Re: [PATCH v3] bcm5974: Set BUTTONPAD property

[Dmitry Torokhov <dmitry.torokhov@gmail.com>]

On Tue, Jan 17, 2012 at 07:25:42PM +0100, Chase Douglas wrote:
> On 01/17/2012 07:06 PM, Dmitry Torokhov wrote:
> > On Thu, Jan 12, 2012 at 11:19:40AM +0100, Chase Douglas wrote:
> >> On 01/12/2012 01:22 AM, Henrik Rydberg wrote:
> >>>> Here's what I believe the meanings should be:
> >>>>
> >>>> Touchpad: pointer, !direct
> >>>> Touchscreen: !pointer, direct
> >>>> Drawing tablet: pointer, direct
> >>>> Magic mouse-like devices: !pointer, !direct
> >>>
> >>> Yes, this is what everyone is saying, except !pointer && !direct means
> >>> "default" or "figure it out some other way".
> >>>
> >>>> However, there is a further problem in that we can't easily support
> >>>> multiple tools with different behavior on the same evdev device. What
> >>>> would you say a bamboo touch+pen is, which I believe is used as an
> >>>> indirect device for touch but a direct device for tools. Thus, in the
> >>>> thread I linked from back in September, Henrik and I agreed that direct
> >>>> should only apply when the tool is touch, and pointer should apply for
> >>>> all other tools. This would result in the following:
> >>>
> >>> To try to move back to a sane track, try this, where the word "apply"
> >>> in the previous paragraph has been changed to "care" instead:
> >>
> >> I am still having trouble understanding what you are saying. If I
> >> literally try to insert "care" into the paragraph, I am confused because
> >> it's not quite correct grammar. I'm really trying to understand though.
> >> Also, maybe a better term than "don't care" is "not applicable"?
> >>
> >> It would help me most if you could explicitly provide your own
> >> definition of the properties.
> >>
> >>>> Touchpad: !pointer, !direct
> >>>
> >>> pointer && !direct, since pointer is "dont care".
> >>
> >> Here you say !direct if "don't care".
> >>
> >>>> Touchscreen: !pointer, direct
> >>>
> >>> Yes, !pointer && direct.
> >>>
> >>>> Drawing tablet (no touch): pointer, !direct
> >>>
> >>> pointer && direct, but the tool is not touch, so direct is "dont care".
> >>
> >> Here you say direct if "don't care".
> >>
> >> Why the difference?
> >>
> >>>> Pen+touch tablet: pointer, direct
> >>>
> >>> Yes, pointer && direct
> >>>
> >>>> Magic mouse-like devices: !pointer, !direct
> >>>
> >>> Both pointer and direct are "dont care", and the device needs to be
> >>> detected some other way. If there ever will be a special driver for
> >>> magic-mouse-like devices, using both relative pointer and touch data,
> >>> it will make sense to add a special property for such devices.
> >>
> >> Right now we are missing a property for a magic-mouse like device. It's
> >> valid to have neither direct nor pointer set from kernels 2.6.38 through
> >> 3.2 (at least).
> >>
> >>> Hopefully the above is showing clearly that what was "documented" in
> >>> the threads enclosing the protocol patches still holds, and that there
> >>> is no use to dwell on it further.
> >>>
> >>>> The properties weren't documented when they were merged, and they
> >>>> obviously aren't clear. However, if either table above is correct, then
> >>>> we can't assume that !pointer && !direct means "unknown".
> >>>
> >>> If all devices fell in the pointer or direct or both categories, we
> >>> could. If not all devices do, the problem is rather that some property
> >>> bits are missing (or excluded) from the description.
> >>
> >> Given my last statement above, we have a problem because previously
> >> released kernels are reporting the magic mouse correctly, and yet we
> >> still can't distinguish it from another device that merely does not have
> >> the property bits set. This is the crux of the issue as I see it. We
> >> cannot differentiate between "unknown" and a specific type of device
> >> given the interfaces from 2.6.38 through 3.2.
> >>
> >>>> There is a way to fix this in a backwards compatible way: add a new
> >>>> property bit called something like "PROPERTIES_AVAILABLE". If any bits
> >>>> are set, then it implies that the properties are available (which covers
> >>>> older kernels). If no bits are set, then the properties are unknown.
> >>>> What do you think?
> >>>
> >>> It is rather the special properties of the magic mouse that are
> >>> missing. All types of devices do not _have_ to use properties; most
> >>> types can be figured out by other means.
> >>>
> >>> Saying "prop == 0" is equivalent to "figure out some other way" makes
> >>> sense, but it is also sensible to say "(prop & some_subclass_of_bits)
