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

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

On Saturday 30 May 2009, Kim Kyuwon wrote:
> On Sat, May 30, 2009 at 8:35 AM, Rafael J. Wysocki <rjw@sisk.pl> wrote:
> > On Monday 25 May 2009, Kim Kyuwon wrote:
> >> Rafael J. Wysocki wrote:
> >> > On Saturday 23 May 2009, Kim Kyuwon wrote:
> >> >> On Sat, May 23, 2009 at 7:29 AM, Rafael J. Wysocki <rjw@sisk.pl> wrote:
> >> >>> On Saturday 23 May 2009, Kim Kyuwon wrote:
> >> > [--snip--]
> >> >>>> You changed the really important part of Linux, which may affect most
> >> >>>> processor architectures. I think you should be careful. If some of
> >> >>>> architectures can't take care of it (they can implement
> >> >>>> disable_irq_wake correctly in H/W level, will you revert your changes?
> >> >>> No, the changes are not going to be reverted.  In fact things should have been
> >> >>> done like this already much earlier.
> >> >>>
> >> >>> Now, do you have any particular example of a problem related to these changes
> >> >>> or is it only a theoretical issue?
> >> >> I'd CCing you when I'm sending a mail for this particular example of a example.
> >> >> http://markmail.org/thread/fvt7d62arofon5xx
> >> >
> >> > Well, as I said above, reverting the changes that introduced
> >> > [suspend|resume]_device_irqs() is not an option, becuase it was the only sane
> >> > way to achieve the goal they were added for.  So, we need to fix the wake-up
> >> > problem on your platform with the assumption that
> >> > [suspend|resume]_device_irqs() are going to stay.
> >> >
> >> > For starters, would it be possible to teach the 'disable' hook of your
> >> > platform's interrupt controller not to mask the IRQs that have both
> >> > IRQ_WAKEUP and IRQ_SUSPENDED set?  That apparently would work around the
> >> > wake-up interrupts problem.
> >>
> >> Thank you for considering this issue and spending your time. In order to
> >> make your idea work, we need to add a dummy 'set_wake' hook which
> >> returns always zero. Anyway, IMO, I think your idea is good to work
> >> around this problem. But Kevin Hilman(OMAP PM Maintainer) would make
> >> final decision.
> >>
> >> Buy the way, how can you handle the problem that a few interrupt are
> >> discarded in a small window? I can be sure they are discarded, because I
> >> have debugged defects which generate in sleep/resume state hundreds of
> >> times on ARM Processors(PXA310, S3C6410, OMAP3430). Wake-up interrupts
> >> are generated as soon as arch_suspend_enable_irqs() invoked.
> >
> > Sorry for the delayed response.
> >
> > If the wake-up interrupts are not masked, they will be delivered to the drivers
> > as soon as arch_suspend_enable_irqs() has run.  So, if the drivers are able to
> > handle them at this point (ie. before resume_device_irqs() is called), they
> > won't be lost.
> 
> Thank you for your response!
> 
> Your suspend_device_irqs() disables all IRQs(except timer IRQ) while
> entering suspend. i.e. Before invoking resume_device_irqs() or
> resume_noirq callback, all IRQs(except timer IRQ) is in IRQ_DISABLED
> status. Right?
> But if an IRQ is in  IRQ_DISABLED status, its interrupt handler can be

s/can be/can't be/ (as you noticed in the follow-up message).

> invoked. (As you know, all IRQs with IRQ_DISABLE are not handled in
> handle_level_irq function). Thus, even if the wake-up interrupts are
> not masked, the drivers are not able to handle interrupts, because the
> interrupt handler can't be invoked due to IRQ_DISABLED set by
> suspend_device_irqs().

The solution to that may be to add some code that will clear the
IRQ_DISABLE flag for the wake-up interrupt that caused the wake-up to happen
in the early resume code of the platform.

> > The only problem I see is that the drivers may expect their
> > ->resume_noirq() callbacks to be executed first.
> 
> resume_noirq() callbacks are also invoked after arch_suspend_enable_irqs().

Yes, they are, and that may be a problem, because a driver may expect
resume_noirq() to be called before it can handle interrupts from the device.

Best,
Rafael