On Thu, 2007-02-08 at 14:14 +0100, Johannes Berg wrote:

On Wed, 2007-02-07 at 13:45 +0100, Johannes Berg wrote:

quoted

Except for the in-irq count hack I'm happy with this. I still haven't found
where the in-hard-irq count is set to 1 in the down path during suspend or
resume and other platforms do similar things so I'm inclined to leave this.

Um, ok, so the hack breaks platforms that don't have paca, e.g. chrp32.

Also, I finally figured out how the in-hard-irq count happens. The thing
is that when I try to turn off the CPU it actually doesn't really turn
off of course, so it ends up doing NAP and taking timer interrupts...
which goes irq_enter() and we happen to kill it afterwards.

I have two ways of fixing this:
 - just ignore it as we do now
 - insert a "if (cpu_dead) return" into the timer interrupt function

I prefer the latter because then we're guaranteed that whatever the
timer interrupt does we don't modify any state for/by the CPU that isn't
supposed to exist.

Agreed.

Ben.