The generic request_irq/setup_irq code should support the SA_PERCPU_IRQ flag.
Not sure why it was left out, but this patch adds that support.
Signed-off-by: Dimitri Sivanich <[email protected]>
Index: linux/kernel/irq/manage.c
===================================================================
--- linux.orig/kernel/irq/manage.c 2006-03-16 14:05:27.957927445 -0600
+++ linux/kernel/irq/manage.c 2006-03-16 14:06:02.283661867 -0600
@@ -201,6 +201,9 @@ int setup_irq(unsigned int irq, struct i
* The following block of code has to be executed atomically
*/
spin_lock_irqsave(&desc->lock,flags);
+ if (new->flags & SA_PERCPU_IRQ) {
+ desc->status |= IRQ_PER_CPU;
+ }
p = &desc->action;
if ((old = *p) != NULL) {
/* Can't share interrupts unless both agree to */
On Thu, Mar 16, 2006 at 06:31:14PM -0600, Dimitri Sivanich wrote:
> The generic request_irq/setup_irq code should support the SA_PERCPU_IRQ flag.
>
> Not sure why it was left out, but this patch adds that support.
>
> Signed-off-by: Dimitri Sivanich <[email protected]>
>
> Index: linux/kernel/irq/manage.c
> ===================================================================
> --- linux.orig/kernel/irq/manage.c 2006-03-16 14:05:27.957927445 -0600
> +++ linux/kernel/irq/manage.c 2006-03-16 14:06:02.283661867 -0600
> @@ -201,6 +201,9 @@ int setup_irq(unsigned int irq, struct i
> * The following block of code has to be executed atomically
> */
> spin_lock_irqsave(&desc->lock,flags);
> + if (new->flags & SA_PERCPU_IRQ) {
> + desc->status |= IRQ_PER_CPU;
> + }
looks good, although the braces aren't really needed for single-line
conditionals.
Dimitri Sivanich <[email protected]> wrote:
>
> he generic request_irq/setup_irq code should support the SA_PERCPU_IRQ flag.
>
> Not sure why it was left out, but this patch adds that support.
>
> Signed-off-by: Dimitri Sivanich <[email protected]>
>
> Index: linux/kernel/irq/manage.c
> ===================================================================
> --- linux.orig/kernel/irq/manage.c 2006-03-16 14:05:27.957927445 -0600
> +++ linux/kernel/irq/manage.c 2006-03-16 14:06:02.283661867 -0600
> @@ -201,6 +201,9 @@ int setup_irq(unsigned int irq, struct i
> * The following block of code has to be executed atomically
> */
> spin_lock_irqsave(&desc->lock,flags);
> + if (new->flags & SA_PERCPU_IRQ) {
> + desc->status |= IRQ_PER_CPU;
> + }
> p = &desc->action;
> if ((old = *p) != NULL) {
> /* Can't share interrupts unless both agree to */
Yes, we're presently going in and needlessly taking desc->lock on every
cpu-local interrupt. However it only appears to affect mmtimer.c at
present.
Shouldn't we also have a check in there, to make sure that all the other
handlers off this IRQ agree about the per-cpuness of this interrupt?
Your patch broke the x86 build, btw..
From: Dimitri Sivanich <[email protected]>
The generic request_irq/setup_irq code should support the SA_PERCPU_IRQ flag.
Also, check that all handlers agree to do per-cpu IRQs.
