Skip to contentSkip to conversation

RT-Preempt: IRQ threads running at prio 0 SCHED_OTHER

BUG+PATCH Remy Bohmer 3 participants 3 messages Dec 19, 2006 IRQ

lore
Share & tools
lore
#0 lore raw

Hello Ingo,

I am using your yum-distributed kernel 2.6.19.1-rt15, and unfortunately I experienced very worse latencies.
It turned out that ALL the IRQ threads were all running at prio 0, SCHED_OTHER.

Looking at the current code in kernel/irq/manage.c, the goal was to put them at MAX_RT_PRIO, but the call to sys_sched_setscheduler() fails with EINVAL. I have attached a patch to set them to (MAX_RT_PRIO-1). This works.

Further I believe that each application of the RT-kernel requires a different configuration of these thread-priorities and I prefer to reconfigure these prios from userland during boot. As these threadnames contain whitespaces in its name, they make the shell-scripts unnecessary complex that I use to reconfigure the thread priorities.
So, I would prefer a slight modification of the names: The attached patch also changes the names from [IRQ nn] to [IRQ-nn]. I hope that you agree with me here. (If not, I stick to do this patch each time myself ;-) )

Kind Regards,

Remy Böhmer

kernel/irq/manage.c renamed from kernel/irq/manage.c.orig +3 -3
@@ -692,9 +692,9 @@
 	current->flags |= PF_NOFREEZE | PF_HARDIRQ;
 
 	/*
-	 * Scale irq thread priorities from prio 50 to prio 25
+	 * Scale irq thread priorities to prio 99
 	 */
-	param.sched_priority = MAX_RT_PRIO;
+	param.sched_priority = MAX_RT_PRIO-1;
 
 	sys_sched_setscheduler(current->pid, SCHED_FIFO, &param);
 
@@ -725,7 +725,7 @@
 	if (desc->thread || !ok_to_create_irq_threads)
 		return 0;
 
-	desc->thread = kthread_create(do_irqd, desc, "IRQ %d", irq);
+	desc->thread = kthread_create(do_irqd, desc, "IRQ-%d", irq);
 	if (!desc->thread) {
 		printk(KERN_ERR "irqd: could not create IRQ thread %d!\n", irq);
 		return -ENOMEM;
#1 lore raw
Quoted 10 lines · Remy Bohmer

Hello Ingo,

I am using your yum-distributed kernel 2.6.19.1-rt15, and unfortunately I experienced very worse latencies. It turned out that ALL the IRQ threads were all running at prio 0, SCHED_OTHER.

Looking at the current code in kernel/irq/manage.c, the goal was to put them at MAX_RT_PRIO, but the call to sys_sched_setscheduler() fails with EINVAL. I have attached a patch to set them to (MAX_RT_PRIO-1). This works.

oops - my intention was to set all IRQs and softirqs to SCHED_FIFO prio
50. I have fixed that now in my tree.

prio 99 is pretty extensive and makes it hard to move tasks 'above' hardirq priority, without setting the priority of /every/ IRQ thread. So i picked SCHED_FIFO:50 - at exact half way.

Quoted 6 lines

Further I believe that each application of the RT-kernel requires a different configuration of these thread-priorities and I prefer to reconfigure these prios from userland during boot. As these threadnames contain whitespaces in its name, they make the shell-scripts unnecessary complex that I use to reconfigure the thread priorities.

ok - lets try it. Clark: does this impact the set_kthread_prio utility? I've changed "IRQ 123" to "IRQ-123" to make pidof & friends work better.

Ingo

#2 lore raw

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Quoted 3 lines · Ingo Molnar

ok - lets try it. Clark: does this impact the set_kthread_prio utility? I've changed "IRQ 123" to "IRQ-123" to make pidof & friends work better.

Yes it does, but I believe I can fix that. Most of the logic is in awk and I believe one call to sub() will handle the change.

Clark
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.6 (GNU/Linux)
Comment: Using GnuPG with Fedora - http://enigmail.mozdev.org

iD8DBQFFiFutHyuj/+TTEp0RAjKlAKCaVK8hf4jY6ZUIZ0Ixc56lhSUcHgCgoqT0 HiDbdVXrwY/LYlZN7AFwHOI=
=1YgX
-----END PGP SIGNATURE-----