[Devel] Re: [PATCH v5 06/14] memcg: kmem controller infrastructure

[Glauber Costa]

On 10/19/2012 01:59 AM, David Rientjes wrote:
> On Thu, 18 Oct 2012, Glauber Costa wrote:
> 
>>>> @@ -2630,6 +2634,171 @@ static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
>>>>  	memcg_check_events(memcg, page);
>>>>  }
>>>>  
>>>> +#ifdef CONFIG_MEMCG_KMEM
>>>> +static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
>>>> +{
>>>> +	return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
>>>> +		(memcg->kmem_accounted & KMEM_ACCOUNTED_MASK);
>>>> +}
>>>> +
>>>> +static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
>>>> +{
>>>> +	struct res_counter *fail_res;
>>>> +	struct mem_cgroup *_memcg;
>>>> +	int ret = 0;
>>>> +	bool may_oom;
>>>> +
>>>> +	ret = res_counter_charge(&memcg->kmem, size, &fail_res);
>>>> +	if (ret)
>>>> +		return ret;
>>>> +
>>>> +	/*
>>>> +	 * Conditions under which we can wait for the oom_killer.
>>>> +	 * We have to be able to wait, but also, if we can't retry,
>>>> +	 * we obviously shouldn't go mess with oom.
>>>> +	 */
>>>> +	may_oom = (gfp & __GFP_WAIT) && !(gfp & __GFP_NORETRY);
>>>
>>> What about gfp & __GFP_FS?
>>>
>>
>> Do you intend to prevent or allow OOM under that flag? I personally
>> think that anything that accepts to be OOM-killed should have GFP_WAIT
>> set, so that ought to be enough.
>>
> 
> The oom killer in the page allocator cannot trigger without __GFP_FS 
> because direct reclaim has little chance of being very successful and 
> thus we end up needlessly killing processes, and that tends to happen 
> quite a bit if we dont check for it.  Seems like this would also happen 
> with memcg if mem_cgroup_reclaim() has a large probability of failing?
> 

I can indeed see tests for GFP_FS in some key locations in mm/ before
calling the OOM Killer.

Should I test for GFP_IO as well? If the idea is preventing OOM to
trigger for allocations that can write their pages back, how would you
feel about the following test:

may_oom = (gfp & GFP_KERNEL) && !(gfp & __GFP_NORETRY) ?

Michal, what is your take in here?