[Libevent-users] Re: shutdown behavior of bufferevents

[Roman Gershman <roman@xxxxxxxxx>]

Following my first question, I think I found a deadlock bug.

During the program shutdown, I call from a thread different from base event thread.

 bufferevent_setcb(bev, NULL, NULL, NULL, NULL);

 bufferevent_disable(bev, EV_READ|EV_WRITE);

Sometimes the call to bufferevent_disable deadlocks at the following line at event.c:2220:

 if (base->current_event == ev && !EVBASE_IN_THREAD(base)) {

         ++base->current_event_waiters;

         EVTHREAD_COND_WAIT(base->current_event_cond, base->th_base_lock);

 }

The other thread that is holding the lock is in bufferevent_readcb() function at bufferevent_sock.c:133

_bufferevent_incref_and_lock(bufev);

it's deadlocked because the first thread is already holding the lock of bev at line bufferevent.c:502 (in function bufferevent_disable).

Cheers,

On Fri, Jun 21, 2013 at 9:23 PM, Roman Gershman <roman@xxxxxxxxx> wrote:

Hi,

I am trying to understand the invariants the libevent library provides when we free its resources. 

Specifically, if I call bufferevent_free(my_bev) in a thread different from the main event_loop thread, is it guaranteed that my_bev's callbacks will never run after bufferevent_free call completes?

Moreover, if the callbacks were running during the call to bufferevent_free(), is it guaranteed that they finished to run as well? Does BEV_OPT_UNLOCK_CALLBACKS affect this behavior?

Note that I do not use BEV_OPT_DEFER_CALLBACKS flag.

I am using 2.0.19-stable version of libevent library. 

Thanks!