Re: [PATCH 1/2] mm: thp: relax __GFP_THISNODE for MADV_HUGEPAGE mappings

[Vlastimil Babka]

On 10/8/18 10:41 PM, David Rientjes wrote:
> +			/*
> +			 * If faulting a hugepage, it is very unlikely that
> +			 * thrashing the zonelist is going to assist compaction
> +			 * in freeing an entire pageblock.  There are no
> +			 * guarantees memory compaction can free an entire
> +			 * pageblock under such memory pressure that it is
> +			 * better to simply fail and fallback to native pages.
> +			 */
> +			if (order == pageblock_order &&
> +					!(current->flags & PF_KTHREAD))
> +				goto nopage;

After we got rid of similar hardcoded heuristics, I would be very
unhappy to start adding them back. A new gfp flag is also unfortunate,
but more acceptable to me.

> +
>  			/*
>  			 * Looks like reclaim/compaction is worth trying, but
>  			 * sync compaction could be very expensive, so keep
>