Byungchul Park <byungchul@xxxxxx> writes:
> On Thu, Feb 22, 2024 at 06:49:00PM +0900, Byungchul Park wrote:
>> On Thu, Feb 22, 2024 at 06:20:42PM +0900, Byungchul Park wrote:
>> > 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;
>> > }
>>
>> Or:
>>
>> diff --git a/mm/vmscan.c b/mm/vmscan.c
>> index bba207f41b14..c54371a398b1 100644
>> --- a/mm/vmscan.c
>> +++ b/mm/vmscan.c
>> @@ -6896,7 +6896,7 @@ static int balance_pgdat(pg_data_t *pgdat, int order, int highest_zoneidx)
>>
>> if (raise_priority || !nr_reclaimed)
>> sc.priority--;
>> - } while (sc.priority >= 1);
>> + } while (sc.priority >= 0);
>>
>> if (!sc.nr_reclaimed)
>> pgdat->kswapd_failures++;
>
> +cc Mel Gorman
>
> I just found this was intended. See commit 9aa41348a8d11 ("mm: vmscan:
> do not allow kswapd to scan at maximum priority"). Mel Gorman didn't want
> to make kswapd too much aggressive. However, does it make sense to stop
> kswapd even if there are plenty cold anon pages to reclaim and make the
> system wait for direct reclaim?
Maybe we can play with cache_trim_mode, for example, if sc.nr_reclaimed
== 0 and sc.cache_trim_mode == true, force disabling cache_trim_mode in
the next round?
--
Best Regards,
Huang, Ying
> Thoughts?
>
> Byungchul
>
>> ---
>>
>> Byungchul
>>
>> > 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