Signed-off-by: Dimitri Sivanich <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
---
kernel/irq/manage.c | 23 +++++++++++++++++++----
1 files changed, 19 insertions(+), 4 deletions(-)
diff -puN kernel/irq/manage.c~generic-irq-add-sa_percpu_irq-flag-support kernel/irq/manage.c
--- 25/kernel/irq/manage.c~generic-irq-add-sa_percpu_irq-flag-support Fri Mar 17 15:06:44 2006
+++ 25-akpm/kernel/irq/manage.c Fri Mar 17 15:23:02 2006
@@ -204,10 +204,14 @@ int setup_irq(unsigned int irq, struct i
p = &desc->action;
if ((old = *p) != NULL) {
/* Can't share interrupts unless both agree to */
- if (!(old->flags & new->flags & SA_SHIRQ)) {
- spin_unlock_irqrestore(&desc->lock,flags);
- return -EBUSY;
- }
+ if (!(old->flags & new->flags & SA_SHIRQ))
+ goto mismatch;
+
+#ifdef ARCH_HAS_IRQ_PER_CPU
+ /* All handlers must agree on per-cpuness */
+ if ((old->flags & IRQ_PER_CPU) != (new->flags & IRQ_PER_CPU))
+ goto mismatch;
+#endif
/* add new interrupt at end of irq queue */
do {
@@ -219,6 +223,11 @@ int setup_irq(unsigned int irq, struct i
*p = new;
+#ifdef ARCH_HAS_IRQ_PER_CPU
+ if (new->flags & SA_PERCPU_IRQ)
+ desc->status |= IRQ_PER_CPU;
+#endif
+
if (!shared) {
desc->depth = 0;
desc->status &= ~(IRQ_DISABLED | IRQ_AUTODETECT |
@@ -236,6 +245,12 @@ int setup_irq(unsigned int irq, struct i
register_handler_proc(irq, new);
return 0;
+
+mismatch:
+ spin_unlock_irqrestore(&desc->lock, flags);
+ printk(KERN_ERR "%s: irq handler mismatch\n", __FUNCTION__);
+ dump_stack();
+ return -EBUSY;
}
#ifdef CONFIG_DEBUG_SHIRQ
_
Andrew Morton <[email protected]> wrote:
>
> +#ifdef ARCH_HAS_IRQ_PER_CPU
> + if (new->flags & SA_PERCPU_IRQ)
> + desc->status |= IRQ_PER_CPU;
> +#endif
OK, five architectures define ARCH_HAS_IRQ_PER_CPU but only one of them
defines SA_PERCPU_IRQ. Giving up.
On Sat, Mar 18, 2006 at 01:49:00AM -0800, Andrew Morton wrote:
> Andrew Morton <[email protected]> wrote:
> >
> > +#ifdef ARCH_HAS_IRQ_PER_CPU
> > + if (new->flags & SA_PERCPU_IRQ)
> > + desc->status |= IRQ_PER_CPU;
> > +#endif
>
> OK, five architectures define ARCH_HAS_IRQ_PER_CPU but only one of them
> defines SA_PERCPU_IRQ. Giving up.
Could we do the following (at least for now)?:
+#if defined(ARCH_HAS_IRQ_PER_CPU) && defined(SA_PERCPU_IRQ)
+ if (new->flags & SA_PERCPU_IRQ)
+ desc->status |= IRQ_PER_CPU;
+#endif
On Mon, Mar 20, 2006 at 08:17:47AM -0600, Dimitri Sivanich wrote:
> On Sat, Mar 18, 2006 at 01:49:00AM -0800, Andrew Morton wrote:
> > Andrew Morton <[email protected]> wrote:
> > >
> > > +#ifdef ARCH_HAS_IRQ_PER_CPU
> > > + if (new->flags & SA_PERCPU_IRQ)
> > > + desc->status |= IRQ_PER_CPU;
> > > +#endif
> >
> > OK, five architectures define ARCH_HAS_IRQ_PER_CPU but only one of them
> > defines SA_PERCPU_IRQ. Giving up.
>
> Could we do the following (at least for now)?:
>
> +#if defined(ARCH_HAS_IRQ_PER_CPU) && defined(SA_PERCPU_IRQ)
> + if (new->flags & SA_PERCPU_IRQ)
> + desc->status |= IRQ_PER_CPU;
> +#endif
Why not just
#ifdef SA_PERCPU_IRQ
if (new->flags & SA_PERCPU_IRQ)
desc->status |= IRQ_PER_CPU;
#endif
On Mon, Mar 20, 2006 at 07:22:07AM -0700, Matthew Wilcox wrote:
> On Mon, Mar 20, 2006 at 08:17:47AM -0600, Dimitri Sivanich wrote:
> > On Sat, Mar 18, 2006 at 01:49:00AM -0800, Andrew Morton wrote:
> > > Andrew Morton <[email protected]> wrote:
> > > >
> > > > +#ifdef ARCH_HAS_IRQ_PER_CPU
> > > > + if (new->flags & SA_PERCPU_IRQ)
> > > > + desc->status |= IRQ_PER_CPU;
> > > > +#endif
> > >
> > > OK, five architectures define ARCH_HAS_IRQ_PER_CPU but only one of them
> > > defines SA_PERCPU_IRQ. Giving up.
> >
> > Could we do the following (at least for now)?:
> >
> > +#if defined(ARCH_HAS_IRQ_PER_CPU) && defined(SA_PERCPU_IRQ)
> > + if (new->flags & SA_PERCPU_IRQ)
> > + desc->status |= IRQ_PER_CPU;
> > +#endif
>
> Why not just
>
> #ifdef SA_PERCPU_IRQ
> if (new->flags & SA_PERCPU_IRQ)
> desc->status |= IRQ_PER_CPU;
> #endif
Fine with me. I was simply maintaining the ARCH_HAS_IRQ_PER_CPU convention.