Re: [PATCHv2 0/7] Support inhibiting input devices

[Andrzej Pietrasiewicz <andrzej.p@collabora.com>]

Hi Hans, hi Dmitry,

W dniu 02.06.2020 o 22:19, Hans de Goede pisze:
> Hi,
> 
> On 6/2/20 8:50 PM, Andrzej Pietrasiewicz wrote:
>> Hi Dmitry,
>>
>> W dniu 02.06.2020 o 19:52, Dmitry Torokhov pisze:
>>> Hi Andrzej,
>>>
>>> On Tue, Jun 02, 2020 at 06:56:40PM +0200, Andrzej Pietrasiewicz wrote:
>>>> Hi Dmitry,
>>>>
>>>> W dniu 27.05.2020 o 08:34, Dmitry Torokhov pisze:
>>>>> That said, I think the way we should handle inhibit/uninhibit, is that
>>>>> if we have the callback defined, then we call it, and only call open and
>>>>> close if uninhibit or inhibit are _not_ defined.
>>>>>
>>>>
>>>> If I understand you correctly you suggest to call either inhibit,
>>>> if provided or close, if inhibit is not provided, but not both,
>>>> that is, if both are provided then on the inhibit path only
>>>> inhibit is called. And, consequently, you suggest to call either
>>>> uninhibit or open, but not both. The rest of my mail makes this
>>>> assumption, so kindly confirm if I understand you correctly.
>>>
>>> Yes, that is correct. If a driver wants really fine-grained control, it
>>> will provide inhibit (or both inhibit and close), otherwise it will rely
>>> on close in place of inhibit.
>>>
>>>>
>>>> In my opinion this idea will not work.
>>>>
>>>> The first question is should we be able to inhibit a device
>>>> which is not opened? In my opinion we should, in order to be
>>>> able to inhibit a device in anticipation without needing to
>>>> open it first.
>>>
>>> I agree.
>>>
>>>>
>>>> Then what does opening (with input_open_device()) an inhibited
>>>> device mean? Should it succeed or should it fail?
>>>
>>> It should succeed.
>>>
>>>> If it is not
>>>> the first opening then effectively it boils down to increasing
>>>> device's and handle's counters, so we can allow it to succeed.
>>>> If, however, the device is being opened for the first time,
>>>> the ->open() method wants to be called, but that somehow
>>>> contradicts the device's inhibited state. So a logical thing
>>>> to do is to either fail input_open_device() or postpone ->open()
>>>> invocation to the moment of uninhibiting - and the latter is
>>>> what the patches in this series currently do.
>>>>
>>>> Failing input_open_device() because of the inhibited state is
>>>> not the right thing to do. Let me explain. Suppose that a device
>>>> is already inhibited and then a new matching handler appears
>>>> in the system. Most handlers (apm-power.c, evbug.c, input-leds.c,
>>>> mac_hid.c, sysrq.c, vt/keyboard.c and rfkill/input.c) don't create
>>>> any character devices (only evdev.c, joydev.c and mousedev.c do),
>>>> so for them it makes no sense to delay calling input_open_device()
>>>> and it is called in handler's ->connect(). If input_open_device()
>>>> now fails, we have lost the only chance for this ->connect() to
>>>> succeed.
>>>>
>>>> Summarizing, IMO the uninhibit path should be calling both
>>>> ->open() and ->uninhibit() (if provided), and conversely, the inhibit
>>>> path should be calling both ->inhibit() and ->close() (if provided).
>>>
>>> So what you are trying to say is that you see inhibit as something that
>>> is done in addition to what happens in close. But what exactly do you
>>> want to do in inhibit, in addition to what close is doing?
>>
>> See below (*).
>>
>>>
>>> In my view, if we want to have a dedicated inhibit callback, then it
>>> will do everything that close does, they both are aware of each other
>>> and can sort out the state transitions between them. For drivers that do
>>> not have dedicated inhibit/uninhibit, we can use open and close
>>> handlers, and have input core sort out when each should be called. That
>>> means that we should not call dev->open() in input_open_device() when
>>> device is inhibited (and same for dev->close() in input_close_device).
>>> And when uninhibiting, we should not call dev->open() when there are no
>>> users for the device, and no dev->close() when inhibiting with no users.
>>>
>>> Do you see any problems with this approach?
>>
>> My concern is that if e.g. both ->open() and ->uninhibit() are provided,
>> then in certain circumstances ->open() won't be called:
>>
>> 1. users == 0
>> 2. inhibit happens
>> 3. input_open_device() happens, ->open() not called
>> 4. uninhibit happens
>> 5. as part of uninhibit ->uninhibit() is only called, but ->open() is not.
>>
>> They way I understand your answer is that we implicitly impose requirements
>> on drivers which choose to implement e.g. both ->open() and ->uninhibit():
>> in such a case ->uninhibit() should be doing exactly the same things as
>> ->open() does. Which leads to a conclusion that in practice no drivers
>> should choose to implement both, otherwise they must be aware that
>> ->uninhibit() can be sometimes called instead of ->open(). Then ->open()
>> becomes synonymous with ->uninhibit(), and ->close() with ->inhibit().
>> Or, maybe, then ->inhibit() can be a superset of ->close() and
>> ->uninhibit() a superset of ->open().
>>
>> If such an approach is ok with you, it is ok with me, too.
>>
>> (*)
>> Calling both ->inhibit() and ->close() (if they are provided) allows
>> drivers to go fancy and fail inhibiting (which is impossible using
>> only ->close() as it does not return a value, but ->inhibit() by design
>> does). Then ->uninhibit() is mostly for symmetry.
> 
> All the complications discussed above are exactly why I still
> believe that there should be only open and close.
> 
> If error propagation on inhibit is considered as something
> really important to have then we can make the input driver close
> callback return an error (*), note I'm talking about the
> driver close callback here, not the system call.
> 
> If the close callback is called for actually closing the fd
> referring to the input node, then the new error return code
> can be ignored, as we already do for errors on close atm
> since the driver close callback returns void.
> 
> I still have not seen a very convincing argument for having
> separate inhibit and close callbacks and as the messy discussion
> above shows, having 2 such very similar yet subtly different
> calls seems like a bad idea...
> 
> Regards,
> 
> Hans
> 
> 
> *) This will require a flag day where "return 0" is added
> to all current close handlers
> 

