Currently, epoll file descriptors or epfds (the fd returned from
epoll_create[1]()) that are added to a shared wakeup source are always
added in a non-exclusive manner. This means that when we have multiple
epfds attached to a shared fd source they are all woken up. This creates
thundering herd type behavior.
Introduce a new 'EPOLLEXCLUSIVE' flag that can be passed as part of the
'event' argument during an epoll_ctl() EPOLL_CTL_ADD operation. This new
flag allows for exclusive wakeups when there are multiple epfds attached
to a shared fd event source.
The implementation walks the list of exclusive waiters, and queues an
event to each epfd, until it finds the first waiter that has threads
blocked on it via epoll_wait(). The idea is to search for threads which
are idle and ready to process the wakeup events. Thus, we queue an event
to at least 1 epfd, but may still potentially queue an event to all epfds
that are attached to the shared fd source.
Performance testing was done by Madars Vitolins using a modified version
of Enduro/X. The use of the 'EPOLLEXCLUSIVE' flag reduce the length of
this particular workload from 860s down to 24s.
Sample epoll_clt text:
EPOLLEXCLUSIVE
Sets an exclusive wakeup mode for the epfd file descriptor that is
being attached to the target file descriptor, fd. Thus, when an event
occurs and multiple epfd file descriptors are attached to the same
target file using EPOLLEXCLUSIVE, one or more epfds will receive an
event with epoll_wait(2). The default in this scenario (when
EPOLLEXCLUSIVE is not set) is for all epfds to receive an event.
EPOLLEXCLUSIVE may only be specified with the op EPOLL_CTL_ADD.
Signed-off-by: Jason Baron <jbaron@akamai.com>
Tested-by: Madars Vitolins <m@silodev.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Al Viro <viro@ftp.linux.org.uk>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Eric Wong <normalperson@yhbt.net>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Hagen Paul Pfeifer <hagen@jauu.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
After waking up a task waiting for an event, we explicitly mark it as
TASK_RUNNING (which is necessary as we do the checks for wakeups as
TASK_INTERRUPTIBLE). Once running and dealing with actually delivering
the events, we're obviously not planning on calling schedule, thus we can
relax the implied barrier and simply update the state with
__set_current_state().
Signed-off-by: Davidlohr Bueso <dbueso@suse.de>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
When calling epoll_ctl with operation EPOLL_CTL_DEL, structure epds is
not initialized but ep_take_care_of_epollwakeup reads its event field.
When this unintialized field has EPOLLWAKEUP bit set, a capability check
is done for CAP_BLOCK_SUSPEND in ep_take_care_of_epollwakeup. This
produces unexpected messages in the audit log, such as (on a system
running SELinux):
type=AVC msg=audit(1408212798.866:410): avc: denied
{ block_suspend } for pid=7754 comm="dbus-daemon" capability=36
scontext=unconfined_u:unconfined_r:unconfined_t
tcontext=unconfined_u:unconfined_r:unconfined_t
tclass=capability2 permissive=1
type=SYSCALL msg=audit(1408212798.866:410): arch=c000003e syscall=233
success=yes exit=0 a0=3 a1=2 a2=9 a3=7fffd4d66ec0 items=0 ppid=1
pid=7754 auid=1000 uid=0 gid=0 euid=0 suid=0 fsuid=0 egid=0 sgid=0
fsgid=0 tty=(none) ses=3 comm="dbus-daemon"
exe="/usr/bin/dbus-daemon"
subj=unconfined_u:unconfined_r:unconfined_t key=(null)
("arch=c000003e syscall=233 a1=2" means "epoll_ctl(op=EPOLL_CTL_DEL)")
Remove use of epds in epoll_ctl when op == EPOLL_CTL_DEL.
Fixes: 4d7e30d989 ("epoll: Add a flag, EPOLLWAKEUP, to prevent suspend while epoll events are ready")
Signed-off-by: Nicolas Iooss <nicolas.iooss_linux@m4x.org>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Arve Hjønnevåg <arve@android.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This fixes use-after-free of epi->fllink.next inside list loop macro.
This loop actually releases elements in the body. The list is
rcu-protected but here we cannot hold rcu_read_lock because we need to
lock mutex inside.
The obvious solution is to use list_for_each_entry_safe(). RCU-ness
isn't essential because nobody can change this list under us, it's final
fput for this file.
