On 03/19/2011 01:51 PM, Thomas Gleixner wrote:

On Fri, 18 Mar 2011, David Daney wrote:

quoted

--- a/include/linux/irqdesc.h
+++ b/include/linux/irqdesc.h

@@ -178,6 +178,12 @@ static inline int irq_has_action(unsigned int irq)
  	return desc->action != NULL;
  }

+/* Test to see if the irq is currently enabled */
+static inline int irq_desc_is_enabled(struct irq_desc *desc)
+{
+	return desc->depth == 0;
+}

That want's to go into kernel/irq/internal.h

I think I need to use this in my irq_chip.irq_unmask method.

Consider this:


CPU0                   CPU1
handle_level_irq
   lock
   mask
   handle_irq_event
     unlock
     .
     .                  disable_irq
     .
     lock
   unmask
   unlock


I need to know in my .unmask method if the interrupt has been disabled. 
  If it has, I will not re-enable (unmask)it.

quoted

  #ifndef CONFIG_GENERIC_HARDIRQS_NO_COMPAT
  static inline int irq_balancing_disabled(unsigned int irq)
  {

diff --git a/kernel/irq/chip.c b/kernel/irq/chip.c
index c9c0601..40736f7 100644
--- a/kernel/irq/chip.c
+++ b/kernel/irq/chip.c

@@ -689,3 +689,38 @@ void irq_modify_status(unsigned int irq, unsigned long clr, unsigned long set)

  	irq_put_desc_unlock(desc, flags);
  }
+
+void irq_cpu_online(unsigned int irq)

Odd function name. It does not reflect that this is for per cpu
interrupts. So something like irq_xxx_per_cpu_irq(irq)
might be a bit more descriptive.

I am using it for per cpu interrupts, but I didn't want to impose that 
policy on others.

quoted

+{

So that's called on the cpu which goes online, right?

Yes.

I wonder whether we can add any sanity check to verify this.

Though I would not worry too much about it. Calling that from a cpu
which is not going offline should have enough nasty side effects that
it's noticed during development. :)

quoted

+	unsigned long flags;
+	struct irq_chip *chip;
+	struct irq_desc *desc = irq_to_desc(irq);

Needs to check !desc

OK.

quoted

+	raw_spin_lock_irqsave(&desc->lock, flags);
+
+	chip = irq_data_get_irq_chip(&desc->irq_data);
+
+	if (chip&&  chip->irq_cpu_online)
+		chip->irq_cpu_online(&desc->irq_data,
+				     irq_desc_is_enabled(desc));
+
+	raw_spin_unlock_irqrestore(&desc->lock, flags);
+}
+
+void irq_cpu_offline(unsigned int irq)
+{
+	unsigned long flags;
+	struct irq_chip *chip;
+	struct irq_desc *desc = irq_to_desc(irq);

See above.

Style nit: I prefer ordering:

+	struct irq_desc *desc = irq_to_desc(irq);
+	struct irq_chip *chip;
+	unsigned long flags;

For some reason, probably because I'm used to it, that's easier to
parse. But don't worry about that, I'll turn it around before sticking
it into git. :)

Otherwise I'm fine with the approach itself.

Though one question remains: should we just iterate over the irq space
and call the online/offline callbacks when available instead of having
the arch code do the iteration.

That would be good I think, especially for sparse irqs.

In the case of the CPU going offline, the .irq_cpu_offline() may need to 
adjust the affinity so that the irq no longer has affinity for the 
off-lined CPU.

This is something needed even for non-per-cpu interrupts.  Also I would 
need a way to call irq_set_affinity() while holding the desc->lock.

David Daney