Re: [PATCH 2/2] mm/page_alloc: only skip kswapd nodes below their kswapd watermark

[Lorenzo Stoakes (ARM)]

On Fri, Nov 06, 2026 at 09:26:19AM -0800, Joshua Hahn wrote:
> Commit 19c5fb83f2a4 ("mm: page_alloc: avoid kswapd thrashing due to
> NUMA restrictions") makes get_page_from_freelist() skip any node whose
> kswapd is active to avoid piling on an already-pressured node that is
> constantly churning to make free space that is immediately consumed.
>
> But kswapd can also be woken up for order 9 requests with THP defrag
> set to defer. This means that a fragmented node with plenty of free
> memory will have a kswapd thread continuously spinning and preventing
> order 0 pages from allocating on the node. This causes the opposite
> problem of piling on pressure on all other nodes.
>
> This behavior was observed on some Meta production workloads running on
> 2-NUMA systems, where processes with a preference for node0 allocations
> continuously allocated order 0 pages on node1, despite node0 having
> hundreds of GB of free memory.
>
> To prevent piling pressure on any one node, only skip a node during
> page allocation if the zone also fails the watermark kswapd is
> reclaiming it up to. This is the high watermark, or the promo watermark
> if NUMA balancing memory tiering is enabled. Make sure to mask the
> reserve flags so that __GFP_HIGH / atomic requests in the slowpath
> do not get a relaxed version of the check.
>
> This preserves the existing fix to prevent piling memory on the local
> node, while also preventing memory pileups on non-local nodes when
> the local node has memory.
>
> On a 2-NUMA VM with a 60G node0 and a 20G node1, with THP enabled=always
> and defrag=defer, I ran a process that prefers allocating on node0.
> With a fragmented node0:
> - Before the patch, 66% of faults were served from node0
> - After the patch, 100% of faults were served from node0
>
> Suggested-by: Daniel Hodges <hodgesd@xxxxxxxx>
> Signed-off-by: Joshua Hahn <joshua.hahnjy@xxxxxxxxx>
> ---
>  mm/page_alloc.c | 16 ++++++++++------
>  1 file changed, 10 insertions(+), 6 deletions(-)
>
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index 5423ff4805f84..8e8a2c8d6e6a4 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -3882,14 +3882,18 @@ get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
>  		}
>
>  		/*
> -		 * If kswapd is already active on a node, keep looking
> -		 * for other nodes that might be idle. This can happen
> -		 * if another process has NUMA bindings and is causing
> -		 * kswapd wakeups on only some nodes. Avoid accidental
> -		 * "node_reclaim_mode"-like behavior in this case.
> +		 * If kswapd is already active on a node and the zone is below
> +		 * its kswapd watermark for this request, keep looking for other
> +		 * nodes that might be idle. This can happen if another process
> +		 * has NUMA bindings and is causing kswapd wakeups on only some
> +		 * nodes. Avoid accidental "node_reclaim_mode" behavior.
>  		 */
>  		if (skip_kswapd_nodes &&
> -		    !waitqueue_active(&zone->zone_pgdat->kswapd_wait)) {
> +		    !waitqueue_active(&zone->zone_pgdat->kswapd_wait) &&
> +		    !zone_watermark_fast(zone, order, kswapd_wmark_pages(zone),
> +					 ac->highest_zoneidx,
> +					 alloc_flags & ~ALLOC_RESERVES,
> +					 gfp_mask)) {

I know this code is hardly the cleanest, but you're now adding a 3rd condition
to a branch that was already complicated enough to have a massive comment and
duplicating a block of code that already dupliclated twice here.

You're also then going against the convention of

	mark = ...
	if (... !zone_watermark_fast(zone, order, mark,
					 ac->highest_zoneidx, alloc_flags,
					 gfp_mask)) {

(Obviously yours varies on alloc_flags too)

And also inlining the mark value.

Could you clean this up a little given this isn't a fix?

In general extending functionality should pay down some technical debt if
possible, it's ok to add a preceding patch in the series to do something like
that.

>  			skipped_kswapd_nodes = true;
>  			continue;
>  		}
> --
> 2.53.0-Meta
>

--
Cheers, Lorenzo