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

[David Rientjes]

On Mon, 24 Oct 2011, Colin Cross wrote:

> Under the following conditions, __alloc_pages_slowpath can loop
> forever:
> gfp_mask & __GFP_WAIT is true
> gfp_mask & __GFP_FS is false
> reclaim and compaction make no progress
> order <= PAGE_ALLOC_COSTLY_ORDER
> 

The oom killer is only called for __GFP_FS because we want to ensure that 
we don't inadvertently kill something if we didn't have a chance to at 
least make a good effort at direct reclaim.  There's a very high liklihood 
that direct reclaim would succeed with __GFP_FS, so we loop endlessly 
waiting for either kswapd to reclaim in the background even though it 
might not be able to because of filesystem locks or another allocation 
happens in a context that allows reclaim to succeed or oom killing.

For low-order allocations (those at or below PAGE_ALLOC_COSTLY_ORDER) 
where fragmentation isn't a huge issue, __GFP_WAIT && !__GFP_FS && 
!did_some_progress makes sense.

> These conditions happen very often during suspend and resume,
> when pm_restrict_gfp_mask() effectively converts all GFP_KERNEL
> allocations into __GFP_WAIT.
> 

This is the problem.  All allocations now have no chance of ever having 
direct reclaim succeed nor the oom killer called.  It seems like you would 
want pm_restrict_gfp_mask() to also include __GFP_NORETRY and ensure it 
can never be called for __GFP_NOFAIL.

> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index fef8dc3..dcd99b3 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -2193,6 +2193,10 @@ rebalance:
>  			}
>  
>  			goto restart;
> +		} else {
> +			/* If we aren't going to try the OOM killer, give up */
> +			if (!(gfp_mask & __GFP_NOFAIL))
> +				goto nopage;
>  		}
>  	}
>  

Nack on this, it is going to cause many very verbose allocation failures 
(if !__GFP_NOWARN) when not using suspend because we're not in a context 
where we can do sensible reclaim or compaction and presently kswapd can 
either reclaim or another allocation will allow low-order amounts of 
memory to be reclaimed or the oom killer to free some memory.  It would 
introduce a regression into page allocation.
--
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/