Re: [RFC v2] Another approach to IR

[Dmitry Torokhov]

On Wed, Dec 02, 2009 at 10:44:58AM -0200, Mauro Carvalho Chehab wrote:
> Dmitry Torokhov wrote:
> > On Tue, Dec 01, 2009 at 07:05:49PM -0200, Mauro Carvalho Chehab wrote:
> >> Dmitry Torokhov wrote:
> >>> On Tue, Dec 01, 2009 at 05:00:40PM -0200, Mauro Carvalho Chehab wrote:
> >>>> Dmitry Torokhov wrote:
> >>>>> On Tue, Dec 01, 2009 at 03:29:44PM -0200, Mauro Carvalho Chehab wrote:
> >>>>>> For sure we need to add an EVIOSETPROTO ioctl to allow the driver 
> >>>>>> to change the protocol in runtime.
> >>>>>>
> >>>>> Mauro,
> >>>>>
> >>>>> I think this kind of confuguration belongs to lirc device space,
> >>>>> not input/evdev. This is the same as protocol selection for psmouse
> >>>>> module: while it is normally auto-detected we have sysfs attribute to
> >>>>> force one or another and it is tied to serio device, not input
> >>>>> device.
> >>>> Dmitry,
> >>>>
> >>>> This has nothing to do with the raw interface nor with lirc. This problem 
> >>>> happens with the evdev interface and already affects the in-kernel drivers.
> >>>>
> >>>> In this case, psmouse is not a good example. With a mouse, when a movement
> >>>> occurs, you'll receive some data from its port. So, a software can autodetect
> >>>> the protocol. The same principle can be used also with a raw pulse/space
> >>>> interface, where software can autodetect the protocol.
> >>> Or, in certain cases, it can not.
> >>>
> >>> [... skipped rationale for adding a way to control protocol (with which
> >>> I agree) ...]
> >>>
> >>>> To solve this, we really need to extend evdev API to do 3 things: enumberate the
> >>>> supported protocols, get the current protocol(s), and select the protocol(s) that
> >>>> will be used by a newer table.
> >>>>
> >>> And here we start disagreeing. My preference would be for adding this
> >>> API on lirc device level (i.e. /syc/class/lirc/lircX/blah namespace),
> >>> since it only applicable to IR, not to input devices in general.
> >>>
> >>> Once you selected proper protocol(s) and maybe instantiated several
> >>> input devices then udev (by examining input device capabilities and
> >>> optionally looking up at the parent device properties) would use
> >>> input evdev API to load proper keymap. Because translation of
> >>> driver-specific codes into standard key definitions is in the input
> >>> realm. Reading these driver-specific codes from hardware is outside of
> >>> input layer domain.
> >>>
> >>> Just as psmouse ability to specify protocol is not shoved into evdev;
> >>> just as atkbd quirks (force release key list and other driver-specific
> >>> options) are not in evdev either; we should not overload evdev interface
> >>> with IR-specific items.
> >> I'm not against mapping those features as sysfs atributes, but they don't belong
> >> to lirc, as far as I understand. From all we've discussed, we'll create a lirc
> >> interface to allow the direct usage of raw IO. However, IR protocol is a property
> >> that is not related to raw IO mode but, instead, to evdev mode.
> >>
> > 
> > Why would protocol relate to evdev node? Evdev does not really care what
> > how the fact that a certain button was pressed was communicated to it.
> > It may be deliveretd through PS/2 port, or maybe it was Bluetooth HID,
> > or USB HID or USB boot protocol or some custom protocol, or RC-5, NEC or
> > some custom IR protocol. It makes no difference _whatsoever_ to evdev
> > nor any users of evdev care about protocol used by underlying hardware
> > device to transmit the data.
> >  
> >> We might add a /sys/class/IR and add IR specific stuff there, but it seems
> >> overkill to me and will hide the fact that those parameters are part of the evdev
> >> interface.
> >>
> >> So, I would just add the IR sysfs parameters at the /sys/class/input, if
> >> the device is an IR (or create it is /sys/class/input/IR).
> >>
> >> I agree that the code to implement the IR specific sysfs parameter should be kept
> >> oustide input core, as they're specific to IR implementations.
> >>
> >> Would this work for you?
> > 
> > I am seeing a little bit differently structured subsystem for IR at the
> > moment. I think we should do something like this:
> > 
> > - receivers create /sys/class/lirc devices. These devices provide API
> >   with a ring buffer (fifo) for the raw data stream coming from (and to)
> >   them.
> 
> The raw interface applies only to the devices that doesn't have a hardware decoder
> (something between 40%-60% of the currently supported devices).

50% is quite a number I think. But if driver does not allow access to
the raw stream - it will refuse binding to lirc_dev interface.

> 
> > - we allow registering several data interfaces/decoders that can be bound
> >   (manually or maybe automatically) to lirc devices. lirc devices may
> >   provide hints as to which interface(s) better suited for handling the
> >   data coming form particular receiver. Several interfaces may be bound
> >   to one device at a time.
> > - one of the interfaces is interface implementing current lirc_dev
> > - other interfaces may be in-kernel RC-5 decoder or other decoders.
> >   decoders will create instances of input devices
> 
> I don't see why having more than one interface, especially for devices with
> hardware decoders.
> 
> On IR remote receivers, internally, there's just one interface per hardware.
> 
> Considering the hardware decoding case, why to artificially create other
> interfaces that can't be used simultaneously? No current hardware
> decoders can do that (or, at least, no current implementation allows).
> We're foreseen some cases where we'll have that (like Patrick's dib0700 driver),
> but for now, it is not possible to offer more than one interface to userspace.
> Creating an arbitrary number of artificial interfaces just to pass a parameter
> to the driver (the protocol), really seems overkill to me.

We need to cater to the future cases as well. I don't want to redesign
it in 2 years. But for devices that have only hardware decoders I
suppose we can short-curcuit "interfaces" and have a library-like module
creating input devices directly.

> 
> In the case of the cheap devices with just raw interfaces, running in-kernel
> decoders, while it will work if you create one interface per protocol
> per IR receiver, this also seems overkill. Why to do that? It sounds that it will
> just create additional complexity at the kernelspace and at the userspace, since
> now userspace programs will need to open more than one device to receive the
> keycodes.

_Yes_!!! You open as many event devices as there are devices you are
interested in receiving data from. Multiplexing devices are bad, bad,
bad. Witness /dev/input/mouse and all the attempts at working around the
fact that if you have a special driver for one of your devices you
receive events from the same device through 2 interfaces and all kind of
"grab", "super-grab", "smart-grab" schemes are born.

> 
> > (for each remote/substream that they can recognize).
> 
> I'm assuming that, by remote, you're referring to a remote receiver (and not to 
> the remote itself), right?

If we could separate by remote transmitter that would be the best I
think, but I understand that it is rarely possible?

-- 
Dmitry
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/