2012-06-05 11:24:21

by Alexander Gordeev

Subject: [PATCH 5/8] x86: apic: Try to spread IRQ vectors to different priority levels

When assigning a new vector it is primarially done by adding 8 to the
previously given out vector number. Hence, two consequently allocated
vector numbers would likely fall into the same priority level. Try to
spread vector numbers to different priority levels better by changing
the step from 8 to 16.

Signed-off-by: Alexander Gordeev <[email protected]>
---
arch/x86/kernel/apic/io_apic.c | 7 +++----
1 files changed, 3 insertions(+), 4 deletions(-)

diff --git a/arch/x86/kernel/apic/io_apic.c b/arch/x86/kernel/apic/io_apic.c
index 1afedef..f043d6a 100644
--- a/arch/x86/kernel/apic/io_apic.c
+++ b/arch/x86/kernel/apic/io_apic.c
@@ -1163,7 +1163,7 @@ __assign_irq_vector(int irq, struct irq_cfg *cfg, const struct cpumask *mask)
* 0x80, because int 0x80 is hm, kind of importantish. ;)
*/
static int current_vector = FIRST_EXTERNAL_VECTOR + VECTOR_OFFSET_START;
- static int current_offset = VECTOR_OFFSET_START % 8;
+ static int current_offset = VECTOR_OFFSET_START % 16;
unsigned int old_vector;
int cpu, err;
cpumask_var_t tmp_mask;
@@ -1195,10 +1195,9 @@ __assign_irq_vector(int irq, struct irq_cfg *cfg, const struct cpumask *mask)
vector = current_vector;
offset = current_offset;
next:
- vector += 8;
+ vector += 16;
if (vector >= first_system_vector) {
- /* If out of vectors on large boxen, must share them. */
- offset = (offset + 1) % 8;
+ offset = (offset + 1) % 16;
vector = FIRST_EXTERNAL_VECTOR + offset;
}
if (unlikely(current_vector == vector))
--
1.7.7.6


--
Regards,
Alexander Gordeev
[email protected]

2012-06-06 19:16:29

by Eric W. Biederman

Subject: Re: [PATCH 5/8] x86: apic: Try to spread IRQ vectors to different priority levels

Alexander Gordeev <[email protected]> writes:

> When assigning a new vector it is primarially done by adding 8 to the
> previously given out vector number. Hence, two consequently allocated
> vector numbers would likely fall into the same priority level. Try to
> spread vector numbers to different priority levels better by changing
> the step from 8 to 16.

A weird goal. Given that linux ignores all priority levels internally
we would be better incrementing by 1. The practical problem is that
some older hardware would fail if you had more than two devices per
priority level.

Eric


> Signed-off-by: Alexander Gordeev <[email protected]>
> ---
> arch/x86/kernel/apic/io_apic.c | 7 +++----
> 1 files changed, 3 insertions(+), 4 deletions(-)
>
> diff --git a/arch/x86/kernel/apic/io_apic.c b/arch/x86/kernel/apic/io_apic.c
> index 1afedef..f043d6a 100644
> --- a/arch/x86/kernel/apic/io_apic.c
> +++ b/arch/x86/kernel/apic/io_apic.c
> @@ -1163,7 +1163,7 @@ __assign_irq_vector(int irq, struct irq_cfg *cfg, const struct cpumask *mask)
> * 0x80, because int 0x80 is hm, kind of importantish. ;)
> */
> static int current_vector = FIRST_EXTERNAL_VECTOR + VECTOR_OFFSET_START;
> - static int current_offset = VECTOR_OFFSET_START % 8;
> + static int current_offset = VECTOR_OFFSET_START % 16;
> unsigned int old_vector;
> int cpu, err;
> cpumask_var_t tmp_mask;
> @@ -1195,10 +1195,9 @@ __assign_irq_vector(int irq, struct irq_cfg *cfg, const struct cpumask *mask)
> vector = current_vector;
> offset = current_offset;
> next:
> - vector += 8;
> + vector += 16;
> if (vector >= first_system_vector) {
> - /* If out of vectors on large boxen, must share them. */
> - offset = (offset + 1) % 8;
> + offset = (offset + 1) % 16;
> vector = FIRST_EXTERNAL_VECTOR + offset;
> }
> if (unlikely(current_vector == vector))
> --
> 1.7.7.6

2012-06-06 19:41:11

by Alexander Gordeev

Subject: Re: [PATCH 5/8] x86: apic: Try to spread IRQ vectors to different priority levels

On Wed, Jun 06, 2012 at 12:16:08PM -0700, Eric W. Biederman wrote:
> Alexander Gordeev <[email protected]> writes:
>
> > When assigning a new vector it is primarially done by adding 8 to the
> > previously given out vector number. Hence, two consequently allocated
> > vector numbers would likely fall into the same priority level. Try to
> > spread vector numbers to different priority levels better by changing
> > the step from 8 to 16.
>
> A weird goal. Given that linux ignores all priority levels internally
> we would be better incrementing by 1. The practical problem is that
> some older hardware would fail if you had more than two devices per
> priority level.

Yes, the patch is aimed at those older hardware. For a newer hardware it will
not bring any change.

But also to stop asking myself: why 8, not 1 or 16? :)

> Eric

--
Regards,
Alexander Gordeev
[email protected]