Re: [PATCH 2/2] mm, oom: do not enfore OOM killer for __GFP_NOFAIL automatically

[Vlastimil Babka]

On 12/06/2016 11:38 AM, Tetsuo Handa wrote:
>>
>> So we are somewhere in the middle between pre-mature and pointless
>> system disruption (GFP_NOFS with a lots of metadata or lowmem request)
>> where the OOM killer even might not help and potential lockup which is
>> inevitable with the current design. Dunno about you but I would rather
>> go with the first option. To be honest I really fail to understand your
>> line of argumentation. We have this
>> 	do {
>> 		cond_resched();
>> 	} while (!(page = alloc_page(GFP_NOFS)));
>> vs.
>> 	page = alloc_page(GFP_NOFS | __GFP_NOFAIL);
>>
>> the first one doesn't invoke OOM killer while the later does. This
>> discrepancy just cannot make any sense... The same is true for
>>
>> 	alloc_page(GFP_DMA) vs alloc_page(GFP_DMA|__GFP_NOFAIL)
>>
>> Now we can discuss whether it is a _good_ idea to not invoke OOM killer
>> for those exceptions but whatever we do __GFP_NOFAIL is not a way to
>> give such a subtle side effect. Or do you disagree even with that?
> 
> "[PATCH 1/2] mm: consolidate GFP_NOFAIL checks in the allocator slowpath"
> silently changes __GFP_NOFAIL vs. __GFP_NORETRY priority.

I guess that wasn't intended?

> Currently, __GFP_NORETRY is stronger than __GFP_NOFAIL; __GFP_NOFAIL
> allocation requests fail without invoking the OOM killer when both
> __GFP_NORETRY and __GFP_NOFAIL are given.
> 
> With [PATCH 1/2], __GFP_NOFAIL becomes stronger than __GFP_NORETRY;
> __GFP_NOFAIL allocation requests will loop forever without invoking
> the OOM killer when both __GFP_NORETRY and __GFP_NOFAIL are given.

Does such combination of flag make sense? Should we warn about it, or
even silently remove __GFP_NORETRY in such case?

> Those callers which prefer lockup over panic can specify both
> __GFP_NORETRY and __GFP_NOFAIL.

What lockup exactly, if __GFP_NORETRY did lead to fail?