cxgb4: add support for vxlan segmentation offload
authorGanesh Goudar <ganeshgr@chelsio.com>
Wed, 10 Jan 2018 12:45:26 +0000 (18:15 +0530)
committerDavid S. Miller <davem@davemloft.net>
Thu, 11 Jan 2018 15:58:56 +0000 (10:58 -0500)
add changes to t4_eth_xmit to enable vxlan segmentation
offload support.

Original work by: Santosh Rastapur <santosh@chelsio.com>
Signed-off-by: Ganesh Goudar <ganeshgr@chelsio.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/ethernet/chelsio/cxgb4/cxgb4.h
drivers/net/ethernet/chelsio/cxgb4/cxgb4_main.c
drivers/net/ethernet/chelsio/cxgb4/sge.c

index f56e893..baa67d3 100644 (file)
@@ -1314,6 +1314,7 @@ void t4_sge_start(struct adapter *adap);
 void t4_sge_stop(struct adapter *adap);
 void cxgb4_set_ethtool_ops(struct net_device *netdev);
 int cxgb4_write_rss(const struct port_info *pi, const u16 *queues);
+enum cpl_tx_tnl_lso_type cxgb_encap_offload_supported(struct sk_buff *skb);
 extern int dbfifo_int_thresh;
 
 #define for_each_port(adapter, iter) \
index 8097762..d0e08f6 100644 (file)
@@ -5210,6 +5210,10 @@ static int init_one(struct pci_dev *pdev, const struct pci_device_id *ent)
                        NETIF_F_RXCSUM | NETIF_F_RXHASH |
                        NETIF_F_HW_VLAN_CTAG_TX | NETIF_F_HW_VLAN_CTAG_RX |
                        NETIF_F_HW_TC;
+
+               if (CHELSIO_CHIP_VERSION(chip) > CHELSIO_T5)
+                       netdev->hw_features |= NETIF_F_GSO_UDP_TUNNEL;
+
                if (highdma)
                        netdev->hw_features |= NETIF_F_HIGHDMA;
                netdev->features |= netdev->hw_features;
index 922f2f9..eab781f 100644 (file)
@@ -770,12 +770,19 @@ static inline unsigned int flits_to_desc(unsigned int n)
  *     Returns whether an Ethernet packet is small enough to fit as
  *     immediate data. Return value corresponds to headroom required.
  */
-static inline int is_eth_imm(const struct sk_buff *skb)
+static inline int is_eth_imm(const struct sk_buff *skb, unsigned int chip_ver)
 {
-       int hdrlen = skb_shinfo(skb)->gso_size ?
-                       sizeof(struct cpl_tx_pkt_lso_core) : 0;
+       int hdrlen = 0;
 
-       hdrlen += sizeof(struct cpl_tx_pkt);
+       if (skb->encapsulation && skb_shinfo(skb)->gso_size &&
+           chip_ver > CHELSIO_T5) {
+               hdrlen = sizeof(struct cpl_tx_tnl_lso);
+               hdrlen += sizeof(struct cpl_tx_pkt_core);
+       } else {
+               hdrlen = skb_shinfo(skb)->gso_size ?
+                        sizeof(struct cpl_tx_pkt_lso_core) : 0;
+               hdrlen += sizeof(struct cpl_tx_pkt);
+       }
        if (skb->len <= MAX_IMM_TX_PKT_LEN - hdrlen)
                return hdrlen;
        return 0;
@@ -788,10 +795,11 @@ static inline int is_eth_imm(const struct sk_buff *skb)
  *     Returns the number of flits needed for a Tx WR for the given Ethernet
  *     packet, including the needed WR and CPL headers.
  */
-static inline unsigned int calc_tx_flits(const struct sk_buff *skb)
+static inline unsigned int calc_tx_flits(const struct sk_buff *skb,
+                                        unsigned int chip_ver)
 {
        unsigned int flits;
-       int hdrlen = is_eth_imm(skb);
+       int hdrlen = is_eth_imm(skb, chip_ver);
 
