Re: [PATCH] PCI / PM: Make PCIe PME interrupts wake up from "freeze" sleep state

["Rafael J. Wysocki" <rjw@rjwysocki.net>]

[-- Attachment #1: Type: text/plain, Size: 3499 bytes --]

On Thursday, July 24, 2014 03:42:41 PM Peter Zijlstra wrote:
> On Wed, Jul 23, 2014 at 10:46:26PM +0200, Rafael J. Wysocki wrote:
> > From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> > 
> > The "freeze" sleep state, also known as suspend-to-idle, is entered
> > without taking nonboot CPUs offline, right after devices have been
> > suspended.  It works by waiting for at least one wakeup source object
> > to become "active" as a result of handling a hardware interrupt.
> > 
> > Of course, interrupts supposed to be able to wake up the system from
> > suspend-to-idle cannot be disabled by suspend_device_irqs() and their
> > interrupt handlers must be able to cope with interrupts coming after
> > all devices have been suspended.  In that case, they only need to
> > call __pm_wakeup_event() for a single wakeup source object without
> > trying to access hardware (that will be resumed later as part of
> > the subsequent system resume).
> > 
> > Make PCIe PME interrupts work this way.
> > 
> > Register an additional wakeup source object for each PCIe PME
> > service device.  That object will be used to generate wakeups from
> > suspend-to-idle.
> > 
> > Add IRQF_NO_SUSPEND to PME interrupt flags.  This will make
> > suspend_device_irqs() to ignore PME interrupts, but that's OK,
> > because the PME interrupt handler is suspend-aware anyway and
> > can cope with interrupts coming during system suspend-resume.
> > 
> > For each PCIe port with PME service during the "prepare" phase of
> > system suspend walk the bus below it and see if any devices on that
> > bus are configured for wakeup.  If so, mark the port as one that can
> > be used for system wakeup signaling and handle it differenty going
> > forward.
> > 
> > Namely, while suspending its PME service, do not disable the PME
> > interrupt, but only set a "suspended" flag for the PME service to
> > make the interrupt handler behave in a special way, which is to call
> > __pm_wakeup_event() with the service's wakeup source object as the
> > first argument whenever the interrupt is triggered.
> > 
> > The "suspended" flag is cleared while resuming the PME service and
> > the "wakeup" flag is cleared at the "complete" stage of system
> > resume.
> > 
> > This change allows Wake-on-LAN to be used for wakeup from
> > suspend-to-idle on my MSI Wind tesbed netbook.
> > 
> > Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> > ---
> 
> > -	ret = request_irq(srv->irq, pcie_pme_irq, IRQF_SHARED, "PCIe PME", srv);
> > +	ret = request_irq(srv->irq, pcie_pme_irq, IRQF_SHARED | IRQF_NO_SUSPEND,
> > +			  "PCIe PME", srv);
> 
> So with this patch on:
> 
>   http://marc.info/?l=linux-kernel&m=140620918218199

ACK for this one (already sent in that thread).

> This will not work on my machine, because aerdrv is requesting the same
> irq.

AER is PCIe error reporting service.

> Now I've not a f'cking clue what aerdrv is, and whether it too wants
> NO_SUSPEND on or not.
> 
> But if I make it also request NO_SUSPEND it all starts working.

I think we should make all PCIe port services (there may be hotplug too)
use IRQF_NO_SUSPEND which will solve this particular problem, but there's
more to that (for example, ACPI SCI which in theory may be shared).

I guess it's time to revisit that thing, but that's a separate issue, of
course.

-- 
I speak only for myself.
Rafael J. Wysocki, Intel Open Source Technology Center.

[-- Attachment #2: This is a digitally signed message part. --]
[-- Type: application/pgp-signature, Size: 836 bytes --]