On Mon 01-07-19 16:04:34, Michal Hocko wrote: > On Mon 01-07-19 22:56:12, Tetsuo Handa wrote: > > On 2019/07/01 22:48, Michal Hocko wrote: > > > On Mon 01-07-19 22:38:58, Tetsuo Handa wrote: > > >> On 2019/07/01 22:17, Michal Hocko wrote: > > >>> On Mon 01-07-19 22:04:22, Tetsuo Handa wrote: > > >>>> But I realized that this patch was too optimistic. We need to wait for mm-less > > >>>> threads until MMF_OOM_SKIP is set if the process was already an OOM victim. > > >>> > > >>> If the process is an oom victim then _all_ threads are so as well > > >>> because that is the address space property. And we already do check that > > >>> before reaching oom_badness IIRC. So what is the actual problem you are > > >>> trying to solve here? > > >> > > >> I'm talking about behavioral change after tsk became an OOM victim. > > >> > > >> If tsk->signal->oom_mm != NULL, we have to wait for MMF_OOM_SKIP even if > > >> tsk->mm == NULL. Otherwise, the OOM killer selects next OOM victim as soon as > > >> oom_unkillable_task() returned true because has_intersects_mems_allowed() returned > > >> false because mempolicy_nodemask_intersects() returned false because all thread's > > >> mm became NULL (despite tsk->signal->oom_mm != NULL). > > > > > > OK, I finally got your point. It was not clear that you are referring to > > > the code _after_ the patch you are proposing. You are indeed right that > > > this would have a side effect that an additional victim could be > > > selected even though the current process hasn't terminated yet. Sigh, > > > another example how the whole thing is subtle so I retract my Ack and > > > request a real life example of where this matters before we think about > > > a proper fix and make the code even more complex. > > > > > > > Instead of checking for mm != NULL, can we move mpol_put_task_policy() from > > do_exit() to __put_task_struct() ? That change will (if it is safe to do) > > prevent exited threads from setting mempolicy = NULL (and confusing > > mempolicy_nodemask_intersects() due to mempolicy == NULL). > > I am sorry but I would have to study it much more and I am not convinced > the time spent on it would be well spent. Thinking about it some more it seems that we can go with your original fix if we also reorder oom_evaluate_task diff --git a/mm/oom_kill.c b/mm/oom_kill.c index f719b64741d6..e5feb0f72e3b 100644 --- a/mm/oom_kill.c +++ b/mm/oom_kill.c @@ -318,9 +318,6 @@ static int oom_evaluate_task(struct task_struct *task, void *arg) struct oom_control *oc = arg; unsigned long points; - if (oom_unkillable_task(task, NULL, oc->nodemask)) - goto next; - /* * This task already has access to memory reserves and is being killed. * Don't allow any other task to have access to the reserves unless @@ -333,6 +330,9 @@ static int oom_evaluate_task(struct task_struct *task, void *arg) goto abort; } + if (oom_unkillable_task(task, NULL, oc->nodemask)) + goto next; + /* * If task is allocating a lot of memory and has been marked to be * killed first if it triggers an oom, then select it. I do not see any strong reason to keep the current ordering. OOM victim check is trivial so it shouldn't add a visible overhead for few unkillable tasks that we might encounter. -- Michal Hocko SUSE Labs
