Byungchul Park <byungchul@xxxxxx> writes:
> On Thu, Feb 29, 2024 at 02:21:15PM +0800, Huang, Ying wrote:
>> Johannes Weiner <hannes@xxxxxxxxxxx> writes:
>>
>> > On Mon, Feb 26, 2024 at 02:06:30PM +0100, Michal Hocko wrote:
>> >> [CC Mel, Vlastimil and Johannes for awareness]
>> >>
>> >> On Fri 23-02-24 14:44:07, Byungchul Park wrote:
>> >> > Changes from v2:
>> >> > 1. Change the condition to stop cache_trim_mode.
>> >> >
>> >> > From - Stop it if it's at high scan priorities, 0 or 1.
>> >> > To - Stop it if it's at high scan priorities, 0 or 1, and
>> >> > the mode didn't work in the previous turn.
>> >> >
>> >> > (feedbacked by Huang Ying)
>> >> >
>> >> > 2. Change the test result in the commit message after testing
>> >> > with the new logic.
>> >> >
>> >> > 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 05846e34bf02ac9b3e254324dc2d7afd97a025d9 Mon Sep 17 00:00:00 2001
>> >> > From: Byungchul Park <byungchul@xxxxxx>
>> >> > Date: Fri, 23 Feb 2024 13:47:16 +0900
>> >> > Subject: [PATCH v3] mm, vmscan: do not turn on cache_trim_mode if it doesn't work
>> >> >
>> >> > 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 leads kswapd_failures to reach MAX_RECLAIM_RETRIES and
>> >> > stopping kswapd from functioning until direct reclaim eventually works
>> >> > to resume kswapd.
>> >> >
>> >> > So do not turn on cache_trim_mode if the mode doesn't work, especially
>> >> > while the sytem is struggling against reclaim.
>> >> >
>> >> > 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 numa balancing stops.
>> >> >
>> >> > With this, you could 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:
>> >> >
>> >> > +-----------------------+-------------------------------+
>> >> > | interesting vmstat | before | after |
>> >> > +-----------------------+-------------------------------+
>> >> > | nr_inactive_anon | 321935 | 1636737 |
>> >> > | nr_active_anon | 1780700 | 465913 |
>> >> > | nr_inactive_file | 30425 | 35711 |
>> >> > | nr_active_file | 14961 | 8698 |
>> >> > | pgpromote_success | 356 | 1267785 |
>> >> > | pgpromote_candidate | 21953245 | 1745631 |
>> >> > | pgactivate | 1844523 | 3309867 |
>> >> > | pgdeactivate | 50634 | 1545041 |
>> >> > | pgfault | 31100294 | 6411036 |
>> >> > | pgdemote_kswapd | 30856 | 2267467 |
>> >> > | pgscan_kswapd | 1861981 | 7729231 |
>> >> > | pgscan_anon | 1822930 | 7667544 |
>> >> > | pgscan_file | 39051 | 61687 |
>> >> > | pgsteal_anon | 386 | 2227217 |
>> >> > | pgsteal_file | 30470 | 40250 |
>> >> > | pageoutrun | 30 | 457 |
>> >> > | numa_hint_faults | 27418279 | 2752289 |
>> >> > | numa_pages_migrated | 356 | 1267785 |
>> >> > +-----------------------+-------------------------------+
>> >> >
>> >> > Signed-off-by: Byungchul Park <byungchul@xxxxxx>
>> >> > ---
>> >> > mm/vmscan.c | 24 +++++++++++++++++++-----
>> >> > 1 file changed, 19 insertions(+), 5 deletions(-)
>> >> >
>> >> > diff --git a/mm/vmscan.c b/mm/vmscan.c
>> >> > index bba207f41b14..f7312d831fed 100644
>> >> > --- a/mm/vmscan.c
>> >> > +++ b/mm/vmscan.c
>> >> > @@ -127,6 +127,9 @@ struct scan_control {
>> >> > /* One of the zones is ready for compaction */
>> >> > unsigned int compaction_ready:1;
>> >> >
>> >> > + /* If the last try was reclaimable */
>> >> > + unsigned int reclaimable:1;
>> >> > +
>> >> > /* There is easily reclaimable cold cache in the current node */
>> >> > unsigned int cache_trim_mode:1;
>> >> >
>> >> > @@ -2266,9 +2269,14 @@ 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.
>> >> > + *
>> >> > + * It doesn't make sense to keep cache_trim_mode on if the mode
>> >> > + * is not working while struggling against reclaim. So do not
>> >> > + * turn it on if so. Note the highest priority of kswapd is 1.
>> >> > */
>> >> > file = lruvec_page_state(target_lruvec, NR_INACTIVE_FILE);
>> >> > - if (file >> sc->priority && !(sc->may_deactivate & DEACTIVATE_FILE))
>> >> > + if (file >> sc->priority && !(sc->may_deactivate & DEACTIVATE_FILE) &&
>> >> > + !(sc->cache_trim_mode && !sc->reclaimable && sc->priority <= 1))
>> >> > sc->cache_trim_mode = 1;
>> >> > else
>> >> > sc->cache_trim_mode = 0;
>> >
>> > The overall goal makes sense to me.
