Re: [PATCH memcg] memcg: prohibit unconditional exceeding the limit of dying tasks

[Michal Hocko]

On Mon 13-09-21 12:37:56, Vasily Averin wrote:
> On 9/13/21 11:39 AM, Michal Hocko wrote:
> > On Mon 13-09-21 10:51:37, Vasily Averin wrote:
> >> On 9/10/21 3:39 PM, Vasily Averin wrote:
> >>> The kernel currently allows dying tasks to exceed the memcg limits.
> >>> The allocation is expected to be the last one and the occupied memory
> >>> will be freed soon.
> >>> This is not always true because it can be part of the huge vmalloc
> >>> allocation. Allowed once, they will repeat over and over again.
> >>
> >>> diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> >>> index 389b5766e74f..67195fcfbddf 100644
> >>> --- a/mm/memcontrol.c
> >>> +++ b/mm/memcontrol.c
> >>> @@ -2622,15 +2625,6 @@ static int try_charge_memcg(struct mem_cgroup *memcg, gfp_t gfp_mask,
> >>>  	if (gfp_mask & __GFP_ATOMIC)
> >>>  		goto force;
> >>>  
> >>> -	/*
> >>> -	 * Unlike in global OOM situations, memcg is not in a physical
> >>> -	 * memory shortage.  Allow dying and OOM-killed tasks to
> >>> -	 * bypass the last charges so that they can exit quickly and
> >>> -	 * free their memory.
> >>> -	 */
> >>> -	if (unlikely(should_force_charge()))
> >>> -		goto force;
> >>> -
> >>
> >> Should we keep current behaviour for (current->flags & PF_EXITING) case perhaps?
> > 
> > Why?
> 
> On this stage task really dies and mostly releases taken resources.
> It can allocate though, and this allocation can reach memcg limit due to the activity
> of parallel memcg threads.
> 
> Noting bad should happen if we reject this allocation,
> because the same thing can happen in non-memcg case too.
> However I doubt misuse is possible here and we have possibility to allow graceful shutdown here.
> 
> In other words: we are not obliged to allow such allocations, but we CAN do it because
> we hope that it is safe and cannot be misused.

This is a lot of hoping that has turned out to be a bad strategy in the
existing code.  So let's stop hoping and if we are shown that an
exit path really benefits from a special treatment then we can add it
with a good reasoning rathat than "we hope it's gonna be ok".
-- 
Michal Hocko
SUSE Labs