On Tue, 3 Jul 2012, Oleg Nesterov wrote:
> > unsigned int points, unsigned long totalpages,
> > struct mem_cgroup *memcg, nodemask_t *nodemask,
> > @@ -454,6 +462,7 @@ void oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
> > */
> > if (p->flags & PF_EXITING) {
> > set_tsk_thread_flag(p, TIF_MEMDIE);
> > + put_task_struct(p);
> > return;
> > }
> >
> > @@ -471,6 +480,7 @@ void oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
> > * parent. This attempts to lose the minimal amount of work done while
> > * still freeing memory.
> > */
> > + read_lock(&tasklist_lock);
> > do {
> > list_for_each_entry(child, &t->children, sibling) {
> > unsigned int child_points;
> > @@ -483,15 +493,26 @@ void oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
> > child_points = oom_badness(child, memcg, nodemask,
> > totalpages);
> > if (child_points > victim_points) {
> > + put_task_struct(victim);
> > victim = child;
> > victim_points = child_points;
> > + get_task_struct(victim);
> > }
> > }
> > } while_each_thread(p, t);
> > + read_unlock(&tasklist_lock);
> >
> > - victim = find_lock_task_mm(victim);
> > - if (!victim)
> > + rcu_read_lock();
> > + p = find_lock_task_mm(victim);
> > + if (!p) {
> > + rcu_read_unlock();
> > + put_task_struct(victim);
> > return;
> > + } else if (victim != p) {
> > + get_task_struct(p);
> > + put_task_struct(victim);
> > + victim = p;
> > + }
> >
> > /* mm cannot safely be dereferenced after task_unlock(victim) */
> > mm = victim->mm;
> > @@ -522,9 +543,11 @@ void oom_kill_process(struct task_struct *p, gfp_t gfp_mask, int order,
> > task_unlock(p);
> > do_send_sig_info(SIGKILL, SEND_SIG_FORCED, p, true);
> > }
> > + rcu_read_unlock();
> >
> > set_tsk_thread_flag(victim, TIF_MEMDIE);
> > do_send_sig_info(SIGKILL, SEND_SIG_FORCED, victim, true);
> > + put_task_struct(victim);
>
> It seems to me we can avoid this get/put dance in oom_kill_process(),
> just you need to extend the rcu-protected area. In this case the caller
> of select_bad_process() does a single put_, and
> sysctl_oom_kill_allocating_task doesn't need get_task_struct(current).
> Look more clean/simple to me.
>
We could grab rcu_read_lock() before the first tasklist scan and hold it
until a process is killed, yes, but there's a higher liklihood that it
will never be dropped for concurrent oom kills in the same way that the
write-side of tasklist_lock is currently starved. On a system with a
large number of cpus this isn't even a rare situation to run into: the
read lock will never be dropped on all cpus. I've attempted to make it as
fine-grained as possible and only hold it when absolutely required and use
task references to keep the selected threads around until they are killed.
Let me know if you have a better solution to rcu read lock starvation.
--
To unsubscribe from this list: send the line "unsubscribe cgroups" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
