Created attachment 72611 [details]
test program to reproduce

If DGA input is activated while a passive grab is active, the release event is lost and the master is left with a stuck button.

http://lists.x.org/archives/xorg-devel/2012-August/033310.html

Analysis copied from the thread above:

DGAProcessPointerEvent() calls UpdateDeviceState() with mouse
button events where ev.detail.key is not set.

As for the slave device, that's a race condition. UDS doesn't get called
for slave devices at all while a DGA handler is active (note that the DGA
extension does not support per-device flags, it only works on masters
anyway).

So the net result is here that the slave device never has any button state
modified, and this should largely be fine. that is, unless you hit the
race condition that MCedit appears to trigger. If you call XDGASetMode()
after receiving a ButtonPress event (and that request gets processed before
the release event), the release event never updates the state of the slave
device, leading to the stuck button.