Comment # 1 on bug 1040381 from Michal Hocko

[   34.041681] [ 1776]     0  1776     2143        0      11       3       46  
          0 oom01
[   34.042871] [ 1783]     0  1783   790624   481065     952       5       48  
          0 oom01
[   34.043962] Out of memory: Kill process 1783 (oom01) score 523 or sacrifice
child
[   34.044934] Killed process 1783 (oom01) total-vm:3162496kB,
anon-rss:1924260kB, file-rss:0kB, shmem-rss:0kB
[   34.046614] oom_reaper: reaped process 1783 (oom01), now anon-rss:1924308kB,
file-rss:0kB, shmem-rss:0kB
[...]

As we can see pid 1783 has been chosen and killed and the oom reaper wasn't
able to reap it. The oom reaper hasn't reported a failure so we can get
mmap_sem lock and there must be other reason why the memory cannot be freed.
Moreover anon-rss even increased.

The task doesn't go away as well
[   34.814903] [ 1784]     0  1783   790624   481077     952       5        0  
          0 oom01

As the oom_reaper is done with it, it is marked as MMF_OOM_SKIP and a different
task has to be killed instead. This is all expected behavior except that it is
strange that oom01 doesn't go away. I will have a look at the test case.