Re: [PATCH 1/2] sched/wait: Break up long wake list walk

[Linus Torvalds]

On Fri, Aug 18, 2017 at 5:23 AM, Mel Gorman <mgorman@xxxxxxxxxxxxxxxxxxx> wrote:
>
>         new_page = alloc_pages_node(node,
> -               (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE),
> +               (GFP_TRANSHUGE_LIGHT | __GFP_THISNODE) & ~__GFP_DIRECT_RECLAIM,
>                 HPAGE_PMD_ORDER);

That can't make any difference. We already have:

  #define GFP_TRANSHUGE_LIGHT     ((GFP_HIGHUSER_MOVABLE | __GFP_COMP | \
                         __GFP_NOMEMALLOC | __GFP_NOWARN) & ~__GFP_RECLAIM)

and that "& ~__GFP_RECLAIM" is removing __GFP_DIRECT_RECLAIM.

So that patch is a no-op, afaik.

Is there something else expensive in there?

It *might* be simply that we have a shit-ton of threads, and the
thread that holds the page lock for migration is just preempted out.
even if it doesn't really do anything particularly expensive.

                           Linus