Re: [RFC][PATCH] PM / Platform: Remove __weak definitions of runtime

[Grant Likely <grant.likely@secretlab.ca>]

On Thu, Apr 07, 2011 at 08:15:41AM +0200, Rafael J. Wysocki wrote:
> On Thursday, April 07, 2011, Grant Likely wrote:
> > On Thu, Apr 07, 2011 at 07:29:45AM +0200, Rafael J. Wysocki wrote:
> > > On Thursday, April 07, 2011, Kevin Hilman wrote:
> > > > Hi Rafael, Magnus,
> > > > 
> > > > "Rafael J. Wysocki" <rjw@sisk.pl> writes:
> > > > 
> > > > > From: Rafael J. Wysocki <rjw@sisk.pl>
> > > > >
> > > > > Remove the __weak definitions of platform bus type runtime PM
> > > > > callbacks, make platform_dev_pm_ops point to the generic routines
> > > > > as appropriate and allow architectures using platform_dev_pm_ops to
> > > > > replace the runtime PM callbacks in that structure with their own
> > > > > set.
> > > > >
> > > > > Convert architectures providing its own definitions of the platform
> > > > > runtime PM callbacks to use the new mechanism.
> > > > >
> > > > > Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
> > > > 
> > > > I dont't think we should be adding yet another new interface for setting
> > > > platform-specific runtime PM ops.
> > > > 
> > > > We now have 3.  Two existing ones:
> > > > 
> > > > 1) new device power domains (presumably preferred)
> > > > 2) platform_bus_set_pm_ops() (disliked by many)
> > > 
> > > Hmm, I wasn't aware of that one, will have a look.
> > > 
> > > > and now the new one you create here
> > > > 
> > > > 3) platform_set_runtime_pm_ops()
> > > > 
> > > > This new one is basically the same as platform_bus_set_pm_ops(), but
> > > > targetted only at runtime PM ops, and also has all the same problems
> > > > that have been discussed before.  Namely, it overrides the pm ops for
> > > > *every* device on the platform_bus, instead of targetting only specific
> > > > devices.
> > > 
> > > This is not a problem for this particular use case.  We really want to
> > > replace the PM ops for all of the platform devices on that platform.
> > 
> > I strongly doubt that you really want to do that.  platform_devices
> > can appear anywhere in the system, and many of them will end up being
> > entirely outside the SoC, and hence outside of any SoC specific
> > behaviour.
> 
> That is a valid observation, but I still think the way Kevin attempted to
> use the power domain callbacks wasn't the right one for addressing this
> particular issue.
> 
> > What is the use case for overriding every platform_device's PM ops?
> 
> The basic idea, which I agree with, is that we should avoid saving device
> registers when the device is not going to be powered down (i.e. we only
> want to gate its clock).  Since the saving of device registers is generally
> done by device drivers' suspend callbacks, it's better to avoid executing
> those callbacks until we know the devices in question are going to be powered
> down.  That, however, is not known to the default platform bus type
> callbacks that automatically invoke the drivers' callbacks if they exist.
> Hence, it's better to replace the default platform bus type callbacks with
> other ones that only disable the devices' clocks and let power domain
> callbacks (that should know whether or not the devices will be powered down)
> handle the rest.

Okay, I think I understand the scenario.  However, replacing the default
behaviour for the entire platform_bus_type I still think is too large
a hammer.  The default behaviour is to assume worst case behaviour for
platform_devices, which means that it doesn't know what the parent of
the device is going to do after the suspend event.  It could do
anything, so platform_bus_type assumes the worst.  Since
platform_devices can turn up anywhere, I think that is the right
behaviour and any override really needs to be per-device.

That said, I'll let you and Kevin work out what the /correct/ approach
for doing those per-device overrides should be. :-)

g.