On Tue, 2020-09-08 at 13:31 -0700, Yang Shi wrote:

On Tue, Sep 8, 2020 at 1:14 PM Julius Hemanth Pitti <jpitti@cisco.com

quoted

wrote:

For non root CG, in try_charge(), we keep trying
to charge until we succeed. On non-preemptive
kernel, when we are OOM, this results in holding
CPU forever.

On SMP systems, this doesn't create a big problem
because oom_reaper get a change to kill victim
and make some free pages. However on a single-core
CPU (or cases where oom_reaper pinned to same CPU
where try_charge is executing), oom_reaper shall
never get scheduled and we stay in try_charge forever.

Steps to repo this on non-smp:
1. mount -t tmpfs none /sys/fs/cgroup
2. mkdir /sys/fs/cgroup/memory
3. mount -t cgroup none /sys/fs/cgroup/memory -o memory
4. mkdir /sys/fs/cgroup/memory/0
5. echo 40M > /sys/fs/cgroup/memory/0/memory.limit_in_bytes
6. echo $$ > /sys/fs/cgroup/memory/0/tasks
7. stress -m 5 --vm-bytes 10M --vm-hang 0

Isn't it the same problem solved by e3336cab2579 ("mm: memcg: fix
memcg reclaim soft lockup")? It has been in Linus's tree.

Yes, indeed.
I just tested with e3336cab2579, and it solved this problem.
Thanks for pointing it out.

quoted

Signed-off-by: Julius Hemanth Pitti <redacted>
Acked-by: Roman Gushchin <redacted>
---

Changes in v2:
 - Added comments.
 - Added "Acked-by: Roman Gushchin [off-list ref]".
---
 mm/memcontrol.c | 9 +++++++++
 1 file changed, 9 insertions(+)

diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index cfa6cbad21d5..4f293bf8c7ed 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c

@@ -2745,6 +2745,15 @@ static int try_charge(struct mem_cgroup

*memcg, gfp_t gfp_mask,
        if (fatal_signal_pending(current))
                goto force;

+       /*
+        * We failed to charge even after retries, give oom_reaper
or
+        * other process a change to make some free pages.
+        *
+        * On non-preemptive, Non-SMP system, this is critical,
else
+        * we keep retrying with no success, forever.
+        */
+       cond_resched();
+
        /*
         * keep retrying as long as the memcg oom killer is able to
make
         * a forward progress or bypass the charge if the oom
killer
--
2.17.1