[ruby-core:23489] [Backport #1471] "Mutual join" deadlock detection faulty in 1.8.6 and 1.8.7

Issue #1471 has been updated by John Carter.


Thanks for the fix...

Unfortunately I think it introduces another bug.

Your patch add invokes rb_thread_set_join from adjust_join from wake_one from unlock_mutex_inner.

Alas, wake_one is _also_ invoked from signal_condvar which is invoked from signal_condvar_call from rb_condvar_signal. Thus all the join pointer of all threads waiting on that conditionvariable will be updated to point at the first thread on the condvar waiting list. Which assumes, perhaps over boldly, that the woken thread will be the one to get the lock.

Now the thread that has just been woken should have be at a ConditionVariable.wait(mutex) point, having come through rb_condvar_wait->wait_condvar->ensure(wait_list,,lock_mutex_call) which will do the right thing.

ie. The call to adjust_join should be moved out of "wake_one" and into "unlock_mutex_inner"

The other place that's going to invoke adjust_join, but shouldn't, is wake_all.
----------------------------------------
http://redmine.ruby-lang.org/issues/show/1471

----------------------------------------
http://redmine.ruby-lang.org