Re: [PATCH/RFC] MMC: remove unbalanced pm_runtime_suspend()

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

On Thursday, April 21, 2011, Alan Stern wrote:
> On Thu, 21 Apr 2011, Rafael J. Wysocki wrote:
> 
> > > > What about making a rule that it is invalid to schedule a future suspend
> > > > or queue a resume request of a device whose driver is being removed?
> > > > 
> > > > Arguably, we can't prevent people from shooting themselves in the foot this
> > > > way or another and I'm not sure if this particular case is worth additional
> > > > handling.
> > > 
> > > After thinking about this, I tend to agree.  The synchronization 
> > > issues, combined with the unknown needs of the driver, make this very 
> > > difficult to handle in the PM core.
> > > 
> > > Here's another possible approach: If a driver wants to leave its device 
> > > in a powered-down state after unbinding then it can invoke its own 
> > > runtime_suspend callback directly, in the following way:
> > > 
> > > 	... unregister all child devices below dev ...
> > > 	pm_runtime_disable(dev);
> > > 	if (dev->power.runtime_status != RPM_SUSPENDED) {
> > > 		pm_set_suspended(dev);
> > > 		my_runtime_suspend_callback(dev);
> > > 	}
> > 
> > I think this would work too, but then possibly many drivers would have to
> > do the same thing in their "remove" routines.
> > 
> > > There may be issues regarding coordination with the subsystem or the
> > > power domain; at the moment it's not clear what should be done.  Maybe
> > > the runtime-PM core should include an API for directly invoking the
> > > appropriate callbacks.
> > 
> > If we choose this approach, then yes, we should provide a suitable API, but
> > I'm still thinking it would be simpler to move the pm_runtime_put_sync() before driver_sysfs_remove() and make the rule as I said previously. :-)
> 
> The problem is synchronization.  At what point is the driver supposed 
> to stop queuing runtime PM requests?  It would have to be sometime 
> before the pm_runtime_barrier() call.  How is the driver supposed to 
> know when that point is reached?  The remove routine isn't called until 
> later.

Executing the driver's callback is not an ideal solution either, because
it simply may be insufficient (it may be necessary to execute the power
domain and/or subsystem callbacks, pretty much what rpm_suspend() does,
but without taking the usage counter into consideration).

Moreover,  if we want the driver's ->remove() to do the cleanup anyway,
there's not much point in doing any cleanup before in the core.  Also,
there's a little problem that the bus ->remove() is called before the
driver's ->remove(), so it may not be entirely possible to power down
the device when the driver's ->remove() is called already.

I think the current code is better than any of the alternatives considered
so far.

Rafael