virtio_net: fix race in RX VQ processing

[ Upstream commit cbdadbbf ]

virtio net called virtqueue_enable_cq on RX path after napi_complete, so
with NAPI_STATE_SCHED clear - outside the implicit napi lock.
This violates the requirement to synchronize virtqueue_enable_cq wrt
virtqueue_add_buf.  In particular, used event can move backwards,
causing us to lose interrupts.
In a debug build, this can trigger panic within START_USE.

Jason Wang reports that he can trigger the races artificially,
by adding udelay() in virtqueue_enable_cb() after virtio_mb().

However, we must call napi_complete to clear NAPI_STATE_SCHED before
polling the virtqueue for used buffers, otherwise napi_schedule_prep in
a callback will fail, causing us to lose RX events.

To fix, call virtqueue_enable_cb_prepare with NAPI_STATE_SCHED
set (under napi lock), later call virtqueue_poll with
NAPI_STATE_SCHED clear (outside the lock).
Reported-by: Jason Wang <jasowang@redhat.com>
Tested-by: Jason Wang <jasowang@redhat.com>
Acked-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

drivers/net/virtio_net.c

@@ -602,7 +602,7 @@ static int virtnet_poll(struct napi_struct *napi, int budget)
 		container_of(napi, struct receive_queue, napi);
 	struct virtnet_info *vi = rq->vq->vdev->priv;
 	void *buf;
-	unsigned int len, received = 0;
+	unsigned int r, len, received = 0;
 
 again:
 	while (received < budget &&
@@ -619,8 +619,9 @@ static int virtnet_poll(struct napi_struct *napi, int budget)
 
 	/* Out of packets? */
 	if (received < budget) {
+		r = virtqueue_enable_cb_prepare(rq->vq);
 		napi_complete(napi);
-		if (unlikely(!virtqueue_enable_cb(rq->vq)) &&
+		if (unlikely(virtqueue_poll(rq->vq, r)) &&
 		    napi_schedule_prep(napi)) {
 			virtqueue_disable_cb(rq->vq);
 			__napi_schedule(napi);