Re: [PATCH RFC] mm/vmscan: try to protect active working set of cgroup from reclaim.

[Johannes Weiner]

On Fri, Mar 01, 2019 at 10:46:34PM +0300, Andrey Ryabinin wrote:
> On 3/1/19 8:49 PM, Johannes Weiner wrote:
> > On Fri, Mar 01, 2019 at 01:38:26PM +0300, Andrey Ryabinin wrote:
> >> On 2/26/19 3:50 PM, Andrey Ryabinin wrote:
> >>> On 2/22/19 10:15 PM, Johannes Weiner wrote:
> >>>> On Fri, Feb 22, 2019 at 08:58:25PM +0300, Andrey Ryabinin wrote:
> >>>>> In a presence of more than 1 memory cgroup in the system our reclaim
> >>>>> logic is just suck. When we hit memory limit (global or a limit on
> >>>>> cgroup with subgroups) we reclaim some memory from all cgroups.
> >>>>> This is sucks because, the cgroup that allocates more often always wins.
> >>>>> E.g. job that allocates a lot of clean rarely used page cache will push
> >>>>> out of memory other jobs with active relatively small all in memory
> >>>>> working set.
> >>>>>
> >>>>> To prevent such situations we have memcg controls like low/max, etc which
> >>>>> are supposed to protect jobs or limit them so they to not hurt others.
> >>>>> But memory cgroups are very hard to configure right because it requires
> >>>>> precise knowledge of the workload which may vary during the execution.
> >>>>> E.g. setting memory limit means that job won't be able to use all memory
> >>>>> in the system for page cache even if the rest the system is idle.
> >>>>> Basically our current scheme requires to configure every single cgroup
> >>>>> in the system.
> >>>>>
> >>>>> I think we can do better. The idea proposed by this patch is to reclaim
> >>>>> only inactive pages and only from cgroups that have big
> >>>>> (!inactive_is_low()) inactive list. And go back to shrinking active lists
> >>>>> only if all inactive lists are low.
> >>>>
> >>>> Yes, you are absolutely right.
> >>>>
> >>>> We shouldn't go after active pages as long as there are plenty of
> >>>> inactive pages around. That's the global reclaim policy, and we
> >>>> currently fail to translate that well to cgrouped systems.
> >>>>
> >>>> Setting group protections or limits would work around this problem,
> >>>> but they're kind of a red herring. We shouldn't ever allow use-once
> >>>> streams to push out hot workingsets, that's a bug.
> >>>>
> >>>>> @@ -2489,6 +2491,10 @@ static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
> >>>>>  
> >>>>>  		scan >>= sc->priority;
> >>>>>  
> >>>>> +		if (!sc->may_shrink_active && inactive_list_is_low(lruvec,
> >>>>> +						file, memcg, sc, false))
> >>>>> +			scan = 0;
> >>>>> +
> >>>>>  		/*
> >>>>>  		 * If the cgroup's already been deleted, make sure to
> >>>>>  		 * scrape out the remaining cache.
> >>>>> @@ -2733,6 +2739,7 @@ static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
> >>>>>  	struct reclaim_state *reclaim_state = current->reclaim_state;
> >>>>>  	unsigned long nr_reclaimed, nr_scanned;
> >>>>>  	bool reclaimable = false;
> >>>>> +	bool retry;
> >>>>>  
> >>>>>  	do {
> >>>>>  		struct mem_cgroup *root = sc->target_mem_cgroup;
> >>>>> @@ -2742,6 +2749,8 @@ static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
> >>>>>  		};
> >>>>>  		struct mem_cgroup *memcg;
> >>>>>  
> >>>>> +		retry = false;
> >>>>> +
> >>>>>  		memset(&sc->nr, 0, sizeof(sc->nr));
> >>>>>  
> >>>>>  		nr_reclaimed = sc->nr_reclaimed;
> >>>>> @@ -2813,6 +2822,13 @@ static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
> >>>>>  			}
> >>>>>  		} while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
> >>>>>  
> >>>>> +		if ((sc->nr_scanned - nr_scanned) == 0 &&
> >>>>> +		     !sc->may_shrink_active) {
> >>>>> +			sc->may_shrink_active = 1;
> >>>>> +			retry = true;
> >>>>> +			continue;
> >>>>> +		}
> >>>>
> >>>> Using !scanned as the gate could be a problem. There might be a cgroup
> >>>> that has inactive pages on the local level, but when viewed from the
> >>>> system level the total inactive pages in the system might still be low
> >>>> compared to active ones. In that case we should go after active pages.
> >>>>
> >>>> Basically, during global reclaim, the answer for whether active pages
> >>>> should be scanned or not should be the same regardless of whether the
> >>>> memory is all global or whether it's spread out between cgroups.
