On Fri, Jan 18, 2019 at 7:35 PM Tetsuo Handa
<penguin-kernel@xxxxxxxxxxxxxxxxxxx> wrote:
>
> On 2019/01/19 9:50, 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.
>
> Good catch. p->usage can become 1 while printk()ing a lot at dump_header().
>
> > @@ -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);
>
> Moving put_task_struct(p) to after read_unlock(&tasklist_lock) will reduce
> latency of a write_lock(&tasklist_lock) waiter.
>
> > read_unlock(&tasklist_lock);
> >
> > /*
> >
>
> By the way, p->usage is already 1 implies that p->mm == NULL due to already
> completed exit_mm(p). Then, process_shares_mm(child, p->mm) might fail to
> return true for some of children. Not critical but might lead to unnecessary
> oom_badness() calls for child selection. Maybe we want to use same logic
> __oom_kill_process() uses (i.e. bail out if find_task_lock_mm(p) failed)?
Thanks for the review. I am thinking of removing the whole children
selection heuristic for now.
Shakeel
