[PATCH] hrtimer: prevent migration for raising softirq

[PATCH] hrtimer: prevent migration for raising softirq

[Steven Rostedt]

Due to a possible deadlock, the waking of the softirq was pushed outside
of the hrtimer base locks. See commit 0c96c5979a522c3323c30a078a70120e29b5bdbc

Unfortunately this allows the task to migrate after setting up the softirq
and raising it. Since softirqs run a queue that is per-cpu we may raise the
softirq on the wrong CPU and this will keep the queued softirq task from
running.

To solve this issue, this patch disables preemption around the releasing
of the hrtimer lock and raising of the softirq.

Signed-off-by: Steven Rostedt <srostedt@redhat.com>

---
 kernel/hrtimer.c |    8 ++++++++
 1 file changed, 8 insertions(+)

Index: linus.git/kernel/hrtimer.c
===================================================================
--- linus.git.orig/kernel/hrtimer.c	2008-05-09 15:11:45.000000000 -0400
+++ linus.git/kernel/hrtimer.c	2008-07-03 14:07:30.000000000 -0400
@@ -1003,10 +1003,18 @@ hrtimer_start(struct hrtimer *timer, kti
 	 */
 	raise = timer->state == HRTIMER_STATE_PENDING;

+	/*
+	 * We use preempt_disable to prevent this task from migrating after
+	 * setting up the softirq and raising it. Otherwise, if me migrate
+	 * we will raise the softirq on the wrong CPU.
+	 */
+	preempt_disable();
+
 	unlock_hrtimer_base(timer, &flags);

 	if (raise)
 		hrtimer_raise_softirq();
+	preempt_enable();

 	return ret;
 }

Re: [PATCH] hrtimer: prevent migration for raising softirq

[Andrew Morton]

On Thu, 3 Jul 2008 14:31:26 -0400 (EDT)
Steven Rostedt <rostedt@goodmis.org> wrote:

> Due to a possible deadlock, the waking of the softirq was pushed outside
> of the hrtimer base locks. See commit 0c96c5979a522c3323c30a078a70120e29b5bdbc

Please quote at least the patch title rather than raw hashes.

afaict that patch isn't in 2.6.25 yet you copied stable@kernel.org. 
Please clarify.

> 
> Unfortunately this allows the task to migrate after setting up the softirq
> and raising it. Since softirqs run a queue that is per-cpu we may raise the
> softirq on the wrong CPU and this will keep the queued softirq task from
> running.
> 
> To solve this issue, this patch disables preemption around the releasing
> of the hrtimer lock and raising of the softirq.
> 

This seems to be 2.6.26 material and is inapplicable to 2.6.25?

Re: [PATCH] hrtimer: prevent migration for raising softirq

[Steven Rostedt]

On Thu, 3 Jul 2008, Andrew Morton wrote:

>
> On Thu, 3 Jul 2008 14:31:26 -0400 (EDT)
> Steven Rostedt <rostedt@goodmis.org> wrote:
>
> > Due to a possible deadlock, the waking of the softirq was pushed outside
> > of the hrtimer base locks. See commit 0c96c5979a522c3323c30a078a70120e29b5bdbc
>
> Please quote at least the patch title rather than raw hashes.

Sorry, will do next time. I'm getting to use to git.

>
> afaict that patch isn't in 2.6.25 yet you copied stable@kernel.org.
> Please clarify.

The reason I did so is because that commit has a Cc: stable@kernel.org.
I figured that if that patch is in queue for stable this one should be in
right afterwards.

>
> >
> > Unfortunately this allows the task to migrate after setting up the softirq
> > and raising it. Since softirqs run a queue that is per-cpu we may raise the
> > softirq on the wrong CPU and this will keep the queued softirq task from
> > running.
> >
> > To solve this issue, this patch disables preemption around the releasing
> > of the hrtimer lock and raising of the softirq.
> >
>
> This seems to be 2.6.26 material and is inapplicable to 2.6.25?
>

[/me goes to look at the 2.6.25.4 he has sitting around]

The above mentioned commit (Title: hrtimer: raise softirq unlocked to
avoid circular lock dependency) is, in fact, in 2.6.25.4.

-- Steve

Re: [PATCH] hrtimer: prevent migration for raising softirq

[Andrew Morton]

On Thu, 3 Jul 2008 21:56:00 -0400 (EDT) Steven Rostedt <rostedt@goodmis.org> wrote:

> 
> 
> On Thu, 3 Jul 2008, Andrew Morton wrote:
> 
> >
> > On Thu, 3 Jul 2008 14:31:26 -0400 (EDT)
> > Steven Rostedt <rostedt@goodmis.org> wrote:
> >
> > > Due to a possible deadlock, the waking of the softirq was pushed outside
> > > of the hrtimer base locks. See commit 0c96c5979a522c3323c30a078a70120e29b5bdbc
> >
> > Please quote at least the patch title rather than raw hashes.
> 
> Sorry, will do next time. I'm getting to use to git.
> 
> >
> > afaict that patch isn't in 2.6.25 yet you copied stable@kernel.org.
> > Please clarify.
> 
> The reason I did so is because that commit has a Cc: stable@kernel.org.
> I figured that if that patch is in queue for stable this one should be in
> right afterwards.
> 
> >
> > >
> > > Unfortunately this allows the task to migrate after setting up the softirq
> > > and raising it. Since softirqs run a queue that is per-cpu we may raise the
> > > softirq on the wrong CPU and this will keep the queued softirq task from
> > > running.
> > >
> > > To solve this issue, this patch disables preemption around the releasing
> > > of the hrtimer lock and raising of the softirq.
> > >
> >
> > This seems to be 2.6.26 material and is inapplicable to 2.6.25?
> >
> 
> [/me goes to look at the 2.6.25.4 he has sitting around]
> 
> The above mentioned commit (Title: hrtimer: raise softirq unlocked to
> avoid circular lock dependency) is, in fact, in 2.6.25.4.
> 

Ah, OK, I'm a dope.

Re: [PATCH] hrtimer: prevent migration for raising softirq

[Steven Rostedt]

On Thu, 3 Jul 2008, Andrew Morton wrote:
> >
> > [/me goes to look at the 2.6.25.4 he has sitting around]
> >
> > The above mentioned commit (Title: hrtimer: raise softirq unlocked to
> > avoid circular lock dependency) is, in fact, in 2.6.25.4.
> >
>
> Ah, OK, I'm a dope.
>

No, you're not. Overworked, perhaps, but definitely not a dope ;-)

-- Steve