>> >
>> > file >> priority is just a heuristic that there are enough potential
>> > candidate pages, not a guarantee that any forward progress will
>> > happen. So it makes sense to retry without before failing.
>> >
>> > The way you wrote this conditional kind of hurts my head,
>> > though. Please don't write negations of complex terms like this.
>> >
>> > It expands to this:
>> >
>> > !sc->cache_trim_mode || sc->reclaimable || sc->priority > 1
>> >
>> > which I'm not sure makes sense. Surely it should be something like
>> >
>> > !sc->cache_trim_mode && sc->reclaimable && sc->priority > 1
>> >
>> > instead?
>> >
>> > Also
>> >
>> > if (!sc->cache_trim_mode)
>> > sc->cache_trim_mode = 1
>> > else
>> > sc->cache_trim_mode = 0
>> >
>> > will toggle on every loop. So if direct reclaim runs through a
>> > zonelist, it'll cache trim every other numa node...?
>> >
>> >> > @@ -5862,7 +5870,6 @@ static void shrink_node(pg_data_t *pgdat, struct scan_control *sc)
>> >> > {
>> >> > unsigned long nr_reclaimed, nr_scanned, nr_node_reclaimed;
>> >> > struct lruvec *target_lruvec;
>> >> > - bool reclaimable = false;
>> >> >
>> >> > if (lru_gen_enabled() && root_reclaim(sc)) {
>> >> > lru_gen_shrink_node(pgdat, sc);
>> >> > @@ -5877,6 +5884,14 @@ static void shrink_node(pg_data_t *pgdat, struct scan_control *sc)
>> >> > nr_reclaimed = sc->nr_reclaimed;
>> >> > nr_scanned = sc->nr_scanned;
>> >> >
>> >> > + /*
>> >> > + * Reset to the default values at the start.
>> >> > + */
>> >> > + if (sc->priority == DEF_PRIORITY) {
>> >> > + sc->reclaimable = 1;
>> >> > + sc->cache_trim_mode = 0;
>> >> > + }
>> >> > +
>> >> > prepare_scan_control(pgdat, sc);
>> >> >
>> >> > shrink_node_memcgs(pgdat, sc);
>> >> > @@ -5890,8 +5905,7 @@ static void shrink_node(pg_data_t *pgdat, struct scan_control *sc)
>> >> > vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
>> >> > sc->nr_scanned - nr_scanned, nr_node_reclaimed);
>> >> >
>> >> > - if (nr_node_reclaimed)
>> >> > - reclaimable = true;
>> >> > + sc->reclaimable = !!nr_node_reclaimed;
>> >
>> > The scope of this doesn't quite make sense. If direct reclaim scans
>> > multiple nodes, reclaim failure on the first node would disable cache
>> > trim mode on the second node, which is totally unrelated.
>> >
>> > I think it needs separate paths for direct reclaim and kswapd. For
>> > direct reclaim, the retry should be before these similar retry catches
>> > in do_try_to_free_pages(), after all zones have been considered:
>> >
>> > /*
>> > * We make inactive:active ratio decisions based on the node's
>> > * composition of memory, but a restrictive reclaim_idx or a
>> > * memory.low cgroup setting can exempt large amounts of
>> > * memory from reclaim. Neither of which are very common, so
>> > * instead of doing costly eligibility calculations of the
>> > * entire cgroup subtree up front, we assume the estimates are
>> > * good, and retry with forcible deactivation if that fails.
>> > */
>> > if (sc->skipped_deactivate) {
>> > sc->priority = initial_priority;
>> > sc->force_deactivate = 1;
>> > sc->skipped_deactivate = 0;
>> > goto retry;
>> > }
>> >
>> > /* Untapped cgroup reserves? Don't OOM, retry. */
>> > if (sc->memcg_low_skipped) {
>> > sc->priority = initial_priority;
>> > sc->force_deactivate = 0;
>> > sc->memcg_low_reclaim = 1;
>> > sc->memcg_low_skipped = 0;
>> > goto retry;
>> > }
>>
>> In get_scan_count(), we have
>>
>> if (!sc->priority && swappiness) {
>> scan_balance = SCAN_EQUAL;
>> goto out;
>> }
>
> Even though this can mitigate the issue for direct reclaim, it's still
> suffering from the problem while direct reclaim goes from DEF_PRIORITY
> to 1. What we need is not a mitigation but making things right.
Whether does it cause issue for you? IMHO, it's the first step to prove
this is a real problem.
--
Best Regards,
Huang, Ying
> Thoughts?
>
> Byungchul
>
>> So, we don't really need the heuristics in direct reclaim path. So, one
>> choice is to constrain this in kswapd reclaim only.
>>
>> --
>> Best Regards,
>> Huang, Ying
>>
>>
>> > and for kswapd it looks like it should be in balance_pgdat(), after
>> > the priority loop, before increasing kswapd_failures.
>> >
>> > Instead of sc->reclaimable, which is very broad, it would be better to
>> > call it sc->may_cache_trim_mode. This is in line with a bunch of other
>> > such mechanisms in scan_control.
