Re: [PATCH 1/2] fs: fsnotify: account fsnotify metadata to kmemcg

[Amir Goldstein <amir73il@gmail.com>]

On Tue, Jun 26, 2018 at 11:05 PM, Shakeel Butt <shakeelb@google.com> wrote:
> On Tue, Jun 26, 2018 at 12:03 PM Johannes Weiner <hannes@cmpxchg.org> wrote:
>>
>> On Mon, Jun 25, 2018 at 04:06:58PM -0700, Shakeel Butt wrote:
>> > @@ -140,8 +141,9 @@ struct fanotify_event_info *fanotify_alloc_event(struct fsnotify_group *group,
>> >                                                struct inode *inode, u32 mask,
>> >                                                const struct path *path)
>> >  {
>> > -     struct fanotify_event_info *event;
>> > +     struct fanotify_event_info *event = NULL;
>> >       gfp_t gfp = GFP_KERNEL;
>> > +     struct mem_cgroup *old_memcg = NULL;
>> >
>> >       /*
>> >        * For queues with unlimited length lost events are not expected and
>> > @@ -151,19 +153,25 @@ struct fanotify_event_info *fanotify_alloc_event(struct fsnotify_group *group,
>> >       if (group->max_events == UINT_MAX)
>> >               gfp |= __GFP_NOFAIL;
>> >
>> > +     /* Whoever is interested in the event, pays for the allocation. */
>> > +     if (group->memcg) {
>> > +             gfp |= __GFP_ACCOUNT;
>> > +             old_memcg = memalloc_use_memcg(group->memcg);
>> > +     }
>>
>> group->memcg is only NULL when memcg is disabled or there is some
>> offlining race. Can you make memalloc_use_memcg(NULL) mean that it
>> should charge root_mem_cgroup instead of current->mm->memcg? That way
>> we can make this site unconditional while retaining the behavior:
>>
>>         gfp_t gfp = GFP_KERNEL | __GFP_ACCOUNT;
>>
>>         memalloc_use_memcg(group->memcg);
>>         kmem_cache_alloc(..., gfp);
>> out:
>>         memalloc_unuse_memcg();
>>
>> (dropping old_memcg and the unuse parameter as per the other mail)
>>
>
> group->memcg is only NULL when memcg is disabled (i.e.
> get_mem_cgroup_from_mm() returns root_mem_cgroup for offlined
> mm->memcg). Though group->memcg can point to an offlined memcg.
>
> If I understand you correctly this is what we want:
>
> 1. If group->memcg is NULL then __GFP_ACCOUNT is a noop i.e. memcg is disabled.
> 2. If group->memcg is root_mem_cgroup, then __GFP_ACCOUNT again is a
> kind of noop (charges to root_mem_cgroups are bypassed).
> 3. If group->memcg is offlined memcg, then make __GFP_ACCOUNT noop by
> returning root_mem_cgroup from get_mem_cgroup_from_current().
> 4. Else charge group->memcg.
>
> This seems reasonable. After your Ack and Amir's or Jan's answer to
> the nesting query, I will resend the next version of this patch
> series.
>
> In future if we find any use-cases of memalloc_use_memcg nesting then
> we can make it work for nesting.
>

For the fsnotify use case memalloc_use_memcg() certainly doesn't
need to nest, but I wonder, if that facility becomes popular among different
subsystems, how exactly do you intend to monitor that it doesn't grow
nested use cases? I would suggest that you at least leave a
WARN_ON_ONCE if memalloc_use_memcg() is called and
active_memcg is already set.

Thanks,
Amir.