The bug was introduced by ae10b2b4eb ("epoll: optimize EPOLL_CTL_DEL
using rcu")
Signed-off-by: Konstantin Khlebnikov <koct9i@gmail.com>
Reported-by: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Stable <stable@vger.kernel.org> # 3.13+
Cc: Sasha Levin <sasha.levin@oracle.com>
Cc: Jason Baron <jbaron@akamai.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The EPOLL_CTL_DEL path of epoll contains a classic, ab-ba deadlock.
That is, epoll_ctl(a, EPOLL_CTL_DEL, b, x), will deadlock with
epoll_ctl(b, EPOLL_CTL_DEL, a, x). The deadlock was introduced with
commmit 67347fe4e6 ("epoll: do not take global 'epmutex' for simple
topologies").
The acquistion of the ep->mtx for the destination 'ep' was added such
that a concurrent EPOLL_CTL_ADD operation would see the correct state of
the ep (Specifically, the check for '!list_empty(&f.file->f_ep_links')
However, by simply not acquiring the lock, we do not serialize behind
the ep->mtx from the add path, and thus may perform a full path check
when if we had waited a little longer it may not have been necessary.
However, this is a transient state, and performing the full loop
checking in this case is not harmful.
The important point is that we wouldn't miss doing the full loop
checking when required, since EPOLL_CTL_ADD always locks any 'ep's that
its operating upon. The reason we don't need to do lock ordering in the
add path, is that we are already are holding the global 'epmutex'
whenever we do the double lock. Further, the original posting of this
patch, which was tested for the intended performance gains, did not
perform this additional locking.
Signed-off-by: Jason Baron <jbaron@akamai.com>
Cc: Nathan Zimmer <nzimmer@sgi.com>
Cc: Eric Wong <normalperson@yhbt.net>
Cc: Nelson Elhage <nelhage@nelhage.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Davide Libenzi <davidel@xmailserver.org>
Cc: "Paul E. McKenney" <paulmck@us.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Merge first patch-bomb from Andrew Morton:
"Quite a lot of other stuff is banked up awaiting further
next->mainline merging, but this batch contains:
- Lots of random misc patches
- OCFS2
- Most of MM
- backlight updates
- lib/ updates
- printk updates
- checkpatch updates
- epoll tweaking
- rtc updates
- hfs
- hfsplus
- documentation
- procfs
- update gcov to gcc-4.7 format
- IPC"
* emailed patches from Andrew Morton <akpm@linux-foundation.org>: (269 commits)
ipc, msg: fix message length check for negative values
ipc/util.c: remove unnecessary work pending test
devpts: plug the memory leak in kill_sb
./Makefile: export initial ramdisk compression config option
init/Kconfig: add option to disable kernel compression
drivers: w1: make w1_slave::flags long to avoid memory corruption
drivers/w1/masters/ds1wm.cuse dev_get_platdata()
drivers/memstick/core/ms_block.c: fix unreachable state in h_msb_read_page()
drivers/memstick/core/mspro_block.c: fix attributes array allocation
drivers/pps/clients/pps-gpio.c: remove redundant of_match_ptr
kernel/panic.c: reduce 1 byte usage for print tainted buffer
gcov: reuse kbasename helper
kernel/gcov/fs.c: use pr_warn()
kernel/module.c: use pr_foo()
gcov: compile specific gcov implementation based on gcc version
gcov: add support for gcc 4.7 gcov format
gcov: move gcov structs definitions to a gcc version specific file
kernel/taskstats.c: return -ENOMEM when alloc memory fails in add_del_listener()
kernel/taskstats.c: add nla_nest_cancel() for failure processing between nla_nest_start() and nla_nest_end()
kernel/sysctl_binary.c: use scnprintf() instead of snprintf()
...
Pull vfs updates from Al Viro:
"All kinds of stuff this time around; some more notable parts:
- RCU'd vfsmounts handling
- new primitives for coredump handling
- files_lock is gone
- Bruce's delegations handling series
- exportfs fixes
plus misc stuff all over the place"
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs: (101 commits)
ecryptfs: ->f_op is never NULL
locks: break delegations on any attribute modification
locks: break delegations on link
locks: break delegations on rename
locks: helper functions for delegation breaking
locks: break delegations on unlink
namei: minor vfs_unlink cleanup
locks: implement delegations
locks: introduce new FL_DELEG lock flag
vfs: take i_mutex on renamed file
vfs: rename I_MUTEX_QUOTA now that it's not used for quotas
vfs: don't use PARENT/CHILD lock classes for non-directories
vfs: pull ext4's double-i_mutex-locking into common code
exportfs: fix quadratic behavior in filehandle lookup
exportfs: better variable name
exportfs: move most of reconnect_path to helper function
exportfs: eliminate unused "noprogress" counter
exportfs: stop retrying once we race with rename/remove
exportfs: clear DISCONNECTED on all parents sooner
exportfs: more detailed comment for path_reconnect
...