        /* If the skb is small enough, we can pump it out as a work request
         * with only immediate data.  In that case we just have to have the
@@ -810,13 +818,20 @@ static inline unsigned int calc_tx_flits(const struct sk_buff *skb)
         * with an embedded TX Packet Write CPL message.
         */
        flits = sgl_len(skb_shinfo(skb)->nr_frags + 1);
-       if (skb_shinfo(skb)->gso_size)
-               flits += (sizeof(struct fw_eth_tx_pkt_wr) +
-                         sizeof(struct cpl_tx_pkt_lso_core) +
-                         sizeof(struct cpl_tx_pkt_core)) / sizeof(__be64);
-       else
+       if (skb_shinfo(skb)->gso_size) {
+               if (skb->encapsulation && chip_ver > CHELSIO_T5)
+                       hdrlen = sizeof(struct fw_eth_tx_pkt_wr) +
+                                sizeof(struct cpl_tx_tnl_lso);
+               else
+                       hdrlen = sizeof(struct fw_eth_tx_pkt_wr) +
+                                sizeof(struct cpl_tx_pkt_lso_core);
+
+               hdrlen += sizeof(struct cpl_tx_pkt_core);
+               flits += (hdrlen / sizeof(__be64));
+       } else {
                flits += (sizeof(struct fw_eth_tx_pkt_wr) +
                          sizeof(struct cpl_tx_pkt_core)) / sizeof(__be64);
+       }
        return flits;
 }
 
@@ -827,9 +842,10 @@ static inline unsigned int calc_tx_flits(const struct sk_buff *skb)
  *     Returns the number of Tx descriptors needed for the given Ethernet
  *     packet, including the needed WR and CPL headers.
  */
-static inline unsigned int calc_tx_descs(const struct sk_buff *skb)
+static inline unsigned int calc_tx_descs(const struct sk_buff *skb,
+                                        unsigned int chip_ver)
 {
-       return flits_to_desc(calc_tx_flits(skb));
+       return flits_to_desc(calc_tx_flits(skb, chip_ver));
 }
 
 /**
@@ -1154,6 +1170,102 @@ cxgb_fcoe_offload(struct sk_buff *skb, struct adapter *adap,
 }
 #endif /* CONFIG_CHELSIO_T4_FCOE */
 
