1 files changed, 51 insertions, 4 deletions

diff --git a/drivers/net/hyperv/netvsc.c b/drivers/net/hyperv/netvsc.c
index c64cc7639c39..9d07c9ce4be2 100644
--- a/drivers/net/hyperv/netvsc.c
+++ b/drivers/net/hyperv/netvsc.c
@@ -31,12 +31,13 @@
  * Switch the data path from the synthetic interface to the VF
  * interface.
  */
-void netvsc_switch_datapath(struct net_device *ndev, bool vf)
+int netvsc_switch_datapath(struct net_device *ndev, bool vf)
 {
 	struct net_device_context *net_device_ctx = netdev_priv(ndev);
 	struct hv_device *dev = net_device_ctx->device_ctx;
 	struct netvsc_device *nv_dev = rtnl_dereference(net_device_ctx->nvdev);
 	struct nvsp_message *init_pkt = &nv_dev->channel_init_pkt;
+	int ret, retry = 0;
 
 	/* Block sending traffic to VF if it's about to be gone */
 	if (!vf)
@@ -51,15 +52,41 @@ void netvsc_switch_datapath(struct net_device *ndev, bool vf)
 		init_pkt->msg.v4_msg.active_dp.active_datapath =
 			NVSP_DATAPATH_SYNTHETIC;
 
+again:
 	trace_nvsp_send(ndev, init_pkt);
 
-	vmbus_sendpacket(dev->channel, init_pkt,
+	ret = vmbus_sendpacket(dev->channel, init_pkt,
 			       sizeof(struct nvsp_message),
-			       (unsigned long)init_pkt,
-			       VM_PKT_DATA_INBAND,
+			       (unsigned long)init_pkt, VM_PKT_DATA_INBAND,
 			       VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
+
+	/* If failed to switch to/from VF, let data_path_is_vf stay false,
+	 * so we use synthetic path to send data.
+	 */
+	if (ret) {
+		if (ret != -EAGAIN) {
+			netdev_err(ndev,
+				   "Unable to send sw datapath msg, err: %d\n",
+				   ret);
+			return ret;
+		}
+
+		if (retry++ < RETRY_MAX) {
+			usleep_range(RETRY_US_LO, RETRY_US_HI);
+			goto again;
+		} else {
+			netdev_err(
+				ndev,
+				"Retry failed to send sw datapath msg, err: %d\n",
+				ret);
+			return ret;
+		}
+	}
+
 	wait_for_completion(&nv_dev->channel_init_wait);
 	net_device_ctx->data_path_is_vf = vf;
+
+	return 0;
 }
 
 /* Worker to setup sub channels on initial setup
@@ -1017,6 +1044,26 @@ static inline void move_pkt_msd(struct hv_netvsc_packet **msd_send,
 }
 
 /* RCU already held by caller */
+/* Batching/bouncing logic is designed to attempt to optimize
+ * performance.
+ *
+ * For small, non-LSO packets we copy the packet to a send buffer
+ * which is pre-registered with the Hyper-V side. This enables the
+ * hypervisor to avoid remapping the aperture to access the packet
+ * descriptor and data.
+ *
+ * If we already started using a buffer and the netdev is transmitting
+ * a burst of packets, keep on copying into the buffer until it is
+ * full or we are done collecting a burst. If there is an existing
+ * buffer with space for the RNDIS descriptor but not the packet, copy
+ * the RNDIS descriptor to the buffer, keeping the packet in place.
+ *
+ * If we do batching and send more than one packet using a single
+ * NetVSC message, free the SKBs of the packets copied, except for the
+ * last packet. This is done to streamline the handling of the case
+ * where the last packet only had the RNDIS descriptor copied to the
+ * send buffer, with the data pointers included in the NetVSC message.
+ */
 int netvsc_send(struct net_device *ndev,
 		struct hv_netvsc_packet *packet,
 		struct rndis_message *rndis_msg,