This patch does two things:
- fixes a lost wakeup introduced by commit 339ddb53d3 ("fs/epoll:
remove unnecessary wakeups of nested epoll")
- improves performance for events delivery.
The description of the problem is the following: if N (>1) threads are
waiting on ep->wq for new events and M (>1) events come, it is quite
likely that >1 wakeups hit the same wait queue entry, because there is
quite a big window between __add_wait_queue_exclusive() and the
following __remove_wait_queue() calls in ep_poll() function.
This can lead to lost wakeups, because thread, which was woken up, can
handle not all the events in ->rdllist. (in better words the problem is
described here: https://lkml.org/lkml/2019/10/7/905)
The idea of the current patch is to use init_wait() instead of
init_waitqueue_entry().
Internally init_wait() sets autoremove_wake_function as a callback,
which removes the wait entry atomically (under the wq locks) from the
list, thus the next coming wakeup hits the next wait entry in the wait
queue, thus preventing lost wakeups.
Problem is very well reproduced by the epoll60 test case [1].
Wait entry removal on wakeup has also performance benefits, because
there is no need to take a ep->lock and remove wait entry from the queue
after the successful wakeup. Here is the timing output of the epoll60
test case:
With explicit wakeup from ep_scan_ready_list() (the state of the
code prior 339ddb53d3):
real 0m6.970s
user 0m49.786s
sys 0m0.113s
After this patch:
real 0m5.220s
user 0m36.879s
sys 0m0.019s
The other testcase is the stress-epoll [2], where one thread consumes
all the events and other threads produce many events:
With explicit wakeup from ep_scan_ready_list() (the state of the
code prior 339ddb53d3):
threads events/ms run-time ms
8 5427 1474
16 6163 2596
32 6824 4689
64 7060 9064
128 6991 18309
After this patch:
threads events/ms run-time ms
8 5598 1429
16 7073 2262
32 7502 4265
64 7640 8376
128 7634 16767
(number of "events/ms" represents event bandwidth, thus higher is
better; number of "run-time ms" represents overall time spent
doing the benchmark, thus lower is better)
[1] tools/testing/selftests/filesystems/epoll/epoll_wakeup_test.c
[2] https://github.com/rouming/test-tools/blob/master/stress-epoll.c
Signed-off-by: Roman Penyaev <rpenyaev@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Jason Baron <jbaron@akamai.com>
Cc: Khazhismel Kumykov <khazhy@google.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Heiher <r@hev.cc>
Cc: <stable@vger.kernel.org>
Link: http://lkml.kernel.org/r/20200430130326.1368509-2-rpenyaev@suse.de
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
In the event that we add to ovflist, before commit 339ddb53d3
("fs/epoll: remove unnecessary wakeups of nested epoll") we would be
woken up by ep_scan_ready_list, and did no wakeup in ep_poll_callback.
With that wakeup removed, if we add to ovflist here, we may never wake
up. Rather than adding back the ep_scan_ready_list wakeup - which was
resulting in unnecessary wakeups, trigger a wake-up in ep_poll_callback.
We noticed that one of our workloads was missing wakeups starting with
339ddb53d3 and upon manual inspection, this wakeup seemed missing to me.
With this patch added, we no longer see missing wakeups. I haven't yet
tried to make a small reproducer, but the existing kselftests in
filesystem/epoll passed for me with this patch.
[khazhy@google.com: use if/elif instead of goto + cleanup suggested by Roman]
Link: http://lkml.kernel.org/r/20200424190039.192373-1-khazhy@google.com
Fixes: 339ddb53d3 ("fs/epoll: remove unnecessary wakeups of nested epoll")
Signed-off-by: Khazhismel Kumykov <khazhy@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Roman Penyaev <rpenyaev@suse.de>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Roman Penyaev <rpenyaev@suse.de>
Cc: Heiher <r@hev.cc>
Cc: Jason Baron <jbaron@akamai.com>
Cc: <stable@vger.kernel.org>
Link: http://lkml.kernel.org/r/20200424025057.118641-1-khazhy@google.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Davidlohr Bueso pointed out that when CONFIG_DEBUG_LOCK_ALLOC is set
ep_poll_safewake() can take several non-raw spinlocks after disabling
interrupts. Since a spinlock can block in the -rt kernel, we can't take a
spinlock after disabling interrupts. So let's re-work how we determine
the nesting level such that it plays nicely with the -rt kernel.
Let's introduce a 'nests' field in struct eventpoll that records the
current nesting level during ep_poll_callback(). Then, if we nest again
we can find the previous struct eventpoll that we were called from and
increase our count by 1. The 'nests' field is protected by
ep->poll_wait.lock.
I've also moved the visited field to reduce the size of struct eventpoll
from 184 bytes to 176 bytes on x86_64 for !CONFIG_DEBUG_LOCK_ALLOC, which
is typical for a production config.
