[patch] testsuite: Fix a race by me - watchthreads-reorder.exp

[patch] testsuite: Fix a race by me - watchthreads-reorder.exp

[Jan Kratochvil]

Hi,

there is a bug explainable by man pthread_cond_signal:
	The [...] pthread_cond_signal() functions shall have no effect if
	there are no threads currently blocked on cond.

meaning a race for the testcase.
	+FAIL: gdb.threads/watchthreads-reorder.exp: reorder1: continue a (timeout)

One can reproduce the race on CVS HEAD by:
#	--- a/gdb/testsuite/gdb.threads/watchthreads-reorder.c
#	+++ b/gdb/testsuite/gdb.threads/watchthreads-reorder.c
#	@@ -89,6 +89,7 @@ thread1_func (void *unused)
#	   int i;
#	   volatile int rwatch_store;
#	 
#	+sleep(1);
#	   thread1_tid = gettid ();
#	   i = pthread_cond_signal (&thread1_tid_cond);
#	   assert (i == 0);
#	@@ -317,6 +318,7 @@ main (int argc, char **argv)
#	 
#	   if (thread1_tid == 0)
#	     {
#	+sleep(2);
#	       i = pthread_cond_wait (&thread1_tid_cond, &thread1_tid_mutex);
#	       assert (i == 0);
#	 

Fixed; gdbstop_mutex got removed as it became redundant there.

Going to check it in as obvious in several days (code is by me + it is just
a testcase).


Sorry,
Jan


gdb/testsuite/
2009-12-17  Jan Kratochvil  <jan.kratochvil@redhat.com>

	* gdb.threads/watchthreads-reorder.c (gdbstop_mutex): Remove.
	(thread1_func): Protect thread1_tid_cond by thread1_tid_mutex.  Remove
	gdbstop_mutex handling.
	(thread2_func): Protect thread2_tid_cond by thread2_tid_mutex.  Remove
	gdbstop_mutex handling.
	(main): Move thread1_tid_mutex and thread2_tid_mutex locks before
	pthread_create.  Remove gdbstop_mutex handling.  New comment.  Remove
	pthread_cond_wait conditionalizations.

--- a/gdb/testsuite/gdb.threads/watchthreads-reorder.c
+++ b/gdb/testsuite/gdb.threads/watchthreads-reorder.c
@@ -34,8 +34,6 @@
    otherwise.  */
 #define TIMEOUT (gettid () == getpid() ? 10 : 15)
 
-static pthread_mutex_t gdbstop_mutex = PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP;
-
 static pid_t thread1_tid;
 static pthread_cond_t thread1_tid_cond = PTHREAD_COND_INITIALIZER;
 static pthread_mutex_t thread1_tid_mutex = PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP;
@@ -90,12 +88,11 @@ thread1_func (void *unused)
   volatile int rwatch_store;
 
   thread1_tid = gettid ();
+
+  timed_mutex_lock (&thread1_tid_mutex);
   i = pthread_cond_signal (&thread1_tid_cond);
   assert (i == 0);
-
-  /* Be sure GDB is already stopped before continuing.  */
-  timed_mutex_lock (&gdbstop_mutex);
-  i = pthread_mutex_unlock (&gdbstop_mutex);
+  i = pthread_mutex_unlock (&thread1_tid_mutex);
   assert (i == 0);
 
   rwatch_store = thread1_rwatch;
@@ -115,12 +112,11 @@ thread2_func (void *unused)
   volatile int rwatch_store;
 
   thread2_tid = gettid ();
+
+  timed_mutex_lock (&thread2_tid_mutex);
   i = pthread_cond_signal (&thread2_tid_cond);
   assert (i == 0);
-
-  /* Be sure GDB is already stopped before continuing.  */
-  timed_mutex_lock (&gdbstop_mutex);
-  i = pthread_mutex_unlock (&gdbstop_mutex);
+  i = pthread_mutex_unlock (&thread2_tid_mutex);
   assert (i == 0);
 
   rwatch_store = thread2_rwatch;
@@ -267,7 +263,8 @@ main (int argc, char **argv)
 
   setbuf (stdout, NULL);
 
-  timed_mutex_lock (&gdbstop_mutex);
+  timed_mutex_lock (&thread1_tid_mutex);
+  timed_mutex_lock (&thread2_tid_mutex);
 
   timed_mutex_lock (&terminate_mutex);
 
@@ -306,30 +303,21 @@ main (int argc, char **argv)
       state_wait (tracer, "T (stopped)");
     }
 
-  timed_mutex_lock (&thread1_tid_mutex);
-  timed_mutex_lock (&thread2_tid_mutex);
-
-  /* Let the threads start.  */
-  i = pthread_mutex_unlock (&gdbstop_mutex);
-  assert (i == 0);
+  /* Threads are now waiting at timed_mutex_lock (thread1_tid_mutex) and so
+     they could not trigger the watchpoints before GDB gets unstopped later.
+     Threads get resumed at pthread_cond_wait below.  */
 
   printf ("Waiting till the threads initialize their TIDs.\n");
 