When calling EPOLL_CTL_ADD for an epoll file descriptor that is attached
directly to a wakeup source, we do not need to take the global 'epmutex',
unless the epoll file descriptor is nested. The purpose of taking the
'epmutex' on add is to prevent complex topologies such as loops and deep
wakeup paths from forming in parallel through multiple EPOLL_CTL_ADD
operations. However, for the simple case of an epoll file descriptor
attached directly to a wakeup source (with no nesting), we do not need to
hold the 'epmutex'.
This patch along with 'epoll: optimize EPOLL_CTL_DEL using rcu' improves
scalability on larger systems. Quoting Nathan Zimmer's mail on SPECjbb
performance:
"On the 16 socket run the performance went from 35k jOPS to 125k jOPS. In
addition the benchmark when from scaling well on 10 sockets to scaling
well on just over 40 sockets.
...
Currently the benchmark stops scaling at around 40-44 sockets but it seems like
I found a second unrelated bottleneck."
[akpm@linux-foundation.org: use `bool' for boolean variables, remove unneeded/undesirable cast of void*, add missed ep_scan_ready_list() kerneldoc]
Signed-off-by: Jason Baron <jbaron@akamai.com>
Tested-by: Nathan Zimmer <nzimmer@sgi.com>
Cc: Eric Wong <normalperson@yhbt.net>
Cc: Nelson Elhage <nelhage@nelhage.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Davide Libenzi <davidel@xmailserver.org>
Cc: "Paul E. McKenney" <paulmck@us.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Nathan Zimmer found that once we get over 10+ cpus, the scalability of
SPECjbb falls over due to the contention on the global 'epmutex', which is
taken in on EPOLL_CTL_ADD and EPOLL_CTL_DEL operations.
Patch #1 removes the 'epmutex' lock completely from the EPOLL_CTL_DEL path
by using rcu to guard against any concurrent traversals.
Patch #2 remove the 'epmutex' lock from EPOLL_CTL_ADD operations for
simple topologies. IE when adding a link from an epoll file descriptor to
a wakeup source, where the epoll file descriptor is not nested.
This patch (of 2):
Optimize EPOLL_CTL_DEL such that it does not require the 'epmutex' by
converting the file->f_ep_links list into an rcu one. In this way, we can
traverse the epoll network on the add path in parallel with deletes.
Since deletes can't create loops or worse wakeup paths, this is safe.
This patch in combination with the patch "epoll: Do not take global 'epmutex'
for simple topologies", shows a dramatic performance improvement in
scalability for SPECjbb.
Signed-off-by: Jason Baron <jbaron@akamai.com>
Tested-by: Nathan Zimmer <nzimmer@sgi.com>
Cc: Eric Wong <normalperson@yhbt.net>
Cc: Nelson Elhage <nelhage@nelhage.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Davide Libenzi <davidel@xmailserver.org>
Cc: "Paul E. McKenney" <paulmck@us.ibm.com>
CC: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This reverts commit 1c441e9212 (epoll: use freezable blocking call)
which is reported to cause user space memory corruption to happen
after suspend to RAM.
Since it appears to be extremely difficult to root cause this
problem, it is best to revert the offending commit and try to address
the original issue in a better way later.
References: https://bugzilla.kernel.org/show_bug.cgi?id=61781
Reported-by: Natrio <natrio@list.ru>
Reported-by: Jeff Pohlmeyer <yetanothergeek@gmail.com>
Bisected-by: Leo Wolf <jclw@ymail.com>
Fixes: 1c441e9212 (epoll: use freezable blocking call)
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Cc: 3.11+ <stable@vger.kernel.org> # 3.11+
Merge first patch-bomb from Andrew Morton:
- various misc bits
- I'm been patchmonkeying ocfs2 for a while, as Joel and Mark have been
distracted. There has been quite a bit of activity.
