Commit db3b6fe5f7 "fallback: change to handle the state at EV_SYN time"
introduced regressions for two types of event sequences.
One is a kernel bug - some devices/drivers like the asus-wireless send a key
press + release within the same event frame which now cancels out and
disappears into the ether. This should be fixed in the kernel drivers but
there appear to be enough of them that we can't just pretend it's an outlier.
The second issue is a libinput bug. If we get two key events in the same frame
(e.g. shift + A) we update the state correctly but the events are sent in the
order of the event codes. KEY_A sorts before KEY_LEFTSHIFT and our shift + A
becomes A + shift.
Fix this by treating key events as before db3b6fe5f7 - by sending them out
as we get them.
https://bugs.freedesktop.org/show_bug.cgi?id=104030
Signed-off-by: Peter Hutterer <peter.hutterer@who-t.net>
The current debouncing code monitors events and switches on when events are
too close together. From then on, any event can be delayed.
Vicente Bergas provided an algorithm that avoids most of these delays:
on a button state change we now forward the change without delay but start a
timer. If the button changes state during that timer, the changes are
ignored. On timer expiry, events are sent to match the hardware state
with the client's view of the device. This is only done if needed.
Thus, a press-release sequence of: PRP sends a single press event, a sequence of
PRPR sends press and then the release at the end of the timeout. The timeout
is short enough that the delay should not be noticeable.
This new mode is called the 'bounce' mode. The old mode is now referred to as
'spurious' mode and only covers the case of a button held down that loses
contact. It works as before, monitoring a button for these spurious contact
losses and switching on. When on, button release events are delayed as before.
The whole button debouncing moves to a state machine which makes debugging a
lot easier. See the accompanying SVG for the diagram.
Signed-off-by: Peter Hutterer <peter.hutterer@who-t.net>
The previous approach was to remember the last event and flush it at the right
time. The new approach is to update the device state during the frame and send
out the events at EV_SYN time.
This gives us two advantages: we are not dependent on the kernel order of how
events come in and we can process events depending on other events in the same
frame. This will come in handy later for button debouncing.
This is also the approach we have in the touchpad and tablet backends.
Two FIXMEs are left in place, the button debouncing code and the lid switch
code. Both need to be handled in future patches.
Signed-off-by: Peter Hutterer <peter.hutterer@who-t.net>
Once the lid is closed, the keyboard event listener is set up to open the lid
for us on keyboard events. With the right sequence, we can trigger the
listener to be added to the list multiple times, triggering an assert in the
list test code (or an infinite loop in the 1.8 branch).
Conditions:
* SW_LID value 1 - sets up the keyboard listener
* keyboard event - sets lid_is_closed to false
* SW_LID value 0 - is ignored because we're already open
* SW_LID value 1 - sets up the keyboard listener again
https://bugs.freedesktop.org/show_bug.cgi?id=103298
Signed-off-by: Peter Hutterer <peter.hutterer@who-t.net>
Because on some devices the keyboard is where the fingers are holding the
device when in tablet mode.
https://bugs.freedesktop.org/show_bug.cgi?id=102729
Signed-off-by: Peter Hutterer <peter.hutterer@who-t.net>
Split out the fallback-specific device handling from the more generic
evdev-specific handling (which is supposed to be available for all devices).
Signed-off-by: Peter Hutterer <peter.hutterer@who-t.net>
