QMutexPool: fix memory order of atomic operations

The array of QAtomicPointer<QMutex> can be initialized using relaxed
stores of nullptr, since nullptr is the whole data. But once we store
an actual QMutex pointer in the array, we need to publish the indirect
data thus created. We did this, with testAndSetRelease(); what was
missing was a corresponding acquire fence on load, without which there
is no happens-before relationship between the writes performed by the
QMutex ctor and the reads performed by a subsequent mutex.lock(), say,
on the same data.

Fix by adding acquire fences to all loads. That includes the dtor,
since mutexes may have been created in different threads, and never
been imported into this_thread before the dtor is running.

As a drive-by, return a new'ed QMutex that was successfully installed
directly to the caller, without again going through a load-acquire.

Fixes: QTBUG-59164
Change-Id: Ia25d205b1127c8c4de0979cef997d1a88123c5c3
Reviewed-by: David Faure <david.faure@kdab.com>
Reviewed-by: Giuseppe D'Angelo <giuseppe.dangelo@kdab.com>
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
(cherry picked from commit 65b8f59e045bb41fef99b1a44f462115de65064a)
(cherry picked from commit da38f0d691d9d7eacfac5fbcbd47b887bd59bd39)

0 files changed, 0 insertions, 0 deletions