Because the close method does not lock and the read method does not check the closed value after locking, it's possible for a deadlock to occur under the following series of events:
Thread 1 (reader): read last element from the channel
Thread 1: call operator>>
Thread 1: check closed(), which is false
Thread 1: lock mutex
Thread 1: check the predicate in wait()
Thread 2 (writer): call close()
Thread 2: set closed=true
Thread 2: call cnd_.notify_all()
Thread 1: calls wait() without predicate (in the implementation of the wait with predicate method)
At this point, Thread 1 waits forever because the notify_all occurred before the actual call to wait().
I don't have a good minimal test case for the issue here, but I have run it for several hours in a loop an integration test that was deadlocking before, so I'm fairly confident the issue is fixed.
Because the close method does not lock and the read method does not check the closed value after locking, it's possible for a deadlock to occur under the following series of events:
Thread 1 (reader): read last element from the channel Thread 1: call operator>> Thread 1: check closed(), which is false Thread 1: lock mutex Thread 1: check the predicate in wait() Thread 2 (writer): call close() Thread 2: set closed=true Thread 2: call cnd_.notify_all() Thread 1: calls wait() without predicate (in the implementation of the wait with predicate method)
At this point, Thread 1 waits forever because the notify_all occurred before the actual call to wait().
I don't have a good minimal test case for the issue here, but I have run it for several hours in a loop an integration test that was deadlocking before, so I'm fairly confident the issue is fixed.