Re: [PATCH 6/9] mm: vmscan: use per memcg nr_deferred of shrinker

[Johannes Weiner <hannes@cmpxchg.org>]

On Wed, Dec 09, 2020 at 09:32:37AM -0800, Yang Shi wrote:
> On Wed, Dec 9, 2020 at 7:42 AM Kirill Tkhai <ktkhai@virtuozzo.com> wrote:
> >
> > On 08.12.2020 20:13, Yang Shi wrote:
> > > On Thu, Dec 3, 2020 at 3:40 AM Kirill Tkhai <ktkhai@virtuozzo.com> wrote:
> > >>
> > >> On 02.12.2020 21:27, Yang Shi wrote:
> > >>> Use per memcg's nr_deferred for memcg aware shrinkers.  The shrinker's nr_deferred
> > >>> will be used in the following cases:
> > >>>     1. Non memcg aware shrinkers
> > >>>     2. !CONFIG_MEMCG
> > >>>     3. memcg is disabled by boot parameter
> > >>>
> > >>> Signed-off-by: Yang Shi <shy828301@gmail.com>
> > >>> ---
> > >>>  mm/vmscan.c | 88 +++++++++++++++++++++++++++++++++++++++++++++++++----
> > >>>  1 file changed, 82 insertions(+), 6 deletions(-)
> > >>>
> > >>> diff --git a/mm/vmscan.c b/mm/vmscan.c
> > >>> index cba0bc8d4661..d569fdcaba79 100644
> > >>> --- a/mm/vmscan.c
> > >>> +++ b/mm/vmscan.c
> > >>> @@ -203,6 +203,12 @@ static DECLARE_RWSEM(shrinker_rwsem);
> > >>>  static DEFINE_IDR(shrinker_idr);
> > >>>  static int shrinker_nr_max;
> > >>>
> > >>> +static inline bool is_deferred_memcg_aware(struct shrinker *shrinker)
> > >>> +{
> > >>> +     return (shrinker->flags & SHRINKER_MEMCG_AWARE) &&
> > >>> +             !mem_cgroup_disabled();
> > >>> +}
> > >>> +
> > >>>  static int prealloc_memcg_shrinker(struct shrinker *shrinker)
> > >>>  {
> > >>>       int id, ret = -ENOMEM;
> > >>> @@ -271,7 +277,58 @@ static bool writeback_throttling_sane(struct scan_control *sc)
> > >>>  #endif
> > >>>       return false;
> > >>>  }
> > >>> +
> > >>> +static inline long count_nr_deferred(struct shrinker *shrinker,
> > >>> +                                  struct shrink_control *sc)
> > >>> +{
> > >>> +     bool per_memcg_deferred = is_deferred_memcg_aware(shrinker) && sc->memcg;
> > >>> +     struct memcg_shrinker_deferred *deferred;
> > >>> +     struct mem_cgroup *memcg = sc->memcg;
> > >>> +     int nid = sc->nid;
> > >>> +     int id = shrinker->id;
> > >>> +     long nr;
> > >>> +
> > >>> +     if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
> > >>> +             nid = 0;
> > >>> +
> > >>> +     if (per_memcg_deferred) {
> > >>> +             deferred = rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_deferred,
> > >>> +                                                  true);
> > >>
> > >> My comment is about both 5/9 and 6/9 patches.
> > >
> > > Sorry for the late reply, I don't know why Gmail filtered this out to spam.
> > >
> > >>
> > >> shrink_slab_memcg() races with mem_cgroup_css_online(). A visibility of CSS_ONLINE flag
> > >> in shrink_slab_memcg()->mem_cgroup_online() does not guarantee that you will see
> > >> memcg->nodeinfo[nid]->shrinker_deferred != NULL in count_nr_deferred(). This may occur
> > >> because of processor reordering on !x86 (there is no a common lock or memory barriers).
> > >>
> > >> Regarding to shrinker_map this is not a problem due to map check in shrink_slab_memcg().
> > >> The map can't be NULL there.
> > >>
> > >> Regarding to shrinker_deferred you should prove either this is not a problem too,
> > >> or to add proper synchronization (maybe, based on barriers) or to add some similar check
> > >> (maybe, in shrink_slab_memcg() too).
> > >
> > > It seems shrink_slab_memcg() might see shrinker_deferred as NULL
> > > either due to the same reason. I don't think there is a guarantee it
> > > won't happen.
> > >
> > > We just need guarantee CSS_ONLINE is seen after shrinker_maps and
> > > shrinker_deferred are allocated, so I'm supposed barriers before
> > > "css->flags |= CSS_ONLINE" should work.
> > >
> > > So the below patch may be ok:
> > >
> > > diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> > > index df128cab900f..9f7fb0450d69 100644
> > > --- a/mm/memcontrol.c
> > > +++ b/mm/memcontrol.c
> > > @@ -5539,6 +5539,12 @@ static int mem_cgroup_css_online(struct
> > > cgroup_subsys_state *css)
> > >                 return -ENOMEM;
> > >         }
> > >
> > >
> > > +       /*
> > > +        * Barrier for CSS_ONLINE, so that shrink_slab_memcg() sees
> > > shirnker_maps
> > > +        * and shrinker_deferred before CSS_ONLINE.
> > > +        */
> > > +       smp_mb();
> > > +
> > >         /* Online state pins memcg ID, memcg ID pins CSS */
> > >         refcount_set(&memcg->id.ref, 1);
> > >         css_get(css);
> >
> > smp barriers synchronize data access from different cpus. They should go in a pair.
> > In case of you add the smp barrier into mem_cgroup_css_online(), we should also
> > add one more smp barrier in another place, which we want to synchonize with this.
> > Also, every place should contain a comment referring to its pair: "Pairs with...".
> 
> Thanks, I think you are correct. Looked into it further, it seems the
> race pattern looks like:
> 
> CPU A                                                                  CPU B
> store shrinker_maps pointer                      load CSS_ONLINE
> store CSS_ONLINE                                   load shrinker_maps pointer
> 
> By checking the memory-barriers document, it seems we need write
> barrier/read barrier pair as below:
> 
> CPU A                                                                  CPU B
> store shrinker_maps pointer                       load CSS_ONLINE
> <write barrier>                                             <read barrier>
> store CSS_ONLINE                                    load shrinker_maps pointer
> 
> 
> So, the patch should look like:
> 
> diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> index df128cab900f..489c0a84f82b 100644
> --- a/mm/memcontrol.c
> +++ b/mm/memcontrol.c
> @@ -5539,6 +5539,13 @@ static int mem_cgroup_css_online(struct
> cgroup_subsys_state *css)
>                 return -ENOMEM;
>         }
> 
> +       /*
> +        * Barrier for CSS_ONLINE, so that shrink_slab_memcg() sees
> shirnker_maps
> +        * and shrinker_deferred before CSS_ONLINE. It pairs with the
> read barrier
> +        * in shrink_slab_memcg().
> +        */
> +       smp_wmb();