Reported-by: Davidlohr Bueso <dbueso@suse.de>
Signed-off-by: Jason Baron <jbaron@akamai.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Davidlohr Bueso <dbueso@suse.de>
Cc: Roman Penyaev <rpenyaev@suse.de>
Cc: Eric Wong <normalperson@yhbt.net>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Link: http://lkml.kernel.org/r/1582739816-13167-1-git-send-email-jbaron@akamai.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This fixes possible lost wakeup introduced by commit a218cc4914.
Originally modifications to ep->wq were serialized by ep->wq.lock, but
in commit a218cc4914 ("epoll: use rwlock in order to reduce
ep_poll_callback() contention") a new rw lock was introduced in order to
relax fd event path, i.e. callers of ep_poll_callback() function.
After the change ep_modify and ep_insert (both are called on epoll_ctl()
path) were switched to ep->lock, but ep_poll (epoll_wait) was using
ep->wq.lock on wqueue list modification.
The bug doesn't lead to any wqueue list corruptions, because wake up
path and list modifications were serialized by ep->wq.lock internally,
but actual waitqueue_active() check prior wake_up() call can be
reordered with modifications of ep ready list, thus wake up can be lost.
And yes, can be healed by explicit smp_mb():
list_add_tail(&epi->rdlink, &ep->rdllist);
smp_mb();
if (waitqueue_active(&ep->wq))
wake_up(&ep->wp);
But let's make it simple, thus current patch replaces ep->wq.lock with
the ep->lock for wqueue modifications, thus wake up path always observes
activeness of the wqueue correcty.
Fixes: a218cc4914 ("epoll: use rwlock in order to reduce ep_poll_callback() contention")
Reported-by: Max Neunhoeffer <max@arangodb.com>
Signed-off-by: Roman Penyaev <rpenyaev@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Tested-by: Max Neunhoeffer <max@arangodb.com>
Cc: Jakub Kicinski <kuba@kernel.org>
Cc: Christopher Kohlhoff <chris.kohlhoff@clearpool.io>
Cc: Davidlohr Bueso <dbueso@suse.de>
Cc: Jason Baron <jbaron@akamai.com>
Cc: Jes Sorensen <jes.sorensen@gmail.com>
Cc: <stable@vger.kernel.org> [5.1+]
Link: http://lkml.kernel.org/r/20200214170211.561524-1-rpenyaev@suse.de
References: https://bugzilla.kernel.org/show_bug.cgi?id=205933
Bisected-by: Max Neunhoeffer <max@arangodb.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Also make it available outside of epoll, along with the helper that
decides if we need to copy the passed in epoll_event.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Currently, ep_poll_safewake() in the CONFIG_DEBUG_LOCK_ALLOC case uses
ep_call_nested() in order to pass the correct subclass argument to
spin_lock_irqsave_nested(). However, ep_call_nested() adds unnecessary
checks for epoll depth and loops that are already verified when doing
EPOLL_CTL_ADD. This mirrors a conversion that was done for
!CONFIG_DEBUG_LOCK_ALLOC in: commit 37b5e5212a ("epoll: remove
ep_call_nested() from ep_eventpoll_poll()")
Link: http://lkml.kernel.org/r/1567628549-11501-1-git-send-email-jbaron@akamai.com
Signed-off-by: Jason Baron <jbaron@akamai.com>
Reviewed-by: Roman Penyaev <rpenyaev@suse.de>
Cc: Davidlohr Bueso <dave@stgolabs.net>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Eric Wong <normalperson@yhbt.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Add an ID and a device pointer to 'struct wakeup_source'. Use them to to
expose wakeup sources statistics in sysfs under
/sys/class/wakeup/wakeup<ID>/*.
Co-developed-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Co-developed-by: Stephen Boyd <swboyd@chromium.org>
Signed-off-by: Stephen Boyd <swboyd@chromium.org>
Signed-off-by: Tri Vo <trong@android.com>
Tested-by: Kalesh Singh <kaleshsingh@google.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
In the sysctl code the proc_dointvec_minmax() function is often used to
validate the user supplied value between an allowed range. This
function uses the extra1 and extra2 members from struct ctl_table as
minimum and maximum allowed value.
On sysctl handler declaration, in every source file there are some
readonly variables containing just an integer which address is assigned
to the extra1 and extra2 members, so the sysctl range is enforced.
The special values 0, 1 and INT_MAX are very often used as range
boundary, leading duplication of variables like zero=0, one=1,
int_max=INT_MAX in different source files:
$ git grep -E '\.extra[12].*&(zero|one|int_max)' |wc -l
248
Add a const int array containing the most commonly used values, some
macros to refer more easily to the correct array member, and use them
instead of creating a local one for every object file.
