[regression] pinch to zoom not working reliably

<bfiller> greyback: easiest way for me to repro is to open a picture in camera photo roll then pinch to zoom and unzoom, double tapping also gets it into the state, usually first time it works then not after that

In my testing it does appear the pinch gesture is failing to be recognized more frequently than before. In camera photo roll, if I zoom in, then try to zoom out immediately, often that zoom-out fails.

Bug more pronounced on Bq than Mako

Might be related to bug 1530940

I can pinch to zoom just fine on my Nexus 7 (r330, flashed today). Trying on krillin...

It's horrible in krillin indeed.
And not so great in the Nexus 7 actually. There's a stark difference compared to doing a pinch to zoom on my touchscreen Xenial laptop running gallery-app. Which points to something in qtbuntu, qtmir or mir (or a combination of those).

Interesting that if you lay a finger, wait a second, then drag with a second finger you always succeed.

And then, in that same touchscreen laptop, running gallery-app in a unity8-mir session yields a bad pinch-to-zoom experience.

From the attachment you can see that mir giving a bogus event stream.
When you lay two fingers on the screen at roughly the same time (like when doing a pinch gesture) it's common for mir to not generate or send the "down" event for the first touch point.

qtmir has a basic event stream sanity check and validation. It detects this problem and transforms the action=change for the first touch event with id=0 into a action=down instead. But then we end up with an input event with *two* pressed touch points, which at least makes sense but I think Qt does not like that. It probably expects a single touch up/down per event, like android. I could make qtmir's event stream validator more sophisticated and have it split those two touch down actions into two separate input events, but mir should really send us a clean event stream in the first place.

My guess is that this regression is related to mir changing from android-input to libinput.

Status changed to 'Confirmed' because the bug affects multiple users.

same on MX4 rc-proposed

dandrader:
It's been a useful quirk of Android input until now that multi-touch gestures get separate events for each finger, but I think we all need to ensure our client code does not assume that's always true.

It wasn't a design choice of Mir's, but just something we inherited from Android. There's a good argument that Mir's semantics should not change, because it annoys people obviously. On the other hand it would be difficult to keep all the quirks of Android input in the transition to libinput. And I think the semantics of receiving a single event where two fingers go down at once is perfectly valid. We should all be writing client code that expects that might happen.

On 12/01/2016 05:39, Daniel van Vugt wrote:
> dandrader:
> It's been a useful quirk of Android input until now that multi-touch gestures get separate events for each finger, but I think we all need to ensure our client code does not assume that's always true.
>
> It wasn't a design choice of Mir's, but just something we inherited from
> Android. There's a good argument that Mir's semantics should not change,
> because it annoys people obviously. On the other hand it would be
> difficult to keep all the quirks of Android input in the transition to
> libinput. And I think the semantics of receiving a single event where
> two fingers go down at once is perfectly valid. We should all be writing
> client code that expects that might happen.
>

If you read my message carefully you will see that Mir is not even doing
that (two fingers down on the same event). It was qtmir that was mending
the event stream to do so. Mir was completely missing the down event for
the first touch. So all of the blue you got a "touch 1 move + touch 2
down" (no touch 1 down beforehand).

Ah yes the android input transport is biting us again. We are still bound to the limitation of that encoding. That encoding means - only one down/up touch id per event. No matter what libinput reads or qtmir sanitizes.. , as long as the transport from server to nested server or from nested server to client chops of the information of every additional down or up the gesture will break. I will keep the fix close to the encoding so we can get rid of it easier...

Fix committed into lp:mir at revision None, scheduled for release in mir, milestone 0.19.0

This bug was fixed in the package mir - 0.18.1+16.04.20160115-0ubuntu1

---------------
mir (0.18.1+16.04.20160115-0ubuntu1) xenial; urgency=medium

  [ Andreas Pokorny ]
  * New upstream release 0.18.1 (https://launchpad.net/mir/+milestone/0.18.1)
    - No ABI changes. Bug fix release only.
    - Bug fixed:
      . [regression] pinch to zoom not working reliably (LP: #1531517)

  [ CI Train Bot ]
  * No-change rebuild.

 -- kevin gunn <email address hidden> Fri, 15 Jan 2016 11:45:05 +0000