[OT] ppc64 serialization problem

[Greg Smith]

Very fair questions!!

Actually the code was

    pthread_mutex_lock(&lock);
    u32 |= bitB;
    TRACE("A", u32, ...);
    TRACE("B", u32, ...);
    pthread_mutex_unlock(&lock);

where TRACE is a function call (entering a trace entry to an in-storage
wrap-around table).  So for the "A" call, u32 could have come directly
from a register and for "B" from the storage location.  I'll have the
user (actually a fellow developer) send me the assembly listing to make
sure.

He has tested SLES9 (kernel 2.6.5, glibc 2.3.3, gcc 3.3.3) and Debian
(kernel 2.6.16, glibc 2.3.6, gcc 4.0.3).

The TRACE occurs while the lock is held.

Now the interesting part.

I suggested he try u64 instead of u32.  That works!!

He is suspecting a recent firmware upgrade may have something to do with
the problem.

Thank you,
Greg Smith

On Wed, 2006-03-29 at 14:11 +1100, Benjamin Herrenschmidt wrote:
> On Tue, 2006-03-28 at 20:58 -0500, Greg Smith wrote:
> > We have a multi-threaded app running on a p520 in 64 bit mode.
> > 
> > Thread A does
> > 
> > pthread_mutex_lock(&lock);
> > u32 &= ~bitA;
> > pthread_mutex_unlock(&lock);
> > 
> > and Thread B does
> > 
> > pthread_mutex_lock(&lock);
> > u32 |= bitB;
> > A = u32;
> > B = u32;
> > pthread_mutex_unlock(&lock);
> > 
> > On rare occasions, values A and B will differ!  In the examples that I
> > have seen, there is contention with `lock'.  This phenomenon does not
> > occur on ppc32 or a number of other architectures that we support.
> 
> How did you actually "look" at A and B ? is that also protected by the
> lock ?
> 
> > I confess I do not know the linux version nor the glibc version nor what
> > pthreads implementation is being used.  I'll find that out shortly.
> 
> That's fairly important to know those yes.
> 
> > What I am curious about is where the problem might lie
> > (kernel/lib/pthreads/app) so I can ask the right people.
> > 
> > Thank you for your patience,
> > Greg Smith