Bug 492621

Created attachment 337476 [details]
Patch to correctly configure the keyboard during installation.

During installation, the rhpl.keyboard.Keyboard activate() method is used to activate the keyboard configuration.  When using the Xen Virtual Keyboard/Mouse combo device we need to ensure that we have the evdev keyboard model to match what the X configuration is using (see other patch).

This patch forces the keyboard model to be "evdev" when needed.

In addition, we need to set the "lv3:switch" xkb option for reasons that aren't entirely clear to me.  There's a bug somewhere that means that the AltGr key (which is normally ISO_Level3_Shift) should, by default, switch to the alternate characters and this option is implicit.   It does no harm to set this option anyway: if the keyboard layout maps the right Alt key to Alt_R (as it does for US keyboards) then this option has no effect.

The bug as reported is about the difficulties to manually configure X
in a certain way, to get proper pointer tracking.  It is certainly
valid.  However, the real problem users care about is that pointer
tracking works very poorly (mouse hits invisible wall).

Upstram, we solved the problem by fixing numerous issues in various
places to make it work out of the box, from Fedora 11 on.

The patches proposed by the reporter solve the problem by making
manual configuration work.  They are unrelated to the upstream
solution.

Like the upstream solution, they're fairly invasive: they touch the
kernel and several user space packages.  This is a significant risk.
Moreover, splitting the device is an ABI change.  The question is
whether there are benefits to justify that.

We intend to put much less invasive changes into 4.9 to improve
pointer tracking (bug 492868).  They fix the "mouse hits invisible
wall" problem.  They still require a pointer grab, which virt-viewer
makes relatively painless.

So, the benefit of the proposed patches over what we've got already is
to enable manual configuration so that the mouse grab can be avoided.
While that's certainly nice, it doesn't justify the risk, in my
opinion.

If you disagree, please reopen the bug.


Some technical background:

There are 3 pieces in the stack: VNC client, QEMU VNC server and the
guest's X.  The old, misbehaving setup has VNC client sending absolute
coordinates, which the VNC server sends through xenkbd to X.  X in
RHEL-4 doesn't have any auto-configuration of input devices, so it
just opens /dev/input/mice, and thus the guest kernel converts from
absolute to relative.  Because the VNC client has no knowledge of this
conversion, you get into situation where you hit an invisible wall in
the client when it thinks it has got to the virtual desktop boundary.

Now, if you had absolute coordinates being passed and used all the way
to X then it would trivially work, but this requires X configuration.
Recent X's do that automatically.  RHEL-4's X simply isn't capable of
that.  The patches proposed by the reporter add manual configuration
steps across a wide range of RPMs.

The fix we did for RHEL-5.4 and plan to backport to 4.9 is to make
xenkbd *not* use absolute coordinates, so the VNC server now sends
relative coordinates.  This removes the broken absolute -> relative
conversion in the guest kernel.  On its own this isn't sufficient,
because you then get broken absolute to relative conversion in the VNC
server instead.  So we also added the VNC relative mouse extension to
QEMU and GTK-VNC.  Now the VNC client is in charge of doing the
absolute -> relative conversion.  This is good, as this is the only
point in the stack capable of doing the conversion correctly, to avoid
hitting an invisible wall.  The only restriction is that you must grab
the host mouse pointer and hide it, so you only see the guest drawn
pointer.

By the way, RHEL-6 pointer tracking will be fine out of the box using absolute coordinates, just like upstream.