On Sat 14-07-18 06:18:58, Tetsuo Handa wrote: > On 2018/07/13 23:26, Michal Hocko wrote: > > On Thu 12-07-18 14:34:00, David Rientjes wrote: > > [...] > >> diff --git a/mm/oom_kill.c b/mm/oom_kill.c > >> index 0fe4087d5151..e6328cef090f 100644 > >> --- a/mm/oom_kill.c > >> +++ b/mm/oom_kill.c > >> @@ -488,9 +488,11 @@ void __oom_reap_task_mm(struct mm_struct *mm) > >> * Tell all users of get_user/copy_from_user etc... that the content > >> * is no longer stable. No barriers really needed because unmapping > >> * should imply barriers already and the reader would hit a page fault > >> - * if it stumbled over a reaped memory. > >> + * if it stumbled over a reaped memory. If MMF_UNSTABLE is already set, > >> + * reaping as already occurred so nothing left to do. > >> */ > >> - set_bit(MMF_UNSTABLE, &mm->flags); > >> + if (test_and_set_bit(MMF_UNSTABLE, &mm->flags)) > >> + return; > > > > This could lead to pre mature oom victim selection > > oom_reaper exiting victim > > oom_reap_task exit_mmap > > __oom_reap_task_mm __oom_reap_task_mm > > test_and_set_bit(MMF_UNSTABLE) # wins the race > > test_and_set_bit(MMF_UNSTABLE) > > set_bit(MMF_OOM_SKIP) # new victim can be selected now. > > > > Besides that, why should we back off in the first place. We can > > race the two without any problems AFAICS. We already do have proper > > synchronization between the two due to mmap_sem and MMF_OOM_SKIP. > > > > diff --git a/mm/mmap.c b/mm/mmap.c > > index fc41c0543d7f..4642964f7741 100644 > > --- a/mm/mmap.c > > +++ b/mm/mmap.c > > @@ -3073,9 +3073,7 @@ void exit_mmap(struct mm_struct *mm) > > * which clears VM_LOCKED, otherwise the oom reaper cannot > > * reliably test it. > > */ > > - mutex_lock(&oom_lock); > > __oom_reap_task_mm(mm); > > - mutex_unlock(&oom_lock); > > > > set_bit(MMF_OOM_SKIP, &mm->flags); > > David and Michal are using different version as a baseline here. > David is making changes using timeout based back off (in linux-next.git) > which is inappropriately trying to use MMF_UNSTABLE for two purposes. > > Michal is making changes using current code (in linux.git) which does not > address David's concern. Yes I have based it on top of Linus tree because the point of this patch is to get rid of the locking which is no longer needed. I do not see what concern are you talking about. > > My version ( https://marc.info/?l=linux-mm&m=153119509215026 ) is > making changes using current code which also provides oom-badness > based back off in order to address David's concern. > > > down_write(&mm->mmap_sem); > > Anyway, I suggest doing > > mutex_lock(&oom_lock); > set_bit(MMF_OOM_SKIP, &mm->flags); > mutex_unlock(&oom_lock); Why do we need it? > like I mentioned at > http://lkml.kernel.org/r/201807130620.w6D6KiAJ093010@xxxxxxxxxxxxxxxxxxx > even if we make changes on top of linux-next's timeout based back off. says : (3) Prevent from selecting new OOM victim when there is an !MMF_OOM_SKIP mm : which current thread should wait for. [...] : Regarding (A), we can reduce the range oom_lock serializes from : "__oom_reap_task_mm()" to "setting MMF_OOM_SKIP", for oom_lock is useful for (3). But why there is a lock needed for this? This doesn't make much sense to me. If we do not have MMF_OOM_SKIP set we still should have mm_is_oom_victim so no new task should be selected. If we race with the oom reaper than ok, we would just not select a new victim and retry later. -- Michal Hocko SUSE Labs
