Re: [PATCH v2] slab,slub: ignore __GFP_WAIT if we're booting or suspending

[Pekka Enberg <penberg@cs.helsinki.fi>]

Hi Linus,

Linus Torvalds wrote:
> 
> On Fri, 12 Jun 2009, Pekka J Enberg wrote:
>>  
>> +	if (system_state != SYSTEM_RUNNING)
>> +		local_flags &= ~__GFP_WAIT;
>> +
>> +	might_sleep_if(local_flags & __GFP_WAIT);
> 
> This is pointless.
> 
> You're doing the "might_sleep_if()" way too late. At that point, you've 
> already lost 99% of all coverage, since now none of the cases of just 
> finding a free slab entry on the list will ever trigger that 
> "might_sleep()" case.
> 
> So you need to do this _early_, at the entry-point, not late, at cache 
> re-fill time.
> 
> So rather than removing the might_sleep_if() at the early point, and then 
> moving it to this late stage (because you only do the local_flags fixups 
> late), you need to move the local-flags fixup early instead, and do the 
> might_sleep_it() there.
> 
> The whole point of "might_sleep()" is that it triggers every time if 
> something is called in the wrong context - not just for the cases where it 
> actually _does_ sleep.

OK, makes sense. So what do you think of this patch then:

http://patchwork.kernel.org/patch/29733/

It's what Ben has been proposing all along in a slightly edited form.

			Pekka