mm: disable top-tier fallback to reclaim on proactive reclaim
authorMina Almasry <almasrymina@google.com>
Thu, 1 Dec 2022 23:33:17 +0000 (15:33 -0800)
committerAndrew Morton <akpm@linux-foundation.org>
Mon, 12 Dec 2022 02:12:19 +0000 (18:12 -0800)
Reclaiming directly from top tier nodes breaks the aging pipeline of
memory tiers.  If we have a RAM -> CXL -> storage hierarchy, we should
demote from RAM to CXL and from CXL to storage.  If we reclaim a page from
RAM, it means we 'demote' it directly from RAM to storage, bypassing
potentially a huge amount of pages colder than it in CXL.

However disabling reclaim from top tier nodes entirely would cause ooms in
edge scenarios where lower tier memory is unreclaimable for whatever
reason, e.g.  memory being mlocked() or too hot to reclaim.  In these
cases we would rather the job run with a performance regression rather
than it oom altogether.

However, we can disable reclaim from top tier nodes for proactive reclaim.
That reclaim is not real memory pressure, and we don't have any cause to
be breaking the aging pipeline.

[akpm@linux-foundation.org: restore comment layout, per Ying Huang]
Link: https://lkml.kernel.org/r/20221201233317.1394958-1-almasrymina@google.com
Signed-off-by: Mina Almasry <almasrymina@google.com>
Reviewed-by: "Huang, Ying" <ying.huang@intel.com>
Reviewed-by: Yang Shi <shy828301@gmail.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Tim Chen <tim.c.chen@linux.intel.com>
Cc: Wei Xu <weixugc@google.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
mm/vmscan.c

index dcd476a..1a59171 100644 (file)
@@ -2088,10 +2088,29 @@ keep:
        nr_reclaimed += demote_folio_list(&demote_folios, pgdat);
        /* Folios that could not be demoted are still in @demote_folios */
        if (!list_empty(&demote_folios)) {
-               /* Folios which weren't demoted go back on @folio_list for retry: */
+               /* Folios which weren't demoted go back on @folio_list */
                list_splice_init(&demote_folios, folio_list);
-               do_demote_pass = false;
-               goto retry;
+
+               /*
+                * goto retry to reclaim the undemoted folios in folio_list if
+                * desired.
+                *
+                * Reclaiming directly from top tier nodes is not often desired
+                * due to it breaking the LRU ordering: in general memory
+                * should be reclaimed from lower tier nodes and demoted from
+                * top tier nodes.
+                *
+                * However, disabling reclaim from top tier nodes entirely
+                * would cause ooms in edge scenarios where lower tier memory
+                * is unreclaimable for whatever reason, eg memory being
+                * mlocked or too hot to reclaim. We can disable reclaim
+                * from top tier nodes in proactive reclaim though as that is
+                * not real memory pressure.
+                */
+               if (!sc->proactive) {
+                       do_demote_pass = false;
+                       goto retry;
+               }
        }
 
        pgactivate = stat->nr_activate[0] + stat->nr_activate[1];