Threads that unlock and lock a mutex at the same time perform the following
operations:
Thread 1 Thread 2
-------- --------
QBasicMutex::lockInternal() QBasicMutex::unlockInternal()
d_ptr.testAndSetOrdered(..., d) d = d_ptr.loadAcquire()
d->waiters.loadRelaxed(); (1)
d->waiters.fetchAndAddRelease() (2)
d_ptr.testAndSetRelease(d, 0) (3)
d->derefWaiters() (4)
d->waiters.testAndSetRelaxed(...) (5)
if (d != d_ptr.loadAcquire()) (6)
d->wait()
The operation (1) isn't serialized with the operation (6) so its memory
effect may be observed before the effect of the operation (1). However,
if memory effects are observed in the following order: (6) -> (1) -> (2)
-> (3) -> (4) -> (5) then Thread 1 doesn't notice that Thread 2 updates
d_ptr and goes to sleep with d pointing to a stale object, this object
isn't reachable since d_ptr is zeroed so Thread 1 cannot be woken up.
The patch adds the "acquire" barrier into the operation (1) so that it
cannot be reordered with the operation (6).
Fixes: QTBUG-88247
Pick-to: 5.15 5.12
Change-Id: I1d0c405c0bf5080ec1815d351b9b4b75efeab21a
Reviewed-by: Olivier Goffart (Woboq GmbH) <ogoffart@woboq.com>
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
d08e3b6de1