On Mon 25-11-19 17:11:14, Yafang Shao wrote:
> On Mon, Nov 25, 2019 at 4:14 PM Michal Hocko <mhocko@xxxxxxxxxx> wrote:
> >
> > On Sat 23-11-19 00:56:42, Yafang Shao wrote:
> > > memory.{min, low} keeps protecting pages in a memcg even if there're no
> > > process running in this memcg. That makes nonsense, because if there're
> > > no processes running in this memcg there may be something wrong happens
> > > and these protected pages are not useful now.
> >
> > Well, the min resp. low limits protect charges rather than tasks. And
> > thre are cases when there is no direct relation between the charged
> > memory and a specific task - e.g. long living kernel objects or page
> > cache. So there is nothing really unusual to have charges in an absence
> > of any tasks. I do not really see any reason why the protection
> > shouldn't apply to them. Such a be behavior would be quite inconsistent
> > btw. Just imagine a case when there is _a_ task but that one doesn't
> > really have any charges in the memcg. Why should we respect reclaim
> > protection in that case?
> >
>
> Per my understanding, the memory.{min, low} is used to protect the workload.
> I don't know in which cases we protect the 'content' only.
I didn't say anything about the content. You are right that those limits
are protecting a working set. I do not think it would be hard to imagine
a workload which has a page cache to be the working set and want to
protect it while tasks are only very short lived and so "no tasks" state
wouldn't be all that unusual.
> If one memcg is used to protect the content only, e.g. some file
> caches, then these file caches will be shared by processes in
> different memcgs, that may cause some unexpected issues, for example,
> how to handle the writeback throttle ?
Yes, sharing might lead to surprising behavior but I do not not
understand how this is really related to the proposed patch.
> I don't know how to use memory.{min, low} to protect long living
> kernel objects neither without any processes.
>
> I would appreciate if you could explain my questions in detail.
See the example above.
> > Btw. offlined memcgs already have no min/low watermarks protect.
> >
> > Do you have any specific usecase that suffers from this behavior?
> >
>
> For example, if oom happens, the processes in this memcg may be killed
> by OOM killer, and if memory.oom.group is set, all processes in it
> will be killed. So these protected pages are useless. In that cases,
> we should clear memory.{min, low} protection.
I do agree that the reclaim protection might lead to an unexpected OOM
handling. Especially when charged objects are not directly bound to the
process life time. But this patch is not going to help here in general.
One has to be really careful when configuring the protection. It is a
hard guarantee after all.
--
Michal Hocko
SUSE Labs
