lkml.org 
[lkml]   [2015]   [Jun]   [29]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [RFC][PATCHv3 7/7] zsmalloc: register a shrinker to trigger auto-compaction
On (06/29/15 22:39), Minchan Kim wrote:
> Hi Sergey,
>
> Sorry for too late reply.
>

Hello Minchan,

Sure, no problem.


Will send out a new patchset later today. Thanks.

-ss

> On Mon, Jun 29, 2015 at 05:57:44PM +0900, Sergey Senozhatsky wrote:
> > Hello,
> >
> > thanks for review.
> >
> > On (06/29/15 16:07), Minchan Kim wrote:
> > [..]
> > > > if (!migrate_zspage(pool, class, &cc))
> > > > - break;
> > > > + goto out;
> > >
> > > It should retry with another target_page instead of going out.
> >
> > yep.
> >
> > [..]
> > > > +static unsigned long zs_shrinker_scan(struct shrinker *shrinker,
> > > > + struct shrink_control *sc)
> > > > +{
> > [..]
> > >
> > > Returns migrated object count.
> > >
> > [..]
> > > > +static unsigned long zs_shrinker_count(struct shrinker *shrinker,
> > > > + struct shrink_control *sc)
> > > > +{
> > [..]
> > >
> > > But it returns wasted_obj / max_obj_per_zspage?
> > >
> >
> > Good catch.
> > So, __zs_compact() and zs_shrinker_count() are ok. Returning
> > "wasted_obj / max_obj_per_zspage" from zs_can_compact() makes
> > sense there. The only place is zs_shrinker_scan()->zs_compact().
>
> I want to make zs_can_compact return freeable page unit.
>
> ie,
>
> return obj_wasted * class->pages_per_zspage;
>
> and let's make __zs_compact returns the number of freed pages.
>
> IOW, I like your (c).
>
> >
> > Hm, I can think of:
> >
> > (a) We can change zs_compact() to return the total number of
> > freed zspages. That will not really change a user visible
> > interface. We export (fwiw) the number of compacted objects
> > in mm_stat. Basically, this is internal zsmalloc() counter and
> > no user space program can ever do anything with that data. From
> > that prospective we will just replace one senseless number with
> > another (equally meaningless) one.
> >
> >
> > (b) replace zs_compact() call in zs_shrinker_scan() with a class loop
> >
> > 1764 int i;
> > 1765 unsigned long nr_migrated = 0;
> > 1766 struct size_class *class;
> > 1767
> > 1768 for (i = zs_size_classes - 1; i >= 0; i--) {
> > 1769 class = pool->size_class[i];
> > 1770 if (!class)
> > 1771 continue;
> > 1772 if (class->index != i)
> > 1773 continue;
> > 1774 nr_migrated += __zs_compact(pool, class);
> > 1775 }
> > 1776
> > 1777 return nr_migrated;
> >
> > But on every iteration increment nr_migrated with
> > "nr_migrated += just_migrated / max_obj_per_zspage"
> >
> > (which will be unnecessary if zs_compact() will return the number of freed
> > zspages).
> >
> > So, (b) is mostly fine, except that we already have several pool->size_class
> > loops, with same `if (!class)' and `if (class->index...)' checks; and it
> > asks for some sort of refactoring or... a tricky for_each_class() define.
> >
> >
> > In both cases, however, we don't tell anything valuable to user space.
> > Thus,
> >
> > (c) Return from zs_compact() the number of pages (PAGE_SIZE) freed.
> > And change compaction to operate in terms of pages (PAGE_SIZE). At
> > least mm_stat::compacted will turn into something useful for user
> > space.
>
> Yes.
>
> Thanks.
>
> >
> > -ss
>
> --
> Kind regards,
> Minchan Kim
>


\
 
 \ /
  Last update: 2015-06-30 02:01    [W:0.053 / U:0.064 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site