[v2,5/9] mm/workingset: use the node counter if memcg is the root memcg

Commit Message

Joonsoo Kim Feb. 20, 2020, 5:11 a.m. UTC

From: Joonsoo Kim <[e-mail]>

In the following patch, workingset detection is implemented for the
swap cache. Swap cache's node is usually allocated by kswapd and it
isn't charged by kmemcg since it is from the kernel thread. So the swap
cache's shadow node is managed by the node list of the list_lru rather
than the memcg specific one.

If counting the shadow node on the root memcg happens to reclaim the slab
object, the shadow node count returns the number of the shadow node on
the node list of the list_lru since root memcg has the kmem_cache_id, -1.

However, the size of pages on the LRU is calculated by using the specific
memcg, so mismatch happens. This causes the number of shadow node not to
be increased to the enough size and, therefore, workingset detection
cannot work correctly. This patch fixes this bug by checking if the memcg
is the root memcg or not. If it is the root memcg, instead of using
the memcg-specific LRU, the system-wide LRU is used to calculate proper
size of the shadow node so that the number of the shadow node can grow
as expected.

Signed-off-by: Joonsoo Kim <[e-mail]>
---
 mm/workingset.c | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)

Patch

diff --git a/mm/workingset.c b/mm/workingset.c
index 5fb8f85..a9f474a 100644
--- a/mm/workingset.c
+++ b/mm/workingset.c
@@ -468,7 +468,13 @@  static unsigned long count_shadow_nodes(struct shrinker *shrinker,
 	 * PAGE_SIZE / xa_nodes / node_entries * 8 / PAGE_SIZE
 	 */
 #ifdef CONFIG_MEMCG
-	if (sc->memcg) {
+	/*
+	 * Kernel allocation on root memcg isn't regarded as allocation of
+	 * specific memcg. So, if sc->memcg is the root memcg, we need to
+	 * use the count for the node rather than one for the specific
+	 * memcg.
+	 */
+	if (sc->memcg && !mem_cgroup_is_root(sc->memcg)) {
 		struct lruvec *lruvec;
 		int i;