On Thu, May 17, 2018 at 12:23:30PM +0200, Michal Hocko wrote:
> On Thu 17-05-18 17:44:43, 禹舟键 wrote:
> > Hi Michal
> > I think the current OOM report is imcomplete. I can get the task which
> > invoked the oom-killer and the task which has been killed by the
> > oom-killer, and memory info when the oom happened. But I cannot infer the
> > certain memcg to which the task killed by oom-killer belongs, because that
> > task has been killed, and the dump_task will print all of the tasks in the
> > system.
> I can see how the origin memcg might be useful, but ...
> > mem_cgroup_print_oom_info will print five lines of content including
> > memcg's name , usage, limit. I don't think five lines of content will cause
> > a big problem. Or it at least prints the memcg's name.
I want only add here that if system-wide OOM is a rare event, you can look
at per-cgroup oom counters to find the cgroup, which contained the killed
task. Not super handy, but might work for debug purposes.
> this is not 5 lines at all. We dump memcg stats for the whole oom memcg
> subtree. For your patch it would be the whole subtree of the memcg of
> the oom victim. With cgroup v1 this can be quite deep as tasks can
> belong to inter-nodes as well. Would be
> pr_info("Task in ");
> pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
> pr_cont(" killed as a result of limit of ");
> part of that output sufficient for your usecase? You will not get memory
> consumption of the group but is that really so relevant when we are
> killing individual tasks? Please note that there are proposals to make
> the global oom killer memcg aware and select by the memcg size rather
> than pick on random tasks
> (http://lkml.kernel.org/r/20171130152824.1591-1-g...@fb.com). Maybe that
> will be more interesting for your container usecase.
Speaking about memcg OOM reports more broadly, IMO
rather than spam with memcg-local OOM dumps to dmesg,
it's better to add a new interface to read memcg-specific OOM reports.
The current dmesg OOM report contains a lot of low-level stuff,
which is handy for debugging system-wide OOM issues,
and memcg-aware stuff too; that makes it bulky.
Anyway, Michal's 1-line proposal looks quite acceptable to me.