[LTP] Question about oom02 testcase

Gou Hao gouhao@uniontech.com
Fri Jun 2 04:47:15 CEST 2023


On 6/1/23 18:50, Li Wang wrote:
> On Thu, Jun 1, 2023 at 5:46 PM Gou Hao <gouhao@uniontech.com> wrote:
>
>> On 6/1/23 16:18, Li Wang wrote:
>>
>> Hi Hao,
>>
>> Thanks for reporting this, comments see below.
>>
>> On Tue, May 30, 2023 at 9:26 AM Gou Hao <gouhao@uniontech.com> wrote:
>>
>>> hello everyone,
>>>
>>> Recently, kernel restarted while I was running oom02.
>>> log:
>>> ```
>>> [480156.950100] Tasks state (memory values in pages):
>>> [480156.950101] [  pid  ]   uid  tgid total_vm      rss pgtables_bytes
>>> swapents oom_score_adj name
>>> [480156.950302] [   2578]    81  2578      523        0 393216
>>> 6          -900 dbus-daemon
>>> [480156.950309] [   2648]   172  2596     2435        0 393216
>>> 5             0 rtkit-daemon
>>> [480156.950322] [   5256]     0  2826    25411        0 589824
>>> 0             0 DetectThread
>>> [480156.950328] [   5404]     0  5404      412        2 393216
>>> 64         -1000 sshd
>>> [480156.950357] [  10518]     0 10518     2586        0 393216
>>> 10             0 at-spi2-registr
>>> [480156.950361] [  10553]     0 10551    10543        0 458752
>>> 9             0 QXcbEventQueue
>>> [480156.950365] [  10867]     0 10567    17579        0 589824
>>> 16             0 QXcbEventQueue
>>> [480156.950370] [  10928]     0 10921     6999        0 458752
>>> 17             0 QXcbEventQueue
>>> [480156.950390] [  11882]     0 11811     7377        0 458752
>>> 10             0 QXcbEventQueue
>>> [480156.950394] [  12052]     0 12052     5823        0 458752
>>> 21             0 fcitx
>>> [480156.950404] [  12115]     0 12114    11678        0 524288
>>> 21             0 QXcbEventQueue
>>> [480156.950408] [ 101558]     0 101558     3549        0 393216
>>> 0             0 runltp
>>> [480156.950486] [1068864]     0 1068864      771        6 327680
>>> 85         -1000 systemd-udevd
>>> [480156.950552] [1035639]     0 1035639       52        0 393216
>>> 14         -1000 oom02
>>> [480156.950556] [1035640]     0 1035640       52        0 393216
>>> 23         -1000 oom02
>>> [480156.950561] [1036065]     0 1036065      493       60 393216
>>> 0          -250 systemd-journal
>>> [480156.950565] [1036087]     0 1036073  6258739  3543942
>>> 37814272        0             0 oom02
>>> [480156.950572] Out of memory and no killable processes...
>>> [480156.950575] Kernel panic - not syncing: System is deadlocked on memory
>>> ```
>>>
>>> oom02-1036073 has been killed before crash.
>>> log:
>>> ```
>>> [480152.242506] [1035177]     0 1035177     4773       20 393216
>>> 115             0 sssd_nss
>>> [480152.242510] [1035376]     0 1035376    25500      391 589824
>>> 602             0 tuned
>>> [480152.242514] [1035639]     0 1035639       52        0 393216
>>> 14         -1000 oom02
>>> [480152.242517] [1035640]     0 1035640       52        0 393216
>>> 19         -1000 oom02
>>> [480152.242522] [1036065]     0 1036065      493      114 393216
>>> 62          -250 systemd-journal
>>> [480152.242525] [1036073]     0 1036073  6258739  3540314 37814272
>>> 104             0 oom02
>>> [480152.242529] Out of memory: Kill process 1036073 (oom02) score 755 or
>>> sacrifice child
>>> [480152.243869] Killed process 1036073 (oom02) total-vm:400559296kB,
>>> anon-rss:226578368kB, file-rss:1728kB, shmem-rss:0kB
>>> [480152.365804] oom_reaper: reaped process 1036073 (oom02), now
>>> anon-rss:226594048kB, file-rss:0kB, shmem-rss:0kB
>>> ```
>>> but its memory can not be reclaimed.I add trace-log to oom_reaper code
>>> in kernel,
>>> I found that there is a large range vma in the memory that cannot be
>>> reclaimed, and the vma has the  `VM_LOCKED` flag, so cannot be reclaimed
>>> immediately.
>>> ```log
>>>         oom_reaper-57    [007] ....   126.063581: __oom_reap_task_mm: gh:
>>> vma is anon:1048691, range=65536
>>>         oom_reaper-57    [007] ....   126.063581: __oom_reap_task_mm: gh:
>>> vma is anon:1048691, range=196608
>>>         oom_reaper-57    [007] ....   126.063582: __oom_reap_task_mm: gh:
>>> vma continue: 1056883, range:3221225472
>>>         oom_reaper-57    [007] ....   126.063583: __oom_reap_task_mm: gh:
>>> vma is anon:112, range=65536
>>>         oom_reaper-57    [007] ....   126.063584: __oom_reap_task_mm: gh:
>>> vma is anon:1048691, range=8388608
>>> ```
>>> `vma continue: 1056883, range:3221225472` is the memory that can not
>>> reclaims. 1057883(0x102073) is vma->vm_flags, it has VM_LOCKED` flag
>>>
>>> oom02 created `nr_cpu` threads and used mmap to allocate memory. mmap
>>> will merge continuous vma into one,
>>> so as long as one thread is still running, the entire vma will not be
>>> released.
>>>
>>> In extreme cases, crashes may occur due to the lack of memory reclamation.
>>>
>>> My question is that the crash in this case is a normal situation or a
>>> bug (kernel or ltp bug) ?
>>>
>>
>> The  ltp-oom test is originally designed to verify OOM mechanism
>> works for memory allocating in three types (normal, mlock, ksm)
>> all as expected.
>>
>> Yes, your analysis is reasonable to some degree, oom_reaper
>> might not reclaim the VMA with locked pages  even after the
>> process termination.
>>
>> But the exact behavior of the oom_reaper and the conditions under
>> which it can or cannot reclaim VMAs may vary depending on the
>> specific kernel version and configuration. So we shouldn't simply
>> regard the system panic as a Kernel or LTP defect.
>> And BTW, what is your tested kernel version?
>>
>> hi Li Wang,
>>
>> Thank you for your reply.
>>
>> My kernel version is 4.19, but it's not a community version.
>>
>> I have only encountered the crash once, and most of the time oom_reaper
>> can handle it well.
>>
>> I tried to find a method or flag to prevent vma merging during mmap, but
>> couldn't find it.
>>
> That also might be related to the value of overcommit_memory,
> if we set 2 (strict mode) to it, the oom_reaper can reclaim VM_LOCKED
> memory more aggressively.
>
> But in oom02 as you can see, it is set to 1 (always mode) for the
> whole test, that might be the reason your system can't recover from
> overcommit and finally crashed.