This is the bloat-o-meter output comparing the old and new binary
compiled with the default Fedora config:
# scripts/bloat-o-meter -d vmlinux.o.old vmlinux.o
add/remove: 2/2 grow/shrink: 0/2 up/down: 24/-188 (-164)
Data old new delta
sysctl_vals - 12 +12
__kstrtab_sysctl_vals - 12 +12
max 14 10 -4
int_max 16 - -16
one 68 - -68
zero 128 28 -100
Total: Before=20583249, After=20583085, chg -0.00%
[mcroce@redhat.com: tipc: remove two unused variables]
Link: http://lkml.kernel.org/r/20190530091952.4108-1-mcroce@redhat.com
[akpm@linux-foundation.org: fix net/ipv6/sysctl_net_ipv6.c]
[arnd@arndb.de: proc/sysctl: make firmware loader table conditional]
Link: http://lkml.kernel.org/r/20190617130014.1713870-1-arnd@arndb.de
[akpm@linux-foundation.org: fix fs/eventpoll.c]
Link: http://lkml.kernel.org/r/20190430180111.10688-1-mcroce@redhat.com
Signed-off-by: Matteo Croce <mcroce@redhat.com>
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Acked-by: Kees Cook <keescook@chromium.org>
Reviewed-by: Aaron Tomlin <atomlin@redhat.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
task->saved_sigmask and ->restore_sigmask are only used in the ret-from-
syscall paths. This means that set_user_sigmask() can save ->blocked in
->saved_sigmask and do set_restore_sigmask() to indicate that ->blocked
was modified.
This way the callers do not need 2 sigset_t's passed to set/restore and
restore_user_sigmask() renamed to restore_saved_sigmask_unless() turns
into the trivial helper which just calls restore_saved_sigmask().
Link: http://lkml.kernel.org/r/20190606113206.GA9464@redhat.com
Signed-off-by: Oleg Nesterov <oleg@redhat.com>
Cc: Deepa Dinamani <deepa.kernel@gmail.com>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Davidlohr Bueso <dave@stgolabs.net>
Cc: Eric Wong <e@80x24.org>
Cc: Jason Baron <jbaron@akamai.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Al Viro <viro@ZenIV.linux.org.uk>
Cc: Eric W. Biederman <ebiederm@xmission.com>
Cc: David Laight <David.Laight@aculab.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This is the minimal fix for stable, I'll send cleanups later.
Commit 854a6ed568 ("signal: Add restore_user_sigmask()") introduced
the visible change which breaks user-space: a signal temporary unblocked
by set_user_sigmask() can be delivered even if the caller returns
success or timeout.
Change restore_user_sigmask() to accept the additional "interrupted"
argument which should be used instead of signal_pending() check, and
update the callers.
Eric said:
: For clarity. I don't think this is required by posix, or fundamentally to
: remove the races in select. It is what linux has always done and we have
: applications who care so I agree this fix is needed.
:
: Further in any case where the semantic change that this patch rolls back
: (aka where allowing a signal to be delivered and the select like call to
: complete) would be advantage we can do as well if not better by using
: signalfd.
:
: Michael is there any chance we can get this guarantee of the linux
: implementation of pselect and friends clearly documented. The guarantee
: that if the system call completes successfully we are guaranteed that no
: signal that is unblocked by using sigmask will be delivered?
Link: http://lkml.kernel.org/r/20190604134117.GA29963@redhat.com
Fixes: 854a6ed568 ("signal: Add restore_user_sigmask()")
Signed-off-by: Oleg Nesterov <oleg@redhat.com>
Reported-by: Eric Wong <e@80x24.org>
Tested-by: Eric Wong <e@80x24.org>
Acked-by: "Eric W. Biederman" <ebiederm@xmission.com>
Acked-by: Arnd Bergmann <arnd@arndb.de>
Acked-by: Deepa Dinamani <deepa.kernel@gmail.com>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Davidlohr Bueso <dave@stgolabs.net>
Cc: Jason Baron <jbaron@akamai.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Al Viro <viro@ZenIV.linux.org.uk>
Cc: David Laight <David.Laight@ACULAB.COM>
Cc: <stable@vger.kernel.org> [5.0+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Based on 1 normalized pattern(s):
this program is free software you can redistribute it and or modify
it under the terms of the gnu general public license as published by
the free software foundation either version 2 of the license or at
your option any later version
extracted by the scancode license scanner the SPDX license identifier
GPL-2.0-or-later
has been chosen to replace the boilerplate/reference in 3029 file(s).
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Allison Randal <allison@lohutok.net>
Cc: linux-spdx@vger.kernel.org
Link: https://lkml.kernel.org/r/20190527070032.746973796@linutronix.de
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The goal of this patch is to reduce contention of ep_poll_callback()
which can be called concurrently from different CPUs in case of high
events rates and many fds per epoll. Problem can be very well
reproduced by generating events (write to pipe or eventfd) from many
threads, while consumer thread does polling. In other words this patch
increases the bandwidth of events which can be delivered from sources to
the poller by adding poll items in a lockless way to the list.
The main change is in replacement of the spinlock with a rwlock, which
is taken on read in ep_poll_callback(), and then by adding poll items to
the tail of the list using xchg atomic instruction. Write lock is taken
everywhere else in order to stop list modifications and guarantee that
list updates are fully completed (I assume that write side of a rwlock
does not starve, it seems qrwlock implementation has these guarantees).
