Re: [PATCH 2/9] perf_counter: fix update_userpage()

[Peter Zijlstra <a.p.zijlstra@chello.nl>]

On Thu, 2009-04-02 at 20:58 +1100, Paul Mackerras wrote:
> Peter Zijlstra writes:
> 
> > > Good point.  This should work, though:
> > > 
> > > 	do {
> > > 		seq = pc->lock;
> > > 		barrier();
> > > 		value = read_pmc(pc->index) + pc->offset;
> > > 		barrier();
> > > 	} while (pc->lock != seq);
> > > 	return value;
> > 
> > I don't think you need the first barrier(), all you need to avoid is it
> > reusing the first pc->lock read, so one should suffice.
> 
> I need it to make sure that the compiler doesn't put the load of
> pc->index or pc->offset before the first load of pc->lock.  The second
> barrier is needed to make sure the compiler puts the second load of
> pc->lock after the loads of pc->index and pc->offset.  So I think I do
> need to barrier()s (but only compiler barriers, not cpu memory
> barriers).

Ah, you're right indeed.

> > Also, you need to handle the !pc->index case.
> 
> Hmmm, yeah.  I claim that read_pmc(0) always returns 0. :)

Hehe :-)

Ok, updated that patch.