When memory.reclaim was introduced, it became the first case where cgroup_reclaim() is true for the root cgroup. Johannes concluded [1] that for most cases this is okay, except for one case. Historically, kswapd would throttle reclaim on a node if a lot of pages marked for reclaim are under writeback (aka the node is congested). This occurred by setting LRUVEC_CONGESTED bit in lruvec->flags. The bit would be cleared when the node is balanced. Similarly, cgroup reclaim would set the same bit when an lruvec is congested, and clear it on the way out of reclaim (to throttle local reclaimers). Before the introduction of memory.reclaim, the root memcg was the only target of kswapd reclaim, and non-root memcgs were the only targets of cgroup reclaim, so they would never interfere. Using the same bit for both was fine. After memory.reclaim, it is possible for cgroup reclaim on the root cgroup to clear the bit set by kswapd. This would result in reclaim on the node to be unthrottled before the node is balanced. Fix this by introducing separate bits for cgroup-level and node-level congestion. kswapd can unthrottle an lruvec that is marked as congested by cgroup reclaim (as the entire node should no longer be congested), but not vice versa (to prevent premature unthrottling before the entire node is balanced). [1]https://lore.kernel.org/lkml/20230405200150.GA35884@xxxxxxxxxxx/ Reported-by: Johannes Weiner <hannes@xxxxxxxxxxx> Closes: https://lore.kernel.org/lkml/20230405200150.GA35884@xxxxxxxxxxx/ Signed-off-by: Yosry Ahmed <yosryahmed@xxxxxxxxxx> --- include/linux/mmzone.h | 18 +++++++++++++++--- mm/vmscan.c | 19 ++++++++++++------- 2 files changed, 27 insertions(+), 10 deletions(-) diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h index 3e822335f214..d863698a84e0 100644 --- a/include/linux/mmzone.h +++ b/include/linux/mmzone.h @@ -293,9 +293,21 @@ static inline bool is_active_lru(enum lru_list lru) #define ANON_AND_FILE 2 enum lruvec_flags { - LRUVEC_CONGESTED, /* lruvec has many dirty pages - * backed by a congested BDI - */ + /* + * An lruvec has many dirty pages backed by a congested BDI: + * 1. LRUVEC_CGROUP_CONGESTED is set by cgroup-level reclaim. + * It can be cleared by cgroup reclaim or kswapd. + * 2. LRUVEC_NODE_CONGESTED is set by kswapd node-level reclaim. + * It can only be cleared by kswapd. + * + * Essentially, kswapd can unthrottle an lruvec throttled by cgroup + * reclaim, but not vice versa. This only applies to the root cgroup. + * The goal is to prevent cgroup reclaim on the root cgroup (e.g. + * memory.reclaim) to unthrottle an unbalanced node (that was throttled + * by kswapd). + */ + LRUVEC_CGROUP_CONGESTED, + LRUVEC_NODE_CONGESTED, }; #endif /* !__GENERATING_BOUNDS_H */ diff --git a/mm/vmscan.c b/mm/vmscan.c index 0dbbf718c53e..c22e4e7368da 100644 --- a/mm/vmscan.c +++ b/mm/vmscan.c @@ -6592,10 +6592,13 @@ static void shrink_node(pg_data_t *pgdat, struct scan_control *sc) * Legacy memcg will stall in page writeback so avoid forcibly * stalling in reclaim_throttle(). */ - if ((current_is_kswapd() || - (cgroup_reclaim(sc) && writeback_throttling_sane(sc))) && - sc->nr.dirty && sc->nr.dirty == sc->nr.congested) - set_bit(LRUVEC_CONGESTED, &target_lruvec->flags); + if (sc->nr.dirty && sc->nr.dirty == sc->nr.congested) { + if (cgroup_reclaim(sc) && writeback_throttling_sane(sc)) + set_bit(LRUVEC_CGROUP_CONGESTED, &target_lruvec->flags); + + if (current_is_kswapd()) + set_bit(LRUVEC_NODE_CONGESTED, &target_lruvec->flags); + } /* * Stall direct reclaim for IO completions if the lruvec is @@ -6605,7 +6608,8 @@ static void shrink_node(pg_data_t *pgdat, struct scan_control *sc) */ if (!current_is_kswapd() && current_may_throttle() && !sc->hibernation_mode && - test_bit(LRUVEC_CONGESTED, &target_lruvec->flags)) + (test_bit(LRUVEC_CGROUP_CONGESTED, &target_lruvec->flags) || + test_bit(LRUVEC_NODE_CONGESTED, &target_lruvec->flags))) reclaim_throttle(pgdat, VMSCAN_THROTTLE_CONGESTED); if (should_continue_reclaim(pgdat, nr_node_reclaimed, sc)) @@ -6862,7 +6866,7 @@ static unsigned long do_try_to_free_pages(struct zonelist *zonelist, lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, zone->zone_pgdat); - clear_bit(LRUVEC_CONGESTED, &lruvec->flags); + clear_bit(LRUVEC_CGROUP_CONGESTED, &lruvec->flags); } } @@ -7251,7 +7255,8 @@ static void clear_pgdat_congested(pg_data_t *pgdat) { struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat); - clear_bit(LRUVEC_CONGESTED, &lruvec->flags); + clear_bit(LRUVEC_NODE_CONGESTED, &lruvec->flags); + clear_bit(LRUVEC_CGROUP_CONGESTED, &lruvec->flags); clear_bit(PGDAT_DIRTY, &pgdat->flags); clear_bit(PGDAT_WRITEBACK, &pgdat->flags); } -- 2.41.0.162.gfafddb0af9-goog
