The server struggles with devices that are both, the protocol (especially XI2)
requires a fairly strict separation of pointer vs keyboard devices. Though the
server has a couple of hacks to route events correctly, mixed
devices still experience bugs like [1].

Instead of advertising the device as a single mixed device, split the device
into two X devices, one with only a pointer/touch component, one with only a
keyboard component. This ensures that the device is effectively attached to
both the VCP and the VCK, something the XI2 protocol doesn't really allow.

This patch drops the keyboard capability on a mixed device, duplicates the
input options and attributes and queues a NewInputDeviceRequest call. The new
device only has the keyboard capability but is otherwise unchanged. The
wacom driver has used this approach for years.

The WorkProc is necessary to avoid inconsistent state, the server doesn't
handle a NewInputDeviceRequest during PreInit well.

The approach:
During pre-init we create a struct xf86libinput_device with the
libinput_device and a unique ID. The child device has that ID added to the
options and will look for the other device during its pre-init. The two
devices then share the xf86libinput_device struct.

We only have a single epollfd for all devices  and the server calls read_input
on the first device in the list with the epollfd as pInfo->fd. That shared
struct is used as the userdata on the libinput_device we get back from the
event, and each device is in the xorg_list device_list of that shared struct.
We loop through those to find the ones with the right capabilities and
post the event through that device.

Since devices can be enabled and disabled independently, the rest of the code
makes sure that we only ever add the device to libinput when the first shared
device is enabled, and remove it accordingly.

The server uses pInfo->major/minor to detect if another device is using the
same path for a logind-controlled fd. If so, it reuses that device's
pInfo->fd and sets the "fd" option to that value. That pInfo->fd is the
libinput epollfd though, not the actual device fd.

This doesn't matter for us, since we manage the fds largely ourselves and the
pInfo->fd we use is the epollfd anyway. On unplug however, the udev code
triggers a device removal for all devices, including the duplicated ones. When
we disable device, we restore the pInfo->fd from the "fd" option so that the
server can request logind to close the fd.

That only works if the "fd" option is correct, otherwise the server asks
logind to close the epollfd and everyone is unhappy.

[1] https://bugs.freedesktop.org/show_bug.cgi?id=49950

https://bugs.freedesktop.org/show_bug.cgi?id=92896



Signed-off-by: Peter Hutterer <peter.hutterer@who-t.net>
Reviewed-by: Hans de Goede <hdegoede@redhat.com>