Re: [PATCH RFC] mm/memcontrol: reclaim severe usage over high limit in get_user_pages loop

[Konstantin Khlebnikov]

On 02.08.2019 12:40, Michal Hocko wrote:
> On Mon 29-07-19 20:55:09, Michal Hocko wrote:
> > On Mon 29-07-19 11:49:52, Johannes Weiner wrote:
> > > On Sun, Jul 28, 2019 at 03:29:38PM +0300, Konstantin Khlebnikov wrote:
> > > > --- a/mm/gup.c
> > > > +++ b/mm/gup.c
> > > > @@ -847,8 +847,11 @@ static long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
> > > >   			ret = -ERESTARTSYS;
> > > >   			goto out;
> > > >   		}
> > > > -		cond_resched();
> > > >   
> > > > +		/* Reclaim memory over high limit before stocking too much */
> > > > +		mem_cgroup_handle_over_high(true);
> > >
> > > I'd rather this remained part of the try_charge() call. The code
> > > comment in try_charge says this:
> > >
> > > 	 * We can perform reclaim here if __GFP_RECLAIM but let's
> > > 	 * always punt for simplicity and so that GFP_KERNEL can
> > > 	 * consistently be used during reclaim.
> > >
> > > The simplicity argument doesn't hold true anymore once we have to add
> > > manual calls into allocation sites. We should instead fix try_charge()
> > > to do synchronous reclaim for __GFP_RECLAIM and only punt to userspace
> > > return when actually needed.
> >
> > Agreed. If we want to do direct reclaim on the high limit breach then it
> > should go into try_charge same way we do hard limit reclaim there. I am
> > not yet sure about how/whether to scale the excess. The only reason to
> > move reclaim to return-to-userspace path was GFP_NOWAIT charges. As you
> > say, maybe we should start by always performing the reclaim for
> > sleepable contexts first and only defer for non-sleeping requests.
>
> In other words. Something like patch below (completely untested). Could
> you give it a try Konstantin?

This should work but also eliminate all benefits from deferred reclaim:
bigger batching and running without of any locks.

After that gap between high and max will work just as reserve for atomic allocations.

> diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> index ba9138a4a1de..53a35c526e43 100644
> --- a/mm/memcontrol.c
> +++ b/mm/memcontrol.c
> @@ -2429,8 +2429,12 @@ static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
>   				schedule_work(&memcg->high_work);
>   				break;
>   			}
> -			current->memcg_nr_pages_over_high += batch;
> -			set_notify_resume(current);
> +			if (gfpflags_allow_blocking(gfp_mask)) {
> +				reclaim_high(memcg, nr_pages, GFP_KERNEL);
> +			} else {
> +				current->memcg_nr_pages_over_high += batch;
> +				set_notify_resume(current);
> +			}
>   			break;
>   		}
>   	} while ((memcg = parent_mem_cgroup(memcg)));