On Mon, 2021-11-29 at 15:01 +0000, Mel Gorman wrote:
> On Sat, Nov 27, 2021 at 01:12:46AM +0900, Alexey Avramov wrote:
> > > After the patch, the test gets killed after roughly 15 seconds which is
> > > the same length of time taken in 5.15.
> >
> > In my tests, the 5.15 still performs much better.
> >
> > New question: is timeout=1 has sense? Will it save CPU?
>
> Ok, the following on top of 5.16-rc1 survived 8 minutes of watching youtube
> on a laptop while "tail /dev/zero" was running within the background. While
> there were some very short glitches, they were no worse than 5.15. I've
> not reproduced your exact test case yet or the memcg ones yet but sending
> now in case I don't complete them before the end of the day.
>
> diff --git a/mm/vmscan.c b/mm/vmscan.c
> index fb9584641ac7..1af12072f40e 100644
> --- a/mm/vmscan.c
> +++ b/mm/vmscan.c
> @@ -1021,6 +1021,39 @@ static void handle_write_error(struct address_space *mapping,
> unlock_page(page);
> }
>
> +bool skip_throttle_noprogress(pg_data_t *pgdat)
> +{
> + int reclaimable = 0, write_pending = 0;
> + int i;
> +
> + /*
> + * If kswapd is disabled, reschedule if necessary but do not
> + * throttle as the system is likely near OOM.
> + */
> + if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES)
> + return true;
> +
> + /*
> + * If there are a lot of dirty/writeback pages then do not
> + * throttle as throttling will occur when the pages cycle
> + * towards the end of the LRU if still under writeback.
> + */
> + for (i = 0; i < MAX_NR_ZONES; i++) {
> + struct zone *zone = pgdat->node_zones + i;
> +
> + if (!populated_zone(zone))
> + continue;
> +
> + reclaimable += zone_reclaimable_pages(zone);
> + write_pending += zone_page_state_snapshot(zone,
> + NR_ZONE_WRITE_PENDING);
> + }
> + if (2 * write_pending <= reclaimable)
That is always true here...
-Mike
