Re: [PATCH v2 0/8] hid driver transport cleanup

[David Herrmann]

Hi

On Mon, Feb 25, 2013 at 4:06 PM, Benjamin Tissoires
<benjamin.tissoires@xxxxxxxxx> wrote:
>> (I did read the Bluetooth HID Profile specification but I have no idea
>> how USB does it, so please bear with me if I mix things up. Maybe some
>> day I will have the time to read the USB specs, too)
>>
>> * There are several drivers using "dev->hid_get_raw_report()" and
>> "dev->hid_output_raw_report()". Any objections to moving these to
>> "hid_ll_driver" and adding wrapper functions
>> hid_hw_get/output_raw_report()?
>
> I can't see any. Those two callbacks are transport dependent, so I
> also think it makes sense to put them into hid_ll_driver.

I will prepare a patch.

> I also wish we have the set/get hid idle callback (it would remove the
> last usb dependency in hid-multitouch).
>
>>
>> * What should the ll->report() callback exactly do? Should it simply
>> send a GET_REPORT or SET_REPORT request depending on the direction
>> (nonblocking)?
>
> Well, given the way it's used in hid drivers and usbhid/i2c-hid:
> - SET_REPORT takes an already parsed report and send it to the device
> -> the usbhid part is nonblocking, i2c-hid is blocking.
> - GET_REPORT asks for the HID command GET_REPORT and once it's done,
> the incoming report is parsed as if it was spontaneously generated
> (meaning that there are no guarantees that the report the caller reads
> after report() is the requested one).
>
> The patch for implementing those functions in i2c-hid is IMHO simpler
> to understand than the usb one as i2c-hid makes no use of dma or async
> calls.
>
> For the nonblocking part, the usb pendant is asynchronous, thus the
> need of the .wait().
> The i2c-hid driver is synchronous, meaning that the call is blocking
> (sorry). But .wait() is not required to be called, and then, we don't
> need to set it.

Then maybe I'll wait for the first driver that needs these and then
decide whether to make the Bluetooth calls blocking/nonblocking. But
see below.

>>
>> * Why don't we do the hid_output_report() call in the hid_hw_report()
>> helper and pass the raw buffers down? It would allow GFP_KERNEL and
>> avoid duplicating code in all four backends.
>
> Well. I continued Henrik's work on that, and the idea was to not break
> any existing device. So, the code path should be the nearest possible
> to the existing one. Now that this part is done, we can look at
> something more efficient in terms of duplication.

Ok, that's fine.

> Actually, the implementation of i2c-hid .request() already uses the
> raw buffers commands, meaning that it's nearly what we need (except
> that we need  to get the size of the buffer in some other way).
>
> However, in usbhid and i2c-hid, .hid_get_raw_report() and
> .hid_output_raw_report() are blocking, meaning that it will change the
> current behavior for drivers.
>
>>
>> * What should ll->wait() exactly do? Block until the last
>> SET/GET_REPORT call has been ack'ed by the device?
>>
>
> It waits for at most 10*HZ until both the CTRL (command queue for
> set_report/get_report and 90% of the hid communication) and the OUTPUT
> queues are empty.
> Honestly, I don't like the way the wait is done: nobody cares about
> the return value, meaning that nobody know why we have finished to
> wait. In 90% of the cases, it's used in conjunction with a .request()
> call.
>
> So to sum up, given its current use, I would say that .wait() "blocks
> until all queues are flushed or for an amount of time that gives the
> caller good chances that his/her previous call has been transferred to
> the device".
>
>> * Bluetooth-HID might return errors on GET/SET_REPORT. Why don't we
>> return these in hid_hw_wait() for the last report? It would allow
>> synchronous calls like:
>> hid_hw_report();
>> ret = hid_hw_wait();
>
> As there are several uses of .wait(), it seems difficult.
> Some are using it to make 'sure' that their previous command has been
> sent -> in this case, I would say, yes, go for it.
> Some others are using it for it's real purpose: flush the queues
> before/after sending new commands -> it's difficult to put the last
> command result as a return value in this case (the latest result does
> not mean anything to the caller).
>
> Maybe turning the hid_hw_report() for these cases into a blocking one
> would solve the problem.
>
>>
>> * Should hid_hw_report() allow multiple pending requests? Or should it
>> return an error if there is another pending report that hasn't been
>> ack'ed, yet?
>
> I would say that it should.
> If you ask for a report, it's then parsed through hid_input_report,
> meaning that you can ask for several reports at the same time. Each of
> them will be parsed in it's own buffer.

I just checked the Bluetooth specs and they forbid multiple pending
requests on the CTRL channel. Hence, I think there is no reason to
make hid_hw_report() non-blocking for BT.

So I think the internal BT design will be a mutex that locks the CTRL
channel and a wait_event_timeout() that waits 5s max for an answer. As
l2cap is reliable and supports re-transmission, the high timeout seems
reasonable.

That would make hid_hw_wait() useless for BT and the design similar to
i2c-hid, I guess.

>>
>> I will try to implement it for Bluetooth-HID and UHID if no-one else
>> wants to do it. Doing it for UHID would make debugging/emulating
>> devices a lot easier.
>
> That would be great. I can't do the bluetooth part (I'm lacking of
> hardware), and for the UHID part... I don't really care about it in my
> tests for now :)

Thanks for clarifying. I don't think I will implement the callbacks,
yet (there's no driver using them). However, it really helps to
improve the internal design.

Thanks
David
--
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/