[PATCH v9 08/16] mm/mglru: lift youngest gen protection for higher priority

["Barry Song (Xiaomi)" <baohua@kernel.org>]

From: Kairui Song <kasong@tencent.com>

The forced protection of younger generations contradicts the
goal of minimizing the I/O cost of reclaim. Under high memory
pressure, MGLRU can repeatedly trigger premature OOMs, while
the classic LRU does not.
The problem is that aging may fail to keep up with reclaim.
When this happens, reclaimers may all abort prematurely
because they see only empty generations or just two
generations remaining.
Proactive aging does not solve this problem either. The
number of threads performing proactive aging is limited,
while the number of concurrent reclaimers can be much
larger. This is a fundamental limitation of MGLRU. One
possible solution is to use an aging lock. However, the
aging lock introduces significant jitter under heavy memory
pressure and has been reported to cause noticeable lags
on Android.
Furthermore, the generation-based reclaim model can
conflict with the goal of swappiness, which is to minimize
I/O cost rather than blindly follow folio hotness.
The previous commit allowed reclaimers running at
non-default priorities to reclaim up to `lrugen->max_seq`,
while reclaimers at the default priority could reclaim only
up to `lrugen->min_seq`.
This patch further refines generation protection based on
reclaim priority. At the default priority, reclaimers can
reclaim generations down to `lrugen->max_seq - MIN_NR_GENS`.
As reclaim priority increases, reclaimers are progressively
allowed to reclaim younger generations.
With this graduated protection, premature OOMs are no
longer observed, while swappiness is also respected.

Signed-off-by: Kairui Song <kasong@tencent.com>
Signed-off-by: Barry Song (Xiaomi) <baohua@kernel.org>
Reviewed-by: Kunwu Chan <kunwu.chan@gmail.com>
---
 mm/vmscan.c | 10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)

diff --git a/mm/vmscan.c b/mm/vmscan.c
index 18cf1c084e78..b538abf02578 100644
--- a/mm/vmscan.c
+++ b/mm/vmscan.c
@@ -4772,12 +4772,14 @@ static int scan_folios(unsigned long nr_to_scan, struct lruvec *lruvec,
 	VM_WARN_ON_ONCE(!list_empty(list));
 
 	/*
-	 * Respect the gen hotness distribution at low priority to avoid
-	 * aging and scan cost, but prefer to satisfy the reclaimers request
-	 * when under pressure to respect swappiness more and reduce IO cost.
+	 * Respect hotness at default priority, and only reach into the
+	 * newest generation when under severe pressure, so hot folios are
+	 * not evicted as soon as priority is raised.
 	 */
 	if (sc->priority == DEF_PRIORITY)
-		seq = lrugen->min_seq[type];
+		seq = lrugen->max_seq - MIN_NR_GENS;
+	else if (sc->priority > DEF_PRIORITY / 2)
+		seq = lrugen->max_seq - 1;
 	else
 		seq = lrugen->max_seq;
 
-- 
2.39.3 (Apple Git-146)