lkml.org 
[lkml]   [2015]   [Mar]   [5]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRe: [PATCH v2 5/6] watchdog: at91sam9: request the irq with IRQF_NO_SUSPEND
Date
On Thursday, March 05, 2015 04:32:27 PM Mark Rutland wrote:
> Hi Rafael,
>
> > enable_irq_wake() has no effect on IRQF_NO_SUSPEND interrupts, so if the
> > driver uses IRQF_NO_SUSPEND, it does not need to use enable_irq_wake()
> > in addition to that.
>
> That's not generally true -- certainly not for irq_chips without the
> IRQCHIP_SKIP_SET_WAKE flag.
>
> Consider systems where the suspended state results in power to the CPU
> being cut, and we rely on an external piece of logic attached to the
> irq_chip to detect wakeup IRQs and restore power.
>
> In those cases irq_chip::irq_set_wake() must be called to ensure that
> the wakeup logic is configured. If the wakeup logic is not configured to
> look out for an IRQ, then when the IRQ is asserted by a device the
> wakeup logic may not trigger. Thus the CPU power never gets restored, so
> the CPU cannot handle the interrupt.
>
> This is handled in enable_irq_wake() -- either the chip has the
> IRQCHIP_SKIP_SET_WAKE flag set or chip->irq_set_wake() must succeed. If
> neither is true enable_irq_wake() will return an error code to indicate
> we can't use the IRQ for wakeup.

Right. I forgot about that part.

> The request_irq path never results in a call to chip->irq_set_wake(),
> even with the IRQF_NO_SUSPEND flag. So requesting an irq with
> IRQF_NO_SUSPEND does not guarantee wakeup; it only guarantees that the
> CPU can take the interrupt _around_ the suspended state, not necessarily
> while _in_ the suspended state.

Right. "Suspended state" meaning full suspend here I suppose?

> > Drivers using IRQF_COND_SUSPEND generally should use enable_irq_wake() too
> > in case they end up in a situation without sharing a NO_SUSPEND interrupt, in
> > which case their interrupt handlers won't be called after suspend_device_irqs(),
> > so they need to rely on the core to do the wakeup.
> >
> > > I agree that if problematic, it's an existing bug. Given Boris's
> > > comments in the other thread this may just a minor semantic issue w.r.t.
> > > IRQF_NO_SUSPEND vs IRQF_COND_SUSPEND.
> >
> > It depends on whether or not the watchdog's interrupt handler has to be
> > called immediately after receiving an interrupt (IRQF_NO_SUSPEND is
> > better then) or it can be deferred till the resume_device_irqs() time.
>
> We seem to be conflating some related properties:
>
> [a] The IRQ will be left unmasked.
> [b] The IRQ will be handled immediately when taken.
> [c] The IRQ will wake the system from suspend.
>
> Requesting an IRQ with IRQF_NO_SUSPEND guarantees [a,b], but does not
> guarantee [c].

That's correct. IRQF_NO_SUSPEND does not guarantee that interrupts from
that IRQ will have any effect after arch_suspend_disable_irqs() in
suspend_enter().

> A successful enable_irq_wake() on an IRQ guarantees [a,c], but usually
> does not give [b] unless the IRQ was requested with IRQF_COND_SUSPEND
> and happens to be shared with an IRQF_NO_SUSPEND user.

That's correct too.

> It sounds like for this kind of watchdog device we want [a,b,c], even if
> the IRQ is not shared with an IRQF_NO_SUSPEND user.

We can't guarantee that, though. arch_suspend_disable_irqs() disables
interrupts on the last working CPU and it won't get any. It may be
brought out of a low-power state by a pending interrupt, but it won't act
upon that interrupt immediately anyway, only after the arch_suspend_enable_irqs()
in suspend_enter(). But then it might as well be deferred until after
resume_device_irqs().

Rafael



\
 
 \ /
  Last update: 2015-03-06 01:21    [W:0.108 / U:0.256 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site