net: thunderx: HW TSO support for pass-2 hardware
authorSunil Goutham <sgoutham@cavium.com>
Thu, 10 Dec 2015 07:55:19 +0000 (13:25 +0530)
committerDavid S. Miller <davem@davemloft.net>
Sat, 12 Dec 2015 04:38:17 +0000 (23:38 -0500)
This adds support for offloading TCP segmentation to HW in pass-2
revision of hardware. Both driver level SW TSO for pass1.x chips
and HW TSO for pass-2 chip will co-exist. Modified SQ descriptor
structures to reflect pass-2 hw implementation.

Signed-off-by: Sunil Goutham <sgoutham@cavium.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/ethernet/cavium/thunder/nic.h
drivers/net/ethernet/cavium/thunder/nic_main.c
drivers/net/ethernet/cavium/thunder/nicvf_main.c
drivers/net/ethernet/cavium/thunder/nicvf_queues.c
drivers/net/ethernet/cavium/thunder/q_struct.h

index 39ca6744a4e68fe317f93e1318e729509c1b978e..688828865c48253d5f96e0bd4cab4c315a635e37 100644 (file)
@@ -265,6 +265,7 @@ struct nicvf {
        u8                      tns_mode:1;
        u8                      sqs_mode:1;
        u8                      loopback_supported:1;
+       bool                    hw_tso;
        u16                     mtu;
        struct queue_set        *qs;
 #define        MAX_SQS_PER_VF_SINGLE_NODE              5
@@ -489,6 +490,11 @@ static inline int nic_get_node_id(struct pci_dev *pdev)
        return ((addr >> NIC_NODE_ID_SHIFT) & NIC_NODE_ID_MASK);
 }
 
+static inline bool pass1_silicon(struct pci_dev *pdev)
+{
+       return pdev->revision < 8;
+}
+
 int nicvf_set_real_num_queues(struct net_device *netdev,
                              int tx_queues, int rx_queues);
 int nicvf_open(struct net_device *netdev);
index 4b7fd63ae57c1d26ab54d9359297c56f35c6105c..9f80de4d501604043a0a2c051847f4eaddaa66bc 100644 (file)
@@ -55,11 +55,6 @@ struct nicpf {
        bool                    irq_allocated[NIC_PF_MSIX_VECTORS];
 };
 
-static inline bool pass1_silicon(struct nicpf *nic)
-{
-       return nic->pdev->revision < 8;
-}
-
 /* Supported devices */
 static const struct pci_device_id nic_id_table[] = {
        { PCI_DEVICE(PCI_VENDOR_ID_CAVIUM, PCI_DEVICE_ID_THUNDER_NIC_PF) },
@@ -123,7 +118,7 @@ static void nic_send_msg_to_vf(struct nicpf *nic, int vf, union nic_mbx *mbx)
         * when PF writes to MBOX(1), in next revisions when
         * PF writes to MBOX(0)
         */
-       if (pass1_silicon(nic)) {
+       if (pass1_silicon(nic->pdev)) {
                /* see the comment for nic_reg_write()/nic_reg_read()
                 * functions above
                 */
@@ -400,7 +395,7 @@ static void nic_config_cpi(struct nicpf *nic, struct cpi_cfg_msg *cfg)
                        padd = cpi % 8; /* 3 bits CS out of 6bits DSCP */
 
                /* Leave RSS_SIZE as '0' to disable RSS */
-               if (pass1_silicon(nic)) {
+               if (pass1_silicon(nic->pdev)) {
                        nic_reg_write(nic, NIC_PF_CPI_0_2047_CFG | (cpi << 3),
                                      (vnic << 24) | (padd << 16) |
                                      (rssi_base + rssi));
@@ -470,7 +465,7 @@ static void nic_config_rss(struct nicpf *nic, struct rss_cfg_msg *cfg)
        }
 
        cpi_base = nic->cpi_base[cfg->vf_id];
-       if (pass1_silicon(nic))
+       if (pass1_silicon(nic->pdev))
                idx_addr = NIC_PF_CPI_0_2047_CFG;
        else
                idx_addr = NIC_PF_MPI_0_2047_CFG;
index dde8dc720cd3f3b7d513776662a4f05b1241486d..c24cb2a86a42f997c31536947ff0ad7f37bd577c 100644 (file)
@@ -525,14 +525,22 @@ static void nicvf_snd_pkt_handler(struct net_device *netdev,
                   __func__, cqe_tx->sq_qs, cqe_tx->sq_idx,
                   cqe_tx->sqe_ptr, hdr->subdesc_cnt);
 
