Re: [PATCH] mm: skip current when memcg reclaim

[Zhaoyang Huang <huangzhaoyang@gmail.com>]

On Wed, Oct 20, 2021 at 4:55 PM Michal Hocko <mhocko@suse.com> wrote:
>
> On Wed 20-10-21 15:33:39, Zhaoyang Huang wrote:
> [...]
> > Do you mean that direct reclaim should succeed for the first round
> > reclaim within which memcg get protected by memory.low and would NOT
> > retry by setting memcg_low_reclaim to true?
>
> Yes, this is the semantic of low limit protection in the upstream
> kernel. Have a look at do_try_to_free_pages and how it sets
> memcg_low_reclaim only if there were no pages reclaimed.
>
> > It is not true in android
> > like system, where reclaim always failed and introduce lmk and even
> > OOM.
>
> I am not familiar with android specific changes to the upstream reclaim
> logic. You should be investigating why the reclaim couldn't make a
> forward progress (aka reclaim pages) from non-protected memcgs. There
> are tracepoints you can use (generally vmscan prefix).
Ok, I am aware of why you get confused now. I think you are analysing
cgroup's behaviour according to a pre-defined workload and memory
pattern, which should work according to the design, such as processes
within root should provide memory before protected memcg get
reclaimed. You can refer [1] as the hierarchy, where effective
userspace workloads locate in protect groups and have rest of
processes be non-grouped. In fact, non-grouped ones can not provide
enough memory as they are kernel threads and the processes with few
pages on LRU(control logic inside). The practical scenario is groupA
launched a high-order kmalloc and introduce reclaiming(kswapd and
direct reclaim). As I said, non-grouped ones can not provide enough
contiguous memory blocks which let direct reclaim quickly fail for the
first round reclaiming. What I am trying to do is that let kswapd try
more for the target. It is also fair if groupA,B,C are trapping in
slow path concurrently.

[1]
root
|                                                       |
|              |
non-grouped processes             groupA    groupB  groupC
>
> --
> Michal Hocko
> SUSE Labs