Re: [patch 108/192] mm: zram: amend SLAB_RECLAIM_ACCOUNT on zspage_cachep

[Zhaoyang Huang]

On Fri, Jul 2, 2021 at 1:47 PM Minchan Kim <minchan@xxxxxxxxxx> wrote:
>
> On Fri, Jul 02, 2021 at 10:45:09AM +0800, Zhaoyang Huang wrote:
> > On Thu, Jul 1, 2021 at 10:56 PM Minchan Kim <minchan@xxxxxxxxxx> wrote:
> > >
> > > On Wed, Jun 30, 2021 at 06:52:58PM -0700, Andrew Morton wrote:
> > > > From: Zhaoyang Huang <zhaoyang.huang@xxxxxxxxxx>
> > > > Subject: mm: zram: amend SLAB_RECLAIM_ACCOUNT on zspage_cachep
> > > >
> > > > Zspage_cachep is found be merged with other kmem cache during test, which
> > > > is not good for debug things (zs_pool->zspage_cachep present to be another
> > > > kmem cache in memory dumpfile).  It is also neccessary to do so as
> > > > shrinker has been registered for zspage.
> > > >
> > > > Amending this flag can help kernel to calculate SLAB_RECLAIMBLE correctly.
> > > >
> > > > Link: https://lkml.kernel.org/r/1623137297-29685-1-git-send-email-huangzhaoyang@xxxxxxxxx
> > > > Signed-off-by: Zhaoyang Huang <zhaoyang.huang@xxxxxxxxxx>
> > > > Cc: Minchan Kim <minchan@xxxxxxxxxx>
> > > > Cc: Sergey Senozhatsky <senozhatsky@xxxxxxxxxxxx>
> > > > Signed-off-by: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
> > >
> > > Sorry for the late. I don't think this is correct.
> > >
> > > It's true "struct zspage" can be freed by zsmalloc's compaction registerred
> > > by slab shrinker so tempted to make it SLAB_RECLAIM_ACCOUNT. However, it's
> > > quite limited to work only when objects in the zspage are heavily fragmented.
> > > Once the compaction is done, zspage are never discardable until objects are
> > > fragmented again. It means it could hurt other reclaimable slab page reclaiming
> > > since the zspage slab object pins the page.
> > IMHO, kmem cache's reclaiming is NOT affected by SLAB_RECLAIM_ACCOUNT
> > . This flag just affects kmem cache merge[1], the slab page's migrate
> > type[2] and the page's statistics. Actually, zspage's cache DO merged
> > with others even without SLAB_RECLAIM_ACCOUNT currently, which maybe
> > cause zspage's object will NEVER be discarded.(SLAB_MERGE_SAME
> > introduce confusions as people believe the cache will merge with
> > others when it set and vice versa)
> >
> > [1]
> >  struct kmem_cache *find_mergeable(size_t size, size_t align, unsigned
> > long flags, const char *name, void (*ctor)(void *))
> > ...
> >     if ((flags & SLAB_MERGE_SAME) != (s->flags & SLAB_MERGE_SAME))
> >      continue;
> >
> > [2]
> > if (s->flags & SLAB_RECLAIM_ACCOUNT)
> >     s->allocflags |= __GFP_RECLAIMABLE;
>
> That's the point here. With SLAB_RECLAIM_ACCOUNT, page allocator
> try to allocate pages from MIGRATE_RECLAIMABLE with belief those
> objects are easily reclaimable. Say a page has object A, B, C, D
> and E. A-D are easily reclaimable but E is hard. What happens is
> VM couldn't reclaim the page in the end due to E even though it
> already reclaimed A-D. And the such fragmenation could be spread
> out entire MIGRATE_RECLAIMABLE pageblocks over time.
> That's why I'd like to put zspage into MIGRATE_UNMOVALBE from the
> beginning since I don't think it's easily reclaimble once compaction
> is done.
The slab page could fallback to any migrate type even allocating with
__GFP_RECLAIMABLE, and there is only one page per slab within zspage's
cache, which will not be affected by compaction, so I think that
doesn't make sense.