Re: [LKP] [mm] c8c06efa8b5: -7.6% unixbench.score

[Davidlohr Bueso <dave@stgolabs.net>]

On Wed, 2015-01-07 at 23:45 -0800, Davidlohr Bueso wrote:
> On Thu, 2015-01-08 at 10:27 +0800, Huang Ying wrote:
> Cc'ing Peter.

Err, resending with the complete msg.

> > FYI, we noticed the below changes on
> > 
> > commit c8c06efa8b552608493b7066c234cfa82c47fcea ("mm: convert i_mmap_mutex to rwsem")
> 
> Same exact everything, except for the lock type. No sharing going on.
> 
> > testbox/testcase/testparams: lituya/unixbench/performance-execl
> > 
> > 83cde9e8ba95d180  c8c06efa8b552608493b7066c2  
> > ----------------  --------------------------  
> >          %stddev     %change         %stddev
> >              \          |                \  
> >     721721 ±  1%    +303.6%    2913110 ±  3%  unixbench.time.voluntary_context_switches
> >      11767 ±  0%      -7.6%      10867 ±  1%  unixbench.score
> 
> And this workload appears to be from execl, right? Make sense with some
> of those numbers!!
> 
> >  2.323e+08 ±  0%      -7.2%  2.157e+08 ±  1%  unixbench.time.minor_page_faults
> >        207 ±  0%      -7.0%        192 ±  1%  unixbench.time.user_time
> >    4923450 ±  0%      -5.7%    4641672 ±  0%  unixbench.time.involuntary_context_switches
> >        584 ±  0%      -5.2%        554 ±  0%  unixbench.time.percent_of_cpu_this_job_got
> >        948 ±  0%      -4.9%        902 ±  0%  unixbench.time.system_time
> >          0 ±  0%      +Inf%     672942 ±  2%  latency_stats.hits.call_rwsem_down_write_failed.vma_adjust.__split_vma.split_vma.mprotect_fixup.SyS_mprotect.system_call_fastpath
> 
> What does this "hits" thing mean exactly? Since I assume both before and
> after runs have the same level of concurrency when pounding on mmap
> operations, I doubt it means that its the amount of calls into the
> slowpath... in addition the lock is obviously contended so we can forget
> about anything in the fastpath.
> 
> So this is a call_rwsem_down_write_failed() vs __mutex_lock_common()
> issue.

It's late, but for some initial thoughts I believe this comes down to
differences in how mutexes and rwsems deal with ultimately blocking (and
based on the nasty sched_debug numbers reported by Huang). We now do in
call_rwsem_down_write_failed:

	/* wait to be given the lock */
	while (true) {
		set_task_state(tsk, TASK_UNINTERRUPTIBLE);
		if (!waiter.task)
			break;
		schedule();
	}

In which that waiter.task check forces us to need the barrier in
set_task_state() -- while in mutexes we get away with it. We could
disable preemption during this whole slowpath step and stabilize things,
just like we do with mutexes.

Thanks,
Davidlohr