Re: [Boost-users] read_write_mutex assertion on multiprocessor Wintel

From: Ben Hutchings (ben.hutchings_at_[hidden])
Date: 2005-01-05 08:18:30

jkharris7_at_[hidden] wrote:
> As I reported earlier, read_write_mutex asserts in debug mode (VC71),
> apparently seeing false wake-ups on a condition variable.
 >
> I was able to do this with just a simple group of threads locking a
> RWM in tight loops, with one writer and a few readers.
> 
> The new information is that  I am runing this on a dual-Xeon with
> hyper-threading enabled (so four apparent CPU's), and I'm thinking
> that maybe boost libs don't get exercised on this kind of hardware
> very much.

I can assure you that they do, since that's a fairly typical 
configuration for our customers.  However we aren't currently using 
read_write_mutex (or large numbers of threads).

> Is is possible that a condition::notify_all()  happens on each
> processor?

Threads are woken by the OS, not the processor.  However, there is no 
guarantee that when a thread notifies a condition and then releases the 
associated mutex that one of the threads waiting on the condition will 
be the next to acquire the mutex.  It is entirely possible that another 
thread waiting on that mutex will acquire it, though that's far more 
likely to happen in a multiprocessor machine.

See <http://www.lambdacs.com/cpt/MFAQ.html#Q94> for more information 
about spurious wakeups.

> Has anyone else seen this?
> 
> For now, I just removed the asserts and rebuilt.  Synchronization
> seems to be unaffected in my real-world stress test of RWM.

Since the assertions say that spurious wakeups don't happen, which is 
neither correct nor necessary for the correctness of the rest of the 
code, you are right to remove them.

Ben.