lkml.org 
[lkml]   [2014]   [Jul]   [25]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [RFC][PATCH] irq: Rework IRQF_NO_SUSPENDED
On Fri, 25 Jul 2014, Rafael J. Wysocki wrote:
> On Thursday, July 24, 2014 11:26:20 PM Peter Zijlstra wrote:
> > Subject: irq: Rework IRQF_NO_SUSPENDED
> > From: Peter Zijlstra <peterz@infradead.org>
> > Date: Thu Jul 24 22:34:50 CEST 2014
> >
> > Typically when devices are suspended they're quiesced such that they
> > will not generate any further interrupts.
> >
> > However some devices should still generate interrupts, even when
> > suspended, typically used to wake the machine back up.
> >
> > Such logic should ideally be contained inside each driver, if it can
> > generate interrupts when suspended, it knows about this and the
> > interrupt handler can deal with it.
> >
> > Except of course for shared interrupts, when such a wakeup device is
> > sharing an interrupt line with a device that does not expect
> > interrupts while suspended things can go funny.
> >
> > This is where IRQF_NO_SUSPEND comes in, the idea is that drivers that
> > have the capability to wake the machine set IRQF_NO_SUSPEND and their
> > handler will still be called, even when the IRQ subsystem is formally
> > suspended. Handlers without IRQF_NO_SUSPEND will not be called.
> >
> > This replaced the prior implementation of IRQF_NO_SUSPEND which had
> > a number of fatal issues in that it didn't actually work for the
> > shared case, exactly the case it should be helping.
> >
> > There is still enable_irq_wake()/IRQD_WAKEUP_STATE that tries to serve
> > a similar purpose but is equially wrecked for shared interrupts,
> > ideally this would be removed.
>
> Let me comment about this particular thing.
>
> I had a discussion with Dmitry about that and his argument was that
> enable_irq_wake() should imply IRQF_NO_SUSPEND, because drivers that
> set up interrupts for system wakeup should expect those interrupts to
> trigger at any time, including system suspend. Hence the patch that
> added the IRQD_WAKEUP_STATE check to __disable_irq().
>
> However, in the face of the problem that is being addressed here I'm
> not really sure that this argument is valid, because if the driver
> calling enable_irq_wake() is sharing the IRQ with another one, the
> other driver may not actually know that the IRQ will be a wakeup one
> and still may not expect interrupts to come to it during system
> suspend/resume.
>
> Yes, drivers using enable_irq_wake() will likely want IRQF_NO_SUSPEND to
> be set for their irqactions, but that should not imply "no suspend" for
> all irqactions sharing the same desc. So I guess it may be better to go
> forth and use a global "interrupts suspended" state variable instead of the
> IRQS_SUSPENDED flag for each desc and throw away the IRQD_WAKEUP_STATE
> check from suspend_device_irqs() entirely.

How should that global state work?

Thanks,

tglx


\
 
 \ /
  Last update: 2014-07-25 12:41    [W:0.275 / U:0.232 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site