On Tue, 2015-09-22 at 11:22 -0400, Alan Stern wrote:

On Tue, 22 Sep 2015, Oliver Neukum wrote:
 

quoted

Cancel, yes, going to low power is a consequence which needn't bother
the power subsystem.

Going to low power needn't involve the power subsystem?  That sounds 
weird.

Think of it like rfkill. It makes sense to suspend an rfkilled device.
It still is the job of the driver to report that its device is idle.

quoted

 You need a callback. If there are spurious
events, the current heuristics will keep devices awake.
You must discard them anyway, as they are spurious. There's no point
in transporting over the bus at all. We can cease IO for input.

quoted

This would create a parallel runtime-PM mechanism which is independent
of the existing one.  Is that really a good idea?

It isn't strictly PM. It helps PM to do a better job, but
conceptually it is independent.

So my next question is: _How_ can this help PM to do a better job?  
That is, what are the mechanisms?

"inhibit" -> driver stops input -> driver sets PM count to zero
-> PM subsystem acts

To go from the first to the second step a callback is needed

One you have already stated: Lack of spurious events will help prevent 
unwanted wakeups (or unwanted failures to go to sleep).

That too. We also save CPU cycles.

But Dmitry made a stronger claim: Inhibiting an input device should 
allow the device to go to low power.  I would like to know how we can 
implement this cleanly.  The most straightforward approach is to use 
runtime PM, but it's not obvious how this can be made to work with the 
current API.

Yes, we can use the current API.
The key is that you think of the mechanism as induced idleness,
not forced suspend. We already have a perfectly working mechanism
for suspending idle devices.

	Regards
		Oliver