Re: getting oom/stalls for ltp test cpuset01 with latest/4.9 kernel

[Ganapatrao Kulkarni]

On Fri, Jan 13, 2017 at 2:36 PM, Vlastimil Babka <vbabka@xxxxxxx> wrote:
> On 01/13/2017 05:35 AM, Ganapatrao Kulkarni wrote:
>> On Thu, Jan 12, 2017 at 4:40 PM, Vlastimil Babka <vbabka@xxxxxxx> wrote:
>>> On 01/11/2017 05:46 PM, Michal Hocko wrote:
>>>>
>>>> On Wed 11-01-17 21:52:29, Ganapatrao Kulkarni wrote:
>>>>
>>>>> [ 2398.169391] Node 1 Normal: 951*4kB (UME) 1308*8kB (UME) 1034*16kB
>>>>> (UME) 742*32kB (UME) 581*64kB (UME) 450*128kB (UME) 362*256kB (UME)
>>>>> 275*512kB (ME) 189*1024kB (UM) 117*2048kB (ME) 2742*4096kB (M) = 12047196kB
>>>>
>>>>
>>>> Most of the memblocks are marked Unmovable (except for the 4MB bloks)
>>>
>>>
>>> No, UME here means that e.g. 4kB blocks are available on unmovable, movable
>>> and reclaimable lists.
>>>
>>>> which shouldn't matter because we can fallback to unmovable blocks for
>>>> movable allocation AFAIR so we shouldn't really fail the request. I
>>>> really fail to see what is going on there but it smells really
>>>> suspicious.
>>>
>>>
>>> Perhaps there's something wrong with zonelists and we are skipping the Node
>>> 1 Normal zone. Or there's some race with cpuset operations (but can't see
>>> how).
>>>
>>> The question is, how reproducible is this? And what exactly the test
>>> cpuset01 does? Is it doing multiple things in a loop that could be reduced
>>> to a single testcase?
>>
>> IIUC, this test does node change to  cpuset.mems in loop in parent
>> process in loop and child processes(equal to no of cpus) keeps on
>> allocation and freeing
>> 10 pages till the execution time is over.
>> more details at
>> https://github.com/linux-test-project/ltp/blob/master/testcases/kernel/mem/cpuset/cpuset01.c
>
> Ah, thanks for explaining. Looks like there might be a race where determining
> ac.preferred_zone using current_mems_allowed as ac.nodemask skips the only zone
> that is allowed after the cpuset.mems update, and we only recalculate
> ac.preferred_zone for allocations that are allowed to escape cpusets/watermarks.
> Thus we see only part of the zonelist, missing the only allowed zone. This would
> be due to commit 682a3385e773 ("mm, page_alloc: inline the fast path of the
> zonelist iterator") and/or some others from that series.
>
> Could you try with the following patch please? It also tries to protect from
> race with last non-root cpuset removal, which could cause cpusets_enable() to
> become false in the middle of the function.
>
> ----8<----
> From 9f041839401681f2678edf5040c851d11963c5fe Mon Sep 17 00:00:00 2001
> From: Vlastimil Babka <vbabka@xxxxxxx>
> Date: Fri, 13 Jan 2017 10:01:26 +0100
> Subject: [PATCH] mm, page_alloc: fix race with cpuset update or removal
>
> Changelog and S-O-B TBD.
> ---
>  mm/page_alloc.c | 10 +++++++++-
>  1 file changed, 9 insertions(+), 1 deletion(-)
>
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index 6de9440e3ae2..c397f146843a 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -3775,9 +3775,17 @@ __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
>         /*
>          * Restore the original nodemask if it was potentially replaced with
>          * &cpuset_current_mems_allowed to optimize the fast-path attempt.
> +        * Also recalculate the starting point for the zonelist iterator or
> +        * we could end up iterating over non-eligible zones endlessly.
>          */
> -       if (cpusets_enabled())
> +       if (unlikely(ac.nodemask != nodemask)) {
>                 ac.nodemask = nodemask;
> +               ac.preferred_zoneref = first_zones_zonelist(ac.zonelist,
> +                                               ac.high_zoneidx, ac.nodemask);
> +               if (!ac.preferred_zoneref)
> +                       goto no_zone;
> +       }
> +
>         page = __alloc_pages_slowpath(alloc_mask, order, &ac);
>
>  no_zone:
> --
> 2.11.0
>

this patch did not fix the issue.
issue still exists!
i did bisect and this test passes in 4.4,4.5 and 4.6
test failing since 4.7-rc1

thanks
Ganapat
>
>
>