[Mesa-dev] [Bug 98172] Concurrent call to glClientWaitSync results in segfault in one of the waiters.

[bugzilla-daemon at freedesktop.org]

https://bugs.freedesktop.org/show_bug.cgi?id=98172

--- Comment #8 from shinji.suzuki at gmail.com ---
Comment on attachment 127267
  --> https://bugs.freedesktop.org/attachment.cgi?id=127267
Lock the shared state mutex and work with a local reference of so->fence

Review of attachment 127267:
-----------------------------------------------------------------

I'm afraid execution of st_fence_sync() can still race.
Thread-A can run upto
  success = screen->fence_finish(screen, pipe, fence, 0);
and then get prempted and Thread-B can run upto the same location.
And then
  screen->fence_reference(screen, &so->fence, NULL);
can still be executed in arbitrary order. If screen->fence_refrence()
is thread-safe and return true only for the first invocation then all is fine
but likely it is not true as otherwise we will not be struggling with this
issue.
I think the gist of it is that checking of so->fence and nullifying of it
should be executed atomically. If "if (success)" is replaced with "if (success
&& so->fence)" then the program may behave correctly but I'm not confortable
about sreen->fence_ference() being called concurrently.
I'm also concerned that mutual exclusion on ctx->Shared->Mutex may introduce
unnecessarily strict serialization. 
Can't we introduce per sync-object mutex so that excution of checking of
so->fence and nullyfying of it happen atomically?
Is that modification too intrusive? (At least it is unnecessary overhead when
st_fence_sync() is not executed concurrently on the same sync object.)

-- 
You are receiving this mail because:
You are the QA Contact for the bug.
You are the assignee for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/mesa-dev/attachments/20161013/aa624907/attachment.html>