[asterisk-bugs] [JIRA] (ASTERISK-29217) LOCK() can grant the same lock to multiple channels spuriously

[Jaco Kroon (JIRA)]

Jaco Kroon created ASTERISK-29217:
-------------------------------------

             Summary: LOCK() can grant the same lock to multiple channels spuriously
                 Key: ASTERISK-29217
                 URL: https://issues.asterisk.org/jira/browse/ASTERISK-29217
             Project: Asterisk
          Issue Type: Bug
      Security Level: None
          Components: Functions/func_lock
    Affects Versions: 18.1.0, 17.9.0, 16.15.0, 13.38.0
            Reporter: Jaco Kroon
            Severity: Critical


Due to a misunderstanding of the wait conditions work in the func/func_lock.c code it's possible that multiple channels will be informed of a successful LOCK(), resulting in bad stuff happening.

Under low lock pressure scenarios this is unlikely to happen (ie, no more than 1 channel should be blocking on LOCK()), but the moment multiple channels are waiting it becomes (but is not guaranteed) possible for multiple channels to be informed they've successfully locked.

Technical details:

>From PTHREAD_COND_BROADCAST(3P):

       On a multi-processor, it may be impossible for an implementation of  pthread_cond_signal()
       to avoid the unblocking of more than one thread blocked on a condition variable. For exam‐
       ple,  consider  the  following   partial   implementation   of   pthread_cond_wait()   and
       pthread_cond_signal(), executed by two threads in the order given. One thread is trying to
       wait on the condition variable, another is concurrently  executing  pthread_cond_signal(),
       while a third thread is already waiting.

I won't bore you with the sample code, but what this says is that multiple threads may be woken.  func_lock assumes that it's the only one and assumes it owns the lock on successful return from ast_cond_timedwait (which wraps the pthread_cond_timedwait).

So let's say two threads gets released here.

At this point the threads will sequentially be released (since the mutex gets grabbed again).

The first thread will set ->owner to itself, and increment ->count to 1.

The second thread will set ->owner to itself, and increment ->count to 2.

When the first thread calls UNLOCK() it will be informed it doens't own the lock.

When the second thread calls UNLOCK() it will decrement ->count to 1, and not release the lock until the channel is destroyed.


The lock_broker thread probably further aggravates the problems here, so my patch will elimate it's use as well.





--
This message was sent by Atlassian JIRA
(v6.2#6252)