+/* Returns tunnel type if hardware supports offloading of the same.
+ * It is called only for T5 and onwards.
+ */
+enum cpl_tx_tnl_lso_type cxgb_encap_offload_supported(struct sk_buff *skb)
+{
+       u8 l4_hdr = 0;
+       enum cpl_tx_tnl_lso_type tnl_type = TX_TNL_TYPE_OPAQUE;
+       struct port_info *pi = netdev_priv(skb->dev);
+       struct adapter *adapter = pi->adapter;
+
+       if (skb->inner_protocol_type != ENCAP_TYPE_ETHER ||
+           skb->inner_protocol != htons(ETH_P_TEB))
+               return tnl_type;
+
+       switch (vlan_get_protocol(skb)) {
+       case htons(ETH_P_IP):
+               l4_hdr = ip_hdr(skb)->protocol;
+               break;
+       case htons(ETH_P_IPV6):
+               l4_hdr = ipv6_hdr(skb)->nexthdr;
+               break;
+       default:
+               return tnl_type;
+       }
+
+       switch (l4_hdr) {
+       case IPPROTO_UDP:
+               if (adapter->vxlan_port == udp_hdr(skb)->dest)
+                       tnl_type = TX_TNL_TYPE_VXLAN;
+               break;
+       default:
+               return tnl_type;
+       }
+
+       return tnl_type;
+}
+
+static inline void t6_fill_tnl_lso(struct sk_buff *skb,
+                                  struct cpl_tx_tnl_lso *tnl_lso,
+                                  enum cpl_tx_tnl_lso_type tnl_type)
+{
+       u32 val;
+       int in_eth_xtra_len;
+       int l3hdr_len = skb_network_header_len(skb);
+       int eth_xtra_len = skb_network_offset(skb) - ETH_HLEN;
+       const struct skb_shared_info *ssi = skb_shinfo(skb);
+       bool v6 = (ip_hdr(skb)->version == 6);
+
+       val = CPL_TX_TNL_LSO_OPCODE_V(CPL_TX_TNL_LSO) |
+             CPL_TX_TNL_LSO_FIRST_F |
+             CPL_TX_TNL_LSO_LAST_F |
+             (v6 ? CPL_TX_TNL_LSO_IPV6OUT_F : 0) |
+             CPL_TX_TNL_LSO_ETHHDRLENOUT_V(eth_xtra_len / 4) |
+             CPL_TX_TNL_LSO_IPHDRLENOUT_V(l3hdr_len / 4) |
+             (v6 ? 0 : CPL_TX_TNL_LSO_IPHDRCHKOUT_F) |
+             CPL_TX_TNL_LSO_IPLENSETOUT_F |
+             (v6 ? 0 : CPL_TX_TNL_LSO_IPIDINCOUT_F);
+       tnl_lso->op_to_IpIdSplitOut = htonl(val);
+
+       tnl_lso->IpIdOffsetOut = 0;
+
+       /* Get the tunnel header length */
+       val = skb_inner_mac_header(skb) - skb_mac_header(skb);
+       in_eth_xtra_len = skb_inner_network_header(skb) -
+                         skb_inner_mac_header(skb) - ETH_HLEN;
+
+       switch (tnl_type) {
+       case TX_TNL_TYPE_VXLAN:
+               tnl_lso->UdpLenSetOut_to_TnlHdrLen =
+                       htons(CPL_TX_TNL_LSO_UDPCHKCLROUT_F |
+                       CPL_TX_TNL_LSO_UDPLENSETOUT_F);
+               break;
+       default:
+               tnl_lso->UdpLenSetOut_to_TnlHdrLen = 0;
+               break;
+       }
+
+       tnl_lso->UdpLenSetOut_to_TnlHdrLen |=
+                htons(CPL_TX_TNL_LSO_TNLHDRLEN_V(val) |
+                      CPL_TX_TNL_LSO_TNLTYPE_V(tnl_type));
+
+       tnl_lso->r1 = 0;
+
+       val = CPL_TX_TNL_LSO_ETHHDRLEN_V(in_eth_xtra_len / 4) |
+             CPL_TX_TNL_LSO_IPV6_V(inner_ip_hdr(skb)->version == 6) |
+             CPL_TX_TNL_LSO_IPHDRLEN_V(skb_inner_network_header_len(skb) / 4) |
+             CPL_TX_TNL_LSO_TCPHDRLEN_V(inner_tcp_hdrlen(skb) / 4);
+       tnl_lso->Flow_to_TcpHdrLen = htonl(val);
+
+       tnl_lso->IpIdOffset = htons(0);
+
+       tnl_lso->IpIdSplit_to_Mss = htons(CPL_TX_TNL_LSO_MSS_V(ssi->gso_size));
+       tnl_lso->TCPSeqOffset = htonl(0);
+       tnl_lso->EthLenOffset_Size = htonl(CPL_TX_TNL_LSO_SIZE_V(skb->len));
+}
+
 /**
  *     t4_eth_xmit - add a packet to an Ethernet Tx queue
  *     @skb: the packet
@@ -1177,6 +1289,9 @@ netdev_tx_t t4_eth_xmit(struct sk_buff *skb, struct net_device *dev)
        bool immediate = false;
        int len, max_pkt_len;
        bool ptp_enabled = is_ptp_enabled(skb, dev);
+       unsigned int chip_ver;
+       enum cpl_tx_tnl_lso_type tnl_type = TX_TNL_TYPE_OPAQUE;
+
 #ifdef CONFIG_CHELSIO_T4_FCOE
        int err;
 #endif /* CONFIG_CHELSIO_T4_FCOE */
@@ -1227,7 +1342,8 @@ out_free: dev_kfree_skb_any(skb);
        }
 #endif /* CONFIG_CHELSIO_T4_FCOE */
 
-       flits = calc_tx_flits(skb);
+       chip_ver = CHELSIO_CHIP_VERSION(adap->params.chip);
+       flits = calc_tx_flits(skb, chip_ver);
        ndesc = flits_to_desc(flits);
        credits = txq_avail(&q->q) - ndesc;
 
@@ -1241,9 +1357,12 @@ out_free:        dev_kfree_skb_any(skb);
                return NETDEV_TX_BUSY;
        }
 
-       if (is_eth_imm(skb))
+       if (is_eth_imm(skb, chip_ver))
                immediate = true;
 
+       if (skb->encapsulation && chip_ver > CHELSIO_T5)
+               tnl_type = cxgb_encap_offload_supported(skb);
+
        if (!immediate &&
            unlikely(map_skb(adap->pdev_dev, skb, addr) < 0)) {
                q->mapping_err++;
@@ -1270,33 +1389,58 @@ out_free:       dev_kfree_skb_any(skb);
                bool v6 = (ssi->gso_type & SKB_GSO_TCPV6) != 0;
                int l3hdr_len = skb_network_header_len(skb);
                int eth_xtra_len = skb_network_offset(skb) - ETH_HLEN;
+               struct cpl_tx_tnl_lso *tnl_lso = (void *)(wr + 1);
+
+               if (tnl_type)
+                       len += sizeof(*tnl_lso);
+               else
+                       len += sizeof(*lso);
 
