eventpoll: wake the next waiter when a woken waiter is killed

Processes sharing an epoll instance can sleep forever while an event is
ready.  An event wakes only one waiter, which then wakes the next one in
ep_done_scan() if events are still ready.  Since commit c257a340ed
("fs, epoll: short circuit fetching events if thread has been killed"),
ep_send_events() returns before the scan when the waiter has a fatal
signal.  If the woken waiter is killed before it runs, nobody wakes the
next one:

  Task A                     Task B                 Task C
  epoll_wait(ep, -1)         epoll_wait(ep, -1)
                                                    write(efd)
                                                      /* wakes A only */
                                                    kill(A, SIGKILL)
  ep_send_events()
    return -EINTR
      /* skips ep_done_scan() */
                             /* still asleep */

B then sleeps until another file on the instance becomes ready, if ever.
This can hit eg a pre-fork server whose worker gets OOM killed.  Commit
7fab29e356 ("fs/epoll: restore waking from ep_done_scan()") fixed the
same kind of hang when the wakeup in ep_done_scan() was removed.

Keep the early return so that a dying task still doesn't take events
nobody will see, but wake the next waiter there if events are ready, as
ep_done_scan() would have.  With the fix, B gets the event.

This bug was identified and confirmed with the help of a deterministic
hypervisor currently under development.  The C reproducer below however
hits it within about 2s.

Fixes: c257a340ed ("fs, epoll: short circuit fetching events if thread has been killed")
Cc: [email protected]
Signed-off-by: Mani Tofigh <[email protected]>
Link: https://patch.msgid.link/[email protected]
Signed-off-by: Christian Brauner (Amutable) <[email protected]>
This commit is contained in:
Mani Tofigh
2026-11-05 18:22:01 +01:00
committed by Christian Brauner
parent 925d238324
commit 9f4413e522
+9 -2
View File
@@ -2144,10 +2144,17 @@ static int ep_send_events(struct eventpoll *ep,
/*
* Always short-circuit for fatal signals to allow threads to make a
* timely exit without the chance of finding more events available and
* fetching repeatedly.
* fetching repeatedly. The wakeup that got us here may have been the
* only one for the events on ->rdllist, so pass it on to the next
* waiter, as ep_done_scan() would have.
*/
if (fatal_signal_pending(current))
if (fatal_signal_pending(current)) {
spin_lock_irq(&ep->lock);
if (!list_empty(&ep->rdllist))
wake_up(&ep->wq);
spin_unlock_irq(&ep->lock);
return -EINTR;
}
init_poll_funcptr(&pt, NULL);