- About half the MM queue
- Some backlight bits
- Various lib/ updates
- checkpatch updates
- zillions more little rtc patches
- ptrace
- signals
- exec
- procfs
- rapidio
- nbd
- aoe
- pps
- memstick
- tools/testing/selftests updates
* emailed patches from Andrew Morton <akpm@linux-foundation.org>: (445 commits)
tools/testing/selftests: don't assume the x bit is set on scripts
selftests: add .gitignore for kcmp
selftests: fix clean target in kcmp Makefile
selftests: add .gitignore for vm
selftests: add hugetlbfstest
self-test: fix make clean
selftests: exit 1 on failure
kernel/resource.c: remove the unneeded assignment in function __find_resource
aio: fix wrong comment in aio_complete()
drivers/w1/slaves/w1_ds2408.c: add magic sequence to disable P0 test mode
drivers/memstick/host/r592.c: convert to module_pci_driver
drivers/memstick/host/jmb38x_ms: convert to module_pci_driver
pps-gpio: add device-tree binding and support
drivers/pps/clients/pps-gpio.c: convert to module_platform_driver
drivers/pps/clients/pps-gpio.c: convert to devm_* helpers
drivers/parport/share.c: use kzalloc
Documentation/accounting/getdelays.c: avoid strncpy in accounting tool
aoe: update internal version number to v83
aoe: update copyright date
aoe: perform I/O completions in parallel
...
Avoid waking up every thread sleeping in an epoll_wait call during
suspend and resume by calling a freezable blocking call. Previous
patches modified the freezer to avoid sending wakeups to threads
that are blocked in freezable blocking calls.
This call was selected to be converted to a freezable call because
it doesn't hold any locks or release any resources when interrupted
that might be needed by another freezing task or a kernel driver
during suspend, and is a common site where idle userspace tasks are
blocked.
Acked-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Colin Cross <ccross@android.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Pull compat cleanup from Al Viro:
"Mostly about syscall wrappers this time; there will be another pile
with patches in the same general area from various people, but I'd
rather push those after both that and vfs.git pile are in."
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/signal:
syscalls.h: slightly reduce the jungles of macros
get rid of union semop in sys_semctl(2) arguments
make do_mremap() static
sparc: no need to sign-extend in sync_file_range() wrapper
ppc compat wrappers for add_key(2) and request_key(2) are pointless
x86: trim sys_ia32.h
x86: sys32_kill and sys32_mprotect are pointless
get rid of compat_sys_semctl() and friends in case of ARCH_WANT_OLD_COMPAT_IPC
merge compat sys_ipc instances
consolidate compat lookup_dcookie()
convert vmsplice to COMPAT_SYSCALL_DEFINE
switch getrusage() to COMPAT_SYSCALL_DEFINE
switch epoll_pwait to COMPAT_SYSCALL_DEFINE
convert sendfile{,64} to COMPAT_SYSCALL_DEFINE
switch signalfd{,4}() to COMPAT_SYSCALL_DEFINE
make SYSCALL_DEFINE<n>-generated wrappers do asmlinkage_protect
make HAVE_SYSCALL_WRAPPERS unconditional
consolidate cond_syscall and SYSCALL_ALIAS declarations
teach SYSCALL_DEFINE<n> how to deal with long long/unsigned long long
get rid of duplicate logics in __SC_....[1-6] definitions
It is common for epoll users to have thousands of epitems, so saving a
cache line on every allocation leads to large memory savings.
Since epitem allocations are cache-aligned, reducing sizeof(struct
epitem) from 136 bytes to 128 bytes will allow it to squeeze under a
cache line boundary on x86_64.
Via /sys/kernel/slab/eventpoll_epi, I see the following changes on my
x86_64 Core2 Duo (which has 64-byte cache alignment):
object_size : 192 => 128
objs_per_slab: 21 => 32
Also, add a BUILD_BUG_ON() to check for future accidental breakage.
[akpm@linux-foundation.org: use __packed, for all architectures]
Signed-off-by: Eric Wong <normalperson@yhbt.net>
Cc: Davide Libenzi <davidel@xmailserver.org>
Cc: Al Viro <viro@ZenIV.linux.org.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Jason Baron
Davidlohr Bueso
Joe Perches
Nicolas Iooss
Konstantin Khlebnikov
Amit Pundir
Linus Torvalds
Rafael J. Wysocki
Al Viro
Eric Dumazet
Oleg Nesterov
Colin Cross
Eric Wong