-               len += sizeof(*lso);
                wr->op_immdlen = htonl(FW_WR_OP_V(FW_ETH_TX_PKT_WR) |
                                       FW_WR_IMMDLEN_V(len));
-               lso->c.lso_ctrl = htonl(LSO_OPCODE_V(CPL_TX_PKT_LSO) |
-                                       LSO_FIRST_SLICE_F | LSO_LAST_SLICE_F |
-                                       LSO_IPV6_V(v6) |
-                                       LSO_ETHHDR_LEN_V(eth_xtra_len / 4) |
-                                       LSO_IPHDR_LEN_V(l3hdr_len / 4) |
-                                       LSO_TCPHDR_LEN_V(tcp_hdr(skb)->doff));
-               lso->c.ipid_ofst = htons(0);
-               lso->c.mss = htons(ssi->gso_size);
-               lso->c.seqno_offset = htonl(0);
-               if (is_t4(adap->params.chip))
-                       lso->c.len = htonl(skb->len);
-               else
-                       lso->c.len = htonl(LSO_T5_XFER_SIZE_V(skb->len));
-               cpl = (void *)(lso + 1);
+               if (tnl_type) {
+                       struct iphdr *iph = ip_hdr(skb);
 
-               if (CHELSIO_CHIP_VERSION(adap->params.chip) <= CHELSIO_T5)
-                       cntrl = TXPKT_ETHHDR_LEN_V(eth_xtra_len);
-               else
-                       cntrl = T6_TXPKT_ETHHDR_LEN_V(eth_xtra_len);
+                       t6_fill_tnl_lso(skb, tnl_lso, tnl_type);
+                       cpl = (void *)(tnl_lso + 1);
+                       /* Driver is expected to compute partial checksum that
+                        * does not include the IP Total Length.
+                        */
+                       if (iph->version == 4) {
+                               iph->check = 0;
+                               iph->tot_len = 0;
+                               iph->check = (u16)(~ip_fast_csum((u8 *)iph,
+                                                                iph->ihl));
+                       }
+                       if (skb->ip_summed == CHECKSUM_PARTIAL)
+                               cntrl = hwcsum(adap->params.chip, skb);
+               } else {
+                       lso->c.lso_ctrl = htonl(LSO_OPCODE_V(CPL_TX_PKT_LSO) |
+                                         LSO_FIRST_SLICE_F | LSO_LAST_SLICE_F |
+                                         LSO_IPV6_V(v6) |
+                                         LSO_ETHHDR_LEN_V(eth_xtra_len / 4) |
+                                         LSO_IPHDR_LEN_V(l3hdr_len / 4) |
+                                         LSO_TCPHDR_LEN_V(tcp_hdr(skb)->doff));
+                       lso->c.ipid_ofst = htons(0);
+                       lso->c.mss = htons(ssi->gso_size);
+                       lso->c.seqno_offset = htonl(0);
+                       if (is_t4(adap->params.chip))
+                               lso->c.len = htonl(skb->len);
+                       else
+                               lso->c.len =
+                                       htonl(LSO_T5_XFER_SIZE_V(skb->len));
+                       cpl = (void *)(lso + 1);
 
-               cntrl |= TXPKT_CSUM_TYPE_V(v6 ?
-                                          TX_CSUM_TCPIP6 : TX_CSUM_TCPIP) |
-                        TXPKT_IPHDR_LEN_V(l3hdr_len);
+                       if (CHELSIO_CHIP_VERSION(adap->params.chip)
+                           <= CHELSIO_T5)
+                               cntrl = TXPKT_ETHHDR_LEN_V(eth_xtra_len);
+                       else
+                               cntrl = T6_TXPKT_ETHHDR_LEN_V(eth_xtra_len);
+
+                       cntrl |= TXPKT_CSUM_TYPE_V(v6 ?
+                                TX_CSUM_TCPIP6 : TX_CSUM_TCPIP) |
+                                TXPKT_IPHDR_LEN_V(l3hdr_len);
+               }
                q->tso++;
                q->tx_cso += ssi->gso_segs;
        } else {