[RT] rt priority losing

[RT] rt priority losing

[Steven Rostedt]

Ingo or Tglx,

It has come to my attention that the dynamic hrtimer softirq can lose a
boosted priority.  That is, if a softirq is running while a timeout
happens, and the call back is of lower priority than the currently
running hrtimer softirq, the timer interrupt will still lower the
hrtimer softirq.

Here's the problem code:

static void wakeup_softirqd_prio(int softirq, int prio)
{
	/* Interrupts are disabled: no need to stop preemption */
	struct task_struct *tsk = __get_cpu_var(ksoftirqd[softirq].tsk);

	if (tsk) {
		if (tsk->normal_prio != prio) {
			struct sched_param param;

			param.sched_priority = MAX_RT_PRIO-1 - prio;
			setscheduler(tsk, -1, SCHED_FIFO, &param);
		}
		if(tsk->state != TASK_RUNNING)
			wake_up_process(tsk);
	}
}


So, tsk could be softirqd-hrmono and we lower the priority. (only
normal_prio is checked versus prio).

So this can be a problem, if the softirq function holds a lock of a high
priority task, and is running boosted.  If another timer goes off with a
lower priority, we can lower the priority of the softirqd and lose the
inherited priority that it was running at.

-- Steve

Re: [RT] rt priority losing

[Steven Rostedt]

On Mon, 2006-07-24 at 18:00 +0100, Esben Nielsen wrote:
> On Mon, 24 Jul 2006, Steven Rostedt wrote:
> 
> > Ingo or Tglx,
> >
> > It has come to my attention that the dynamic hrtimer softirq can lose a
> > boosted priority.  That is, if a softirq is running while a timeout
> > happens, and the call back is of lower priority than the currently
> > running hrtimer softirq, the timer interrupt will still lower the
> > hrtimer softirq.
> >
> > Here's the problem code:
> >
> > static void wakeup_softirqd_prio(int softirq, int prio)
> > {
> > 	/* Interrupts are disabled: no need to stop preemption */
> > 	struct task_struct *tsk = __get_cpu_var(ksoftirqd[softirq].tsk);
> >
> > 	if (tsk) {
> > 		if (tsk->normal_prio != prio) {
> > 			struct sched_param param;
> >
> > 			param.sched_priority = MAX_RT_PRIO-1 - prio;
> > 			setscheduler(tsk, -1, SCHED_FIFO, &param);
> > 		}
> > 		if(tsk->state != TASK_RUNNING)
> > 			wake_up_process(tsk);
> > 	}
> > }
> >
> >
> > So, tsk could be softirqd-hrmono and we lower the priority. (only
> > normal_prio is checked versus prio).
> >
> > So this can be a problem, if the softirq function holds a lock of a high
> > priority task, and is running boosted.  If another timer goes off with a
> > lower priority, we can lower the priority of the softirqd and lose the
> > inherited priority that it was running at.
> 
> There is a check for that inside setscheduler():
>  	p->prio = rt_mutex_getprio(p);

OK, you are right about this.  The PI chain should not be affected.  But
this could still be a problem if the softirq was running at a high prio
for a task when a lower prio callback needs to be made.  It looks like
timer is removed from the base before the function runs.  So when the
interrupt looks at the base to determine the priority to set it at, it
might actually lower the priority of a running hrtimer thread.

-- Steve

Re: [RT] rt priority losing

[Thomas Gleixner]

At Mon, 24 Jul 2006 12:37:22 -0400,
Steven Rostedt wrote:
> > > So this can be a problem, if the softirq function holds a lock of a high
> > > priority task, and is running boosted.  If another timer goes off with a
> > > lower priority, we can lower the priority of the softirqd and lose the
> > > inherited priority that it was running at.
> > 
> > There is a check for that inside setscheduler():
> >  	p->prio = rt_mutex_getprio(p);
> 
> OK, you are right about this.  The PI chain should not be affected.  But
> this could still be a problem if the softirq was running at a high prio
> for a task when a lower prio callback needs to be made.  It looks like
> timer is removed from the base before the function runs.  So when the
> interrupt looks at the base to determine the priority to set it at, it
> might actually lower the priority of a running hrtimer thread.

The correct solution is to run the callback in the context of
the thread which will receive it and get rid of the softirq.

I experimented with this already, but its more than a saturday
afternoon project.

	tglx

Re: [RT] rt priority losing

[Esben Nielsen]

On Mon, 24 Jul 2006, Steven Rostedt wrote:

> Ingo or Tglx,
>
> It has come to my attention that the dynamic hrtimer softirq can lose a
> boosted priority.  That is, if a softirq is running while a timeout
> happens, and the call back is of lower priority than the currently
> running hrtimer softirq, the timer interrupt will still lower the
> hrtimer softirq.
>
> Here's the problem code:
>
> static void wakeup_softirqd_prio(int softirq, int prio)
> {
> 	/* Interrupts are disabled: no need to stop preemption */
> 	struct task_struct *tsk = __get_cpu_var(ksoftirqd[softirq].tsk);
>
> 	if (tsk) {
> 		if (tsk->normal_prio != prio) {
> 			struct sched_param param;
>
> 			param.sched_priority = MAX_RT_PRIO-1 - prio;
> 			setscheduler(tsk, -1, SCHED_FIFO, &param);
> 		}
> 		if(tsk->state != TASK_RUNNING)
> 			wake_up_process(tsk);
> 	}
> }
>
>
> So, tsk could be softirqd-hrmono and we lower the priority. (only
> normal_prio is checked versus prio).
>
> So this can be a problem, if the softirq function holds a lock of a high
> priority task, and is running boosted.  If another timer goes off with a
> lower priority, we can lower the priority of the softirqd and lose the
> inherited priority that it was running at.

