socket file descriptor garbage collection code, which is intended to
detect and clear cycles of orphaned file descriptors that are "in-flight"
in a socket when that socket is closed before they are received. The
algorithm present was both run at poor times (resulting in recursion and
reentrance), and also buggy in the presence of parallelism. In order to
fix these problems, make the following changes:
- When there are in-flight sockets and a UNIX domain socket is destroyed,
asynchronously schedule the garbage collector, rather than running it
synchronously in the current context. This avoids lock order issues
when the garbage collection code reenters the UNIX domain socket code,
avoiding lock order reversals, deadlocks, etc. Run the code
asynchronously in a task queue.
- In the garbage collector, when skipping file descriptors that have
entered a closing state (i.e., have f_count == 0), re-test the FDEFER
flag, and decrement unp_defer. As file descriptors can now transition
to a closed state, while the garbage collector is running, it is no
longer the case that unp_defer will remain an accurate count of
deferred sockets in the mark portion of the GC algorithm. Otherwise,
the garbage collector will loop waiting waiting for unp_defer to reach
zero, which it will never do as it is skipping file descriptors that
were marked in an earlier pass, but now closed.
- Acquire the UNIX domain socket subsystem lock in unp_discard() when
modifying the unp_rights counter, or a read/write race is risked with
other threads also manipulating the counter.
While here:
- Remove #if 0'd code regarding acquiring the socket buffer sleep lock in
the garbage collector, this is not required as we are able to use the
socket buffer receive lock to protect scanning the receive buffer for
in-flight file descriptors on the socket buffer.
- Annotate that the description of the garbage collector implementation
is increasingly inaccurate and needs to be updated.
- Add counters of the number of deferred garbage collections and recycled
file descriptors. This will be removed and is here temporarily for
debugging purposes.
With these changes in place, the unp_passfd regression test now appears
to be passed consistently on UP and SMP systems for extended runs,
whereas before it hung quickly or panicked, depending on which bug was
triggered.
Reported by: Philip Kizer <pckizer at nostrum dot com>
MFC after: 2 weeks
- Prefer '_' to ' ', as it results in more easily parsed results in
memory monitoring tools such as vmstat.
- Remove punctuation that is incompatible with using memory type names
as file names, such as '/' characters.
- Disambiguate some collisions by adding subsystem prefixes to some
memory types.
- Generally prefer lower case to upper case.
- If the same type is defined in multiple architecture directories,
attempt to use the same name in additional cases.
Not all instances were caught in this change, so more work is required to
finish this conversion. Similar changes are required for UMA zone names.
following the protocol pru_listen() call to solisten_proto(), so
that it occurs under the socket lock acquisition that also sets
SO_ACCEPTCONN. This requires passing the new backlog parameter
to the protocol, which also allows the protocol to be aware of
changes in queue limit should it wish to do something about the
new queue limit. This continues a move towards the socket layer
acting as a library for the protocol.
Bump __FreeBSD_version due to a change in the in-kernel protocol
interface. This change has been tested with IPv4 and UNIX domain
sockets, but not other protocols.
- Add unp_addsockcred() (for LOCAL_CREDS).
- Add an argument to unp_connect2() to differentiate between
PRU_CONNECT and PRU_CONNECT2. (for LOCAL_CONNWAIT)
Obtained from: NetBSD (with some changes)
a socket from a regular socket to a listening socket able to accept new
connections. As part of this state transition, solisten() calls into the
protocol to update protocol-layer state. There were several bugs in this
implementation that could result in a race wherein a TCP SYN received
in the interval between the protocol state transition and the shortly
following socket layer transition would result in a panic in the TCP code,
as the socket would be in the TCPS_LISTEN state, but the socket would not
have the SO_ACCEPTCONN flag set.
This change does the following:
- Pushes the socket state transition from the socket layer solisten() to
to socket "library" routines called from the protocol. This permits
the socket routines to be called while holding the protocol mutexes,
preventing a race exposing the incomplete socket state transition to TCP
after the TCP state transition has completed. The check for a socket
layer state transition is performed by solisten_proto_check(), and the
actual transition is performed by solisten_proto().
- Holds the socket lock for the duration of the socket state test and set,
and over the protocol layer state transition, which is now possible as
the socket lock is acquired by the protocol layer, rather than vice
versa. This prevents additional state related races in the socket
layer.
This permits the dual transition of socket layer and protocol layer state
to occur while holding locks for both layers, making the two changes
atomic with respect to one another. Similar changes are likely require
elsewhere in the socket/protocol code.
Reported by: Peter Holm <peter@holm.cc>
Review and fixes from: emax, Antoine Brodin <antoine.brodin@laposte.net>
Philosophical head nod: gnn
of the global UNIX domain socket mutex: no protection is needed that
early in the setup of the UNIX domain socket and socket structures.
MFC after: 3 days
occur between a reader and a writer that results in a panic upon close,
e.g.,
"panic: sbflush_locked: cc 4 || mb 0xffffff0052afa400 || mbcnt 0"
Reviewed by: rwatson@
MFC after: 2 weeks
(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>
UNIX domain socket garbage collection implementation, as that risks
holding the mutex over potentially sleeping operations (as well as
introducing some nasty lock order issues, etc). unp_gc() will hold
the lock long enough to do necessary deferal checks and set that it's
running, but then release it until it needs to reset the gc state.
RELENG_5 candidate.
Discussed with: alfred
lock is not held.
Rather than annotating that the lock is released after calls to
unp_detach() with a comment, annotate with an assertion.
