On Thu 25-10-18 10:24:00, Michal Hocko wrote: > The previous version of this RFC has been posted here [1]. I have fixed > few issues spotted during the review and by 0day bot. I have also reworked > patch 2 to be ratio rather than an absolute number based. > > With this series applied the locking protocol between the oom_reaper and > the exit path is as follows. > > All parts which cannot race should use the exclusive lock on the exit > path. Once the exit path has passed the moment when no blocking locks > are taken then it clears mm->mmap under the exclusive lock. oom_reaper > checks for this and sets MMF_OOM_SKIP only if the exit path is not guaranteed > to finish the job. This is patch 3 so see the changelog for all the details. > > I would really appreciate if David could give this a try and see how > this behaves in workloads where the oom_reaper falls flat now. I have > been playing with sparsely allocated memory with a high pte/real memory > ratio and large mlocked processes and it worked reasonably well. Does this help workloads you were referring to earlier David? > There is still some room for tuning here of course. We can change the > number of retries for the oom_reaper as well as the threshold when the > keep retrying. > > Michal Hocko (3): > mm, oom: rework mmap_exit vs. oom_reaper synchronization > mm, oom: keep retrying the oom_reap operation as long as there is substantial memory left > mm, oom: hand over MMF_OOM_SKIP to exit path if it is guranteed to finish > > Diffstat: > include/linux/oom.h | 2 -- > mm/internal.h | 3 +++ > mm/memory.c | 28 ++++++++++++++-------- > mm/mmap.c | 69 +++++++++++++++++++++++++++++++++-------------------- > mm/oom_kill.c | 45 ++++++++++++++++++++++++---------- > 5 files changed, 97 insertions(+), 50 deletions(-) > > [1] http://lkml.kernel.org/r/20180910125513.311-1-mhocko@xxxxxxxxxx > -- Michal Hocko SUSE Labs
