Re: [PATCH 1/3] lockdep: Move lock events under lockdep recursion protection

[Jens Axboe <jens.axboe@oracle.com>]

On Wed, Mar 10 2010, Frederic Weisbecker wrote:
> On Tue, Mar 09, 2010 at 09:34:10AM +0100, Jens Axboe wrote:
> > I went to try this on 2.6.34-rc1 to see how much it would improve things
> > here. With 2.6.34-rc1, a
> > 
> > $ time sudo perf lock record ls /dev/shm
> > 
> > essentially hangs the box, it ends up crashing hard (not just live
> > locked like before). With the patch in place, it does eventually finish:
> > 
> > real    0m21.301s
> > user    0m0.030s
> > sys     0m21.040s
> > 
> > The directory is empty.
> 
> 
> Hehe :-)
> 
> That said you are probably missing a part of the puzzle.
> This patch avoids the scary recursions we had. But there
> is another separate patch that fixes the buffers multiplexing
> dependency we had. Buffering is now done per cpu.
> 
> You need this patch:
> 
> 	commit b67577dfb45580c498bfdb1bc76c00c3b2ad6310
> 	Author: Frederic Weisbecker <fweisbec@gmail.com>
> 	Date:   Wed Feb 3 09:09:33 2010 +0100
> 
> 	perf lock: Drop the buffers multiplexing dependency
> 
> But this is not in -rc1, you need to checkout tip/perf/core
> 
> Well, the best would be you actually pull this on top of -rc1:
> 
> git://git.kernel.org/pub/scm/linux/kernel/git/frederic/random-tracing.git
> 	perf/core
> 
> so that you get the buffers multiplexing dependency that is in -tip,
> and you also get the recursions fixes, not yet in -tip (but should make
> it soon).
> 
> I bet you still won't have magic results after that, but still it should
> be better.

OK, I will give it a spin and see what happens. I had seen the
de-multiplexing patch, but just assumed it was in -rc1 already.

> We still have further projects in mind to improve the scalability,
> like the events injection thing (that avoids the string copy)
> 
> Again, you are my eyes on this, I'm still blind with my poor dual
> laptop or my atom testbox.
> 
> Thanks for your testing.

Not a problem, thank you for trying to get some nicer lock profiling in
the kernel :-)

-- 
Jens Axboe