lkml.org 
[lkml]   [2022]   [Aug]   [4]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] mm: vmscan: fix extreme overreclaim and swap floods
On Tue, Aug 02, 2022 at 12:28:11PM -0400, Johannes Weiner wrote:
> During proactive reclaim, we sometimes observe severe overreclaim,
> with several thousand times more pages reclaimed than requested.
>
> This trace was obtained from shrink_lruvec() during such an instance:
>
> prio:0 anon_cost:1141521 file_cost:7767
> nr_reclaimed:4387406 nr_to_reclaim:1047 (or_factor:4190)
> nr=[7161123 345 578 1111]
>
> While he reclaimer requested 4M, vmscan reclaimed close to 16G, most
> of it by swapping. These requests take over a minute, during which the
> write() to memory.reclaim is unkillably stuck inside the kernel.
>
> Digging into the source, this is caused by the proportional reclaim
> bailout logic. This code tries to resolve a fundamental conflict: to
> reclaim roughly what was requested, while also aging all LRUs fairly
> and in accordance to their size, swappiness, refault rates etc. The
> way it attempts fairness is that once the reclaim goal has been
> reached, it stops scanning the LRUs with the smaller remaining scan
> targets, and adjusts the remainder of the bigger LRUs according to how
> much of the smaller LRUs was scanned. It then finishes scanning that
> remainder regardless of the reclaim goal.
>
> This works fine if priority levels are low and the LRU lists are
> comparable in size. However, in this instance, the cgroup that is
> targeted by proactive reclaim has almost no files left - they've
> already been squeezed out by proactive reclaim earlier - and the
> remaining anon pages are hot. Anon rotations cause the priority level
> to drop to 0, which results in reclaim targeting all of anon (a lot)
> and all of file (almost nothing). By the time reclaim decides to bail,
> it has scanned most or all of the file target, and therefor must also
> scan most or all of the enormous anon target. This target is thousands
> of times larger than the reclaim goal, thus causing the overreclaim.
>
> The bailout code hasn't changed in years, why is this failing now?
> The most likely explanations are two other recent changes in anon
> reclaim:
>
> 1. Before the series starting with commit 5df741963d52 ("mm: fix LRU
> balancing effect of new transparent huge pages"), the VM was
> overall relatively reluctant to swap at all, even if swap was
> configured. This means the LRU balancing code didn't come into play
> as often as it does now, and mostly in high pressure situations
> where pronounced swap activity wouldn't be as surprising.
>
> 2. For historic reasons, shrink_lruvec() loops on the scan targets of
> all LRU lists except the active anon one, meaning it would bail if
> the only remaining pages to scan were active anon - even if there
> were a lot of them.
>
> Before the series starting with commit ccc5dc67340c ("mm/vmscan:
> make active/inactive ratio as 1:1 for anon lru"), most anon pages
> would live on the active LRU; the inactive one would contain only a
> handful of preselected reclaim candidates. After the series, anon
> gets aged similarly to file, and the inactive list is the default
> for new anon pages as well, making it often the much bigger list.
>
> As a result, the VM is now more likely to actually finish large
> anon targets than before.
>
> Change the code such that only one SWAP_CLUSTER_MAX-sized nudge toward
> the larger LRU lists is made before bailing out on a met reclaim goal.
>
> This fixes the extreme overreclaim problem.
>
> Fairness is more subtle and harder to evaluate. No obvious misbehavior
> was observed on the test workload, in any case. Conceptually, fairness
> should primarily be a cumulative effect from regular, lower priority
> scans. Once the VM is in trouble and needs to escalate scan targets to
> make forward progress, fairness needs to take a backseat. This is also
> acknowledged by the myriad exceptions in get_scan_count(). This patch
> makes fairness decrease gradually, as it keeps fairness work static
> over increasing priority levels with growing scan targets. This should
> make more sense - although we may have to re-visit the exact values.
>
> Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>

Acked-by: Mel Gorman <mgorman@techsingularity.net>

--
Mel Gorman
SUSE Labs

\
 
 \ /
  Last update: 2022-08-04 11:41    [W:0.262 / U:0.236 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site