Re: [LKP] [mm] 9bc8039e71: will-it-scale.per_thread_ops -64.1% regression

[Yang Shi]

On 11/5/18 10:35 AM, Linus Torvalds wrote:
> On Mon, Nov 5, 2018 at 10:28 AM Yang Shi <yang.shi@xxxxxxxxxxxxxxxxx> wrote:
> > Actually, the commit is mainly for optimizing the long stall time caused
> > by holding mmap_sem by write when unmapping or shrinking large mapping.
> > It downgrades write mmap_sem to read when zapping pages. So, it looks
> > the downgrade incurs more context switches. This is kind of expected.
> >
> > However, the test looks just shrink the mapping with one normal 4K page
> > size. It sounds the overhead of context switches outpace the gain in
> > this case at the first glance.
> I'm not seeing why there should be a context switch in the first place.
>
> Even if you have lots of concurrent brk() users, they should all block
> exactly the same way as before (a write lock blocks against a write
> lock, but it *also* blocks against a downgraded read lock).

Yes, it is true. The brk() users will not get waken up. What I can think 
of for now is there might be other helper processes and/or kernel 
threads are waiting for read mmap_sem. They might get waken up by the 
downgrade.

But, I also saw huge increase in cpu idle time and sched_goidle events. 
Not have clue yet for why idle goes up.

20610709 Â 15%   +2376.0%  5.103e+08 Â 34%  cpuidle.C1.time
28753819 Â 39%   +1054.5%  3.319e+08 Â 49%  cpuidle.C3.time

175049 Â 72%    +840.7%    1646720 Â 72%  sched_debug.cpu.sched_goidle.stddev


Thanks,
Yang

> So no, I don't want just some limit to hide this problem for that
> particular test. There's something else going on.
>
>                   Linus