-  if (thread1_tid == 0)
-    {
-      i = pthread_cond_wait (&thread1_tid_cond, &thread1_tid_mutex);
-      assert (i == 0);
+  i = pthread_cond_wait (&thread1_tid_cond, &thread1_tid_mutex);
+  assert (i == 0);
 
-      assert (thread1_tid > 0);
-    }
+  assert (thread1_tid > 0);
 
-  if (thread2_tid == 0)
-    {
-      i = pthread_cond_wait (&thread2_tid_cond, &thread2_tid_mutex);
-      assert (i == 0);
+  i = pthread_cond_wait (&thread2_tid_cond, &thread2_tid_mutex);
+  assert (i == 0);
 
-      assert (thread2_tid > 0);
-    }
+  assert (thread2_tid > 0);
 
   printf ("Thread 1 TID = %lu, thread 2 TID = %lu, PID = %lu.\n",
 	  (unsigned long) thread1_tid, (unsigned long) thread2_tid,

Re: [patch] testsuite: Fix a race by me - watchthreads-reorder.exp

[Tom Tromey]

>>>>> "Jan" == Jan Kratochvil <jan.kratochvil@redhat.com> writes:

Jan> there is a bug explainable by man pthread_cond_signal:
Jan> 	The [...] pthread_cond_signal() functions shall have no effect if
Jan> 	there are no threads currently blocked on cond.

Jan> +  i = pthread_cond_wait (&thread1_tid_cond, &thread1_tid_mutex);
Jan> +  assert (i == 0);

pthread_cond_wait can also spuriously wake up.  The usual thing to do is
call it in a loop that checks some condition.  Then, have the signalling
thread set the condition before calling pthread_cond_signal.  Something
like:

while (thread1_tid == 0)
  pthread_cond_wait (...);

This is race-free as long as the signalling thread also acquires the
mutex associated with the condition.

Is there some reason not to do this in this test case?

Tom

Re: [patch] testsuite: Fix a race by me - watchthreads-reorder.exp

[Jan Kratochvil]

On Thu, 17 Dec 2009 21:27:24 +0100, Tom Tromey wrote:
> >>>>> "Jan" == Jan Kratochvil <jan.kratochvil@redhat.com> writes:
> Jan> +  i = pthread_cond_wait (&thread1_tid_cond, &thread1_tid_mutex);
> Jan> +  assert (i == 0);
> 
> pthread_cond_wait can also spuriously wake up.  The usual thing to do is
> call it in a loop that checks some condition.  Then, have the signalling
> thread set the condition before calling pthread_cond_signal.  Something
> like:
> 
> while (thread1_tid == 0)
>   pthread_cond_wait (...);
> 
> This is race-free as long as the signalling thread also acquires the
> mutex associated with the condition.
> 
> Is there some reason not to do this in this test case?

Thanks, I was not proficient in these functions (and have not read the whole
manual).

OK to check-in?


Thanks,
Jan


gdb/testsuite/
2010-01-19  Jan Kratochvil  <jan.kratochvil@redhat.com>

	* gdb.threads/watchthreads-reorder.c (gdbstop_mutex): Remove.
	(thread1_func): Protect thread1_tid with thread1_tid_cond by
	thread1_tid_mutex.  Remove gdbstop_mutex handling.
	(thread2_func): Protect thread2_tid with thread2_tid_cond by
	thread2_tid_mutex.  Remove gdbstop_mutex handling.
	(main): Move thread1_tid_mutex and thread2_tid_mutex locks before
	pthread_create.  Remove gdbstop_mutex handling.  New comment.  Replace
	pthread_cond_wait conditionalizations by while loops.

--- a/gdb/testsuite/gdb.threads/watchthreads-reorder.c
+++ b/gdb/testsuite/gdb.threads/watchthreads-reorder.c
@@ -34,8 +34,6 @@
    otherwise.  */
 #define TIMEOUT (gettid () == getpid() ? 10 : 15)
 
-static pthread_mutex_t gdbstop_mutex = PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP;
-
 static pid_t thread1_tid;
 static pthread_cond_t thread1_tid_cond = PTHREAD_COND_INITIALIZER;
 static pthread_mutex_t thread1_tid_mutex = PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP;
@@ -89,13 +87,14 @@ thread1_func (void *unused)
   int i;
   volatile int rwatch_store;
 
+  timed_mutex_lock (&thread1_tid_mutex);
+
+  /* THREAD1_TID_MUTEX must be already locked to avoid race.  */
   thread1_tid = gettid ();
+
   i = pthread_cond_signal (&thread1_tid_cond);
   assert (i == 0);
-
-  /* Be sure GDB is already stopped before continuing.  */
-  timed_mutex_lock (&gdbstop_mutex);
-  i = pthread_mutex_unlock (&gdbstop_mutex);
+  i = pthread_mutex_unlock (&thread1_tid_mutex);
   assert (i == 0);
 
   rwatch_store = thread1_rwatch;
@@ -114,13 +113,14 @@ thread2_func (void *unused)
   int i;
   volatile int rwatch_store;
 
