linux-kernel - Re: [PATCH] mm/oom_kill: break evaluation when a task has been selected

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Date:   Mon, 6 Jun 2022 10:33:29 +0200
From:   Michal Hocko <mhocko@...e.com>
To:     Zackary Liu <zackary.liu.pro@...il.com>
Cc:     "akpm@...ux-foundation.org" <akpm@...ux-foundation.org>,
        "linux-mm@...ck.org" <linux-mm@...ck.org>,
        "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH] mm/oom_kill: break evaluation when a task has been
 selected

On Sat 04-06-22 18:35:19, Zackary Liu wrote:
> 
> On Jun 1 2022, at 3:45 pm, Michal Hocko <mhocko@...e.com> wrote:
> 
> > On Sat 14-05-22 15:52:28, Zhaoyu Liu wrote:
> >> oom points no longer need to be calculated if a task is oom_task_origin(),
> >> so return 1 to stop the oom_evaluate_task().
> > 
> > This doesn't really explain why this is really desired. Is this a fix,
> > optimization?
> > 
> > Please also note that this change has some side effects. For one, the
> > task marked as oom origin will get killed even if there is still a
> > pending oom victim which hasn't been fully dismantled. Is this
> > intentional?
> 
> Thank you very much for reminding.
> 
> From my point of view, the victim was marked in the last oom, and now it
> has entered the oom again, which means that the system still has no
> deprecated memory available.

This is not an unusual situation. OOM victims can take some time to die
and release their memory. The oom_reaper is there to fast forward that
process and guarantee a forward progress. But this can still take some
time. Our general policy is to back off when there is an alive oom
victim encountered. Have a look at the tsk_is_oom_victim test in
oom_evaluate_task. For that heuristic to be effective the whole task
list (wether the global one or memcg) has to be evaluated.

> In order to ensure that the system can
> return to normal as soon as possible, killing the origin task
> immediately should be A good choice, and the role of this patch is to
> end oom_evaluate_task and return true as soon as the origin task is found.

Could you be more specific how does this patch guarantees a forward
progress? What is the actual usecase that benefits from this change?

These are all important information for future reference. Please note I
am not saying the patch is wrong. I just still do not see why it is
useful.

> Maybe this patch is not the optimal solution, it has a trade-off.

If there are trade-offs, please document them in the changelog.

The way I see it is that oom_task_origin heuristic has been introduced
to help killing swapoff operation because the swapped out memory doesn't
fit into memory. This is a very reasonable thing to do in general but it
also represents an early failure visible to the userspace. If there is a
pre-existing oom victim then I would argue that we should try to avoid
the failure.

-- 
Michal Hocko
SUSE Labs