Is there a reason why the shrinker allocations aren't done in
.css_alloc()? That would take care of all necessary ordering:

      #0
      css = ->css_alloc()
      list_add_tail_rcu(css, parent->children)
        rcu_assign_pointer()
      ->css_online(css)
      css->flags |= CSS_ONLINE

      #1
      memcg = mem_cgroup_iter()
        list_entry_rcu()
	  rcu_dereference()
      shrink_slab(.., memcg)

RCU ensures that once the cgroup shows up in the reclaim cgroup it's
also fully allocated.

>         /* Online state pins memcg ID, memcg ID pins CSS */
>         refcount_set(&memcg->id.ref, 1);
>         css_get(css);
> diff --git a/mm/vmscan.c b/mm/vmscan.c
> index 9d2a6485e982..fc9bda576d98 100644
> --- a/mm/vmscan.c
> +++ b/mm/vmscan.c
> @@ -603,13 +603,15 @@ static unsigned long shrink_slab_memcg(gfp_t
> gfp_mask, int nid,
>         if (!mem_cgroup_online(memcg))
>                 return 0;

...then we should be able to delete this online check here entirely:

A not-yet online cgroup is guaranteed to have a shrinker map, just
with no bits set. The shrinker code handles that just fine.

An offlined cgroup will eventually have an empty bitmap as the called
shrinkers return SHRINK_EMPTY. This could also be shortcut by clearing
the bit in memcg_drain_list_lru_node() the same way we set it in the
parent when we move all objects upward, but seems correct as-is.