x86: skip migrating IRQF_PER_CPU irq in fixup_irqs
IRQF_PER_CPU marks a irq binding to a specific cpu, and can never be
moved away from that cpu. So it shouldn't be migrated when fixup irqs
to offline a cpu. Xen pvops guest is one source using IRQF_PER_CPU
on a set of virtual interrupts. Previously no error is observed
because Xen event chip silently fails the set_affinity ops, and
logically IRQF_PER_CPU should be recognized here.
Signed-off-by: Fengzhe Zhang <[email protected]>
Signed-off-by: Kevin Tian <[email protected]>
CC: Thomas Gleixner <[email protected]>
CC: Ingo Molnar <[email protected]>
CC: H. Peter Anvin <[email protected]>
CC: Ian Campbell <[email protected]>
CC: Jan Beulich <[email protected]>
--- linux-2.6.39-rc6.orig/arch/x86/kernel/irq.c 2011-05-04 10:59:13.000000000 +0800
+++ linux-2.6.39-rc6/arch/x86/kernel/irq.c 2011-05-06 09:20:25.563963000 +0800
@@ -249,7 +250,7 @@ void fixup_irqs(void)
data = irq_desc_get_irq_data(desc);
affinity = data->affinity;
- if (!irq_has_action(irq) ||
+ if (!irq_has_action(irq) || irqd_is_per_cpu(data) ||
cpumask_subset(affinity, cpu_online_mask)) {
raw_spin_unlock(&desc->lock);
continue;
On Fri, May 06, 2011 at 02:43:36PM +0800, Tian, Kevin wrote:
> x86: skip migrating IRQF_PER_CPU irq in fixup_irqs
>
> IRQF_PER_CPU marks a irq binding to a specific cpu, and can never be
> moved away from that cpu. So it shouldn't be migrated when fixup irqs
> to offline a cpu. Xen pvops guest is one source using IRQF_PER_CPU
^- are called
> on a set of virtual interrupts. Previously no error is observed
^^- was
Which ones? Can you be more specific here of which type of virtual interrupts?
spinlock? timer?
> because Xen event chip silently fails the set_affinity ops, and
> logically IRQF_PER_CPU should be recognized here.
OK, so what if the set_affinity ops was implemented?
>
> Signed-off-by: Fengzhe Zhang <[email protected]>
> Signed-off-by: Kevin Tian <[email protected]>
> CC: Thomas Gleixner <[email protected]>
> CC: Ingo Molnar <[email protected]>
> CC: H. Peter Anvin <[email protected]>
> CC: Ian Campbell <[email protected]>
> CC: Jan Beulich <[email protected]>
>
> --- linux-2.6.39-rc6.orig/arch/x86/kernel/irq.c 2011-05-04 10:59:13.000000000 +0800
> +++ linux-2.6.39-rc6/arch/x86/kernel/irq.c 2011-05-06 09:20:25.563963000 +0800
> @@ -249,7 +250,7 @@ void fixup_irqs(void)
>
> data = irq_desc_get_irq_data(desc);
> affinity = data->affinity;
> - if (!irq_has_action(irq) ||
> + if (!irq_has_action(irq) || irqd_is_per_cpu(data) ||
> cpumask_subset(affinity, cpu_online_mask)) {
> raw_spin_unlock(&desc->lock);
> continue;
> --
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to [email protected]
> More majordomo info at http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at http://www.tux.org/lkml/
> From: Konrad Rzeszutek Wilk [mailto:[email protected]]
> Sent: Friday, May 06, 2011 9:58 PM
>
> On Fri, May 06, 2011 at 02:43:36PM +0800, Tian, Kevin wrote:
> > x86: skip migrating IRQF_PER_CPU irq in fixup_irqs
> >
> > IRQF_PER_CPU marks a irq binding to a specific cpu, and can never be
> > moved away from that cpu. So it shouldn't be migrated when fixup irqs
> > to offline a cpu. Xen pvops guest is one source using IRQF_PER_CPU
> ^- are called
> > on a set of virtual interrupts. Previously no error is observed
> ^^- was
> Which ones? Can you be more specific here of which type of virtual interrupts?
