Re: [RFC PATCH] mm: don't reclaim inodes with many attached pages

[Roman Gushchin]

On Fri, Oct 26, 2018 at 09:33:03AM +0200, Michal Hocko wrote:
> On Thu 25-10-18 20:32:47, Roman Gushchin wrote:
> > On Thu, Oct 25, 2018 at 04:20:14PM -0400, Sasha Levin wrote:
> > > On Thu, Oct 25, 2018 at 12:44:42PM -0700, Andrew Morton wrote:
> > > > On Thu, 25 Oct 2018 11:23:52 +0200 Michal Hocko <mhocko@xxxxxxxxxx> wrote:
> > > > 
> > > > > On Wed 24-10-18 15:19:50, Andrew Morton wrote:
> > > > > > On Tue, 23 Oct 2018 16:43:29 +0000 Roman Gushchin <guro@xxxxxx> wrote:
> > > > > >
> > > > > > > Spock reported that the commit 172b06c32b94 ("mm: slowly shrink slabs
> > > > > > > with a relatively small number of objects") leads to a regression on
> > > > > > > his setup: periodically the majority of the pagecache is evicted
> > > > > > > without an obvious reason, while before the change the amount of free
> > > > > > > memory was balancing around the watermark.
> > > > > > >
> > > > > > > The reason behind is that the mentioned above change created some
> > > > > > > minimal background pressure on the inode cache. The problem is that
> > > > > > > if an inode is considered to be reclaimed, all belonging pagecache
> > > > > > > page are stripped, no matter how many of them are there. So, if a huge
> > > > > > > multi-gigabyte file is cached in the memory, and the goal is to
> > > > > > > reclaim only few slab objects (unused inodes), we still can eventually
> > > > > > > evict all gigabytes of the pagecache at once.
> > > > > > >
> > > > > > > The workload described by Spock has few large non-mapped files in the
> > > > > > > pagecache, so it's especially noticeable.
> > > > > > >
> > > > > > > To solve the problem let's postpone the reclaim of inodes, which have
> > > > > > > more than 1 attached page. Let's wait until the pagecache pages will
> > > > > > > be evicted naturally by scanning the corresponding LRU lists, and only
> > > > > > > then reclaim the inode structure.
> > > > > >
> > > > > > Is this regression serious enough to warrant fixing 4.19.1?
> > > > > 
> > > > > Let's not forget about stable tree(s) which backported 172b06c32b94. I
> > > > > would suggest reverting there.
> > > > 
> > > > Yup.  Sasha, can you please take care of this?
> > > 
> > > Sure, I'll revert it from current stable trees.
> > > 
> > > Should 172b06c32b94 and this commit be backported once Roman confirms
> > > the issue is fixed? As far as I understand 172b06c32b94 addressed an
> > > issue FB were seeing in their fleet and needed to be fixed.
> > 
> > The memcg leak was also independently reported by several companies,
> > so it's not only about our fleet.
> 
> By memcg leak you mean a lot of dead memcgs with small amount of memory
> which are staying behind and the global memory pressure removes them
> only very slowly or almost not at all, right?

Right.

> 
> I have avague recollection that systemd can trigger a pattern which
> makes this "leak" noticeable. Is that right? If yes what would be a
> minimal and safe fix for the stable tree? "mm: don't miss the last page
> because of round-off error" would sound like the candidate but I never
> got around to review it properly.

Yes, systemd can create and destroy a ton of cgroups under some circumstances,
but there is nothing systemd-specific here. It's quite typical to run services
in new cgroups, so with time the number of dying cgroups tends to grow.

I've listed all necessary patches, it's the required set (except the last patch,
but it has to be squashed). f2e821fc8c63 can be probably skipped, but I haven't
tested without it, and it's the most straightforward patch from the set.

Daniel McGinnes has reported the same issue in the cgroups@ mailing list,
and he confirmed that this patchset solved the problem for him.

> > The memcg css leak is fixed by a series of commits (as in the mm tree):
> >   37e521912118 math64: prevent double calculation of DIV64_U64_ROUND_UP() arguments
> >   c6be4e82b1b3 mm: don't miss the last page because of round-off error
> >   f2e821fc8c63 mm: drain memcg stocks on css offlining
> >   03a971b56f18 mm: rework memcg kernel stack accounting
> 
> btw. none of these sha are refering to anything in my git tree. They all
> seem to be in the next tree though.

Yeah, they all are in the mm tree, and hashes are from Johannes's git.

> 
> >   172b06c32b94 mm: slowly shrink slabs with a relatively small number of objects
> > 
> > The last one by itself isn't enough, and it makes no sense to backport it
> > without all other patches. So, I'd either backport them all (including
> > 47036ad4032e ("mm: don't reclaim inodes with many attached pages"),
> > either just revert 172b06c32b94.
> > 
> > Also 172b06c32b94 ("mm: slowly shrink slabs with a relatively small number of objects")
> > by itself is fine, but it reveals an independent issue in inode reclaim code,
> > which 47036ad4032e ("mm: don't reclaim inodes with many attached pages") aims to fix.
> 
> To me it sounds it needs much more time to settle before it can be
> considered safe for the stable tree. Even if the patch itself is correct
> it seems too subtle and reveal a behavior which was not anticipated and
> that just proves it is far from straightforward.

Absolutely. I'm not pushing this to stable at all, that single patch
was an accident.

Thanks!