Re: [PATCH memcg 0/1] false global OOM triggered by memcg-limited task

[Michal Hocko <mhocko@suse.com>]

On Thu 21-10-21 11:03:43, Vasily Averin wrote:
> On 18.10.2021 12:04, Michal Hocko wrote:
> > On Mon 18-10-21 11:13:52, Vasily Averin wrote:
> > [...]
> >> How could this happen?
> >>
> >> User-space task inside the memcg-limited container generated a page fault,
> >> its handler do_user_addr_fault() called handle_mm_fault which could not
> >> allocate the page due to exceeding the memcg limit and returned VM_FAULT_OOM.
> >> Then do_user_addr_fault() called pagefault_out_of_memory() which executed
> >> out_of_memory() without set of memcg.
> 
> > I will be honest that I am not really happy about pagefault_out_of_memory.
> > I have tried to remove it in the past. Without much success back then,
> > unfortunately[1]. 
> > 
> > [1] I do not have msg-id so I cannot provide a lore link but google
> > pointed me to https://www.mail-archive.com/linux-kernel@vger.kernel.org/msg1400402.html
> 
> I re-read this discussion and in general I support your position.
> As far as I understand your opponents cannot explain why "random kill" is mandatory here,
> they are just afraid that it might be useful here and do not want to remove it completely.

That aligns with my recollection.

> Ok, let's allow him to do it. Moreover I'm ready to keep it as default behavior.
> 
> However I would like to have some choice in this point.
> 
> In general we can:
> - continue to use "random kill" and rely on the wisdom of the ancestors.

I do not follow. Does that mean to preserve existing oom killer from
#PF?

> - do nothing, repeat #PF and rely on fate: "nothing bad will happen if we do it again".
> - add some (progressive) killable delay, rely on good will of (unkillable) neighbors and wait for them to release required memory.

Again, not really sure what you mean

> - mark the current task as cycled in #PF and somehow use this mark in allocator

How?

> - make sure that the current task is really cycled, have no progress, send him fatal signal to kill it and break the cycle.

No! We cannot really kill the task if we could we would have done it by
the oom killer already

> - implement any better ideas,
> - use any combination of previous points
> 
> We can select required strategy for example via sysctl.

Absolutely no! How can admin know any better than the kernel?

> For me "random kill" is worst choice, 
> Why can't we just kill the looped process instead?

See above.

> It can be marked as oom-unkillable, so OOM-killer was unable to select it.
> However I doubt it means "never kill it", for me it is something like "last possible victim" priority.

It means never kill it because of OOM. If it is retrying because of OOM
then it is effectively the same thing.

The oom killer from the #PF doesn't really provide any clear advantage
these days AFAIK. On the other hand it allows for a very disruptive
behavior. In a worst case it can lead to a system panic if the
VM_FAULT_OOM is not really caused by a memory shortage but rather a
wrong error handling. If a task is looping there without any progress
then it is still kilallable which is a much saner behavior IMHO.
-- 
Michal Hocko
SUSE Labs