The following are some microbenchmark results based on the test [1]
which starts threads which generate N events each. The test ends when
all events are successfully fetched by the poller thread:
spinlock
========
threads events/ms run-time ms
8 6402 12495
16 7045 22709
32 7395 43268
rwlock + xchg
=============
threads events/ms run-time ms
8 10038 7969
16 12178 13138
32 13223 24199
According to the results bandwidth of delivered events is significantly
increased, thus execution time is reduced.
This patch was tested with different sort of microbenchmarks and
artificial delays (e.g. "udelay(get_random_int() & 0xff)") introduced
in kernel on paths where items are added to lists.
[1] https://github.com/rouming/test-tools/blob/master/stress-epoll.c
Link: http://lkml.kernel.org/r/20190103150104.17128-5-rpenyaev@suse.de
Signed-off-by: Roman Penyaev <rpenyaev@suse.de>
Cc: Davidlohr Bueso <dbueso@suse.de>
Cc: Jason Baron <jbaron@akamai.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Original comment "Activate ep->ws since epi->ws may get deactivated at
any time" indeed sounds loud, but it is incorrect, because the path
where we check epi->ws is a path where insert to ovflist happens, i.e.
ep_scan_ready_list() has taken ep->mtx and waits for this callback to
finish, thus ep_modify() (which unregisters wakeup source) waits for
ep_scan_ready_list().
Here in this patch I simply call ep_pm_stay_awake_rcu(), which is a bit
extra for this path (indirectly protected by main ep->mtx, so even rcu
is not needed), but I do not want to create another naked
__ep_pm_stay_awake() variant only for this particular case, so rcu variant
is just better for all the cases.
Link: http://lkml.kernel.org/r/20190103150104.17128-4-rpenyaev@suse.de
Signed-off-by: Roman Penyaev <rpenyaev@suse.de>
Cc: Davidlohr Bueso <dbueso@suse.de>
Cc: Jason Baron <jbaron@akamai.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Patch series "use rwlock in order to reduce ep_poll_callback()
contention", v3.
The last patch targets the contention problem in ep_poll_callback(),
which can be very well reproduced by generating events (write to pipe or
eventfd) from many threads, while consumer thread does polling.
The following are some microbenchmark results based on the test [1]
which starts threads which generate N events each. The test ends when
all events are successfully fetched by the poller thread:
spinlock
========
threads events/ms run-time ms
8 6402 12495
16 7045 22709
32 7395 43268
rwlock + xchg
=============
threads events/ms run-time ms
8 10038 7969
16 12178 13138
32 13223 24199
According to the results bandwidth of delivered events is significantly
increased, thus execution time is reduced.
This patch (of 4):
All coming events are stored in FIFO order and this is also should be
applicable to ->ovflist, which originally is stack, i.e. LIFO.
Thus to keep correct FIFO order ->ovflist should reversed by adding
elements to the head of the read list but not to the tail.
Link: http://lkml.kernel.org/r/20190103150104.17128-2-rpenyaev@suse.de
Signed-off-by: Roman Penyaev <rpenyaev@suse.de>
Reviewed-by: Davidlohr Bueso <dbueso@suse.de>
Cc: Jason Baron <jbaron@akamai.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Merge more updates from Andrew Morton:
- procfs updates
- various misc bits
- lib/ updates
- epoll updates
- autofs
- fatfs
- a few more MM bits
* emailed patches from Andrew Morton <akpm@linux-foundation.org>: (58 commits)
mm/page_io.c: fix polled swap page in
checkpatch: add Co-developed-by to signature tags
docs: fix Co-Developed-by docs
drivers/base/platform.c: kmemleak ignore a known leak
fs: don't open code lru_to_page()
fs/: remove caller signal_pending branch predictions
mm/: remove caller signal_pending branch predictions
arch/arc/mm/fault.c: remove caller signal_pending_branch predictions
kernel/sched/: remove caller signal_pending branch predictions
kernel/locking/mutex.c: remove caller signal_pending branch predictions
mm: select HAVE_MOVE_PMD on x86 for faster mremap
mm: speed up mremap by 20x on large regions
mm: treewide: remove unused address argument from pte_alloc functions
initramfs: cleanup incomplete rootfs
scripts/gdb: fix lx-version string output
kernel/kcov.c: mark write_comp_data() as notrace
kernel/sysctl: add panic_print into sysctl
panic: add options to print system info when panic happens
bfs: extra sanity checking and static inode bitmap
exec: separate MM_ANONPAGES and RLIMIT_STACK accounting
...
There is no reason why we rearm the waitiqueue upon every fetch_events
retry (for when events are found yet send_events() fails). If nothing
else, this saves four lock operations per retry, and furthermore reduces
the scope of the lock even further.
[akpm@linux-foundation.org: restore code to original position, fix and reflow comment]
Link: http://lkml.kernel.org/r/20181114182532.27981-2-dave@stgolabs.net
Signed-off-by: Davidlohr Bueso <dbueso@suse.de>
Cc: Jason Baron <jbaron@akamai.com>
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>
It is currently called check_events because it, well, did exactly that.