Assert that the UNIX domain socket subsystem lock is not held when
unp_externalize() and unp_internalize() are called.
before dereferencing sotounpcb() and checking its value, as so_pcb
is protected by protocol locking, not subsystem locking. This
prevents races during close() by one thread and use of ths socket
in another.
unp_bind() now assert the UNP lock, and uipc_bind() now acquires
the lock around calls to unp_bind().
connect to, re-check that the local UNIX domain socket hasn't been
closed while we slept, and if so, return EINVAL. This affects the
system running both with and without Giant over the network stack,
and recent ULE changes appear to cause it to trigger more
frequently than previously under load. While here, improve catching
of possibly closed UNIX domain sockets in one or two additional
circumstances. I have a much larger set of related changes in
Perforce, but they require more testing before they can be merged.
One debugging printf is left in place to indicate when such a race
takes place: this is typically triggered by a buggy application
that simultaenously connect()'s and close()'s a UNIX domain socket
file descriptor. I'll remove this at some point in the future, but
am interested in seeing how frequently this is reported. In the
case of Martin's reported problem, it appears to be a result of a
non-thread safe syslog() implementation in the C library, which
does not synchronize access to its logging file descriptor.
Reported by: mbr
earlier in unp_connect() so that vp->v_socket can't change between
our copying its value to a local variable and later use of that
variable. This may have been responsible for a panic during
shutdown that I experienced where simultaneous closing of a listen
socket by rpcbind and a new connection being made to rpcbind by
mountd.
associated with performing a wakeup on the socket buffer:
- When performing an sbappend*() followed by a so[rw]wakeup(), explicitly
acquire the socket buffer lock and use the _locked() variants of both
calls. Note that the _locked() sowakeup() versions unlock the mutex on
return. This is done in uipc_send(), divert_packet(), mroute
socket_send(), raw_append(), tcp_reass(), tcp_input(), and udp_append().
- When the socket buffer lock is dropped before a sowakeup(), remove the
explicit unlock and use the _locked() sowakeup() variant. This is done
in soisdisconnecting(), soisdisconnected() when setting the can't send/
receive flags and dropping data, and in uipc_rcvd() which adjusting
back-pressure on the sockets.
For UNIX domain sockets running mpsafe with a contention-intensive SMP
mysql benchmark, this results in a 1.6% query rate improvement due to
reduce mutex costs.
- 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
The big lines are:
NODEV -> NULL
NOUDEV -> NODEV
udev_t -> dev_t
udev2dev() -> findcdev()
Various minor adjustments including handling of userland access to kernel
space struct cdev etc.
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.
SOCK_LOCK(so):
- Hold socket lock over calls to MAC entry points reading or
manipulating socket labels.
- Assert socket lock in MAC entry point implementations.
- When externalizing the socket label, first make a thread-local
copy while holding the socket lock, then release the socket lock
to externalize to userspace.
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
global and allocated variables. This strategy is derived from work
originally developed by BSDi for BSD/OS, and applied to FreeBSD by Sam
Leffler:
- Add unp_mtx, a global mutex which will protect all UNIX domain socket
related variables, structures, etc.
- Add UNP_LOCK(), UNP_UNLOCK(), UNP_LOCK_ASSERT() macros.
- Acquire unp_mtx on entering most UNIX domain socket code,
drop/re-acquire around calls into VFS, and release it on return.
- Avoid performing sodupsockaddr() while holding the mutex, so in general
move to allocating storage before acquiring the mutex to copy the data.
- Make a stack copy of the xucred rather than copying out while holding
unp_mtx. Copy the peer credential out after releasing the mutex.
- Add additional assertions of vnode locks following VOP_CREATE().
A few notes:
- Use of an sx lock for the file list mutex may cause problems with regard
to unp_mtx when garbage collection passed file descriptors.
- The locking in unp_pcblist() for sysctl monitoring is correct subject to
the unpcb zone not returning memory for reuse by other subsystems
(consistent with similar existing concerns).
- Sam's version of this change, as with the BSD/OS version, made use of
both a global lock and per-unpcb locks. However, in practice, the
global lock covered all accesses, so I have simplified out the unpcb
locks in the interest of getting this merged faster (reducing the
overhead but not sacrificing granularity in most cases). We will want
to explore possibilities for improving lock granularity in this code in
the future.
Submitted by: sam
Sponsored by: FreeBSD Foundatiuon
Obtained from: BSD/OS 5 snapshot provided by BSDi
and consume that interface in portalfs and fifofs instead. In the
new world order, unp_connect2() assumes that the unpcb mutex is
held, whereas uipc_connect2() validates that the passed sockets are
UNIX domain sockets, then grabs the mutex.
NB: the portalfs and fifofs code gets down and dirty with UNIX domain
sockets. Maybe this is a bad thing.
functions in kern_socket.c.
Rename the "canwait" field to "mflags" and pass M_WAITOK and M_NOWAIT
in from the caller context rather than "1" or "0".
Correct mflags pass into mac_init_socket() from previous commit to not
include M_ZERO.
Submitted by: sam
in OpenBSD by Niels Provos. The patch introduces a bitmap of allocated
file descriptors which is used to locate available descriptors when a new
one is needed. It also moves the task of growing the file descriptor table
out of fdalloc(), reducing complexity in both fdalloc() and do_dup().
Debts of gratitude are owed to tjr@ (who provided the original patch on
which this work is based), grog@ (for the gdb(4) man page) and rwatson@
(for assistance with pxeboot(8)).
