[LTP] [PATCH v4 4/5] controllers/memcg: increase memory limit in subgroup charge

[Richard Palethorpe <rpalethorpe@suse.de>]

Hello,

Krzysztof Kozlowski <krzysztof.kozlowski@canonical.com> writes:

> On 13/07/2021 11:40, Richard Palethorpe wrote:
>> Hello Krzysztof,
>> 
>> Krzysztof Kozlowski <krzysztof.kozlowski@canonical.com> writes:
>> 
>>> The memcg_subgroup_charge was failing on kernel v5.8 in around 10% cases
>>> with:
>>>
>>>     memcg_subgroup_charge 1 TINFO: Running memcg_process --mmap-anon -s 135168
>>>     memcg_subgroup_charge 1 TINFO: Warming up pid: 19289
>>>     memcg_subgroup_charge 1 TINFO: Process is still here after warm up: 19289
>>>     memcg_subgroup_charge 1 TFAIL: rss is 0, 135168 expected
>>>     memcg_subgroup_charge 1 TPASS: rss is 0 as expected
>>>
>>> In dmesg one could see that OOM killer killed the process even though
>>> group memory limit was matching the usage:
>>>
>>>     memcg_process invoked oom-killer: gfp_mask=0xcc0(GFP_KERNEL), order=0, oom_score_adj=0
>>>     CPU: 4 PID: 19289 Comm: memcg_process Not tainted 5.8.0-1031-oracle #32~20.04.2-Ubuntu
>>>     Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.4.1 12/03/2020
>>>     ...
>>>     memory: usage 132kB, limit 132kB, failcnt 9
>>>     memory+swap: usage 132kB, limit 9007199254740988kB, failcnt 0
>>>     kmem: usage 4kB, limit 9007199254740988kB, failcnt 0
>>>     ...
>>>     Tasks state (memory values in pages):
>>>     [  pid  ]   uid  tgid total_vm      rss pgtables_bytes swapents oom_score_adj name
>>>     [  19289]     0 19289      669      389    40960        0             0 memcg_process
>>>     oom-kill:constraint=CONSTRAINT_MEMCG,nodemask=(null),cpuset=/,mems_allowed=0,oom_memcg=/ltp_19257,task_memcg=/ltp_19257,task=memcg_process,pid=19289,uid=0
>>>     Memory cgroup out of memory: Killed process 19289 (memcg_process) total-vm:2676kB, anon-rss:84kB, file-rss:1468kB, shmem-rss:4kB, UID:0 pgtables:40kB oom_score_adj:0
>>>     oom_reaper: reaped process 19289 (memcg_process), now anon-rss:0kB, file-rss:0kB, shmem-rss:4kB
>>>
>>> It seems actual kernel memory usage by a given group depends on number
>>> of CPUs, where at least one page is allocated per-cpu beside regular
>>> (expected) allocation.  Fix the test on machines with more CPUs by
>>> including this per-CPU memory in group limits, plus some slack of 4
>>> PAGES.  Increase also memory allocation from 32 to 64 pages to be more
>>> distinctive from kernel per-CPU memory.
>> 
>> Actually I think it is 66 pages? Because PAGESIZES=pagesize*33.
>>
>
> Yes, right. Maybe this could be fixed when applying - no need for resend.
>
>
> Best regards,
> Krzysztof

Well I am more than happy with the patchset.

Acked-by: Richard Palethorpe <rpalethorpe@suse.com>

-- 
Thank you,
Richard.