I'm taking one step back and looking at the ->open() and ->close()
driver callbacks. They are called from input_open_device() and
input_close_device(), respectively:

input_open_device():
"This function should be called by input handlers when they
want to start receive events from given input device."

->open() callback:
"this method is called when the very first user calls
input_open_device(). The driver must prepare the device to start
generating events (start polling thread, request an IRQ, submit
URB, etc.)"

input_close_device():
"This function should be called by input handlers when they
want to stop receive events from given input device."

->close() callback:
"this method is called when the very last user calls
input_close_device()"

It seems to me that the callback names do not reflect their
purpose: their meaning is not to "open" or to "close" but to
give drivers a chance to control when they start or stop
providing events to the input core.

What would you say about changing the callbacks' names?
I'd envsion: ->provide_events() instead of ->open() and
->stop_events() instead of ->close(). Of course drivers can
exploit the fact of knowing that nobody wants any events
from them and do whatever they consider appropriate, for
example go into a low power mode - but the latter is beyond
the scope of the input subsystem and is driver-specific.

With such a naming change in mind let's consider inhibiting.
We want to be able to control when to disregard events from
a given device. It makes sense to do it at device level, otherwise
such an operation would have to be invoked in all associated
handlers (those that have an open handle associating them with
the device in question). But of course we can do better than
merely ignoring the events received: we can tell the drivers
that we don't want any events from them, and later, at uninhibit
time, tell them to start providing the events again. Conceptually,
the two operations (provide or don't provide envents) are exactly
the same thing we want to be happening at input_open_device() and
input_close_device() time. To me, changing the names of
->open() and ->close() exposes this fact very well.

Consequently, ->inhibit() and ->uninhibit() won't be needed,
and drivers which already implement ->provide_events() (formerly
->open()) and ->stop_events() (formerly ->close()) will receive
full inhibit/uninhibit support for free (subject to how well they
implement ->provide_events()/->stop_events()). Unless we can come
up with what the drivers might be doing on top of ->stop_events()
and ->provide_events() when inhibiting/uninhibiting, but it seems
to me we can't. Can we?

Optionally ->close() (only the callback, not input_close_device())
can be made return a value, just as Hans suggests. The value
can be ignored in input_close_device() but used in input_inhibit().
No strong opinion here, though. (btw it seems to me that
input_inhibit() should be renamed to input_inhibit_device()).

Regards,

Andrzej