Re: [PATCH] PM: Prevent dpm_prepare() from returning errors unnecessarily

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

On Monday, November 29, 2010, Alan Stern wrote:
> On Sun, 28 Nov 2010, Rafael J. Wysocki wrote:
> 
> > On Sunday, November 28, 2010, Alan Stern wrote:
> > > On Sun, 28 Nov 2010, Rafael J. Wysocki wrote:
> > > 
> > > > From: Rafael J. Wysocki <rjw@sisk.pl>
> > > > 
> > > > Currently dpm_prepare() returns error code if it finds that a device
> > > > being suspended has a pending runtime resume request.  However, it
> > > > should not do that if the checking for wakeup events is not enabled.
> > > 
> > > It doesn't.  The line you changed _does_ check device_may_wakeup().
> > 
> > That's not the point.  The problem is that it shouldn't abort suspend
> > when events_check_enabled is unset.
> 
> Oh, I see.  This is a tricky issue.  Every driver for a device that can
> have wakeup-enabled children needs to worry about the race between
> suspending the device and receiving a wakeup request from a child.  
> For example, in drivers/usb/core/hcd-pci.c, the suspend_common()
> routine goes out of its way to return -EBUSY if device_may_wakeup() is
> true and the controller's root hub has a pending wakeup request.
> 
> How should drivers handle this in general?  Should we make an effort to
> convert them to use the wakeup framework so they they can let the PM
> core take care of these races?

I think so.

We also need to put a pm_check_wakeup_events() check into dpm_suspend() IMO,
so that we abort the suspending of devices as soon as a wakeup event is
reported.

> Do we have to consider similar races during runtime suspend?

Ideally, yes, but I'm not sure if that's generally possible.  IMO, it won't be
a big deal if a parent device is suspended and immediately resumed occasionally
due to a pending wakeup signal from one of its children.  It may be a problem
if that happens too often, though.

> > > > On the other hand, if the checking for wakeup events is enabled, it
> > > > can return error when a wakeup event is detected, regardless of its
> > > > source.
> > > 
> > > Will adding this call to pm_wakeup_event() end up double-counting some 
> > > events?
> > 
> > Yes, it will, if the event has already been reported by the subsystem or driver.
> > 
> > I don't think it's a very big issue and I'm not sure trying to avoid it is
> > worth the effort (we can check if the device's wakeup source object is active
> > and skip reporting the wakeup event in that case, but that doesn't guarantee
> > that the event won't be counted twice anyway).
> 
> I agree that it's not a big issue.  Wakeups reported twice because they 
> occur just before a system sleep won't cause serious accounting 
> problems and probably won't happen very often anyway.  I just wanted to 
> make sure that the issue wasn't being ignored by mistake.

OK

Does it mean you're fine with the patch?

Rafael