There is a check for that inside setscheduler():
 	p->prio = rt_mutex_getprio(p);

Esben

>
> -- Steve
>
> -
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at  http://www.tux.org/lkml/
>

[PATCH -rt] Don't let raise_softirq_prio lower the prio (was: [RT] rt priority losing)

[Steven Rostedt]

On Mon, 2006-07-24 at 18:54 +0100, Esben Nielsen wrote:

> >
> > OK, you are right about this.  The PI chain should not be affected.  But
> > this could still be a problem if the softirq was running at a high prio
> > for a task when a lower prio callback needs to be made.  It looks like
> > timer is removed from the base before the function runs.  So when the
> > interrupt looks at the base to determine the priority to set it at, it
> > might actually lower the priority of a running hrtimer thread.
> >
> 
> That is a simple bug which ought to be simple fixable.

I guess the simple fix is not to allow the interrupt to lower the
priority. But simple fixes do not always handle all the cases.

Thomas G., see any side effects with this patch?

-- Steve

Signed-off-by: Steven Rostedt <rostedt@goodmis.org>

Index: linux-2.6.17-rt7/kernel/softirq.c
===================================================================
--- linux-2.6.17-rt7.orig/kernel/softirq.c	2006-07-24 13:08:53.000000000 -0400
+++ linux-2.6.17-rt7/kernel/softirq.c	2006-07-24 13:10:13.000000000 -0400
@@ -108,7 +108,12 @@ static void wakeup_softirqd_prio(int sof
 	struct task_struct *tsk = __get_cpu_var(ksoftirqd[softirq].tsk);
 
 	if (tsk) {
-		if (tsk->normal_prio != prio) {
+		/*
+		 * The lower the prio, the higher the priority.
+		 * This can only raise the priority but it can
+		 * not lower it.
+		 */
+		if (tsk->normal_prio > prio) {
 			struct sched_param param;
 
 			param.sched_priority = MAX_RT_PRIO-1 - prio;

Re: [RT] rt priority losing

[Esben Nielsen]

On Mon, 24 Jul 2006, Steven Rostedt wrote:

> On Mon, 2006-07-24 at 18:00 +0100, Esben Nielsen wrote:
>> On Mon, 24 Jul 2006, Steven Rostedt wrote:
>>
>>> Ingo or Tglx,
>>>
>>> It has come to my attention that the dynamic hrtimer softirq can lose a
>>> boosted priority.  That is, if a softirq is running while a timeout
>>> happens, and the call back is of lower priority than the currently
>>> running hrtimer softirq, the timer interrupt will still lower the
>>> hrtimer softirq.
>>>
>>> Here's the problem code:
>>>
>>> static void wakeup_softirqd_prio(int softirq, int prio)
>>> {
>>> 	/* Interrupts are disabled: no need to stop preemption */
>>> 	struct task_struct *tsk = __get_cpu_var(ksoftirqd[softirq].tsk);
>>>
>>> 	if (tsk) {
>>> 		if (tsk->normal_prio != prio) {
>>> 			struct sched_param param;
>>>
>>> 			param.sched_priority = MAX_RT_PRIO-1 - prio;
>>> 			setscheduler(tsk, -1, SCHED_FIFO, &param);
>>> 		}
>>> 		if(tsk->state != TASK_RUNNING)
>>> 			wake_up_process(tsk);
>>> 	}
>>> }
>>>
>>>
>>> So, tsk could be softirqd-hrmono and we lower the priority. (only
>>> normal_prio is checked versus prio).
>>>
>>> So this can be a problem, if the softirq function holds a lock of a high
>>> priority task, and is running boosted.  If another timer goes off with a
>>> lower priority, we can lower the priority of the softirqd and lose the
>>> inherited priority that it was running at.
>>
>> There is a check for that inside setscheduler():
>>  	p->prio = rt_mutex_getprio(p);
>
> OK, you are right about this.  The PI chain should not be affected.  But
> this could still be a problem if the softirq was running at a high prio
> for a task when a lower prio callback needs to be made.  It looks like
> timer is removed from the base before the function runs.  So when the
> interrupt looks at the base to determine the priority to set it at, it
> might actually lower the priority of a running hrtimer thread.
>

That is a simple bug which ought to be simple fixable.

Esben
> -- Steve
>
>

Re: [PATCH -rt] Don't let raise_softirq_prio lower the prio (was: [RT] rt priority losing)

[Ingo Molnar]

* Steven Rostedt <rostedt@goodmis.org> wrote:

> I guess the simple fix is not to allow the interrupt to lower the 
> priority. But simple fixes do not always handle all the cases.
> 
> Thomas G., see any side effects with this patch?
> 
> -- Steve
> 
> Signed-off-by: Steven Rostedt <rostedt@goodmis.org>

thanks, applied. Thomas, fine with you too?

	Ingo