netfront: plug a theoretical leak on setup_device()'s error path

netfront: plug a theoretical leak on setup_device()'s error path

Suppose the frontend is in XenbusStateInitialising and the backend
advances to XenbusStateInitWait. backend_changed() will call
network_connect() -> talk_to_backend() -> setup_device().

If bind_listening_port_to_irqhandler() fails (due to event channel
allocation failure, dynamic IRQ allocation failure etc), then the
grant references and the shared ring pages remain allocated. The error
percolates back to backend_changed() without any cleanup.
backend_changed() will not change the frontend state; it will stay in
XenbusStateInitialising. If the backend retries XenbusStateInitWait
(possibly through a fake, no-op mediate state, like
XenbusStateClosed), then the frontend repeats the above call chain and
leaks the previously unreleased ring pages and grant references.

If setup_device() fails to grant access to the RX page before it tries
to call bind_listening_port_to_irqhandler(), then only the TX page /
grant reference are leaked in the next round.

I'm not sure what happens when the RX page allocation fails:
xenbus_dev_fatal() is called then, which changes the frontend's state
and seems to preclude a second immediate call to network_connect(). I
think this patch shouldn't hurt in that case either (or after
earlier failures in setup_device() for that matter).

Signed-off-by: Laszlo Ersek <lersek@redhat.com>