hv_netvsc: Add support for LRO/RSC in the vSwitch

LRO/RSC in the vSwitch is a feature available in Windows Server 2019
hosts and later. It reduces the per packet processing overhead by
coalescing multiple TCP segments when possible. This patch adds netvsc
driver support for this feature.

Signed-off-by: Haiyang Zhang <[e-mail]>
Signed-off-by: David S. Miller <[e-mail]>

drivers/net/hyperv/hyperv_net.h

···
 /* Interface */
 struct rndis_message;
 struct netvsc_device;
+struct netvsc_channel;
 struct net_device_context;
 
 extern u32 netvsc_ring_bytes;
···
 				struct rndis_message *resp);
 int netvsc_recv_callback(struct net_device *net,
 			 struct netvsc_device *nvdev,
-			 struct vmbus_channel *channel,
-			 void  *data, u32 len,
-			 const struct ndis_tcp_ip_checksum_info *csum_info,
-			 const struct ndis_pkt_8021q_info *vlan);
+			 struct netvsc_channel *nvchan);
 void netvsc_channel_cb(void *context);
 int netvsc_poll(struct napi_struct *napi, int budget);
 
···
 			       const u8 *key);
 int rndis_filter_receive(struct net_device *ndev,
 			 struct netvsc_device *net_dev,
-			 struct vmbus_channel *channel,
+			 struct netvsc_channel *nvchan,
 			 void *data, u32 buflen);
 
 int rndis_filter_set_device_mac(struct netvsc_device *ndev,
···
 			u64 ieee8021q:1;
 			u64 correlation_id:1;
 			u64 teaming:1;
+			u64 vsubnetid:1;
+			u64 rsc:1;
 		};
 	};
 } __packed;
···
 
 #define NETVSC_SUPPORTED_HW_FEATURES (NETIF_F_RXCSUM | NETIF_F_IP_CSUM | \
 				      NETIF_F_TSO | NETIF_F_IPV6_CSUM | \
-				      NETIF_F_TSO6)
+				      NETIF_F_TSO6 | NETIF_F_LRO)
 
 #define VRSS_SEND_TAB_SIZE 16  /* must be power of 2 */
 #define VRSS_CHANNEL_MAX 64
···
 	struct recv_comp_data *slots;
 	u32 first;	/* first data entry */
 	u32 next;	/* next entry for writing */
+};
+
+#define NVSP_RSC_MAX 562 /* Max #RSC frags in a vmbus xfer page pkt */
+
+struct nvsc_rsc {
+	const struct ndis_pkt_8021q_info *vlan;
+	const struct ndis_tcp_ip_checksum_info *csum_info;
+	u8 is_last; /* last RNDIS msg in a vmtransfer_page */
+	u32 cnt; /* #fragments in an RSC packet */
+	u32 pktlen; /* Full packet length */
+	void *data[NVSP_RSC_MAX];
+	u32 len[NVSP_RSC_MAX];
 };
 
 struct netvsc_stats {
···
 	struct multi_send_data msd;
 	struct multi_recv_comp mrc;
 	atomic_t queue_sends;
+	struct nvsc_rsc rsc;
 
 	struct netvsc_stats tx_stats;
 	struct netvsc_stats rx_stats;
···
 /* Packet extension field contents associated with a Data message. */
 struct rndis_per_packet_info {
 	u32 size;
-	u32 type;
+	u32 type:31;
+	u32 internal:1;
 	u32 ppi_offset;
 };
 
···
 	CACHED_NET_BUFLIST,
 	SHORT_PKT_PADINFO,
 	MAX_PER_PKT_INFO
+};
+
+enum rndis_per_pkt_info_interal_type {
+	RNDIS_PKTINFO_ID = 1,
+	/* Add more memebers here */
+
+	RNDIS_PKTINFO_MAX
+};
+
+#define RNDIS_PKTINFO_SUBALLOC BIT(0)
+#define RNDIS_PKTINFO_1ST_FRAG BIT(1)
+#define RNDIS_PKTINFO_LAST_FRAG BIT(2)
+
+#define RNDIS_PKTINFO_ID_V1 1
+
+struct rndis_pktinfo_id {
+	u8 ver;
+	u8 flag;
+	u16 pkt_id;
 };
 
 struct ndis_pkt_8021q_info {

drivers/net/hyperv/netvsc.c

···
 		init_packet->msg.v2_msg.send_ndis_config.capability.teaming = 1;
 	}
 
+	if (nvsp_ver >= NVSP_PROTOCOL_VERSION_61)
+		init_packet->msg.v2_msg.send_ndis_config.capability.rsc = 1;
+
 	trace_nvsp_send(ndev, init_packet);
 
 	ret = vmbus_sendpacket(device->channel, init_packet,
···
 
 static int netvsc_receive(struct net_device *ndev,
 			  struct netvsc_device *net_device,
-			  struct vmbus_channel *channel,
+			  struct netvsc_channel *nvchan,
 			  const struct vmpacket_descriptor *desc,
 			  const struct nvsp_message *nvsp)
 {
 	struct net_device_context *net_device_ctx = netdev_priv(ndev);
+	struct vmbus_channel *channel = nvchan->channel;
 	const struct vmtransfer_page_packet_header *vmxferpage_packet
 		= container_of(desc, const struct vmtransfer_page_packet_header, d);
 	u16 q_idx = channel->offermsg.offer.sub_channel_index;
···
 		int ret;
 
 		if (unlikely(offset + buflen > net_device->recv_buf_size)) {
+			nvchan->rsc.cnt = 0;
 			status = NVSP_STAT_FAIL;
 			netif_err(net_device_ctx, rx_err, ndev,
 				  "Packet offset:%u + len:%u too big\n",
···
 
 		data = recv_buf + offset;
 
+		nvchan->rsc.is_last = (i == count - 1);
+
 		trace_rndis_recv(ndev, q_idx, data);
 
 		/* Pass it to the upper layer */
 		ret = rndis_filter_receive(ndev, net_device,
-					   channel, data, buflen);
+					   nvchan, data, buflen);
 
 		if (unlikely(ret != NVSP_STAT_SUCCESS))
 			status = NVSP_STAT_FAIL;
···
 }
 
 static int netvsc_process_raw_pkt(struct hv_device *device,
-				  struct vmbus_channel *channel,
+				  struct netvsc_channel *nvchan,
 				  struct netvsc_device *net_device,
 				  struct net_device *ndev,
 				  const struct vmpacket_descriptor *desc,
 				  int budget)
 {
+	struct vmbus_channel *channel = nvchan->channel;
 	const struct nvsp_message *nvmsg = hv_pkt_data(desc);
 
 	trace_nvsp_recv(ndev, channel, nvmsg);
···
 		break;
 
 	case VM_PKT_DATA_USING_XFER_PAGES:
-		return netvsc_receive(ndev, net_device, channel,
+		return netvsc_receive(ndev, net_device, nvchan,
 				      desc, nvmsg);
 		break;
 
···
 		nvchan->desc = hv_pkt_iter_first(channel);
 
 	while (nvchan->desc && work_done < budget) {
-		work_done += netvsc_process_raw_pkt(device, channel, net_device,
+		work_done += netvsc_process_raw_pkt(device, nvchan, net_device,
 						    ndev, nvchan->desc, budget);
 		nvchan->desc = hv_pkt_iter_next(channel, nvchan->desc);
 	}

drivers/net/hyperv/netvsc_drv.c

···
 }
 
 static struct sk_buff *netvsc_alloc_recv_skb(struct net_device *net,
-					     struct napi_struct *napi,
-					     const struct ndis_tcp_ip_checksum_info *csum_info,
-					     const struct ndis_pkt_8021q_info *vlan,
-					     void *data, u32 buflen)
+					     struct netvsc_channel *nvchan)
 {
+	struct napi_struct *napi = &nvchan->napi;
+	const struct ndis_pkt_8021q_info *vlan = nvchan->rsc.vlan;
+	const struct ndis_tcp_ip_checksum_info *csum_info =
+						nvchan->rsc.csum_info;
 	struct sk_buff *skb;
+	int i;
 
-	skb = napi_alloc_skb(napi, buflen);
+	skb = napi_alloc_skb(napi, nvchan->rsc.pktlen);
 	if (!skb)
 		return skb;
 
