Re: [PATCH] fs, mm: account filp and names caches to kmemcg

[Michal Hocko]

On Tue 10-10-17 15:21:53, Shakeel Butt wrote:
[...]
> On Tue, Oct 10, 2017 at 2:10 AM, Michal Hocko <mhocko@xxxxxxxxxx> wrote:
> > On Mon 09-10-17 20:17:54, Michal Hocko wrote:
> >> the primary concern for this patch was whether we really need/want to
> >> charge short therm objects which do not outlive a single syscall.
> >
> > Let me expand on this some more. What is the benefit of kmem accounting
> > of such an object? It cannot stop any runaway as a syscall lifetime
> > allocations are bound to number of processes which we kind of contain by
> > other means.
> 
> We can contain by limited the number of processes or thread but for us
> applications having thousands of threads is very common. So, limiting
> the number of threads/processes will not work.

Well, the number of tasks is already contained in a way because we do
account each task (kernel) stack AFAIR.

> > If we do account then we put a memory pressure due to
> > something that cannot be reclaimed by no means. Even the memcg OOM
> > killer would simply kick a single path while there might be others
> > to consume the same type of memory.
> >
> > So what is the actual point in accounting these? Does it help to contain
> > any workload better? What kind of workload?
> >
> 
> I think the benefits will be isolation and more accurate billing. As I
> have said before we have observed 100s of MiBs in names_cache on many
> machines and cumulative amount is not something we can ignore as just
> memory overhead.

I do agree with Al arguing this is rather dubious and it can add an
overhead without a good reason.

> > Or am I completely wrong and name objects can outlive a syscall
> > considerably?
> >
> 
> No, I didn't find any instance of the name objects outliving the syscall.
> 
> Anyways, we can discuss more on names_cache, do you have any objection
> regarding charging filp?

I think filep makes more sense. But let's drop the names for now. I am
not really convinced this is a good idea.

-- 
Michal Hocko
SUSE Labs