On Wed, Dec 09, 2009 at 09:31:00AM +0200, Artem Bityutskiy wrote: > On Tue, 2009-12-08 at 09:42 -0800, Dmitry Torokhov wrote: > > > > I also see that gpio-keys is quite different in the sence that it can > > > > shut off buttons selectively. I fact, at the moment every button can be > > > > considered a separate device... But that would be too much overhead. > > > > > > > > They could probably split the keys into 2 groups (critical that should > > > > be always active) and not critical, that could be shut off, but I think > > > > they want teh flexibility of controlling this at runtime instead of > > > > doing it in board data. > > > > > > I suggested including this into the "abstract input device" model, but > > > you refuse this. But I still think it is a good idea. > > > > > > Indeed, if we look at an input device as at something abstract which has > > > many keys, why we cannot assume that separate keys can be > > > enabled/disabled? Just imagine you have a very advanced keybord :-) And > > > we simply implement an ioctl which enables/disables a specific key. The > > > generic layers just pass this ioctl down to the lower lever drivers. If > > > the specific input device or driver support it - fine, if not - it > > > returns -EINVAL or something like that. > > > > I refuse it because it will be supported by exactly 1 driver in the > > kernel - gpio-keys. It is the only driver that allows shut half of the > > "device" (because in reality it is a group of disjoint devices). It is > > the only case when "muting" a button means that IRQ is shut off abnd > > thus CPU can continue to sleep if that button is pressed. For all other > > devices that have 1 inettrupt per device, you still have to wake up, > > because you don't know whether the button that generated event is > > "important" or not. > > Fair enough. > > > Now, there is a issue of waking up userspace task, additional scheduling > > and keeping CPU running longer than necessary for "uninteresting" keys. > > This can be solved by implementing a subscription model which allows > > filtering uninteresing events on a per-client basis at evdev level. > > Right. And for gpio_keys, this would be dine on the driver level. But the semantics are different - if done on driver level you'd be affecting _all_ consumers of the device; what I want to be done only affects owner of the file descriptor. -- Dmitry _______________________________________________ linux-pm mailing list linux-pm@xxxxxxxxxxxxxxxxxxxxxxxxxx https://lists.linux-foundation.org/mailman/listinfo/linux-pm
