Re: [PATCH] PM: suspend_device_irqs(): don't disable wakeup IRQs

["Rafael J. Wysocki" <rjw@sisk.pl>]

On Wednesday 06 May 2009, Kevin Hilman wrote:
> Kevin Hilman <khilman@deeprootsystems.com> writes:
> 
> > "Rafael J. Wysocki" <rjw@sisk.pl> writes:
> >
> >> On Wednesday 06 May 2009, Kevin Hilman wrote:
> 
> [...]
> 
> >>> 
> >>> [...]
> >>> 
> >>> >>>
> >>> >>> If this fixes some bug then please provide a description of that bug?
> >>> >>
> >>> >> The bug is that on TI OMAP, interrupts that are used for wakeup events
> >>> >> are disabled by this code causing the system to no longer wake up.
> >>> >
> >>> > What do you do if the interrupt triggers right after your driver has
> >>> > returned from its late suspend hook?  
> >>> 
> >>> If it's a wakeup IRQ, I assume you want it to prevent suspend.
> >>> 
> >>> But I don't see how that can happen in the current code. IIUC, by the
> >>> time your late suspend hook is run, your device IRQ is already
> >>> disabled, so it won't trigger an interrupt that will be caught by
> >>> check_wakeup_irqs() anyways.
> >>
> >> My understanding of __disable_irq() was that it didn't actually disable the
> >> IRQ at the hardware level, allowing the CPU to actually receive the interrupt
> >> and acknowledge it, but preventing the device driver for receiving it.  
> >
> >> Does it work differently on the affected systems?
> >
> > Yes.
> >
> > __disable_irq() calls the irq_chip's disable method which is platform
> > specific.  On OMAP, this masks the IRQ at the hardware level
> > preventing the CPU from seeing the interrupt.
> 
> Looking at x86, the i8259 disable hook also seems to mask the IRQ at
> the PIC level.
> 
> The various IO-APIC irq_chips do not have a disable hook so the
> __disable_irq() here is a NOP.

Except that it sets IRQ_DISABLED.

All right there.

We can either avoid disabling wake-up interrupts, in which case we should
drop check_wakeup_irqs() IMO, or rework things so that check_wakeup_irqs() will
catch them.  Doing both doesn't seem to make sense to me.

Which one would be the right approach, then?

Rafael