[PATCH v2 0/7] Improve latency of IR decoding

[Sean Young]

On Tue, Apr 10, 2018 at 09:24:19PM +0200, Matthias Reichl wrote:
> On Tue, Apr 10, 2018 at 07:39:43PM +0100, Sean Young wrote:
> > On Tue, Apr 10, 2018 at 07:53:43PM +0200, Matthias Reichl wrote:
> > > Hi Sean,
> > > 
> > > On Sun, Apr 08, 2018 at 10:19:35PM +0100, Sean Young wrote:
> > > > The current IR decoding is much too slow. Many IR protocols rely on
> > > > a trailing space for decoding (e.g. rc-6 needs to know when the bits
> > > > end). The trailing space is generated by the IR timeout, and if this
> > > > is longer than required, buttons can feel slow to respond.
> > > > 
> > > > The other issue is the keyup timer. IR has no concept of a keyup message,
> > > > this is implied by the absence of IR. So, minimising the timeout for
> > > > this makes buttons less "sticky"; the are released much quicker.
> > > > 
> > > > With these patches in place, using IR with the builtin decoders is much
> > > > improved and feels very snappy.
> > > > 
> > > > Changes since v1:
> > > >  - lost more testing
> > > >  - fixed various issues with mce decoder
> > > >  - fixed mceusb so it can use better timeout too
> > > 
> > > thanks, this version is working fine with meson-ir and gpio-rc-recv
> > > (latter on RPi). I mainly tested it with rc-5 remotes so far, more
> > > will follow and I'll update LibreELEC in a day or two to include
> > > the v2 series.
> > 
> > Thanks again for testing!
> > 
> > > Also thanks a lot for the mce_kbd fixes, I was just going to dig
> > > into the decoder (we received a bug report about stuck keys with
> > > mce_kbd last week), your patches can in just at the right time :)
> > > 
> > > I had a closer look at ir-mce_kbd-decoder.c and noticed 2 things
> > > (which can be handled separately):
> > > 
> > > It looks like the input_sync call in the state machine error
> > > path is not necessary:
> > > 
> > > out:
> > > 	dev_dbg(&dev->dev, "failed at state %i (%uus %s)\n",
> > > 		data->state, TO_US(ev.duration), TO_STR(ev.pulse));
> > > 	data->state = STATE_INACTIVE;
> > > 	input_sync(data->idev);
> > > 	return -EINVAL;
> > > 
> > > If I followed the code paths correctly states that generate
> > > input events won't go to out, only paths were no events are
> > > generated jump to out - but better verify that, I could have
> > > missed something.
> > 
> > The input_sync() patch is in mce_kbd_rx_timeout() code path, which
> > doesn't go through this. Hopefully it should fix the keys stuck
> > issue.
> 
> Yes, I saw that and the input_sync() in mce_kbd_rx_timeout() fixed
> this (I could verify that locally).
> 
> What I meant is that the input_sync() in the snippet above, which
> isn't touched by your patches, doesn't seem necessary, as the
> code paths that lead to there don't submit input events. The
> happy paths (successful key/mouse event and keyup) call input_sync()
> at the end of STATE_FINISHED:
> 
> 		lsc.scancode = scancode;
> 		ir_lirc_scancode_event(dev, &lsc);
> 		data->state = STATE_INACTIVE;
> 		input_event(data->idev, EV_MSC, MSC_SCAN, scancode);
> 		input_sync(data->idev);
> 		return 0;

Yes, you're completely right.

> > > The other thing I noticed is that there's no spinlock synchronizing
> > > the events from the timeout callback with the ones from the state
> > > machine - like keylock in rc-main. So the code could potentially
> > > be racy (if the timeout fires while the state machine is outputting
> > > events).
> > 
> > This timeout is for sending keyups in case no keyup is received from
> > the keyboard via input_report_key(). The input system does locking
> > for us (see drivers/input/input.c:436).
> 
> The synchronization code in rc-main looks like it's used to make
> code blocks that eg call input_report_key and input_sync do that
> in an atomic way (plus there's the special handling of keyup_jiffies
> to prevent a race betwen the keyup timeout callback and new
> decoded scancodes.
> 
> As both the timeout callback and the state machine send out
> multiple key events and end them with input_sync it looks to me
> like we'd need to make that atomic as well. Otherwise we could
> have eg timeout running, starting to report a bunch of shift
> and other keys as up, then interruption by a new key/scancode
> calling input_report_key down plus input_sync and then timeout
> resuming and reporting the rest of the keys down and again
> do input_sync().

Again I stand corrected. I'm just testing some patches now for this.

Thanks for being patient and explaining that again :)

Sean