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

[Colin Cross]

On Tue, Oct 25, 2011 at 3:18 PM, David Rientjes <rientjes@xxxxxxxxxx> wrote:
> On Tue, 25 Oct 2011, Mel Gorman wrote:
>
>> That said, it will be difficult to remember why checking __GFP_NOFAIL in
>> this case is necessary and someone might "optimitise" it away later. It
>> would be preferable if it was self-documenting. Maybe something like
>> this? (This is totally untested)
>>
>
> __GFP_NOFAIL _should_ be optimized away in this case because all he's
> passing is __GFP_WAIT | __GFP_NOFAIL.  That doesn't make any sense unless
> all you want to do is livelock.

__GFP_NOFAIL is not set in the case that I care about.  If my change
is hit, no forward progress has been made, so I agree it should not
honor __GFP_NOFAIL.

> __GFP_NOFAIL doesn't mean the page allocator would infinitely loop in all
> conditions.  That's why GFP_ATOMIC | __GFP_NOFAIL actually fails, and I
> would argue that __GFP_WAIT | __GFP_NOFAIL should fail as well since it's
> the exact same condition except doesn't have access to the extra memory
> reserves.
>
> Suspend needs to either set __GFP_NORETRY to avoid the livelock if it's
> going to disable all means of memory reclaiming or freeing in the page
> allocator.  Or, better yet, just make it GFP_NOWAIT.
>

It would be nice to give compaction and the slab shrinker a chance to
recover a few pages, both methods will work fine in suspend.
GFP_NOWAIT will prevent them from ever running, and __GFP_NORETRY will
give up even if they are making progress but haven't recovered enough
pages.

Converting suspend to GFP_NOWAIT would simply be ~GFP_KERNEL instead
of ~GFP_IOFS in pm_restrict_gfp_mask().
--
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/