I do a oom02-test according to your suggestion: set overcommit_memory to 2,

most of the time, returning ENOMEM from the mmap() directly, the oom-kill is

only triggered approximately once, and the memory cannot reclaimed 
quickly by

oom-reaper.

```
Jun  2 10:24:51 ltptest kernel: [  71588]     0 71588 792      244   
393216        0             0 sshd
Jun  2 10:24:51 ltptest kernel: [  71590]     0 71590 792      150   
393216        0             0 sshd
Jun  2 10:24:51 ltptest kernel: [  71591]     0 71591 3565      109   
393216        0             0 bash
Jun  2 10:24:51 ltptest kernel: [  72118]     0 72118 3364       17   
458752        0             0 sleep
Jun  2 10:24:51 ltptest kernel: [  72134]     0 72134 3364       17   
393216        0             0 tail
Jun  2 10:24:51 ltptest kernel: [  72157]     0 72157 52       25   
393216        0         -1000 oom02
Jun  2 10:24:51 ltptest kernel: [  72158]     0 72158 52       14   
393216        0         -1000 oom02
Jun  2 10:24:51 ltptest kernel: [  72203]     0 72203   295609 244870  
2359296        0             0 oom02
Jun  2 10:24:51 ltptest kernel: Out of memory: Kill process 72203 
(oom02) score 373 or sacrifice child

Jun  2 10:24:51 ltptest kernel: Killed process 72203 (oom02) 
total-vm:18918976kB, anon-rss:15671680kB, file-rss:0kB, shmem-rss:0kB


Jun  2 10:24:51 ltptest kernel: oom_reaper: reaped process 72203 
(oom02), now anon-rss:15681280kB, file-rss:0kB, shmem-rss:0kB

```

-- 
thanks,
Gou Hao <gouhao@uniontech.com>



More information about the ltp mailing list