On Thu, Dec 10, 2020 at 7:36 AM Johannes Weiner <hannes@xxxxxxxxxxx> wrote:
>
> On Wed, Dec 02, 2020 at 10:27:21AM -0800, Yang Shi wrote:
> > @@ -504,6 +577,34 @@ int memcg_expand_shrinker_maps(int new_id)
> > return ret;
> > }
> >
> > +int memcg_expand_shrinker_deferred(int new_id)
> > +{
> > + int size, old_size, ret = 0;
> > + struct mem_cgroup *memcg;
> > +
> > + size = (new_id + 1) * sizeof(atomic_long_t);
> > + old_size = memcg_shrinker_deferred_size;
> > + if (size <= old_size)
> > + return 0;
> > +
> > + mutex_lock(&memcg_shrinker_mutex);
>
> The locking is somewhat confusing. I was wondering why we first read
> memcg_shrinker_deferred_size "locklessly", then change it while
> holding the &memcg_shrinker_mutex.
>
> memcg_shrinker_deferred_size only changes under shrinker_rwsem(write),
> correct? This should be documented in a comment, IMO.
Yes, it is correct.
>
> memcg_shrinker_mutex looks superfluous then. The memcg allocation path
> is the read-side of memcg_shrinker_deferred_size, and so simply needs
> to take shrinker_rwsem(read) to lock out shrinker (de)registration.
I see you point. Yes, it seems shrinker_{maps|deferred} allocation
could be synchronized with shrinker registration by shrinker_rwsem.
memcg_shrinker_mutex is just renamed from memcg_shrinker_map_mutex
which was introduced by shrinker_maps patchset. I'm not quite sure why
this mutex was introduced at the first place, I guess the main purpose
is to *not* exacerbate the contention of shrinker_rwsem?
If that contention is not a concern, we could remove that dedicated mutex.
>
> Also, isn't memcg_shrinker_deferred_size just shrinker_nr_max? And
No, it is variable. It is nr * sizeof(atomit_long_t). The nr is the
current last shrinker ID. If a new shrinker is registered, the nr may
grow.
> memcg_expand_shrinker_deferred() is only called when size >= old_size
> in the first place (because id >= shrinker_nr_max)?
Yes.