However, since the lockless ep_events_available() call, the label no
longer checks, but just sends the events. Rename as such.
Link: http://lkml.kernel.org/r/20181114182532.27981-1-dave@stgolabs.net
Signed-off-by: Davidlohr Bueso <dbueso@suse.de>
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Jason Baron <jbaron@akamai.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Upon timeout, we can just exit out of the loop, without the cost of the
changing the task's state with an smp_store_mb call. Just exit out of
the loop and be done - setting the task state afterwards will be, of
course, redundant.
[dave@stgolabs.net: forgotten fixlets]
Link: http://lkml.kernel.org/r/20181109155258.jxcr4t2pnz6zqct3@linux-r8p5
Link: http://lkml.kernel.org/r/20181108051006.18751-7-dave@stgolabs.net
Signed-off-by: Davidlohr Bueso <dave@stgolabs.net>
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Davidlohr Bueso <dbueso@suse.de>
Cc: Jason Baron <jbaron@akamai.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This patch aims at reducing ep wq.lock hold times in epoll_wait(2). For
the blocking case, there is no need to constantly take and drop the
spinlock, which is only needed to manipulate the waitqueue.
The call to ep_events_available() is now lockless, and only exposed to
benign races. Here, if false positive (returns available events and
does not see another thread deleting an epi from the list) we call into
send_events and then the list's state is correctly seen. Otoh, if a
false negative and we don't see a list_add_tail(), for example, from irq
callback, then it is rechecked again before blocking, which will see the
correct state.
In order for more accuracy to see concurrent list_del_init(), use the
list_empty_careful() variant -- of course, this won't be safe against
insertions from wakeup.
For the overflow list we obviously need to prevent load/store tearing as
we don't want to see partial values while the ready list is disabled.
[dave@stgolabs.net: forgotten fixlets]
Link: http://lkml.kernel.org/r/20181109155258.jxcr4t2pnz6zqct3@linux-r8p5
Link: http://lkml.kernel.org/r/20181108051006.18751-6-dave@stgolabs.net
Signed-off-by: Davidlohr Bueso <dbueso@suse.de>
Suggested-by: Jason Baron <jbaron@akamai.com>
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>
Insted of just commenting how important it is, lets make it more robust
and add a lockdep_assert_held() call.
Link: http://lkml.kernel.org/r/20181108051006.18751-5-dave@stgolabs.net
Signed-off-by: Davidlohr Bueso <dbueso@suse.de>
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Jason Baron <jbaron@akamai.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The ep->ovflist is a secondary ready-list to temporarily store events
that might occur when doing sproc without holding the ep->wq.lock. This
accounts for every time we check for ready events and also send events
back to userspace; both callbacks, particularly the latter because of
copy_to_user, can account for a non-trivial time.
As such, the unlikely() check to see if the pointer is being used, seems
both misleading and sub-optimal. In fact, we go to an awful lot of
trouble to sync both lists, and populating the ovflist is far from an
uncommon scenario.
For example, profiling a concurrent epoll_wait(2) benchmark, with
CONFIG_PROFILE_ANNOTATED_BRANCHES shows that for a two threads a 33%
incorrect rate was seen; and when incrementally increasing the number of
epoll instances (which is used, for example for multiple queuing load
balancing models), up to a 90% incorrect rate was seen.
Similarly, by deleting the prediction, 3% throughput boost was seen
across incremental threads.
Link: http://lkml.kernel.org/r/20181108051006.18751-4-dave@stgolabs.net
Signed-off-by: Davidlohr Bueso <dbueso@suse.de>
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Jason Baron <jbaron@akamai.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The current logic is a bit convoluted. Lets simplify this with a
standard list_for_each_entry_safe() loop instead and just break out
after maxevents is reached.
While at it, remove an unnecessary indentation level in the loop when
there are in fact ready events.
Link: http://lkml.kernel.org/r/20181108051006.18751-3-dave@stgolabs.net
Signed-off-by: Davidlohr Bueso <dbueso@suse.de>
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Jason Baron <jbaron@akamai.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Patch series "epoll: some miscellaneous optimizations".
The following are some incremental optimizations on some of the epoll
core. Each patch has the details, but together, the series is seen to
shave off measurable cycles on a number of systems and workloads.
For example, on a 40-core IB, a pipetest as well as parallel
epoll_wait() benchmark show around a 20-30% increase in raw operations
per second when the box is fully occupied (incremental thread counts),
and up to 15% performance improvement with lower counts.
Passes ltp epoll related testcases.
This patch(of 6):
All callers pass the EP_MAX_NESTS constant already, so lets simplify
this a tad and get rid of the redundant parameter for nested eventpolls.
Link: http://lkml.kernel.org/r/20181108051006.18751-2-dave@stgolabs.net
Signed-off-by: Davidlohr Bueso <dbueso@suse.de>
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Jason Baron <jbaron@akamai.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Nobody has actually used the type (VERIFY_READ vs VERIFY_WRITE) argument
of the user address range verification function since we got rid of the
old racy i386-only code to walk page tables by hand.
