[Devel] [PATCH v2 3/7] memcg, slab: separate memcg vs root cache creation paths

[Vladimir Davydov]

On 02/06/2014 10:17 PM, Michal Hocko wrote:
> On Thu 06-02-14 21:12:51, Vladimir Davydov wrote:
>> On 02/06/2014 08:41 PM, Michal Hocko wrote:
> [...]
>>>> +int kmem_cache_create_memcg(struct mem_cgroup *memcg, struct kmem_cache *cachep)
>>>>  {
>>>> -	return kmem_cache_create_memcg(NULL, name, size, align, flags, ctor, NULL);
>>>> +	struct kmem_cache *s;
>>>> +	int err;
>>>> +
>>>> +	get_online_cpus();
>>>> +	mutex_lock(&slab_mutex);
>>>> +
>>>> +	/*
>>>> +	 * Since per-memcg caches are created asynchronously on first
>>>> +	 * allocation (see memcg_kmem_get_cache()), several threads can try to
>>>> +	 * create the same cache, but only one of them may succeed.
>>>> +	 */
>>>> +	err = -EEXIST;
>>> Does it make any sense to report the error here? If we are racing then at
>>> least on part wins and the work is done.
>> Yeah, you're perfectly right. It's better to return 0 here.
> Why not void?

Yeah, better to make it void for now, just to keep it clean. I guess if
one day we need an error code there (for accounting of error reporting),
we'll add it then, but currently there is no point in that.

>
>>> We should probably warn about errors which prevent from accounting but
>>> I do not think there is much more we can do so returning an error code
>>> from this function seems pointless. memcg_create_cache_work_func ignores
>>> the return value anyway.
>> I do not think warnings are appropriate here, because it is not actually
>> an error if we are short on memory and can't do proper memcg accounting
>> due to this. Perhaps, we'd better add fail counters for memcg cache
>> creations and/or accounting to the root cache instead of memcg's one.
>> That would be useful for debugging. I'm not sure though.
> warn on once per memcg would be probably sufficient but it would still
> be great if an admin could see that a memcg is not accounted although it
> is supposed to be. Scanning all the memcgs might be really impractical.
> We do not fail allocations needed for those object in the real life now
> but we shouldn't rely on that.

Hmm, an alert in dmesg first time kmem_cache_create_memcg() fails for a
particular memcg, just to draw attention, plus accounting of total
number of failures for each memcg so that admin could check if it's a
real problem... Sounds reasonable to me. I guess I'll handle it in a
separate patch a bit later.

Thanks.