Mutex locking issue on Solaris 10

Question

We are having a problem with mutex locking.
On one scenario we are running our application
On a v245 running Solaris 10 6/06 update 2 patch 118833-24
On our 5120 we are running Solaris 10 8/07 update 4patch 137111-05.
Our application creates several thousand simultaneous mutex inits (pthread_mutex_init). Approx 12 are locked at any one time. Our problem is after several min the application fails because the mutex destroy does not appear to destroy the mutex. This is true only on our 5120 but not on our v245. As a result our application runs fine on the v245 but fails quickly on the 5120.
Do you have any knowledge of why this might be happening?
Any help you can provide in this will be greatly appreciated.
If this is not the right group to post this messages please let me know.
Thanks!

Answer 1

How many threads do you have performing these mutex operations? The 5120 probably runs a lot more concurrent threads than the v245 does, so it's likely you're running into a race condition that's only apparent when the number of concurrent threads is large enough.
What kind of mutexes are they? Recursive, process-shared, "normal"? Where are they located? Dynamically-allocated heap memory, local stack variables, mmap()'d memory? If they're recursive or "normal" mutexes in heap memory, you've most likely got a race condition in your application, because the source code at opensolaris.org for pthread_mutex_destroy() is pretty much nothing but "memset( &mutex, 0, sizeof( mutex ) )" for normal mutexes. It's hard to imagine how that could fail.
You can use the race-condition tools that are part of Sun Studio 12 or another memory-checking application such as Purify to find out where your error may be. Given the timing issues apparently involved, those are probably your best options. The overhead of tools such as watchmalloc or even the run-time memory checker in Sun Studio may be too much and mask your timing issues. Assuming that's what the root cause of your problem is.
I didn't look at what's involved in destroying a process-shared mutex, though. That's a lot more complicated.
On edit: How do you know that pthread_mutex_destroy() is not destroying the mutex?
Edited by: AndrewHenle on Jun 16, 2009 3:46 PM