From: Vlastimil Babka <vbabka@suse.cz>

Subject: prevent active file list thrashing due to refault detection

Patch-mainline: Never, discussing proper upstream solution

References: VM Performance, bsc#1156286

In bsc#1156286 we found that 12SP4 kernel regression compared to 12SP3 is due

to commit 2a2e48854d70 ("mm: vmscan: fix IO/refault regression in cache

workingset transition") causing active file list thrashing, as the refault

counter may increase since the last snapshot between kswapd runs and cause

the second kswapd run to focus all reclaim on the active list.

Proper upstreamable solution needs to be discussed, but we need to fix the

regression meanwhile, so effectively disabling commit 2a2e48854d70 is the

simplest option. There has been positive feedback from the customer, and

performance team found no obvious regressions from this change.

Signed-off-by: Vlastimil Babka <vbabka@suse.cz>

---

 mm/vmscan.c |    8 ++++++++

 1 file changed, 8 insertions(+)

--- a/mm/vmscan.c

+++ b/mm/vmscan.c

@@ -2061,6 +2061,12 @@ static bool inactive_list_is_low(struct

 	unsigned long refaults;

 	unsigned long gb;

+	/*

+	 * spurious refault detection results in active list thrashing,

+	 * disable it - bsc#1156286

+	 */

+	actual_reclaim = false;

+

 	/*

 	 * If we don't have swap space, anonymous page deactivation

 	 * is pointless.

@@ -2091,6 +2097,8 @@ static bool inactive_list_is_low(struct

 			inactive_ratio = 1;

 	}

+	/* bsc#1156286 - don't lose the tracepoint */

+	actual_reclaim = true;

 	if (actual_reclaim)

 		trace_mm_vmscan_inactive_list_is_low(pgdat->node_id, sc->reclaim_idx,

 			lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive,