Patchwork [1/1] irq: Maintain user set affinity

login
register
mail settings
Submitter Kumar Gala
Date Nov. 4, 2008, 4:31 p.m.
Message ID <Pine.LNX.4.64.0811041030520.25054@blarg.am.freescale.net>
Download mbox | patch
Permalink /patch/7134/
State Superseded
Headers show

Comments

Kumar Gala - Nov. 4, 2008, 4:31 p.m.
From 819ad3ed4660f4238e053728a8b5aa93d22b13d7 Mon Sep 17 00:00:00 2001
From: Kumar Gala <galak@kernel.crashing.org>
Date: Tue, 4 Nov 2008 10:13:53 -0600
Subject: [PATCH 1/1] irq: Maintain user set affinity

This addresses a regression in that if a user specified an affinity
for an interrupt that affinity information would get reset between
a request_irq(), free_irq() and request_irq() for the same irq.

On example of this is bringing up an ethernet interface, taking
it down and bringing it back up again.

Instead of always using the irq_default_affinity we use the
previous setting of the IRQs affinity with the assumption that
at boot the IRQs affinity will be set to irq_default_affinity.

Signed-off-by: Kumar Gala <galak@kernel.crashing.org>
---
 kernel/irq/manage.c |    7 +++++--
 1 files changed, 5 insertions(+), 2 deletions(-)
Andrew Morton - Nov. 6, 2008, 9:35 p.m.
On Tue, 4 Nov 2008 10:31:05 -0600 (CST)
Kumar Gala <galak@kernel.crashing.org> wrote:

> >From 819ad3ed4660f4238e053728a8b5aa93d22b13d7 Mon Sep 17 00:00:00 2001
> From: Kumar Gala <galak@kernel.crashing.org>
> Date: Tue, 4 Nov 2008 10:13:53 -0600
> Subject: [PATCH 1/1] irq: Maintain user set affinity
> 
> This addresses a regression in that if a user specified an affinity
> for an interrupt that affinity information would get reset between
> a request_irq(), free_irq() and request_irq() for the same irq.
> 
> On example of this is bringing up an ethernet interface, taking
> it down and bringing it back up again.
> 
> Instead of always using the irq_default_affinity we use the
> previous setting of the IRQs affinity with the assumption that
> at boot the IRQs affinity will be set to irq_default_affinity.
> 
> Signed-off-by: Kumar Gala <galak@kernel.crashing.org>
> ---
>  kernel/irq/manage.c |    7 +++++--
>  1 files changed, 5 insertions(+), 2 deletions(-)
> 
> diff --git a/kernel/irq/manage.c b/kernel/irq/manage.c
> index c498a1b..ea7d11b 100644
> --- a/kernel/irq/manage.c
> +++ b/kernel/irq/manage.c
> @@ -115,9 +115,12 @@ int irq_select_affinity(unsigned int irq)
>  	if (!irq_can_set_affinity(irq))
>  		return 0;
> 
> -	cpus_and(mask, cpu_online_map, irq_default_affinity);
> -
>  	desc = irq_to_desc(irq);
> +
> +	mask = desc->affinity;
> +
> +	cpus_and(mask, cpu_online_map, mask);
> +
>  	desc->affinity = mask;
>  	desc->chip->set_affinity(irq, mask);
> 

Are you able to identify when this regression occurred?  It isn't
immediately obvious to me.

If the regression is present in 2.6.27 or earlier, do you think we
should backport the fix?  If so, a different patch would be needed due
to the presence of

commit 08678b0841267c1d00d771fe01548d86043d065e
Author: Yinghai Lu <yhlu.kernel@gmail.com>
Date:   Tue Aug 19 20:50:05 2008 -0700

    generic: sparse irqs: use irq_desc() together with dyn_array, instead of irq_desc[]
Kumar Gala - Nov. 6, 2008, 11:23 p.m.
On Nov 6, 2008, at 3:35 PM, Andrew Morton wrote:

> On Tue, 4 Nov 2008 10:31:05 -0600 (CST)
> Kumar Gala <galak@kernel.crashing.org> wrote:
>
>>> From 819ad3ed4660f4238e053728a8b5aa93d22b13d7 Mon Sep 17 00:00:00  
>>> 2001
>> From: Kumar Gala <galak@kernel.crashing.org>
>> Date: Tue, 4 Nov 2008 10:13:53 -0600
>> Subject: [PATCH 1/1] irq: Maintain user set affinity
>>
>> This addresses a regression in that if a user specified an affinity
>> for an interrupt that affinity information would get reset between
>> a request_irq(), free_irq() and request_irq() for the same irq.
>>
>> On example of this is bringing up an ethernet interface, taking
>> it down and bringing it back up again.
>>
>> Instead of always using the irq_default_affinity we use the
>> previous setting of the IRQs affinity with the assumption that
>> at boot the IRQs affinity will be set to irq_default_affinity.
>>
>> Signed-off-by: Kumar Gala <galak@kernel.crashing.org>
>> ---
>> kernel/irq/manage.c |    7 +++++--
>> 1 files changed, 5 insertions(+), 2 deletions(-)
>>
>> diff --git a/kernel/irq/manage.c b/kernel/irq/manage.c
>> index c498a1b..ea7d11b 100644
>> --- a/kernel/irq/manage.c
>> +++ b/kernel/irq/manage.c
>> @@ -115,9 +115,12 @@ int irq_select_affinity(unsigned int irq)
>> 	if (!irq_can_set_affinity(irq))
>> 		return 0;
>>
>> -	cpus_and(mask, cpu_online_map, irq_default_affinity);
>> -
>> 	desc = irq_to_desc(irq);
>> +
>> +	mask = desc->affinity;
>> +
>> +	cpus_and(mask, cpu_online_map, mask);
>> +
>> 	desc->affinity = mask;
>> 	desc->chip->set_affinity(irq, mask);
>>
>
> Are you able to identify when this regression occurred?  It isn't
> immediately obvious to me.
>
> If the regression is present in 2.6.27 or earlier, do you think we
> should backport the fix?  If so, a different patch would be needed due
> to the presence of
>
> commit 08678b0841267c1d00d771fe01548d86043d065e
> Author: Yinghai Lu <yhlu.kernel@gmail.com>
> Date:   Tue Aug 19 20:50:05 2008 -0700
>
>    generic: sparse irqs: use irq_desc() together with dyn_array,  
> instead of irq_desc[]

It didn't creep in during 2.6.27.  I will back port a patch to  
2.6.27.  Thomas has provided me a better patch that handles a few  
critical issues that mine doesn't.  I'll be testing it this evening  
and will backport to 2.6.27 if it works out.

- k
Yinghai Lu - Nov. 8, 2008, 10:12 p.m.
On Tue, Nov 4, 2008 at 8:31 AM, Kumar Gala <galak@kernel.crashing.org> wrote:
> From 819ad3ed4660f4238e053728a8b5aa93d22b13d7 Mon Sep 17 00:00:00 2001
> From: Kumar Gala <galak@kernel.crashing.org>
> Date: Tue, 4 Nov 2008 10:13:53 -0600
> Subject: [PATCH 1/1] irq: Maintain user set affinity
>
> This addresses a regression in that if a user specified an affinity
> for an interrupt that affinity information would get reset between
> a request_irq(), free_irq() and request_irq() for the same irq.
>
> On example of this is bringing up an ethernet interface, taking
> it down and bringing it back up again.
>
> Instead of always using the irq_default_affinity we use the
> previous setting of the IRQs affinity with the assumption that
> at boot the IRQs affinity will be set to irq_default_affinity.

why ?, the default_affinity is expected behavior...

for example: when MSI is used, second driver could get same irq if
first driver is unloaded.

YH
Yinghai Lu - Nov. 8, 2008, 10:13 p.m.
On Thu, Nov 6, 2008 at 1:35 PM, Andrew Morton <akpm@linux-foundation.org> wrote:
> On Tue, 4 Nov 2008 10:31:05 -0600 (CST)
> Kumar Gala <galak@kernel.crashing.org> wrote:
>
>> >From 819ad3ed4660f4238e053728a8b5aa93d22b13d7 Mon Sep 17 00:00:00 2001
>> From: Kumar Gala <galak@kernel.crashing.org>
>> Date: Tue, 4 Nov 2008 10:13:53 -0600
>> Subject: [PATCH 1/1] irq: Maintain user set affinity
>>
>> This addresses a regression in that if a user specified an affinity
>> for an interrupt that affinity information would get reset between
>> a request_irq(), free_irq() and request_irq() for the same irq.
>>
...
>
> Are you able to identify when this regression occurred?  It isn't
> immediately obvious to me.
>
> If the regression is present in 2.6.27 or earlier, do you think we
> should backport the fix?  If so, a different patch would be needed due
> to the presence of

it seems that is not a regression...

YH
Benjamin Herrenschmidt - Nov. 8, 2008, 10:17 p.m.
On Sat, 2008-11-08 at 14:12 -0800, Yinghai Lu wrote:
> On Tue, Nov 4, 2008 at 8:31 AM, Kumar Gala <galak@kernel.crashing.org> wrote:
> > From 819ad3ed4660f4238e053728a8b5aa93d22b13d7 Mon Sep 17 00:00:00 2001
> > From: Kumar Gala <galak@kernel.crashing.org>
> > Date: Tue, 4 Nov 2008 10:13:53 -0600
> > Subject: [PATCH 1/1] irq: Maintain user set affinity
> >
> > This addresses a regression in that if a user specified an affinity
> > for an interrupt that affinity information would get reset between
> > a request_irq(), free_irq() and request_irq() for the same irq.
> >
> > On example of this is bringing up an ethernet interface, taking
> > it down and bringing it back up again.
> >
> > Instead of always using the irq_default_affinity we use the
> > previous setting of the IRQs affinity with the assumption that
> > at boot the IRQs affinity will be set to irq_default_affinity.
> 
> why ?, the default_affinity is expected behavior...
> 
> for example: when MSI is used, second driver could get same irq if
> first driver is unloaded.

The default affinity should be set when the irq is allocated / assigned
to a device, not when it's requested.

Ben.
Yinghai Lu - Nov. 9, 2008, 12:01 a.m.
On Sat, Nov 8, 2008 at 2:17 PM, Benjamin Herrenschmidt
<benh@kernel.crashing.org> wrote:
> On Sat, 2008-11-08 at 14:12 -0800, Yinghai Lu wrote:
>> On Tue, Nov 4, 2008 at 8:31 AM, Kumar Gala <galak@kernel.crashing.org> wrote:
>> > From 819ad3ed4660f4238e053728a8b5aa93d22b13d7 Mon Sep 17 00:00:00 2001
>> > From: Kumar Gala <galak@kernel.crashing.org>
>> > Date: Tue, 4 Nov 2008 10:13:53 -0600
>> > Subject: [PATCH 1/1] irq: Maintain user set affinity
>> >
>> > This addresses a regression in that if a user specified an affinity
>> > for an interrupt that affinity information would get reset between
>> > a request_irq(), free_irq() and request_irq() for the same irq.
>> >
>> > On example of this is bringing up an ethernet interface, taking
>> > it down and bringing it back up again.
>> >
>> > Instead of always using the irq_default_affinity we use the
>> > previous setting of the IRQs affinity with the assumption that
>> > at boot the IRQs affinity will be set to irq_default_affinity.
>>
>> why ?, the default_affinity is expected behavior...
>>
>> for example: when MSI is used, second driver could get same irq if
>> first driver is unloaded.
>
> The default affinity should be set when the irq is allocated / assigned
> to a device, not when it's requested.

ok, you are right.

for MSI, affinity is set two times.

also in dynamic_irq_init is setting with
#ifdef CONFIG_SMP
        cpus_setall(desc->affinity);
#endif

it seems should use irq_default_affinity instead of CPU_MASK_ALL directly there.

YH
Benjamin Herrenschmidt - Nov. 9, 2008, 2:17 a.m.
> for MSI, affinity is set two times.
> 
> also in dynamic_irq_init is setting with
> #ifdef CONFIG_SMP
>         cpus_setall(desc->affinity);
> #endif
> 
> it seems should use irq_default_affinity instead of CPU_MASK_ALL directly there.

That sounds more like it.

Cheers,
Ben.

Patch

diff --git a/kernel/irq/manage.c b/kernel/irq/manage.c
index c498a1b..ea7d11b 100644
--- a/kernel/irq/manage.c
+++ b/kernel/irq/manage.c
@@ -115,9 +115,12 @@  int irq_select_affinity(unsigned int irq)
 	if (!irq_can_set_affinity(irq))
 		return 0;

-	cpus_and(mask, cpu_online_map, irq_default_affinity);
-
 	desc = irq_to_desc(irq);
+
+	mask = desc->affinity;
+
+	cpus_and(mask, cpu_online_map, mask);
+
 	desc->affinity = mask;
 	desc->chip->set_affinity(irq, mask);