This test runs fine almost all of the time on systems with 1
processor, which were the norm when the test was written and are still
the way that the Qt Continuous Integration system works as of
today. But it falls flatly on multi-processor systems.
The root of the problem is that QSystemSemaphore recreates the
semaphore if it disappears underneath it. However, the recreation
process is not thread-safe at all: if two threads race to recreate it,
weird things might happen. strace on Linux shows that a thread got
stuck trying to acquire the semaphore:
<... nanosleep resumed> NULL) = 0
stat("/tmp/qipc_systemsem_market5c9f73af73334ffe350c60ec076e5744db0ecda3", {st_mode=S_IFREG|0640, st_size=0, ...}) = 0
stat("/tmp/qipc_systemsem_market5c9f73af73334ffe350c60ec076e5744db0ecda3", {st_mode=S_IFREG|0640, st_size=0, ...}) = 0
semget(0x51001388, 1, IPC_CREAT|IPC_EXCL|0600) = -1 EEXIST (File exists)
semget(0x51001388, 1, IPC_CREAT|0600) = 114786308
semop(114786308, {{0, -1, SEM_UNDO}}, 1 <unfinished ...>
This problem does not happen if the creation and destruction of the
QSharedMemory (which uses QSystemSemaphore) does not race with other
threads or processes attaching and detaching. For the threads test
it's easy. For the processes, we use stdin and stdout as a
communication channel.
Change-Id: Ie11b135431d4abfc59234654848b67f622eb03c9
Reviewed-by: Richard J. Moore <rich@kde.org>
60fc88a09c