xen-netfront: bug with napi_synchronize()

["Charles (Chas) Williams" <3chas3@gmail.com>]

I have seen some hangs in recent versions of the xen-netfront driver.
If I unload or unbind the driver immediately befor bringing up the
interface, I get a hang as show in the traceback here.


...
[  194.037351]  [<ffffffff810ae1a7>] ? msleep+0x27/0x30
[  194.037358]  [<ffffffffa004b48a>] ? xennet_disconnect_backend+0xfa/0x390 [xen_netfront]
[  194.037364]  [<ffffffffa004b789>] ? xennet_remove+0x19/0x60 [xen_netfront]
[  194.037368]  [<ffffffff8132971b>] ? xenbus_dev_remove+0x4b/0xa0
[  194.037375]  [<ffffffff8137bf41>] ? __device_release_driver+0x81/0x120
[  194.037378]  [<ffffffff8137bffd>] ? device_release_driver+0x1d/0x30
[  194.037382]  [<ffffffff8137b340>] ? unbind_store+0xf0/0x150
...

This offset in xennet_disconnect_backend is approximately in the area
of napi_synchronize().  napi_synchronize() does contain an msleep()
and nothing else would really sleep in this routine.

                queue->tx_evtchn = queue->rx_evtchn = 0;
                queue->tx_irq = queue->rx_irq = 0;

                napi_synchronize(&queue->napi);

                xennet_release_tx_bufs(queue);
                xennet_release_rx_bufs(queue);

Since I never started the interface, I know napi_enable() was never
called on queue->napi.  However, what is the expected behavior for
napi_synchronize() on something that was never enabled/disabled?