> >>> == 0", since some properties are bound to describe totally different
> >>> things. This is what we did with "!direct && !pointer".
> >>
> >> This may work, but we need to document the classes. The next time any
> >> properties are added the documentation must be included :).
> >>
> >>>>> The same is applicable to other properties as well. If device is telling
> >>>>> you that it is a "buttonpad" you can trust it, but if it does not you
> >>>>> need to decide for yourself how to treat it.
> >>>
> >>> Yes, and this will always be true. Old devices or systems that become
> >>> used in new ways cannot always adapt to a "if property not present
> >>> then dont use that way" policy.
> >>>
> >>>> No, in kernels previous to 2.6.38 it's clearly unknown. My problem is
> >>>> that I believe there was no way to determine unknown properties. If
> >>>> unknown properties is equivalent to magic-mouse like devices, then we're
> >>>> going to treat a lot of devices wrong. Or, we have to use heuristics to
> >>>> determine what a device is, like no properties and MT and REL_{X,Y} ==
> >>>> magic-mouse like. Properties was supposed to resolve this once and for
> >>>> all, so we didn't need heuristics.
> >>>
> >>> Properties were added to be able to distinguish usecases that could
> >>> not be distinguished at all before. It was never meant to replace
> >>> everything else.
> >>
> >> Why shouldn't we use it for that? The code in evdev for determining the
> >> type of device is just a big hack. We'll obviously need it for a while
> >> since we don't have all drivers with all necessary properties set, but
> >> it seems a waste to have the interface and not fully use it.
> >>
> >>>>>> Henrik, can you comment on the documentation patches? You wrote the
> >>>>>> patch, so you hopefully know what's going on :).
> >>>
> >>> I wasn't copied in on the conversation, but they seem fairly well
> >>> commented on already.
> >>
> >> It's still not clear to me what the definitions are. It seems it won't
> >> be clear until either you or Dmitry give your own definitions in an
> >> explicit manner (something that could be copied into the formal
> >> documentation). Please help me out :).
> >>
> > 
> > OK, so how about this:
> > 
> > INPUT_PROP_DIRECT:
> > 
> > This property idicates that device coordinates can be directly mapped to
> > screen coordinates (not taking into account trivial transformations,
> > such as scaling, flipping and rotating). Non-direct input devices
> > require non-trivial transformation, such as absolute to relative
> > transformation for touchpads. Typical direct input devices:
> > touchscreens, drawing tablets; non-direct devices: touchpads, mice.
> > 
> > INPUT_PROP_POINTER:
> > 
> > This property indicates that the device is not transposed on the screen
> > and thus requires use of an on-screen pointer to trace user's movements.
> > Typical pointer devices: touchpads, tablets, mice; non-pointer device:
> > touchscreen.
> > 
> > How does this sound?
> 
> These definitions sound fine to me. We also need definitions of what it
> means when a bit is set versus when it is not set. Does an unset bit
> mean "unknown"? As stated before, I don't like that definition because:
> 
> * It means we can never get away from device type heuristics in user-space.
> * There's no negative version of the properties. For example, there's no
> way to say "this device is indirect" because it looks the same as "unknown".
> 

However treating unset as negative will not allow us introducing new
properties without doing mass-updates of all in-kernel (and out of tree)
drivers.

> Imagine a tablet driver only has INPUT_PROP_DIRECT set. According to the
> "unknown" definition, it's ok to leave INPUT_PROP_POINTER as unset. But
> then userspace will end up treating it like a touchscreen instead of a
> tablet.

That would be a not too smart on part of tablet driver, as these 2
properties were designed to work in tandem.

> 
> If you are unwilling to backport property bits to stable kernels, then I
> don't think we have any other choice than to leave the definition as
> "unset bits are unknown", but it clearly puts userspace in a bind.
> Because the evdev event codes are clear (BTN_TOUCH means touchscreen,
> BTN_TOOL_FINGER means touchpad) and these property bits are not, I doubt
> userspace will ever rely on the direct and pointer property bits.
> 

Right, since we can not commit that we'll never add more properties and
you have to support both newer and older kernels there will always be a
certain level of heuristic involved.

Thanks.

-- 
Dmitry