On Fri, Jan 18, 2019 at 11:09 PM Michal Hocko <mhocko@xxxxxxxxxx> wrote:
>
> On Fri 18-01-19 16:50:22, Shakeel Butt wrote:
> [...]
> > On looking further it seems like the process selected to be oom-killed
> > has exited even before reaching read_lock(&tasklist_lock) in
> > oom_kill_process(). More specifically the tsk->usage is 1 which is due
> > to get_task_struct() in oom_evaluate_task() and the put_task_struct
> > within for_each_thread() frees the tsk and for_each_thread() tries to
> > access the tsk. The easiest fix is to do get/put across the
> > for_each_thread() on the selected task.
>
> Very well spotted! The code seems safe because we are careful to
> transfer the victim along with reference counting but I've totally
> missed that the loop itself needs a reference. It seems that this has
> been broken since the heuristic has been introduced. But I haven't
> checked it closely. I am still on vacation.
>
> > Now the next question is should we continue with the oom-kill as the
> > previously selected task has exited? However before adding more
> > complexity and heuristics, let's answer why we even look at the
> > children of oom-kill selected task?
>
> The objective was the work protection assuming that children did less
> work than their parrent. I find this argument a bit questionable because
> it highly depends a specific workload while it opens doors for
> problematic behavior at the same time. If you have a fork bomb like
> workload then it is basically hard to resolve the OOM condition as
> children have barely any memory so we keep looping killing tasks which
> will not free up much. So I am all for removing this heuristic.
>
> > The select_bad_process() has already
> > selected the worst process in the system/memcg. Due to race, the
> > selected process might not be the worst at the kill time but does that
> > matter matter?
>
> No, we don't I believe. The aim of the oom killer is to kill something.
> We will never be ideal here because this is a land of races.
>
> > The userspace can play with oom_score_adj to prefer
> > children to be killed before the parent. I looked at the history but it
> > seems like this is there before git history.
> >
> > Signed-off-by: Shakeel Butt <shakeelb@xxxxxxxxxx>
>
> Fixes: 5e9d834a0e0c ("oom: sacrifice child with highest badness score for parent")
> Cc: stable
>
> Acked-by: Michal Hocko <mhocko@xxxxxxxx>
>
> Thanks!
Thanks for the review. I will keep this for the stable branches and
for the next release I will remove this whole children selection
heuristic.
Shakeel
> > ---
> > mm/oom_kill.c | 8 ++++++++
> > 1 file changed, 8 insertions(+)
> >
> > diff --git a/mm/oom_kill.c b/mm/oom_kill.c
> > index 0930b4365be7..1a007dae1e8f 100644
> > --- a/mm/oom_kill.c
> > +++ b/mm/oom_kill.c
> > @@ -981,6 +981,13 @@ static void oom_kill_process(struct oom_control *oc, const char *message)
> > * still freeing memory.
> > */
> > read_lock(&tasklist_lock);
> > +
> > + /*
> > + * The task 'p' might have already exited before reaching here. The
> > + * put_task_struct() will free task_struct 'p' while the loop still try
> > + * to access the field of 'p', so, get an extra reference.
> > + */
> > + get_task_struct(p);
> > for_each_thread(p, t) {
> > list_for_each_entry(child, &t->children, sibling) {
> > unsigned int child_points;
> > @@ -1000,6 +1007,7 @@ static void oom_kill_process(struct oom_control *oc, const char *message)
> > }
> > }
> > }
> > + put_task_struct(p);
> > read_unlock(&tasklist_lock);
> >
> > /*
> > --
> > 2.20.1.321.g9e740568ce-goog
>
> --
> Michal Hocko
> SUSE Labs
