Re: [PATCH] memcg: Remove memcg_cgroup::id from IDR on mem_cgroup_css_alloc() failure

From: Vladimir Davydov
Date: Wed Aug 01 2018 - 12:22:42 EST


On Wed, Aug 01, 2018 at 11:55:52AM -0400, Johannes Weiner wrote:
> On Tue, Jul 31, 2018 at 04:39:08PM -0700, Andrew Morton wrote:
> > On Mon, 30 Jul 2018 11:31:13 -0400 Johannes Weiner <hannes@xxxxxxxxxxx> wrote:
> >
> > > Subject: [PATCH] mm: memcontrol: simplify memcg idr allocation and error
> > > unwinding
> > >
> > > The memcg ID is allocated early in the multi-step memcg creation
> > > process, which needs 2-step ID allocation and IDR publishing, as well
> > > as two separate IDR cleanup/unwind sites on error.
> > >
> > > Defer the IDR allocation until the last second during onlining to
> > > eliminate all this complexity. There is no requirement to have the ID
> > > and IDR entry earlier than that. And the root reference to the ID is
> > > put in the offline path, so this matches nicely.
> >
> > This patch isn't aware of Kirill's later "mm, memcg: assign memcg-aware
> > shrinkers bitmap to memcg", which altered mem_cgroup_css_online():
> >
> > @@ -4356,6 +4470,11 @@ static int mem_cgroup_css_online(struct
> > {
> > struct mem_cgroup *memcg = mem_cgroup_from_css(css);
> >
> > + if (memcg_alloc_shrinker_maps(memcg)) {
> > + mem_cgroup_id_remove(memcg);
> > + return -ENOMEM;
> > + }
> > +
> > /* Online state pins memcg ID, memcg ID pins CSS */
> > atomic_set(&memcg->id.ref, 1);
> > css_get(css);
> >
>
> Hm, that looks out of place too. The bitmaps are allocated for the
> entire lifetime of the css, not just while it's online.
>
> Any objections to the following fixup to that patch?

That would be incorrect. Memory cgroups that haven't been put online
are invisible to for_each_mem_cgroup(), which is used for expanding
shrinker maps of all cgroups - see memcg_expand_shrinker_maps(). So if
memcg_expand_shrinker_maps() is called between css_alloc and css_online,
it will miss this cgroup and its shrinker_map won't be reallocated to
fit the new id. Allocating the shrinker map in css_online guarantees
that it won't happen. Looks like this code lacks a comment...