On Thu, Dec 02, 2021 at 08:30:51AM -0800, Shakeel Butt wrote:

Hi Mel,

On Thu, Dec 2, 2021 at 7:07 AM Mel Gorman [off-list ref] wrote:

quoted

Mike Galbraith, Alexey Avramov and Darrick Wong all reported similar
problems due to reclaim throttling for excessive lengths of time.
In Alexey's case, a memory hog that should go OOM quickly stalls for
several minutes before stalling. In Mike and Darrick's cases, a small
memcg environment stalled excessively even though the system had enough
memory overall.

Commit 69392a403f49 ("mm/vmscan: throttle reclaim when no progress is being
made") introduced the problem although commit a19594ca4a8b ("mm/vmscan:
increase the timeout if page reclaim is not making progress") made it
worse. Systems at or near an OOM state that cannot be recovered must
reach OOM quickly and memcg should kill tasks if a memcg is near OOM.

Is there a reason we can't simply revert 69392a403f49 instead of adding
more code/heuristics? Looking more into 69392a403f49, I don't think the
code and commit message are in sync.

For the memcg reclaim, instead of just removing congestion_wait or
replacing it with schedule_timeout in mem_cgroup_force_empty(), why
change the behavior of all memcg reclaim. Also this patch effectively
reverts that behavior of 69392a403f49.

It doesn't fully revert it but I did consider reverting it. The reason
why I preserved it because the intent originally was to throttle somewhat
when progress is not being made to avoid a premature OOM and I wanted to
preserve that charactersistic. Right now, this is the least harmful way
of doing it.

As more memcg, I removed the NOTHROTTLE because the primary reason why a
memcg might fail to make progress is excessive writeback and that should
still throttle. Completely failing to make progress in a memcg is most
likely due to a memcg-OOM.

For direct reclaimers under global pressure, why is page allocator a bad
place for stalling on no progress reclaim? IMHO the callers of the
reclaim should decide what to do if reclaim is not making progress.

Because it's a layering violation and the caller has little direct control
over the reclaim retry logic. The page allocator has no visibility on
why reclaim failed only that it did fail.

-- 
Mel Gorman
SUSE Labs