due to the
packing/alignment of shorts in the socket buffer structure.
reference count:
- Assert SOCK_LOCK(so) macros that directly manipulate so_count:
soref(), sorele().
- Assert SOCK_LOCK(so) in macros/functions that rely on the state of
so_count: sofree(), sotryfree().
- Acquire SOCK_LOCK(so) before calling these functions or macros in
various contexts in the stack, both at the socket and protocol
layers.
- In some cases, perform soisdisconnected() before sotryfree(), as
this could result in frobbing of a non-present socket if
sotryfree() actually frees the socket.
- Note that sofree()/sotryfree() will release the socket lock even if
they don't free the socket.
Submitted by: sam
Sponsored by: FreeBSD Foundation
Obtained from: BSD/OS
protect fields in the socket buffer. Add accessor macros to use the
mutex (SOCKBUF_*()). Initialize the mutex in soalloc(), and destroy
it in sodealloc(). Add addition, add SOCK_*() access macros which
will protect most remaining fields in the socket; for the time being,
use the receive socket buffer mutex to implement socket level locking
to reduce memory overhead.
Submitted by: sam
Sponosored by: FreeBSD Foundation
Obtained from: BSD/OS
global mutex, accept_mtx, which serializes access to the following
fields across all sockets:
so_qlen so_incqlen so_qstate
so_comp so_incomp so_list
so_head
While providing only coarse granularity, this approach avoids lock
order issues between sockets by avoiding ownership of the fields
by a specific socket and its per-socket mutexes.
While here, rewrite soclose(), sofree(), soaccept(), and
sonewconn() to add assertions, close additional races and address
lock order concerns. In particular:
- Reorganize the optimistic concurrency behavior in accept1() to
always allocate a file descriptor with falloc() so that if we do
find a socket, we don't have to encounter the "Oh, there wasn't
a socket" race that can occur if falloc() sleeps in the current
code, which broke inbound accept() ordering, not to mention
requiring backing out socket state changes in a way that raced
with the protocol level. We may want to add a lockless read of
the queue state if polling of empty queues proves to be important
to optimize.
- In accept1(), soref() the socket while holding the accept lock
so that the socket cannot be free'd in a race with the protocol
layer. Likewise in netgraph equivilents of the accept1() code.
- In sonewconn(), loop waiting for the queue to be small enough to
insert our new socket once we've committed to inserting it, or
races can occur that cause the incomplete socket queue to
overfill. In the previously implementation, it was sufficient
to simply tested once since calling soabort() didn't release
synchronization permitting another thread to insert a socket as
we discard a previous one.
- In soclose()/sofree()/et al, it is the responsibility of the
caller to remove a socket from the incomplete connection queue
before calling soabort(), which prevents soabort() from having
to walk into the accept socket to release the socket from its
queue, and avoids races when releasing the accept mutex to enter
soabort(), permitting soabort() to avoid lock ordering issues
with the caller.
- Generally cluster accept queue related operations together
throughout these functions in order to facilitate locking.
Annotate new locking in socketvar.h.
the socket is on an accept queue of a listen socket. This change
renames the flags to SQ_COMP and SQ_INCOMP, and moves them to a new
state field on the socket, so_qstate, as the locking for these flags
is substantially different for the locking on the remainder of the
flags in so_state.
them to behave the same as if the SS_NBIO socket flag had been set
for this call. The SS_NBIO flag for ordinary sockets is set by
fcntl(fd, F_SETFL, O_NONBLOCK).
Pass the MSG_NBIO flag to the soreceive() and sosend() calls in
fifo_read() and fifo_write() instead of frobbing the SS_NBIO flag
on the underlying socket for each I/O operation. The O_NONBLOCK
flag is a property of the descriptor, and unlike ordinary sockets,
fifos may be referenced by multiple descriptors.
mbuma is an Mbuf & Cluster allocator built on top of a number of
extensions to the UMA framework, all included herein.
Extensions to UMA worth noting:
- Better layering between slab <-> zone caches; introduce
Keg structure which splits off slab cache away from the
zone structure and allows multiple zones to be stacked
on top of a single Keg (single type of slab cache);
perhaps we should look into defining a subset API on
top of the Keg for special use by malloc(9),
for example.
- UMA_ZONE_REFCNT zones can now be added, and reference
counters automagically allocated for them within the end
of the associated slab structures. uma_find_refcnt()
does a kextract to fetch the slab struct reference from
the underlying page, and lookup the corresponding refcnt.
mbuma things worth noting:
- integrates mbuf & cluster allocations with extended UMA
and provides caches for commonly-allocated items; defines
several zones (two primary, one secondary) and two kegs.
- change up certain code paths that always used to do:
m_get() + m_clget() to instead just use m_getcl() and
try to take advantage of the newly defined secondary
Packet zone.
- netstat(1) and systat(1) quickly hacked up to do basic
stat reporting but additional stats work needs to be
done once some other details within UMA have been taken
care of and it becomes clearer to how stats will work
within the modified framework.
From the user perspective, one implication is that the
NMBCLUSTERS compile-time option is no longer used. The
maximum number of clusters is still capped off according
to maxusers, but it can be made unlimited by setting
the kern.ipc.nmbclusters boot-time tunable to zero.
Work should be done to write an appropriate sysctl
handler allowing dynamic tuning of kern.ipc.nmbclusters
at runtime.
Additional things worth noting/known issues (READ):
- One report of 'ips' (ServeRAID) driver acting really
slow in conjunction with mbuma. Need more data.
Latest report is that ips is equally sucking with
and without mbuma.
- Giant leak in NFS code sometimes occurs, can't
reproduce but currently analyzing; brueffer is
able to reproduce but THIS IS NOT an mbuma-specific
problem and currently occurs even WITHOUT mbuma.
- Issues in network locking: there is at least one
code path in the rip code where one or more locks
are acquired and we end up in m_prepend() with
M_WAITOK, which causes WITNESS to whine from within
UMA. Current temporary solution: force all UMA
allocations to be M_NOWAIT from within UMA for now
to avoid deadlocks unless WITNESS is defined and we
can determine with certainty that we're not holding
any locks when we're M_WAITOK.
- I've seen at least one weird socketbuffer empty-but-
mbuf-still-attached panic. I don't believe this
to be related to mbuma but please keep your eyes
open, turn on debugging, and capture crash dumps.
This change removes more code than it adds.
A paper is available detailing the change and considering
various performance issues, it was presented at BSDCan2004:
http://www.unixdaemons.com/~bmilekic/netbuf_bmilekic.pdf
Please read the paper for Future Work and implementation
details, as well as credits.
Testing and Debugging:
rwatson,
brueffer,
Ketrien I. Saihr-Kesenchedra,
...
Reviewed by: Lots of people (for different parts)