> spinlock? timer?
all of them: spinlock, timer, resched, callfunc, ...
> > because Xen event chip silently fails the set_affinity ops, and
> > logically IRQF_PER_CPU should be recognized here.
>
> OK, so what if the set_affinity ops was implemented?
it was implemented: (drivers/xen/event.c, rebind_irq_to_cpu)
/*
* If this fails, it usually just indicates that we're dealing with a
* virq or IPI channel, which don't actually need to be rebound. Ignore
* it, but don't do the xenlinux-level rebind in that case.
*/
if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_vcpu, &bind_vcpu) >= 0)
bind_evtchn_to_cpu(evtchn, tcpu);
Hypervisor doesn't allow to change affinity for virq and ipi.
Thanks,
Kevin
> >
> > Signed-off-by: Fengzhe Zhang <[email protected]>
> > Signed-off-by: Kevin Tian <[email protected]>
> > CC: Thomas Gleixner <[email protected]>
> > CC: Ingo Molnar <[email protected]>
> > CC: H. Peter Anvin <[email protected]>
> > CC: Ian Campbell <[email protected]>
> > CC: Jan Beulich <[email protected]>
> >
> > --- linux-2.6.39-rc6.orig/arch/x86/kernel/irq.c 2011-05-04
> 10:59:13.000000000 +0800
> > +++ linux-2.6.39-rc6/arch/x86/kernel/irq.c 2011-05-06 09:20:25.563963000
> +0800
> > @@ -249,7 +250,7 @@ void fixup_irqs(void)
> >
> > data = irq_desc_get_irq_data(desc);
> > affinity = data->affinity;
> > - if (!irq_has_action(irq) ||
> > + if (!irq_has_action(irq) || irqd_is_per_cpu(data) ||
> > cpumask_subset(affinity, cpu_online_mask)) {
> > raw_spin_unlock(&desc->lock);
> > continue;
> > --
> > To unsubscribe from this list: send the line "unsubscribe
> > linux-kernel" in the body of a message to [email protected]
> > More majordomo info at http://vger.kernel.org/majordomo-info.html
> > Please read the FAQ at http://www.tux.org/lkml/
On Fri, 6 May 2011, Konrad Rzeszutek Wilk wrote:
> On Fri, May 06, 2011 at 02:43:36PM +0800, Tian, Kevin wrote:
> > because Xen event chip silently fails the set_affinity ops, and
> > logically IRQF_PER_CPU should be recognized here.
>
> OK, so what if the set_affinity ops was implemented?
An interrupt chip which has a set_affinity op should not mark
something per cpu, which implies that the irq CANNOT be moved.
Thanks,
tglx
>>> On 09.05.11 at 14:39, Thomas Gleixner <[email protected]> wrote:
> On Fri, 6 May 2011, Konrad Rzeszutek Wilk wrote:
>> On Fri, May 06, 2011 at 02:43:36PM +0800, Tian, Kevin wrote:
>> > because Xen event chip silently fails the set_affinity ops, and
>> > logically IRQF_PER_CPU should be recognized here.
>>
>> OK, so what if the set_affinity ops was implemented?
>
> An interrupt chip which has a set_affinity op should not mark
> something per cpu, which implies that the irq CANNOT be moved.
Why shouldn't it be possible o use the same "chip" for both per-CPU
and "normal" IRQs?
Jan
> From: Thomas Gleixner [mailto:[email protected]]
> Sent: Monday, May 09, 2011 8:39 PM
>
> On Fri, 6 May 2011, Konrad Rzeszutek Wilk wrote:
> > On Fri, May 06, 2011 at 02:43:36PM +0800, Tian, Kevin wrote:
> > > because Xen event chip silently fails the set_affinity ops, and
> > > logically IRQF_PER_CPU should be recognized here.
> >
> > OK, so what if the set_affinity ops was implemented?
>
> An interrupt chip which has a set_affinity op should not mark something per
> cpu, which implies that the irq CANNOT be moved.
>
If this is the hard requirement, why not throwing out an error when a chip
is registered?
Thanks
Kevin