···
 	 * Copy to skb. This copy is needed here since the memory pointed by
 	 * hv_netvsc_packet cannot be deallocated
 	 */
-	skb_put_data(skb, data, buflen);
+	for (i = 0; i < nvchan->rsc.cnt; i++)
+		skb_put_data(skb, nvchan->rsc.data[i], nvchan->rsc.len[i]);
 
 	skb->protocol = eth_type_trans(skb, net);
 
···
  */
 int netvsc_recv_callback(struct net_device *net,
 			 struct netvsc_device *net_device,
-			 struct vmbus_channel *channel,
-			 void  *data, u32 len,
-			 const struct ndis_tcp_ip_checksum_info *csum_info,
-			 const struct ndis_pkt_8021q_info *vlan)
+			 struct netvsc_channel *nvchan)
 {
 	struct net_device_context *net_device_ctx = netdev_priv(net);
+	struct vmbus_channel *channel = nvchan->channel;
 	u16 q_idx = channel->offermsg.offer.sub_channel_index;
-	struct netvsc_channel *nvchan = &net_device->chan_table[q_idx];
 	struct sk_buff *skb;
 	struct netvsc_stats *rx_stats;
 
···
 		return NVSP_STAT_FAIL;
 
 	/* Allocate a skb - TODO direct I/O to pages? */
-	skb = netvsc_alloc_recv_skb(net, &nvchan->napi,
-				    csum_info, vlan, data, len);
+	skb = netvsc_alloc_recv_skb(net, nvchan);
+
 	if (unlikely(!skb)) {
 		++net_device_ctx->eth_stats.rx_no_memory;
 		rcu_read_unlock();
···
 	rx_stats = &nvchan->rx_stats;
 	u64_stats_update_begin(&rx_stats->syncp);
 	rx_stats->packets++;
-	rx_stats->bytes += len;
+	rx_stats->bytes += nvchan->rsc.pktlen;
 
 	if (skb->pkt_type == PACKET_BROADCAST)
 		++rx_stats->broadcast;

drivers/net/hyperv/rndis_filter.c

···
  * Get the Per-Packet-Info with the specified type
  * return NULL if not found.
  */
-static inline void *rndis_get_ppi(struct rndis_packet *rpkt, u32 type)
+static inline void *rndis_get_ppi(struct rndis_packet *rpkt,
+				  u32 type, u8 internal)
 {
 	struct rndis_per_packet_info *ppi;
 	int len;
···
 	len = rpkt->per_pkt_info_len;
 
 	while (len > 0) {
-		if (ppi->type == type)
+		if (ppi->type == type && ppi->internal == internal)
 			return (void *)((ulong)ppi + ppi->ppi_offset);
 		len -= ppi->size;
 		ppi = (struct rndis_per_packet_info *)((ulong)ppi + ppi->size);
···
 	return NULL;
 }
 
+static inline
+void rsc_add_data(struct netvsc_channel *nvchan,
+		  const struct ndis_pkt_8021q_info *vlan,
+		  const struct ndis_tcp_ip_checksum_info *csum_info,
+		  void *data, u32 len)
+{
+	u32 cnt = nvchan->rsc.cnt;
+
+	if (cnt) {
+		nvchan->rsc.pktlen += len;
+	} else {
+		nvchan->rsc.vlan = vlan;
+		nvchan->rsc.csum_info = csum_info;
+		nvchan->rsc.pktlen = len;
+	}
+
+	nvchan->rsc.data[cnt] = data;
+	nvchan->rsc.len[cnt] = len;
+	nvchan->rsc.cnt++;
+}
+
 static int rndis_filter_receive_data(struct net_device *ndev,
 				     struct netvsc_device *nvdev,
-				     struct vmbus_channel *channel,
+				     struct netvsc_channel *nvchan,
 				     struct rndis_message *msg,
 				     u32 data_buflen)
 {
 	struct rndis_packet *rndis_pkt = &msg->msg.pkt;
 	const struct ndis_tcp_ip_checksum_info *csum_info;
 	const struct ndis_pkt_8021q_info *vlan;
+	const struct rndis_pktinfo_id *pktinfo_id;
 	u32 data_offset;
 	void *data;
+	bool rsc_more = false;
+	int ret;
 
