Re: [PATCH 1/2] mm/vmscan: consolidate promo vs. high watermark checks

[sashiko-bot@kernel.org]

Thank you for your contribution! Sashiko AI review found no new issues
in this patch, but noticed 1 related pre-existing issue it does not
introduce:
- [High] mm: livelock in throttle_direct_reclaim() [1]
--

commit 577a3c6c5ce30c67e4a789965d099d1ca4f2b341
Author: Joshua Hahn <joshua.hahnjy@gmail.com>

mm/vmscan: consolidate promo vs. high watermark checks

This commit introduces kswapd_wmark_pages() to consolidate the promo vs.
high watermark selection in should_abort_scan() and pgdat_balanced().

> diff --git a/mm/vmscan.c b/mm/vmscan.c
> index 071a6eaef1a1a..97c5843e14f15 100644
> --- a/mm/vmscan.c
> +++ b/mm/vmscan.c
[ ... ]
> @@ -6966,10 +6973,7 @@ static bool pgdat_balanced(pg_data_t *pgdat, int order, int highest_zoneidx)
>  		enum zone_stat_item item;
>  		unsigned long free_pages;
>  
> -		if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING)
> -			mark = promo_wmark_pages(zone);
> -		else
> -			mark = high_wmark_pages(zone);
> +		mark = kswapd_wmark_pages(zone);

[Severity: High]
This problem wasn't introduced by this patch, but can pgdat_balanced()
livelock with allow_direct_reclaim() in throttle_direct_reclaim()
[1] when a lower zone such as DMA has free pages above its watermark
but no reclaimable pages while higher zones are exhausted?

allow_direct_reclaim() skips zones that have free pages but no reclaimable
pages:

mm/vmscan.c:allow_direct_reclaim() {
    ...
	for_each_managed_zone_pgdat(zone, pgdat, i, ZONE_NORMAL) {
		if (!zone_reclaimable_pages(zone) &&
		    zone_page_state_snapshot(zone, NR_FREE_PAGES))
			continue;
    ...
}

When all other zones are exhausted, allow_direct_reclaim() returns false
and throttle_direct_reclaim() wakes kswapd and sleeps on
pgdat->reclaim_wait.  However, pgdat_balanced() does not skip that zone,
sees its watermark satisfied, and returns true.  balance_pgdat() then exits
without reclaiming or incrementing pgdat->kswapd_failures, so
kswapd_test_hopeless() never triggers and direct reclaimers and kswapd
wake each other in a tight loop.

[1] https://sashiko.dev/#/bug/linux-b88b79bf-7e7f-40df-b7e7-7afea197c220

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20261106172620.318113-1-joshua.hahnjy@gmail.com?part=1