public inbox for libc-hacker@sourceware.org
 help / color / mirror / Atom feed
* [PATCH] Avoid pthread_cond_destroy blocking for too long
@ 2004-09-02 21:50 Jakub Jelinek
  0 siblings, 0 replies; only message in thread
From: Jakub Jelinek @ 2004-09-02 21:50 UTC (permalink / raw)
  To: Ulrich Drepper; +Cc: Glibc hackers

Hi!

This patch adds a mutex wakeup if there are still waiters at
pthread_cond_destroy time.
The reason is to make sure pthread_cond_destroy won't block for too long.
If some threads are blocked on the pthread_mutex_t's __lock, it is
under application control when (if ever) they will be woken up and
pthread_cond_destroy would block for that whole time.
By waking all mutex waiters we are just waiting until the scheduler
gives all threads enough timeslice to acquire the condvar internal lock,
hop through the short critical section and release the lock (last thread after
waking up pthread_cond_destroy).

In most programs nwaiters will be < (1 << COND_CLOCK_BITS) at
pthread_cond_destroy time and thus this patch shouldn't cause performance
regressions.

2004-09-02  Jakub Jelinek  <jakub@redhat.com>

	* pthread_cond_destroy.c (__pthread_cond_destroy): If there are
	waiters, awake all waiters on the associated mutex.

--- libc/nptl/pthread_cond_destroy.c.jj	2004-09-02 22:27:56.000000000 +0200
+++ libc/nptl/pthread_cond_destroy.c	2004-09-02 23:33:43.278763736 +0200
@@ -44,15 +44,35 @@ __pthread_cond_destroy (cond)
      broadcasted, but still are using the pthread_cond_t structure,
      pthread_cond_destroy needs to wait for them.  */
   unsigned int nwaiters = cond->__data.__nwaiters;
-  while (nwaiters >= (1 << COND_CLOCK_BITS))
+
+  if (nwaiters >= (1 << COND_CLOCK_BITS))
     {
-      lll_mutex_unlock (cond->__data.__lock);
+      /* Wake everybody on the associated mutex in case there are
+         threads that have been requeued to it.
+         Without this, pthread_cond_destroy could block potentially
+         for a long time or forever, as it would depend on other
+         thread's using the mutex.
+         When all threads waiting on the mutex are woken up, pthread_cond_wait
+         only waits for threads to acquire and release the internal
+         condvar lock.  */
+      if (cond->__data.__mutex != NULL
+	  && cond->__data.__mutex != (void *) ~0l)
+	{
+	  pthread_mutex_t *mut = (pthread_mutex_t *) cond->__data.__mutex;
+	  lll_futex_wake (&mut->__data.__lock, INT_MAX);
+	}
+
+      do
+	{
+	  lll_mutex_unlock (cond->__data.__lock);
 
-      lll_futex_wait (&cond->__data.__nwaiters, nwaiters);
+	  lll_futex_wait (&cond->__data.__nwaiters, nwaiters);
 
-      lll_mutex_lock (cond->__data.__lock);
+	  lll_mutex_lock (cond->__data.__lock);
 
-      nwaiters = cond->__data.__nwaiters;
+	  nwaiters = cond->__data.__nwaiters;
+	}
+      while (nwaiters >= (1 << COND_CLOCK_BITS));
     }
 
   return 0;

	Jakub

^ permalink raw reply	[flat|nested] only message in thread

only message in thread, other threads:[~2004-09-02 21:50 UTC | newest]

Thread overview: (only message) (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2004-09-02 21:50 [PATCH] Avoid pthread_cond_destroy blocking for too long Jakub Jelinek

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).