Re: [PATCH v3 1/2] mm: memcg: remote memcg charging for kmem allocations

[Shakeel Butt]

On Thu, Mar 15, 2018 at 10:49 AM, Michal Hocko <mhocko@xxxxxxxxxx> wrote:
> On Tue 13-03-18 10:55:18, Shakeel Butt wrote:
>> On Tue, Mar 13, 2018 at 6:49 AM, Michal Hocko <mhocko@xxxxxxxxxx> wrote:
>> > On Wed 21-02-18 14:37:56, Shakeel Butt wrote:
>> > [...]
>> >> +#ifdef CONFIG_MEMCG
>> >> +static inline struct mem_cgroup *memalloc_memcg_save(struct mem_cgroup *memcg)
>> >> +{
>> >> +     struct mem_cgroup *old_memcg = current->target_memcg;
>> >> +     current->target_memcg = memcg;
>> >> +     return old_memcg;
>> >> +}
>> >
>> > So you are relying that the caller will handle the reference counting
>> > properly? I do not think this is a good idea.
>>
>> For the fsnotify use-case, this assumption makes sense as fsnotify has
>> an abstraction of fsnotify_group which is created by the
>> person/process interested in the events and thus can be used to hold
>> the reference to the person/process's memcg.
>
> OK, but there is not any direct connection between fsnotify_group and
> task_struct lifetimes, is it? This makes the API suspectible to
> use-after-free bugs.
>

For fsnotify, whoever is calling [fanotify|inotify]_handle_event()
will have a stable reference to fsnotify_group and fsnotify_group has
reference to memcg. These allocations happen within
[fanotify|inotify]_handle_event(), so, for fsnotify I don't think
there will be use-after-free bugs.

Basically whoever is calling memcg variant of kmem_cache_alloc or
kmalloc should either have stable direct or indirect reference to the
memcg.

>> Another use-case I have
>> in mind is the filesystem mount. Basically attaching a mount with a
>> memcg and thus all user pages and kmem allocations (inodes, dentries)
>> for that mount will be charged to the attached memcg.
>
> So you charge page cache to the origin task but metadata to a different
> memcg?
>

No, both page cache and metadata to a different memcg.

>> In this use-case
>> the super_block is the perfect structure to hold the reference to the
>> memcg.
>>
>> If in future we find a use-case where this assumption does not make
>> sense we can evolve the API and since this is kernel internal API, it
>> should not be hard to evolve.
>>
>> > Also do we need some kind
>> > of debugging facility to detect unbalanced save/restore scopes?
>> >
>>
>> I am not sure, I didn't find other similar patterns (like PF_MEMALLOC)
>> having debugging facility.
>
> Maybe we need something more generic here.
>

Please do let me know if you have something in mind.

>> Maybe we can add such debugging facility
>> when we find more users other than kmalloc & kmem_cache_alloc. Vmalloc
>> may be one but I could not think of a use-case for vmalloc for remote
>> charging, so, no need to add more code at this time.
>>
>> > [...]
>> >> @@ -2260,7 +2269,10 @@ struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep)
>> >>       if (current->memcg_kmem_skip_account)
>> >>               return cachep;
>> >>
>> >> -     memcg = get_mem_cgroup_from_mm(current->mm);
>> >> +     if (current->target_memcg)
>> >> +             memcg = get_mem_cgroup(current->target_memcg);
>> >> +     if (!memcg)
>> >> +             memcg = get_mem_cgroup_from_mm(current->mm);
>> >>       kmemcg_id = READ_ONCE(memcg->kmemcg_id);
>> >>       if (kmemcg_id < 0)
>> >>               goto out;
>> >
>> > You are also adding one branch for _each_ charge path even though the
>> > usecase is rather limited.
>> >
>>
>> I understand the concern but the charging path, IMO, is much complex
>> than just one or couple of additional branches. I can run a simple
>> microbenchmark to see if there is anything noticeable here.
>
> Charging path is still a _hot path_. Especially when the kmem accounting
> is enabled by default. You cannot simply downplay the overhead. We have
> _one_ user but all users should pay the price. This is simply hard to
> justify. Maybe we can thing of something that would put the burden on
> the charging context?
>

I will see if I can find out a way for that.