It existed because the original 80386 would not honor the write protect
bit when in kernel mode, so you had to do COW by hand before doing any
user access. But we haven't supported that in a long time, and these
days the 'type' argument is a purely historical artifact.
A discussion about extending 'user_access_begin()' to do the range
checking resulted this patch, because there is no way we're going to
move the old VERIFY_xyz interface to that model. And it's best done at
the end of the merge window when I've done most of my merges, so let's
just get this done once and for all.
This patch was mostly done with a sed-script, with manual fix-ups for
the cases that weren't of the trivial 'access_ok(VERIFY_xyz' form.
There were a couple of notable cases:
- csky still had the old "verify_area()" name as an alias.
- the iter_iov code had magical hardcoded knowledge of the actual
values of VERIFY_{READ,WRITE} (not that they mattered, since nothing
really used it)
- microblaze used the type argument for a debug printout
but other than those oddities this should be a total no-op patch.
I tried to fix up all architectures, did fairly extensive grepping for
access_ok() uses, and the changes are trivial, but I may have missed
something. Any missed conversion should be trivially fixable, though.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Refactor the logic to restore the sigmask before the syscall
returns into an api.
This is useful for versions of syscalls that pass in the
sigmask and expect the current->sigmask to be changed during
the execution and restored after the execution of the syscall.
With the advent of new y2038 syscalls in the subsequent patches,
we add two more new versions of the syscalls (for pselect, ppoll
and io_pgetevents) in addition to the existing native and compat
versions. Adding such an api reduces the logic that would need to
be replicated otherwise.
Signed-off-by: Deepa Dinamani <deepa.kernel@gmail.com>
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Refactor reading sigset from userspace and updating sigmask
into an api.
This is useful for versions of syscalls that pass in the
sigmask and expect the current->sigmask to be changed during,
and restored after, the execution of the syscall.
With the advent of new y2038 syscalls in the subsequent patches,
we add two more new versions of the syscalls (for pselect, ppoll,
and io_pgetevents) in addition to the existing native and compat
versions. Adding such an api reduces the logic that would need to
be replicated otherwise.
Note that the calls to sigprocmask() ignored the return value
from the api as the function only returns an error on an invalid
first argument that is hardcoded at these call sites.
The updated logic uses set_current_blocked() instead.
Signed-off-by: Deepa Dinamani <deepa.kernel@gmail.com>
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Instead of having each caller pass the rdllink explicitly, just have
ep_is_linked() pass it while the callers just need the epi pointer. This
helper is all about the rdllink, and this change, furthermore, improves
the function's self documentation.
Link: http://lkml.kernel.org/r/20180727053432.16679-3-dave@stgolabs.net
Signed-off-by: Davidlohr Bueso <dbueso@suse.de>
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Jason Baron <jbaron@akamai.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Similar to other calls, ep_poll() is not called with interrupts disabled,
and we can therefore avoid the irq save/restore dance and just disable
local irqs. In fact, the call should never be called in irq context at
all, considering that the only path is
epoll_wait(2) -> do_epoll_wait() -> ep_poll().
When running on a 2 socket 40-core (ht) IvyBridge a common pipe based
epoll_wait(2) microbenchmark, the following performance improvements are
seen:
# threads vanilla dirty
1 1805587 2106412
2 1854064 2090762
4 1805484 2017436
8 1751222 1974475
16 1725299 1962104
32 1378463 1571233
64 787368 900784
Which is a pretty constantly near 15%.
Also add a lockdep check such that we detect any mischief before
deadlocking.
Link: http://lkml.kernel.org/r/20180727053432.16679-2-dave@stgolabs.net
Signed-off-by: Davidlohr Bueso <dbueso@suse.de>
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Jason Baron <jbaron@akamai.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
... 'tis easier on the eye.
[akpm@linux-foundation.org: use inlines rather than macros]
Link: http://lkml.kernel.org/r/20180725185620.11020-1-dave@stgolabs.net
Signed-off-by: Davidlohr Bueso <dbueso@suse.de>
Cc: Jason Baron <jbaron@akamai.com>
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>
Sprinkle lockdep_assert_irqs_enabled() checks in the functions that do not
save and restore interrupts when dealing with the ep->wq.lock. These are
ep_scan_ready_list() and those called by epoll_ctl(): ep_insert, ep_modify
and ep_remove.
[akpm@linux-foundation.org: remove too-obvious comments]
Link: http://lkml.kernel.org/r/20180721183127.3busfa335zlcjeox@linux-r8p5
Signed-off-by: Davidlohr Bueso <dbueso@suse.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Both functions are similar to the context of ep_modify(), called via
epoll_ctl(2). Just like ep_modify(), saving and restoring interrupts is
an overkill in these calls as it will never be called with irqs disabled.
While ep_remove() can be called directly from EPOLL_CTL_DEL, it can also
be called when releasing the file, but this also complies with the above.
Link: http://lkml.kernel.org/r/20180720172956.2883-3-dave@stgolabs.net
Signed-off-by: Davidlohr Bueso <dbueso@suse.de>
Cc: Jason Baron <jbaron@akamai.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Patch series "fs/epoll: loosen irq safety when possible".
Both patches replace saving+restoring interrupts when taking the ep->lock
(now the waitqueue lock), with just disabling local irqs. This shows
immediate performance benefits in patch 1 for an epoll workload running on
Xen. The main concern we need to have with this sort of changes in epoll
is the ep_poll_callback() which is passed to the wait queue wakeup and is
done very often under irq context, this patch does not touch this call.
Patches have been tested pretty heavily with the customer workload,
microbenchmarks, ltp testcases and two high level workloads that use epoll
under the hood: nginx and libevent benchmarks.
This patch (of 2):
Saving and restoring interrupts in ep_scan_ready_list() is an
overkill as it is never called with interrupts disabled. Loosen
this to simply disabling local irqs such that archs where managing
irqs is expensive or virtual environments. This patch yields
some throughput improvements on a workload that is epoll intensive
running on a single Xen DomU.
1 Job 7500 --> 8800 enq/s (+17%)
2 Jobs 14000 --> 15200 enq/s (+8%)
3 Jobs 20500 --> 22300 enq/s (+8%)
4 Jobs 25000 --> 28000 enq/s (+8-12)%
On bare metal:
For a 2-socket 40-core (ht) IvyBridge on a few workloads, unfortunately I
don't have a xen environment and the results for Xen I do have (which
numbers are in patch 1) I don't have the actual workload, so cannot
compare them directly.
1) Different configurations were used for a epoll_wait (pipes io)
microbench (http://linux-scalability.org/epoll/epoll-test.c) and shows
around a 7-10% improvement in overall total number of times the
epoll_wait() loops when using both regular and nested epolls, so very
raw numbers, but measurable nonetheless.
# threads vanilla dirty
1 1677717 1805587
2 1660510 1854064
4 1610184 1805484
8 1577696 1751222
16 1568837 1725299
32 1291532 1378463
64 752584 787368
Note that stddev is pretty small.
2) Another pipe test, which shows no real measurable improvement.
(http://www.xmailserver.org/linux-patches/pipetest.c)
Link: http://lkml.kernel.org/r/20180720172956.2883-2-dave@stgolabs.net
Signed-off-by: Davidlohr Bueso <dbueso@suse.de>
Cc: Jason Baron <jbaron@akamai.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Patch series "waitqueue lockdep annotation", v3.
This series adds a strategic lockdep_assert_held to __wake_up_common to
ensure callers really do hold the wait_queue_head lock when calling the
unlocked wake_up variants. It turns out epoll did not do this for a
fairly common path (hit all the time by systemd during bootup), so the
second patch fixed this instance as well.
This patch (of 3):
The epoll code currently uses the unlocked waitqueue helpers for managing
ep->wq, but instead of holding the waitqueue lock around these calls, it
uses its own ep->lock spinlock. Given that the waitqueue is not exposed
to the rest of the kernel this actually works ok at the moment, but
prevents the epoll locking rules from being enforced using lockdep.
Remove ep->lock and use the waitqueue lock to not only reduce the size of
struct eventpoll but also to make sure we can assert locking invariants in
the waitqueue code.
Link: http://lkml.kernel.org/r/20171214152344.6880-2-hch@lst.de
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Jason Baron <jbaron@akamai.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Mike Rapoport <rppt@linux.vnet.ibm.com>
Cc: Jason Baron <jbaron@akamai.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Davidlohr Bueso <dave@stgolabs.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The poll() changes were not well thought out, and completely
unexplained. They also caused a huge performance regression, because
"->poll()" was no longer a trivial file operation that just called down
to the underlying file operations, but instead did at least two indirect
calls.
Indirect calls are sadly slow now with the Spectre mitigation, but the
performance problem could at least be largely mitigated by changing the
"->get_poll_head()" operation to just have a per-file-descriptor pointer
to the poll head instead. That gets rid of one of the new indirections.
But that doesn't fix the new complexity that is completely unwarranted
for the regular case. The (undocumented) reason for the poll() changes
was some alleged AIO poll race fixing, but we don't make the common case
slower and more complex for some uncommon special case, so this all
really needs way more explanations and most likely a fundamental
redesign.
[ This revert is a revert of about 30 different commits, not reverted
individually because that would just be unnecessarily messy - Linus ]
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Christoph Hellwig <hch@lst.de>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
These abstract out calls to the poll method in preparation for changes
in how we poll.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Using the helper functions do_epoll_create() and do_epoll_wait() allows us
to remove in-kernel calls to the related syscall functions.
This patch is part of a series which removes in-kernel calls to syscalls.
On this basis, the syscall entry path can be streamlined. For details, see
http://lkml.kernel.org/r/20180325162527.GA17492@light.dominikbrodowski.net
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Dominik Brodowski <linux@dominikbrodowski.net>
This is the mindless scripted replacement of kernel use of POLL*
variables as described by Al, done by this script:
for V in IN OUT PRI ERR RDNORM RDBAND WRNORM WRBAND HUP RDHUP NVAL MSG; do
L=`git grep -l -w POLL$V | grep -v '^t' | grep -v /um/ | grep -v '^sa' | grep -v '/poll.h$'|grep -v '^D'`
for f in $L; do sed -i "-es/^\([^\"]*\)\(\<POLL$V\>\)/\\1E\\2/" $f; done
done
with de-mangling cleanups yet to come.
