epoll: atomically remove wait entry on wake up
authorRoman Penyaev <rpenyaev@suse.de>
Fri, 8 May 2020 01:36:16 +0000 (18:36 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Fri, 8 May 2020 02:27:21 +0000 (19:27 -0700)
commit412895f03cbf9633298111cb4dfde13b7720e2c5
tree7e604e64f5f70225195b69898567de8fb2bb7fc3
parent474328c06e3ee75bb6b92826fec90fdc8ef3c573
epoll: atomically remove wait entry on wake up

This patch does two things:

 - fixes a lost wakeup introduced by commit 339ddb53d373 ("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 339ddb53d373):

    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 339ddb53d373):

    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>
fs/eventpoll.c