Re: [PATCH v3 15/16] memcg/sl[au]b: shrink dead caches

From: JoonSoo Kim
Date: Fri Sep 21 2012 - 05:27:55 EST


Hi, Glauber.

>> 2012/9/18 Glauber Costa <glommer@xxxxxxxxxxxxx>:
>>> diff --git a/mm/slub.c b/mm/slub.c
>>> index 0b68d15..9d79216 100644
>>> --- a/mm/slub.c
>>> +++ b/mm/slub.c
>>> @@ -2602,6 +2602,7 @@ redo:
>>> } else
>>> __slab_free(s, page, x, addr);
>>>
>>> + kmem_cache_verify_dead(s);
>>> }
>>
>> As far as u know, I am not a expert and don't know anything about memcg.
>> IMHO, this implementation may hurt system performance in some case.
>>
>> In case of memcg is destoried, remained kmem_cache is marked "dead".
>> After it is marked,
>> every free operation to this "dead" kmem_cache call
>> kmem_cache_verify_dead() and finally call kmem_cache_shrink().
>
> As long as it is restricted to that cache, this is a non issue.
> dead caches are exactly what they name imply: dead.
>
> Means that we actively want them to go away, and just don't kill them
> right away because they have some inflight objects - which we expect not
> to be too much.

Hmm.. I don't think so.
We can destroy memcg whenever we want, is it right?
If it is right, there is many inflight objects when we destory memcg.
If there is so many inflight objects, performance of these processes
can be hurt too much.

>> And, I found one case that destroying memcg's kmem_cache don't works properly.
>> If we destroy memcg after all object is freed, current implementation
>> doesn't destroy kmem_cache.
>> kmem_cache_destroy_work_func() check "cachep->memcg_params.nr_pages == 0",
>> but in this case, it return false, because kmem_cache may have
>> cpu_slab, and cpu_partials_slabs.
>> As we already free all objects, kmem_cache_verify_dead() is not invoked forever.
>> I think that we need another kmem_cache_shrink() in
>> kmem_cache_destroy_work_func().
>
> I'll take a look here. What you describe makes sense, and can
> potentially happen. I tried to handle this case with care in
> destroy_all_caches, but I may have always made a mistake...
>
> Did you see this actively happening, or are you just assuming this can
> happen from your read of the code?

Just read of the code.

Thanks.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/