Re: regression: mm: vmscan: - size XL irqoff time increase v6.10+

[Mike Galbraith <efault@gmx.de>]

On Tue, 2024-10-29 at 13:03 +0100, Mike Galbraith wrote:
> Greetings,
>
> I've noticed size XL cyclictest -Smp99 wakeup latency spikes for some
> time now, but due to bandwidth etc (ie laziness), never got around to
> taking a peek until today.

isolate_lru_folios() dropped off the irqsoff radar with the below.

mm: vmscan: Limit isolate_lru_folios() IRQ holdoff latency impact.

While examining an RT latency regression beginning at v6.10, it was
discovered that isolate_lru_folios() runtimes could cause IRQ holdoff
to easily span several ticks.

Given it's called with IRQs disabled, cap its runtme to around a ms.

Signed-off-by: Mike Galbraith <efault@gmx.de>
---
 mm/vmscan.c |   12 ++++++++++++
 1 file changed, 12 insertions(+)

--- a/mm/vmscan.c
+++ b/mm/vmscan.c
@@ -63,6 +63,7 @@
 #include <linux/swapops.h>
 #include <linux/balloon_compaction.h>
 #include <linux/sched/sysctl.h>
+#include <linux/sched/clock.h>

 #include "internal.h"
 #include "swap.h"
@@ -1656,6 +1657,8 @@ static unsigned long isolate_lru_folios(
 	unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
 	unsigned long skipped = 0;
 	unsigned long scan, total_scan, nr_pages;
+	int cpu = raw_smp_processor_id(), iter = 0;
+	u64 then = cpu_clock(cpu);
 	LIST_HEAD(folios_skipped);

 	total_scan = 0;
@@ -1709,6 +1712,15 @@ static unsigned long isolate_lru_folios(
 		move_to = dst;
 move:
 		list_move(&folio->lru, move_to);
+
+		/*
+		 * IRQs are disabled, cap holdoff at a millisecond or so.
+		 */
+		if (iter++ < 1000)
+			continue;
+		if (cpu_clock(cpu) - then > NSEC_PER_MSEC)
+			break;
+		iter = 0;
 	}

 	/*