+  timed_mutex_lock (&thread2_tid_mutex);
+
+  /* THREAD2_TID_MUTEX must be already locked to avoid race.  */
   thread2_tid = gettid ();
+
   i = pthread_cond_signal (&thread2_tid_cond);
   assert (i == 0);
-
-  /* Be sure GDB is already stopped before continuing.  */
-  timed_mutex_lock (&gdbstop_mutex);
-  i = pthread_mutex_unlock (&gdbstop_mutex);
+  i = pthread_mutex_unlock (&thread2_tid_mutex);
   assert (i == 0);
 
   rwatch_store = thread2_rwatch;
@@ -267,7 +267,8 @@ main (int argc, char **argv)
 
   setbuf (stdout, NULL);
 
-  timed_mutex_lock (&gdbstop_mutex);
+  timed_mutex_lock (&thread1_tid_mutex);
+  timed_mutex_lock (&thread2_tid_mutex);
 
   timed_mutex_lock (&terminate_mutex);
 
@@ -306,29 +307,23 @@ main (int argc, char **argv)
       state_wait (tracer, "T (stopped)");
     }
 
-  timed_mutex_lock (&thread1_tid_mutex);
-  timed_mutex_lock (&thread2_tid_mutex);
-
-  /* Let the threads start.  */
-  i = pthread_mutex_unlock (&gdbstop_mutex);
-  assert (i == 0);
+  /* Threads are now waiting at timed_mutex_lock (thread1_tid_mutex) and so
+     they could not trigger the watchpoints before GDB gets unstopped later.
+     Threads get resumed at pthread_cond_wait below.  Use `while' loops for
+     protection against spurious pthread_cond_wait wakeups.  */
 
   printf ("Waiting till the threads initialize their TIDs.\n");
 
-  if (thread1_tid == 0)
+  while (thread1_tid == 0)
     {
       i = pthread_cond_wait (&thread1_tid_cond, &thread1_tid_mutex);
       assert (i == 0);
-
-      assert (thread1_tid > 0);
     }
 
-  if (thread2_tid == 0)
+  while (thread2_tid == 0)
     {
       i = pthread_cond_wait (&thread2_tid_cond, &thread2_tid_mutex);
       assert (i == 0);
-
-      assert (thread2_tid > 0);
     }
 
   printf ("Thread 1 TID = %lu, thread 2 TID = %lu, PID = %lu.\n",

Re: [patch] testsuite: Fix a race by me - watchthreads-reorder.exp

[Tom Tromey]

>>>>> "Jan" == Jan Kratochvil <jan.kratochvil@redhat.com> writes:

Jan> 2010-01-19  Jan Kratochvil  <jan.kratochvil@redhat.com>
Jan> 	* gdb.threads/watchthreads-reorder.c (gdbstop_mutex): Remove.
Jan> 	(thread1_func): Protect thread1_tid with thread1_tid_cond by
Jan> 	thread1_tid_mutex.  Remove gdbstop_mutex handling.
Jan> 	(thread2_func): Protect thread2_tid with thread2_tid_cond by
Jan> 	thread2_tid_mutex.  Remove gdbstop_mutex handling.
Jan> 	(main): Move thread1_tid_mutex and thread2_tid_mutex locks before
Jan> 	pthread_create.  Remove gdbstop_mutex handling.  New comment.  Replace
Jan> 	pthread_cond_wait conditionalizations by while loops.

Ok.  Thanks.

Tom

Re: [patch] testsuite: Fix a race by me - watchthreads-reorder.exp

[Jan Kratochvil]

On Wed, 20 Jan 2010 22:01:22 +0100, Tom Tromey wrote:
> >>>>> "Jan" == Jan Kratochvil <jan.kratochvil@redhat.com> writes:
> 
> Jan> 2010-01-19  Jan Kratochvil  <jan.kratochvil@redhat.com>
> Jan> 	* gdb.threads/watchthreads-reorder.c (gdbstop_mutex): Remove.
> Jan> 	(thread1_func): Protect thread1_tid with thread1_tid_cond by
> Jan> 	thread1_tid_mutex.  Remove gdbstop_mutex handling.
> Jan> 	(thread2_func): Protect thread2_tid with thread2_tid_cond by
> Jan> 	thread2_tid_mutex.  Remove gdbstop_mutex handling.
> Jan> 	(main): Move thread1_tid_mutex and thread2_tid_mutex locks before
> Jan> 	pthread_create.  Remove gdbstop_mutex handling.  New comment.  Replace
> Jan> 	pthread_cond_wait conditionalizations by while loops.
> 
> Ok.  Thanks.

Checked-in:
	http://sourceware.org/ml/gdb-cvs/2010-01/msg00182.html


Thanks,
Jan