Re: aim7 performance regression by commit 5a50508 report from LKP

[Yuanhan Liu <yuanhan.liu@linux.intel.com>]

On Tue, Jan 29, 2013 at 10:12:45AM +0100, Ingo Molnar wrote:
> 
> * Yuanhan Liu <yuanhan.liu@linux.intel.com> wrote:
> 
> > On Tue, Jan 29, 2013 at 09:44:00AM +0100, Ingo Molnar wrote:
> > > 
> > > * Yuanhan Liu <yuanhan.liu@linux.intel.com> wrote:
> > > 
> > > > [...]
> > > 
> > > Very nice measurements and analysis, thanks!
> > > 
> > > > As stated above, anybody can have a chance to own the lock in 
> > > > mutex once somebody release the lock. Well, there is only one 
> > > > to own the lock in rwsem write lock, and the one is known 
> > > > already: the one in the head of wait list. That would result 
> > > > to more contention in rwsem write lock case, especially if the 
> > > > one _will_ own the lock is not running now.
> > > 
> > > I think we should allow lock-steal between rwsem writers - that 
> > > will not hurt fairness as most rwsem fairness concerns relate to 
> > > reader vs. writer fairness.
> > 
> > Agreed, and I'm sure this will improve performance and may 
> > make this performance regression go away.
> > 
> > David, is that Ok to you? If so, I may have a try.
> 
> I'm not David but please try it :-)
> 
> Making rwsem behavior and scalability similar to mutexes would 
> have numerous advantages.
> 
> > > Am I correct to assume that all relevant users in this 
> > > workload are down_write() users?
> > 
> > Yes, as commit 5a50508 just convert all mutex to down_write.
> 
> A second track of inquiry would be to see whether any of the key 
> usage sites could be converted to down_read()

I tried before, and seems I didn't find one.

Well, I did find an irrelevant one: vma_lock_anon_vma at validate_mm()
in mm/mmap.c. That function is _real_ only if CONFIG_DEBUG_VM_RB is
set, and there is no vma_lock_anon_vma_read() or something similar,
thus I guess it's may not worthy to turn it.

> or whether the 
> lock hold times could be reduced drastically

I also found one, but it doesn't sound like the one will reduce lock
hole times drastically:
   vma_lock_anon_vma() seems covered too much code at
   expand_up/downwards.

Well, again, it's quite a tiny optimization for reducing the coverage.

Thanks.

	--yliu



> - but I doubt 
> that's really possible on such heavily forking workloads.
> 
> Thanks,
> 
> 	Ingo