lkml.org 
[lkml]   [2017]   [Nov]   [15]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 1/2] mm,vmscan: Kill global shrinker lock.
On Wed 15-11-17 09:56:25, Michal Hocko wrote:
> On Wed 15-11-17 09:56:02, Minchan Kim wrote:
> > On Tue, Nov 14, 2017 at 06:37:42AM +0900, Tetsuo Handa wrote:
> > > When shrinker_rwsem was introduced, it was assumed that
> > > register_shrinker()/unregister_shrinker() are really unlikely paths
> > > which are called during initialization and tear down. But nowadays,
> > > register_shrinker()/unregister_shrinker() might be called regularly.
> > > This patch prepares for allowing parallel registration/unregistration
> > > of shrinkers.
> > >
> > > Since do_shrink_slab() can reschedule, we cannot protect shrinker_list
> > > using one RCU section. But using atomic_inc()/atomic_dec() for each
> > > do_shrink_slab() call will not impact so much.
> > >
> > > This patch uses polling loop with short sleep for unregister_shrinker()
> > > rather than wait_on_atomic_t(), for we can save reader's cost (plain
> > > atomic_dec() compared to atomic_dec_and_test()), we can expect that
> > > do_shrink_slab() of unregistering shrinker likely returns shortly, and
> > > we can avoid khungtaskd warnings when do_shrink_slab() of unregistering
> > > shrinker unexpectedly took so long.
> > >
> > > Signed-off-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
> >
> > Before reviewing this patch, can't we solve the problem with more
> > simple way? Like this.
> >
> > Shakeel, What do you think?
> >
> > diff --git a/mm/vmscan.c b/mm/vmscan.c
> > index 13d711dd8776..cbb624cb9baa 100644
> > --- a/mm/vmscan.c
> > +++ b/mm/vmscan.c
> > @@ -498,6 +498,14 @@ static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
> > sc.nid = 0;
> >
> > freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible);
> > + /*
> > + * bail out if someone want to register a new shrinker to prevent
> > + * long time stall by parallel ongoing shrinking.
> > + */
> > + if (rwsem_is_contended(&shrinker_rwsem)) {
> > + freed = 1;
> > + break;
> > + }
>
> So you want to do only partial slab shrinking if we have more contending
> direct reclaimers? This would just make a larger pressure on those on
> the list head rather than the tail. I do not think this is a good idea.

Scratch that. rwsem_is_contended is true only if there is at least one
writer. So the regular reclaim path will be OK. (Un)Register will
shortcut the reclaim. This is acceptable and actually much more simple
than the complex locking proposed in the patch. So it looks good to me.
--
Michal Hocko
SUSE Labs

\
 
 \ /
  Last update: 2017-11-15 10:20    [W:0.086 / U:0.364 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site