 	/* Remove the rndis header and pass it back up the stack */
 	data_offset = RNDIS_HEADER_SIZE + rndis_pkt->data_offset;
···
 		return NVSP_STAT_FAIL;
 	}
 
-	vlan = rndis_get_ppi(rndis_pkt, IEEE_8021Q_INFO);
+	vlan = rndis_get_ppi(rndis_pkt, IEEE_8021Q_INFO, 0);
 
-	csum_info = rndis_get_ppi(rndis_pkt, TCPIP_CHKSUM_PKTINFO);
+	csum_info = rndis_get_ppi(rndis_pkt, TCPIP_CHKSUM_PKTINFO, 0);
+
+	pktinfo_id = rndis_get_ppi(rndis_pkt, RNDIS_PKTINFO_ID, 1);
 
 	data = (void *)msg + data_offset;
 
-	/*
-	 * Remove the rndis trailer padding from rndis packet message
+	/* Identify RSC frags, drop erroneous packets */
+	if (pktinfo_id && (pktinfo_id->flag & RNDIS_PKTINFO_SUBALLOC)) {
+		if (pktinfo_id->flag & RNDIS_PKTINFO_1ST_FRAG)
+			nvchan->rsc.cnt = 0;
+		else if (nvchan->rsc.cnt == 0)
+			goto drop;
+
+		rsc_more = true;
+
+		if (pktinfo_id->flag & RNDIS_PKTINFO_LAST_FRAG)
+			rsc_more = false;
+
+		if (rsc_more && nvchan->rsc.is_last)
+			goto drop;
+	} else {
+		nvchan->rsc.cnt = 0;
+	}
+
+	if (unlikely(nvchan->rsc.cnt >= NVSP_RSC_MAX))
+		goto drop;
+
+	/* Put data into per channel structure.
+	 * Also, remove the rndis trailer padding from rndis packet message
 	 * rndis_pkt->data_len tell us the real data length, we only copy
 	 * the data packet to the stack, without the rndis trailer padding
 	 */
-	return netvsc_recv_callback(ndev, nvdev, channel,
-				    data, rndis_pkt->data_len,
-				    csum_info, vlan);
+	rsc_add_data(nvchan, vlan, csum_info, data, rndis_pkt->data_len);
+
+	if (rsc_more)
+		return NVSP_STAT_SUCCESS;
+
+	ret = netvsc_recv_callback(ndev, nvdev, nvchan);
+	nvchan->rsc.cnt = 0;
+
+	return ret;
+
+drop:
+	/* Drop incomplete packet */
+	nvchan->rsc.cnt = 0;
+	return NVSP_STAT_FAIL;
 }
 
 int rndis_filter_receive(struct net_device *ndev,
 			 struct netvsc_device *net_dev,
-			 struct vmbus_channel *channel,
+			 struct netvsc_channel *nvchan,
 			 void *data, u32 buflen)
 {
 	struct net_device_context *net_device_ctx = netdev_priv(ndev);
···
 
 	switch (rndis_msg->ndis_msg_type) {
 	case RNDIS_MSG_PACKET:
-		return rndis_filter_receive_data(ndev, net_dev, channel,
+		return rndis_filter_receive_data(ndev, net_dev, nvchan,
 						 rndis_msg, buflen);
 	case RNDIS_MSG_INIT_C:
 	case RNDIS_MSG_QUERY_C:
···
 			offloads.udp_ip_v6_csum = NDIS_OFFLOAD_PARAMETERS_TX_RX_ENABLED;
 			net_device_ctx->tx_checksum_mask |= TRANSPORT_INFO_IPV6_UDP;
 		}
+	}
+
+	if (hwcaps.rsc.ip4 && hwcaps.rsc.ip6) {
+		net->hw_features |= NETIF_F_LRO;
+
+		offloads.rsc_ip_v4 = NDIS_OFFLOAD_PARAMETERS_RSC_ENABLED;
+		offloads.rsc_ip_v6 = NDIS_OFFLOAD_PARAMETERS_RSC_ENABLED;
 	}
 
 	/* In case some hw_features disappeared we need to remove them from