Re: [PATCH v2] mm, vmscan: don't turn on cache_trim_mode at high scan priorities

From: Byungchul Park
Date: Thu Feb 22 2024 - 04:21:15 EST


On Thu, Feb 22, 2024 at 04:37:16PM +0800, Huang, Ying wrote:
> Byungchul Park <byungchul@xxxxxx> writes:
>
> > Changes from v1:
> > 1. Add a comment describing why this change is necessary in code
> > and rewrite the commit message with how to reproduce and what
> > the result is using vmstat. (feedbacked by Andrew Morton and
> > Yu Zhao)
> > 2. Change the condition to avoid cache_trim_mode from
> > 'sc->priority != 1' to 'sc->priority > 1' to reflect cases
> > where the priority goes to zero all the way. (feedbacked by
> > Yu Zhao)
> >
> > --->8---
> > From 07e0baab368160e50b6ca35d95745168aa60e217 Mon Sep 17 00:00:00 2001
> > From: Byungchul Park <byungchul@xxxxxx>
> > Date: Thu, 22 Feb 2024 14:50:17 +0900
> > Subject: [PATCH v2] mm, vmscan: don't turn on cache_trim_mode at high scan priorities
> >
> > With cache_trim_mode on, reclaim logic doesn't bother reclaiming anon
> > pages. However, it should be more careful to turn on the mode because
> > it's going to prevent anon pages from being reclaimed even if there are
> > a huge number of anon pages that are cold and should be reclaimed. Even
> > worse, that can lead kswapd_failures to reach MAX_RECLAIM_RETRIES and
> > stopping kswapd until direct reclaim eventually works to resume kswapd.
> > So this is more like a bug fix than a performance improvement.
> >
> > The problematic behavior can be reproduced by:
> >
> > CONFIG_NUMA_BALANCING enabled
> > sysctl_numa_balancing_mode set to NUMA_BALANCING_MEMORY_TIERING
> >
> > numa node0 (8GB local memory, 16 CPUs)
> > numa node1 (8GB slow tier memory, no CPUs)
> >
> > Sequence:
> >
> > 1) echo 3 > /proc/sys/vm/drop_caches
> > 2) To emulate the system with full of cold memory in local DRAM, run
> > the following dummy program and never touch the region:
> >
> > mmap(0, 8 * 1024 * 1024 * 1024, PROT_READ | PROT_WRITE,
> > MAP_ANONYMOUS | MAP_PRIVATE | MAP_POPULATE, -1, 0);
> >
> > 3) Run any memory intensive work e.g. XSBench.
> > 4) Check if numa balancing is working e.i. promotion/demotion.
> > 5) Iterate 1) ~ 4) until kswapd stops.
> >
> > With this, you could eventually see that promotion/demotion are not
> > working because kswapd has stopped due to ->kswapd_failures >=
> > MAX_RECLAIM_RETRIES.
> >
> > Interesting vmstat delta's differences between before and after are like:
> >
> > -nr_inactive_anon 321935
> > -nr_active_anon 1780700
> > -nr_inactive_file 30425
> > -nr_active_file 14961
> > -pgpromote_success 356
> > -pgpromote_candidate 21953245
> > -pgactivate 1844523
> > -pgdeactivate 50634
> > -pgfault 31100294
> > -pgdemote_kswapd 30856
> > -pgscan_kswapd 1861981
> > -pgscan_anon 1822930
> > -pgscan_file 39051
> > -pgsteal_anon 386
> > -pgsteal_file 30470
> > -pageoutrun 30
> > -numa_hint_faults 27418279
> > -numa_pages_migrated 356
> >
> > +nr_inactive_anon 1662306
> > +nr_active_anon 440303
> > +nr_inactive_file 27669
> > +nr_active_file 1654
> > +pgpromote_success 1314102
> > +pgpromote_candidate 1892525
> > +pgactivate 3284457
> > +pgdeactivate 1527504
> > +pgfault 6847775
> > +pgdemote_kswapd 2142047
> > +pgscan_kswapd 7496588
> > +pgscan_anon 7462488
> > +pgscan_file 34100
> > +pgsteal_anon 2115661
> > +pgsteal_file 26386
> > +pageoutrun 378
> > +numa_hint_faults 3220891
> > +numa_pages_migrated 1314102
> >
> > where -: before this patch, +: after this patch
> >
> > Signed-off-by: Byungchul Park <byungchul@xxxxxx>
> > ---
> > mm/vmscan.c | 10 +++++++++-
> > 1 file changed, 9 insertions(+), 1 deletion(-)
> >
> > diff --git a/mm/vmscan.c b/mm/vmscan.c
> > index bba207f41b14..6eda59fce5ee 100644
> > --- a/mm/vmscan.c
> > +++ b/mm/vmscan.c
> > @@ -2266,9 +2266,17 @@ static void prepare_scan_control(pg_data_t *pgdat, struct scan_control *sc)
> > * If we have plenty of inactive file pages that aren't
> > * thrashing, try to reclaim those first before touching
> > * anonymous pages.
> > + *
> > + * However, the condition 'sc->cache_trim_mode == 1' all through
> > + * the scan priorties might lead reclaim failure. If it keeps
> > + * MAX_RECLAIM_RETRIES times, then kswapd would get stopped even
> > + * if there are still plenty anon pages to reclaim, which is not
> > + * desirable. So do not use cache_trim_mode when reclaim is not
> > + * smooth e.i. high scan priority.
> > */
> > file = lruvec_page_state(target_lruvec, NR_INACTIVE_FILE);
> > - if (file >> sc->priority && !(sc->may_deactivate & DEACTIVATE_FILE))
> > + if (sc->priority > 1 && file >> sc->priority &&
> > + !(sc->may_deactivate & DEACTIVATE_FILE))
> > sc->cache_trim_mode = 1;
> > else
> > sc->cache_trim_mode = 0;
>
> In get_scan_count(), there's following code,
>
> /*
> * Do not apply any pressure balancing cleverness when the
> * system is close to OOM, scan both anon and file equally
> * (unless the swappiness setting disagrees with swapping).
> */
> if (!sc->priority && swappiness) {
> scan_balance = SCAN_EQUAL;
> goto out;
> }
>
> So, swappiness is 0 in you system? Please check it. If it's not 0,
> please check why this doesn't help.

Nice information! Then the change should be:

diff --git a/mm/vmscan.c b/mm/vmscan.c
index bba207f41b14..91f9bab86e92 100644
--- a/mm/vmscan.c
+++ b/mm/vmscan.c
@@ -2357,7 +2357,7 @@ static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
* system is close to OOM, scan both anon and file equally
* (unless the swappiness setting disagrees with swapping).
*/
- if (!sc->priority && swappiness) {
+ if (sc->priority <= 1 && swappiness) {
scan_balance = SCAN_EQUAL;
goto out;
}

Worth noting that the priority goes from DEF_PRIORITY to 1 in
balance_pgdat() of kswapd. I will change how to fix to this if this
looks more reasonable.

Byungchul