Description of problem:
Fixed in xen-unstable cset 14926.  Quoting its log message:

 1. XendDomainInfo._createDevices() gets a list of devices to be
    created from XendConfig.ordered_device_refs().
    On a simple guest, this has 4 devices, vfb, vbd, vif, vkbd - in
    that order.

 2. It iterates over those devices, creating the appropriate
    DevController subclass instance, and then calling createDevice()
    on that object.

 3. When createDevice() is called on the vfb  device, it spawns
    xen-vncfb daemon.

 4. During startup xen-vncfb writes into the backend paths
            /local/domain/0/backend/vfb/0
     And
            /local/domain/0/backend/vkbd/0

 5. When createDevice() is called on the vkbd device in XenD, if the
    2nd xenstore path write from step 4 has occurred, then you'll hit
    the 'Device 0 (vkbd) is already connected' error. If the 2nd path
    write didn't complete yet then everything is fine.

I think the reason it often works once after boot is that loading
xen-vncfb from disk the first time around is just enough of a slow
down to ensure step 5 occurs before the 2nd xenstore write in step 4
has occurred.

The key seems to be to ensure the vkbd device is initialized in
xenstore before the vfb device - this ensures all the xenstored setup
from XenD is complete before the xen-vncfb daemon starts. I'm now able
to create & destroy a domain many times over with this patch & never
hit the error message any more.

End quote.

The order in which devices are created is non-random but unpredictable; small,
innocent-looking changes to the xend code can change the order, as can a Python
upgrade.  With the current xend code and Python, vkbd is created before vfb, and
therefore the bug can't bite.

How reproducible:
Not reproducible with the current version.

Actual results:
If it were reproducible, xend would complain:
Device 0 (vkbd) is already connected
and refuse to create the vkbd device, which breaks the pvfb.

Expected results:
Device is created, pvfb works.

Additional info: