Re: [PATCH] mm: memcg: fix over reclaiming mem cgroup

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Mon, Jan 23, 2012 at 5:02 AM, Michal Hocko <mhocko@xxxxxxx> wrote:
> On Sat 21-01-12 22:49:23, Hillf Danton wrote:
>> In soft limit reclaim, overreclaim occurs when pages are reclaimed from mem
>> group that is under its soft limit, or when more pages are reclaimd than the
>> exceeding amount, then performance of reclaimee goes down accordingly.
>
> First of all soft reclaim is more a help for the global memory pressure
> balancing rather than any guarantee about how much we reclaim for the
> group.
> We need to do more changes in order to make it a guarantee.
> For example you implementation will cause severe problems when all
> cgroups are soft unlimited (default conf.) or when nobody is above the
> limit but the total consumption triggers the global reclaim. Therefore
> nobody is in excess and you would skip all groups and only bang on the
> root memcg.
>
> Ying Han has a patch which basically skips all cgroups which are under
> its limit until we reach a certain reclaim priority but even for this we
> need some additional changes - e.g. reverse the current default setting
> of the soft limit.
>
> Anyway, I like the nr_to_reclaim reduction idea because we have to do
> this in some way because the global reclaim starts with ULONG
> nr_to_scan.

Agree with Michal where there are quite a lot changes we need to get
in for soft limit before any further optimization.

Hillf, please refer to the patch from Johannes
https://lkml.org/lkml/2012/1/13/99 which got quite a lot recent
discussions. I am expecting to get that in before further soft limit
changes.

Thanks

--Ying



>
>> A helper function is added to compute the number of pages that exceed the soft
>> limit of given mem cgroup, then the excess pages are used when every reclaimee
>> is reclaimed to avoid overreclaim.
>>
>> Signed-off-by: Hillf Danton <dhillf@xxxxxxxxx>
>> ---
>>
>> --- a/mm/memcontrol.c Tue Jan 17 20:41:36 2012
>> +++ b/mm/memcontrol.c Sat Jan 21 21:18:46 2012
>> @@ -1662,6 +1662,21 @@ static int mem_cgroup_soft_reclaim(struc
>>       return total;
>>  }
>>
>> +unsigned long mem_cgroup_excess_pages(struct mem_cgroup *memcg)
>> +{
>> +     unsigned long pages;
>> +
>> +     if (mem_cgroup_disabled())
>> +             return 0;
>> +     if (!memcg)
>> +             return 0;
>> +     if (mem_cgroup_is_root(memcg))
>> +             return 0;
>> +
>> +     pages = res_counter_soft_limit_excess(&memcg->res) >> PAGE_SHIFT;
>> +     return pages;
>> +}
>> +
>>  /*
>>   * Check OOM-Killer is already running under our hierarchy.
>>   * If someone is running, return false.
>> --- a/mm/vmscan.c     Sat Jan 14 14:02:20 2012
>> +++ b/mm/vmscan.c     Sat Jan 21 21:30:06 2012
>> @@ -2150,8 +2150,34 @@ static void shrink_zone(int priority, st
>>                       .mem_cgroup = memcg,
>>                       .zone = zone,
>>               };
>> +             unsigned long old;
>> +             bool clobbered = false;
>> +
>> +             if (memcg != NULL) {
>> +                     unsigned long excess;
>> +
>> +                     excess = mem_cgroup_excess_pages(memcg);
>> +                     /*
>> +                      * No bother reclaiming pages from mem cgroup that
>> +                      * is under soft limit
>> +                      */
>> +                     if (!excess)
>> +                             goto next;
>> +                     /*
>> +                      * And reclaim no more pages than excess
>> +                      */
>> +                     if (excess < sc->nr_to_reclaim) {
>> +                             old = sc->nr_to_reclaim;
>> +                             sc->nr_to_reclaim = excess;
>> +                             clobbered = true;
>> +                     }
>> +             }
>>
>>               shrink_mem_cgroup_zone(priority, &mz, sc);
>> +
>> +             if (clobbered)
>> +                     sc->nr_to_reclaim = old;
>> +next:
>>               /*
>>                * Limit reclaim has historically picked one memcg and
>>                * scanned it with decreasing priority levels until
>> --- a/include/linux/memcontrol.h      Thu Jan 19 22:03:14 2012
>> +++ b/include/linux/memcontrol.h      Sat Jan 21 21:35:50 2012
>> @@ -161,6 +161,7 @@ unsigned long mem_cgroup_soft_limit_recl
>>                                               gfp_t gfp_mask,
>>                                               unsigned long *total_scanned);
>>  u64 mem_cgroup_get_limit(struct mem_cgroup *memcg);
>> +unsigned long mem_cgroup_excess_pages(struct mem_cgroup *memcg);
>>
>>  void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx);
>>  #ifdef CONFIG_TRANSPARENT_HUGEPAGE
>> @@ -376,6 +377,11 @@ unsigned long mem_cgroup_soft_limit_recl
>>
>>  static inline
>>  u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
>> +{
>> +     return 0;
>> +}
>> +
>> +static inline unsigned long mem_cgroup_excess_pages(struct mem_cgroup *memcg)
>>  {
>>       return 0;
>>  }
>> --
>> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
>> the body of a message to majordomo@xxxxxxxxxxxxxxx
>> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>> Please read the FAQ at  http://www.tux.org/lkml/
>
> --
> Michal Hocko
> SUSE Labs
> SUSE LINUX s.r.o.
> Lihovarska 1060/12
> 190 00 Praha 9
> Czech Republic

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href


[Index of Archives]     [Linux ARM Kernel]     [Linux ARM]     [Linux Omap]     [Fedora ARM]     [IETF Annouce]     [Bugtraq]     [Linux]     [Linux OMAP]     [Linux MIPS]     [ECOS]     [Asterisk Internet PBX]     [Linux API]