On Friday, August 01, 2014 03:43:21 PM Thomas Gleixner wrote:

On Fri, 1 Aug 2014, Rafael J. Wysocki wrote:

quoted

OK, I guess "IRQ_HANDLED from a wakeup interrupt" may be interpreted as
IRQ_HANDLED_PMWAKE.  On the other hand, if that's going to be handled in
handle_irq_event_percpu(), then using a special return code would save us
a brach for IRQ_HANDLED interrupts.  We could convert it to IRQ_HANDLED
immediately then.

We can handle it at the end of the function by calling
note_interrupt() unconditionally do the following there:

      if (suspended) {
      	 if (ret == IRQ_NONE) {
	    if (shared)
	       yell_and_abort_or_resume();
         } else {
	    abort_or_resume();
         }
      }
      if (noirqdebug)
      	 return;

I see.

quoted

OK, I'll take a stab at the IRQF_SHARED thing if you don't mind.

Definitely not :)

quoted

Here's my current understanding of what can be done for IRQF_NO_SUSPEND.

In suspend_device_irqs():

(1) If all actions in the list have the same setting (eg. IRQF_NO_SUSPEND unset),
    keep the current behavior.
(2) If the actions have different settings:
    - Actions with IRQF_NO_SUSPEND set are not modified.
    - Actions with IRQF_NO_SUSPEND unset are switched over to a stub handler.
    - IRQS_SUSPEND_MODE (new flag) is set for the IRQ.

Can we please do that in setup_irq() and let the shared ones always
run through the stub? That keeps suspend/resume_device_irqs() simple.

OK

Rafael