Fix a deadlock related to event-base notification. Diagnosed by Zhou Li, Avi Bab...

The problem was that the thread doing the notification could block on
write in evthread_notify_base_default while holding the th_base_lock.
The main thread would never drain th_notify_fd[0], since it would need
th_base_lock to actually trigger events.

diff --git a/event.c b/event.c
index eecabbe221124bd424ae3f8c23a4231600b07693..f68d345c284646510270c36476f3912276afd86b 100644 (file)
--- a/event.c
+++ b/event.c
@@ -1835,7 +1835,7 @@ evthread_notify_base_default(struct event_base *base)
 #else
        r = write(base->th_notify_fd[1], buf, 1);
 #endif
-       return (r < 0) ? -1 : 0;
+       return (r < 0 && errno != EAGAIN) ? -1 : 0;
 }
 
 #if defined(_EVENT_HAVE_EVENTFD) && defined(_EVENT_HAVE_SYS_EVENTFD_H)
@@ -2594,10 +2594,15 @@ evthread_make_base_notifiable(struct event_base *base)
        base->th_notify_fn = notify;
 
        /*
-         This can't be right, can it?  We want writes to this socket to
-         just succeed.
-         evutil_make_socket_nonblocking(base->th_notify_fd[1]);
+         Making the second socket nonblocking is a bit subtle, given that we
+         ignore any EAGAIN returns when writing to it, and you don't usally
+         do that for a nonblocking socket. But if the kernel gives us EAGAIN,
+         then there's no need to add any more data to the buffer, since
+         the main thread is already either about to wake up and drain it,
+         or woken up and in the process of draining it.
        */
+       if (base->th_notify_fd[1] > 0)
+               evutil_make_socket_nonblocking(base->th_notify_fd[1]);
 
        /* prepare an event that we can use for wakeup */
        event_assign(&base->th_notify, base, base->th_notify_fd[0],