> >>>>
> >>>> The reason this isn't the case is because we're checking the ratio at
> >>>> the lruvec level - which is the highest level (and identical to the
> >>>> node counters) when memory is global, but it's at the lowest level
> >>>> when memory is cgrouped.
> >>>>
> >>>> So IMO what we should do is:
> >>>>
> >>>> - At the beginning of global reclaim, use node_page_state() to compare
> >>>>   the INACTIVE_FILE:ACTIVE_FILE ratio and then decide whether reclaim
> >>>>   can go after active pages or not. Regardless of what the ratio is in
> >>>>   individual lruvecs.
> >>>>
> >>>> - And likewise at the beginning of cgroup limit reclaim, walk the
> >>>>   subtree starting at sc->target_mem_cgroup, sum up the INACTIVE_FILE
> >>>>   and ACTIVE_FILE counters, and make inactive_is_low() decision on
> >>>>   those sums.
> >>>>
> >>>
> >>> Sounds reasonable.
> >>>
> >>
> >> On the second thought it seems to be better to keep the decision on lru level.
> >> There are couple reasons for this:
> >>
> >> 1) Using bare node_page_state() (or sc->targe_mem_cgroup's total_[in]active counters) would be wrong.
> >>  Because some cgroups might have protection set (memory.low) and we must take it into account. Also different
> >> cgroups have different available swap space/memory.swappiness and it must be taken into account as well to.
> >>
> >> So it has to be yet another full memcg-tree iteration.
> > 
> > It should be possible to take that into account on the first iteration
> > and adjust the inactive/active counters in proportion to how much of
> > the cgroup's total memory is exempt by memory.low or min, right?
> > 
> 
> Should be possible, more complexity though to this subtle code.
> 
> 
> >> 2) Let's consider simple case. Two cgroups, one with big 'active' set of pages the other allocates one-time used pages.
> >> So the total inactive is low, thus checking inactive ratio on higher level will result in reclaiming pages.
> >> While with check on lru-level only inactive will be reclaimed.
> > 
> > It's the other way around. Let's say you have two cgroups, A and B:
> > 
> >         A:  500M inactive   10G active -> inactive is low
> >         B:   10G inactive  500M active -> inactive is NOT low
> >    ----------------------------------------------------------
> >    global: 10.5G inactive 10.5G active -> inactive is NOT low
> > 
> > Checking locally will scan active pages from A.
> 
> No, checking locally will not scan active from A. Initial state of
> sc->may_shrink_active = 0, so A group will be skipped completely,
> and will reclaim from B. Since overall reclaim was successful,
> sc->may_shrink_active remain 0 and A will be protected as long as B
> supply enough inactive pages.

Oh, this was a misunderstanding. When you wrote "on second thought it
seems to be better to keep the decision at the lru level", I assumed
you were arguing for keeping the current code as-is and abandoning
your patch.

But that leaves my questions from above unanswered. Consider the
following situation:

  A: 50M inactive   0 active
  B:   0 inactive 20G active

If the processes in A and B were not cgrouped, these pages would be on
a single LRU and we'd go after B's active pages.

But with your patches, we'd reclaim only A's inactive pages.

What's the justification for that unfairness?

Keep in mind that users also enable the memory controller on cgroups
simply to keep track of how much memory different process groups are
using.  Merely enabling the controller without otherwise configuring
any protections shouldn't change the aging behavior for that memory.

And that USED to be the case. We USED to have a physical global LRU
list that held all the cgroup pages, and then a per-cgroup LRU list
that was only for limit reclaim.

We got rid of the physical global LRU list for cgrouped memory with
the idea that we can emulate the global aging. Hence the round-robin
tree iteration in shrink_node().

We just need to fix the inactive_list_is_low() check to also emulate
the global LRU behavior.

That would fix your problem of scanning active pages when there are
already plenty of inactive pages in the system, but without the risk
of severely overreclaiming the inactive pages of a small group.