[01/12] mm: oom_kill: remove unnecessary locking in oom_enable()

Commit Message

Michal Hocko March 26, 2015, 11:51 a.m. UTC

On Wed 25-03-15 17:51:31, David Rientjes wrote:
> On Wed, 25 Mar 2015, Johannes Weiner wrote:
> 
> > Setting oom_killer_disabled to false is atomic, there is no need for
> > further synchronization with ongoing allocations trying to OOM-kill.
> > 
> > Signed-off-by: Johannes Weiner <[e-mail]>
> > ---
> >  mm/oom_kill.c | 2 --
> >  1 file changed, 2 deletions(-)
> > 
> > diff --git a/mm/oom_kill.c b/mm/oom_kill.c
> > index 2b665da1b3c9..73763e489e86 100644
> > --- a/mm/oom_kill.c
> > +++ b/mm/oom_kill.c
> > @@ -488,9 +488,7 @@ bool oom_killer_disable(void)
> >   */
> >  void oom_killer_enable(void)
> >  {
> > -	down_write(&oom_sem);
> >  	oom_killer_disabled = false;
> > -	up_write(&oom_sem);
> >  }
> >  
> >  #define K(x) ((x) << (PAGE_SHIFT-10))
> 
> I haven't looked through the new disable-oom-killer-for-pm patchset that 
> was merged, but this oom_killer_disabled thing already looks improperly 
> handled.  I think any correctness or cleanups in this area would be very 
> helpful.
> 
> I think mark_tsk_oom_victim() in mem_cgroup_out_of_memory() is just 
> luckily not racing with a call to oom_killer_enable() and triggering the 
                                    ^^^^^^^^^^
                                    oom_killer_disable?

> WARN_ON(oom_killer_disabled) since there's no "oom_sem" held here, and 
> it's an improper context based on the comment of mark_tsk_oom_victim().

OOM killer is disabled only _after_ all user tasks have been frozen. So
we cannot get any page fault and a race. So the semaphore is not needed
in this path although the comment says otherwise. I can add a comment
clarifying this...
---

> There might be something else that is intended but not implemented 
> correctly that I'm unaware of, but I know of no reason why setting of 
> oom_killer_disabled would need to take a semaphore?
> 
> I'm thinking it has something to do with the remainder of that comment, 
> specifically the "never after oom has been disabled already."
> 
> Michal?

Comments

Michal Hocko March 26, 2015, 1:18 p.m. UTC | #1

On Thu 26-03-15 12:51:40, Michal Hocko wrote:
> On Wed 25-03-15 17:51:31, David Rientjes wrote:
> > On Wed, 25 Mar 2015, Johannes Weiner wrote:
> > 
> > > Setting oom_killer_disabled to false is atomic, there is no need for
> > > further synchronization with ongoing allocations trying to OOM-kill.
> > > 
> > > Signed-off-by: Johannes Weiner <[e-mail]>
> > > ---
> > >  mm/oom_kill.c | 2 --
> > >  1 file changed, 2 deletions(-)
> > > 
> > > diff --git a/mm/oom_kill.c b/mm/oom_kill.c
> > > index 2b665da1b3c9..73763e489e86 100644
> > > --- a/mm/oom_kill.c
> > > +++ b/mm/oom_kill.c
> > > @@ -488,9 +488,7 @@ bool oom_killer_disable(void)
> > >   */
> > >  void oom_killer_enable(void)
> > >  {
> > > -	down_write(&oom_sem);
> > >  	oom_killer_disabled = false;
> > > -	up_write(&oom_sem);
> > >  }
> > >  
> > >  #define K(x) ((x) << (PAGE_SHIFT-10))
> > 
> > I haven't looked through the new disable-oom-killer-for-pm patchset that 
> > was merged, but this oom_killer_disabled thing already looks improperly 
> > handled.  I think any correctness or cleanups in this area would be very 
> > helpful.
> > 
> > I think mark_tsk_oom_victim() in mem_cgroup_out_of_memory() is just 
> > luckily not racing with a call to oom_killer_enable() and triggering the 
>                                     ^^^^^^^^^^
>                                     oom_killer_disable?
> 
> > WARN_ON(oom_killer_disabled) since there's no "oom_sem" held here, and 
> > it's an improper context based on the comment of mark_tsk_oom_victim().
> 
> OOM killer is disabled only _after_ all user tasks have been frozen. So
> we cannot get any page fault and a race. So the semaphore is not needed
> in this path although the comment says otherwise. I can add a comment
> clarifying this...

I am wrong here! pagefault_out_of_memory takes the lock and so the whole
mem_cgroup_out_of_memory is called under the same lock.

David Rientjes March 26, 2015, 7:30 p.m. UTC | #2

On Thu, 26 Mar 2015, Michal Hocko wrote:

> I am wrong here! pagefault_out_of_memory takes the lock and so the whole
> mem_cgroup_out_of_memory is called under the same lock.

If all userspace processes are frozen by the time oom_killer_disable() is 
called, then there shouldn't be any race with the android lmk calling 
mark_tsk_oom_victim() either, so I assume that you're acking this patch?
--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to [e-mail]
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Patch

diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index 14c2f2017e37..20828ecaf3ba 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -1536,6 +1536,11 @@  static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
 	 * quickly exit and free its memory.
 	 */
 	if (fatal_signal_pending(current) || task_will_free_mem(current)) {
+		/*
+		 * We do not hold oom_sem in this path because we know
+		 * we cannot race with oom_kill_disable(). No user runable
+		 * tasks are allowed at the time oom_kill_disable is called.
+		 */
 		mark_tsk_oom_victim(current);
 		return;
 	}