NOTE! On almost all architectures, the EPOLL* constants have the same
values as the POLL* constants do. But they keyword here is "almost".
For various bad reasons they aren't the same, and epoll() doesn't
actually work quite correctly in some cases due to this on Sparc et al.
The next patch from Al will sort out the final differences, and we
should be all done.
Scripted-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
two callers that do so don't need to bother - we'd already
masked it with epi->event.events, which
* couldn't have changed since we are holding ->mtx
* had been set to event->events
* is still equal to event->events, since *event is never
changed by anything.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
__poll_t is also used as wait key in some waitqueues.
Verify that wait_..._poll() gets __poll_t as key and
provide a helper for wakeup functions to get back to
that __poll_t value.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
The use of ep_call_nested() in ep_eventpoll_poll(), which is the .poll
routine for an epoll fd, is used to prevent excessively deep epoll
nesting, and to prevent circular paths.
However, we are already preventing these conditions during
EPOLL_CTL_ADD. In terms of too deep epoll chains, we do in fact allow
deep nesting of the epoll fds themselves (deeper than EP_MAX_NESTS),
however we don't allow more than EP_MAX_NESTS when an epoll file
descriptor is actually connected to a wakeup source. Thus, we do not
require the use of ep_call_nested(), since ep_eventpoll_poll(), which is
called via ep_scan_ready_list() only continues nesting if there are
events available.
Since ep_call_nested() is implemented using a global lock, applications
that make use of nested epoll can see large performance improvements
with this change.
Davidlohr said:
: Improvements are quite obscene actually, such as for the following
: epoll_wait() benchmark with 2 level nesting on a 80 core IvyBridge:
:
: ncpus vanilla dirty delta
: 1 2447092 3028315 +23.75%
: 4 231265 2986954 +1191.57%
: 8 121631 2898796 +2283.27%
: 16 59749 2902056 +4757.07%
: 32 26837 2326314 +8568.30%
: 64 12926 1341281 +10276.61%
:
: (http://linux-scalability.org/epoll/epoll-test.c)
Link: http://lkml.kernel.org/r/1509430214-5599-1-git-send-email-jbaron@akamai.com
Signed-off-by: Jason Baron <jbaron@akamai.com>
Cc: Davidlohr Bueso <dave@stgolabs.net>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Salman Qazi <sqazi@google.com>
Cc: Hou Tao <houtao1@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
ep_poll_safewake() is used to wakeup potentially nested epoll file
descriptors. The function uses ep_call_nested() to prevent entering the
same wake up queue more than once, and to prevent excessively deep
wakeup paths (deeper than EP_MAX_NESTS). However, this is not necessary
since we are already preventing these conditions during EPOLL_CTL_ADD.
This saves extra function calls, and avoids taking a global lock during
the ep_call_nested() calls.
I have, however, left ep_call_nested() for the CONFIG_DEBUG_LOCK_ALLOC
case, since ep_call_nested() keeps track of the nesting level, and this
is required by the call to spin_lock_irqsave_nested(). It would be nice
to remove the ep_call_nested() calls for the CONFIG_DEBUG_LOCK_ALLOC
case as well, however its not clear how to simply pass the nesting level
through multiple wake_up() levels without more surgery. In any case, I
don't think CONFIG_DEBUG_LOCK_ALLOC is generally used for production.
This patch, also apparently fixes a workload at Google that Salman Qazi
reported by completely removing the poll_safewake_ncalls->lock from
wakeup paths.
Link: http://lkml.kernel.org/r/1507920533-8812-1-git-send-email-jbaron@akamai.com
Signed-off-by: Jason Baron <jbaron@akamai.com>
Acked-by: Davidlohr Bueso <dbueso@suse.de>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Salman Qazi <sqazi@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
A userspace application can directly trigger the allocations from
eventpoll_epi and eventpoll_pwq slabs. A buggy or malicious application
can consume a significant amount of system memory by triggering such
allocations. Indeed we have seen in production where a buggy
application was leaking the epoll references and causing a burst of
eventpoll_epi and eventpoll_pwq slab allocations. This patch opt-in the
charging of eventpoll_epi and eventpoll_pwq slabs.
There is a per-user limit (~4% of total memory if no highmem) on these
caches. I think it is too generous particularly in the scenario where
jobs of multiple users are running on the system and the administrator
is reducing cost by overcomitting the memory. This is unaccounted
kernel memory and will not be considered by the oom-killer. I think by
accounting it to kmemcg, for systems with kmem accounting enabled, we
can provide better isolation between jobs of different users.
Link: http://lkml.kernel.org/r/20171003021519.23907-1-shakeelb@google.com
Signed-off-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>