SyncMutexEvent(), SyncMutexEvent_r()

Arguments:

sync

A pointer to the synchronization object for the mutex that you want to attach an event to.

event

A pointer to the sigevent structure that describes the event that you want to attach, or NULL if you want to detach the currently registered event.

Library:

libc

Use the -l c option to qcc to link against this library. This library is usually included automatically.

Description:

The SyncMutexEvent() and SyncMutexEvent_r() kernel calls attach the specified event to a mutex pointed to by sync. This event is delivered when the mutex enters the DEAD state, which happens when a process holding the mutex removes its last mapping of the mutex without unlocking it (including when the process dies while holding the mutex in shared memory).

SyncMutexEvent() and SyncMutexEvent_r() are similar, except for the way they indicate errors. See the Returns section for details.

When you're notified that a mutex has been put into the DEAD state, you can revive it with SyncMutexRevive().

If you call SyncMutexEvent() with a NULL event, the function deletes any existing event registration.

Managing the death of a mutex

Robust mutexes automate the notification about the termination of the current mutex owner, but the recovery code must be spread over multiple functions, as explained in the pthread_mutexattr_setrobust() reference. It may be easier to use a non-robust mutex with a sigevent because it provides centralized recovery handling. Here's how you can use a non-robust mutex in shared memory between threads in two or more processes and deal with the death of one of the processes while it's holding the mutex.

1. Have a thread in one process create and initialize the mutex, and possibly attach an event using SyncMutexEvent().
2. This thread marks the shared memory area and the mutex as usable.
3. Everyone goes ahead and uses the mutex and shared memory area; everything goes well until a process dies.
4. If the process that dies has a thread that holds the mutex (i.e., it's the current owner), then the kernel marks the mutex as DEAD and sends the event registered with SyncMutexEvent().

   If the event target happens to be the process that's dying, the event is not sent; you can use procmgr_guardian() to try to handle this case. If you don't designate a guardian or if you don't register an event, the mutex is marked as DESTROYED. In both these cases, the kernel sends a SIGDEADLK signal to all threads that are waiting without a timeout on the mutex. For any of these threads that have not blocked SIGDEADLK, their pthread_mutex_lock() calls return with EINVAL. For any threads waiting with a timeout on the mutex, their pthread_mutex_timedlock*() calls return with EINVAL.

5. If an event was attached and the target process is still running (i.e., it was not the one that died), threads that are already waiting on the dead mutex continue to wait on it. The thread that received the event calls SyncMutexRevive() on the dead mutex and revives and acquires the mutex in one atomic operation.

If you call pthread_mutex_lock():

• on a destroyed mutex, the function returns EINVAL
• on a dead mutex, your thread will be blocked until the mutex is revived and released, or the process that has the SyncMutexEvent() event registered dies, in which case the lock operation returns EINVAL

Returns:

The only difference between these functions is the way they indicate errors:

SyncMutexEvent()

If an error occurs, the function returns -1 and sets errno. Any other value returned indicates success.

SyncMutexEvent_r()

Returns EOK on success. This function does NOT set errno. If an error occurs, the function returns any value listed in the Errors section.

Errors:

EAGAIN

All kernel synchronization event objects are in use.

EFAULT

A fault occurred when the kernel tried to access sync.

EINVAL

The synchronization object pointed to by sync doesn't exist, or it's a robust mutex.

EPERM

The calling process doesn't have the required permission; see procmgr_ability().

Classification:

QNX Neutrino