[PATCH] memcg: flush nr_charged_bytes when the obj stock trylock fails

[Shakeel Butt <shakeel.butt@linux.dev>]

When interrupt code, such as an RCU callback in softirq, runs on top of
a task that holds the obj stock lock, its trylock fails and the bytes go
to objcg->nr_charged_bytes. Nothing empties that counter while every CPU
using the objcg keeps it in its stock, so it can grow without limit.

Users see memory.current and the "kernel" field of memory.stat keep
growing while the cgroup's real usage does not. With memory.high or
memory.max set, the cgroup is reclaimed, throttled or even OOM-killed
for memory it does not use. The counter is 32-bit, so after 4 GiB it
wraps, and those 4 GiB stay charged to the cgroup and its ancestors
until reboot, even after the cgroup is removed.

Fix it by uncharging the whole pages in nr_charged_bytes whenever the
trylock fails.

Fixes: 200577f69f29 ("memcg: objcg stock trylock without irq disabling")
Reported-by: Chris Mason <mason@kernel.org>
Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
Cc: stable@vger.kernel.org
---
 mm/memcontrol.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index f5f567f7d4a9..091401cecf39 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -3553,6 +3553,8 @@ static void __refill_obj_stock(struct obj_cgroup *objcg,
 	int i, slot = -1, empty_slot = -1;
 
 	if (!stock) {
+		nr_bytes += atomic_read(&objcg->nr_charged_bytes)
+			? atomic_xchg(&objcg->nr_charged_bytes, 0) : 0;
 		nr_pages = nr_bytes >> PAGE_SHIFT;
 		nr_bytes = nr_bytes & (PAGE_SIZE - 1);
 		atomic_add(nr_bytes, &objcg->nr_charged_bytes);
-- 
2.53.0-Meta