Re: [PATCH] mm/oom: Add oom_score_adj value to oom Killed process message

[Edward Chron <echron@arista.com>]

On Thu, Aug 22, 2019 at 12:21 AM Michal Hocko <mhocko@kernel.org> wrote:
>
> On Wed 21-08-19 16:12:08, Edward Chron wrote:
> [...]
> > Additionally (which you know, but mentioning for reference) the OOM
> > output used to look like this:
> >
> > Nov 14 15:23:48 oldserver kernel: [337631.991218] Out of memory: Kill
> > process 19961 (python) score 17 or sacrifice child
> > Nov 14 15:23:48 oldserver kernel: [337631.991237] Killed process 31357
> > (sh) total-vm:5400kB, anon-rss:252kB, file-rss:4kB, shmem-rss:0kB
> >
> > It now looks like this with 5.3.0-rc5 (minus the oom_score_adj):
> >
> > Jul 22 10:42:40 newserver kernel:
> > oom-kill:constraint=CONSTRAINT_NONE,nodemask=(null),cpuset=/,mems_allowed=0,global_oom,task_memcg=/user.slice/user-10383.slice/user@10383.service,task=oomprocs,pid=3035,uid=10383
> > Jul 22 10:42:40 newserver kernel: Out of memory: Killed process 3035
> > (oomprocs) total-vm:1056800kB, anon-rss:8kB, file-rss:4kB,
> > shmem-rss:0kB
> > Jul 22 10:42:40 newserver kernel: oom_reaper: reaped process 3035
> > (oomprocs), now anon-rss:0kB, file-rss:0kB, shmem-rss:0kB
> >
> > The old output did explain that a oom_score of 17 must have either
> > tied for highest or was the highest.
> > This did document why OOM selected the process it did, even if ends up
> > killing the related sh process.
> >
> > With the newer format that added constraint message, it does provide
> > uid which can be helpful and
> > the oom_reaper showing that the memory was reclaimed is certainly reassuring.
> >
> > My understanding now is that printing the oom_score is discouraged.
> > This seems unfortunate.  The oom_score_adj can be adjusted
> > appropriately if oom_score is known.
> > So It would be useful to have both.
>
> As already mentioned in our previous discussion I am really not happy
> about exporting oom_score withtout a larger context - aka other tasks
> scores to have something to compare against. Other than that the value
> is an internal implementation detail and it is meaningless without
> knowing the exact algorithm which can change at any times so no
> userspace should really depend on it. All important metrics should be
> displayed by the oom report message already.

The oom_score is no longer displayed any where in the OOM output with 5.3
so there isn't anything to compare against any way with the current OOM
per process output and for the killed process.

I understand the reasoning for this from your discussion.
Thanks for explaining the rational.

>
> --
> Michal Hocko
> SUSE Labs