Re: [PATCH] mm: avoid livelock on !__GFP_FS allocations

[Colin Cross]

On Tue, Oct 25, 2011 at 11:33 PM, David Rientjes <rientjes@xxxxxxxxxx> wrote:
> On Tue, 25 Oct 2011, Colin Cross wrote:
>
>> Makes sense.  What about this?  Official patch to follow.
>>
>> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
>> index fef8dc3..59cd4ff 100644
>> --- a/mm/page_alloc.c
>> +++ b/mm/page_alloc.c
>> @@ -1786,6 +1786,13 @@ should_alloc_retry(gfp_t gfp_mask, unsigned int order,
>>                 return 0;
>>
>>         /*
>> +        * If PM has disabled I/O, OOM is disabled and reclaim is unlikely
>> +        * to make any progress.  To prevent a livelock, don't retry.
>> +        */
>> +       if (!(gfp_allowed_mask & __GFP_FS))
>> +               return 0;
>> +
>> +       /*
>>          * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
>>          * means __GFP_NOFAIL, but that may not be true in other
>>          * implementations.
>
> Eek, this is precisely what we don't want and is functionally the same as
> what you initially proposed except it doesn't care about __GFP_NOFAIL.

This is checking against gfp_allowed_mask, not gfp_mask.

> You're trying to address a suspend issue where nothing on the system can
> logically make progress because __GFP_FS seriously restricts the ability
> of reclaim to do anything useful if it doesn't succeed the first time and
> kswapd isn't effective.  That's why I suggested a hook into
> pm_restrict_gfp_mask() to set a variable and then treat it exactly as
> __GFP_NORETRY in should_alloc_retry().
>
> Consider if nobody is using suspend and they are allocating with GFP_NOFS.
> There's potentially a lot of candidates:
>
>        $ grep -r GFP_NOFS * | wc -l
>        1016
>
> and now we've just introduced a regression where the allocation would
> eventually succeed because of either kswapd, a backing device that is no
> longer congested, or an allocation on another cpu in a context where
> direct reclaim can be more aggressive or the oom killer can at least free
> some memory.
>
> So you definitely want to localize your change to only suspend and
> pm_restrict_gfp_mask() is a very easy way to do it.  So I'd suggest adding
> a static bool that can be tested in should_alloc_retry() and identify such
> situations and tag it as __read_mostly.
>
--
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/