-       nicvf_put_sq_desc(sq, hdr->subdesc_cnt + 1);
        nicvf_check_cqe_tx_errs(nic, cq, cqe_tx);
        skb = (struct sk_buff *)sq->skbuff[cqe_tx->sqe_ptr];
-       /* For TSO offloaded packets only one head SKB needs to be freed */
+       /* For TSO offloaded packets only one SQE will have a valid SKB */
        if (skb) {
+               nicvf_put_sq_desc(sq, hdr->subdesc_cnt + 1);
                prefetch(skb);
                dev_consume_skb_any(skb);
                sq->skbuff[cqe_tx->sqe_ptr] = (u64)NULL;
+       } else {
+               /* In case of HW TSO, HW sends a CQE for each segment of a TSO
+                * packet instead of a single CQE for the whole TSO packet
+                * transmitted. Each of this CQE points to the same SQE, so
+                * avoid freeing same SQE multiple times.
+                */
+               if (!nic->hw_tso)
+                       nicvf_put_sq_desc(sq, hdr->subdesc_cnt + 1);
        }
 }
 
@@ -1549,6 +1557,9 @@ static int nicvf_probe(struct pci_dev *pdev, const struct pci_device_id *ent)
 
        netdev->vlan_features = NETIF_F_SG | NETIF_F_IP_CSUM | NETIF_F_TSO;
 
+       if (!pass1_silicon(nic->pdev))
+               nic->hw_tso = true;
+
        netdev->netdev_ops = &nicvf_netdev_ops;
        netdev->watchdog_timeo = NICVF_TX_TIMEOUT;
 
index 1fbd9084333e89e542e00bf592747b7000301a1d..b11fc094f7694524424b4fc756e47a86b19c30a7 100644 (file)
@@ -925,7 +925,7 @@ static int nicvf_sq_subdesc_required(struct nicvf *nic, struct sk_buff *skb)
 {
        int subdesc_cnt = MIN_SQ_DESC_PER_PKT_XMIT;
 
-       if (skb_shinfo(skb)->gso_size) {
+       if (skb_shinfo(skb)->gso_size && !nic->hw_tso) {
                subdesc_cnt = nicvf_tso_count_subdescs(skb);
                return subdesc_cnt;
        }
@@ -940,7 +940,7 @@ static int nicvf_sq_subdesc_required(struct nicvf *nic, struct sk_buff *skb)
  * First subdescriptor for every send descriptor.
  */
 static inline void
-nicvf_sq_add_hdr_subdesc(struct snd_queue *sq, int qentry,
+nicvf_sq_add_hdr_subdesc(struct nicvf *nic, struct snd_queue *sq, int qentry,
                         int subdesc_cnt, struct sk_buff *skb, int len)
 {
        int proto;
@@ -976,6 +976,15 @@ nicvf_sq_add_hdr_subdesc(struct snd_queue *sq, int qentry,
                        break;
                }
        }
+
+       if (nic->hw_tso && skb_shinfo(skb)->gso_size) {
+               hdr->tso = 1;
+               hdr->tso_start = skb_transport_offset(skb) + tcp_hdrlen(skb);
+               hdr->tso_max_paysize = skb_shinfo(skb)->gso_size;
+               /* For non-tunneled pkts, point this to L2 ethertype */
+               hdr->inner_l3_offset = skb_network_offset(skb) - 2;
+               nic->drv_stats.tx_tso++;
+       }
 }
 
 /* SQ GATHER subdescriptor
@@ -1045,7 +1054,7 @@ static int nicvf_sq_append_tso(struct nicvf *nic, struct snd_queue *sq,
                        data_left -= size;
                        tso_build_data(skb, &tso, size);
                }
-               nicvf_sq_add_hdr_subdesc(sq, hdr_qentry,
+               nicvf_sq_add_hdr_subdesc(nic, sq, hdr_qentry,
                                         seg_subdescs - 1, skb, seg_len);
                sq->skbuff[hdr_qentry] = (u64)NULL;
                qentry = nicvf_get_nxt_sqentry(sq, qentry);
@@ -1098,11 +1107,12 @@ int nicvf_sq_append_skb(struct nicvf *nic, struct sk_buff *skb)
        qentry = nicvf_get_sq_desc(sq, subdesc_cnt);
 
        /* Check if its a TSO packet */
-       if (skb_shinfo(skb)->gso_size)
+       if (skb_shinfo(skb)->gso_size && !nic->hw_tso)
                return nicvf_sq_append_tso(nic, sq, sq_num, qentry, skb);
 
        /* Add SQ header subdesc */
-       nicvf_sq_add_hdr_subdesc(sq, qentry, subdesc_cnt - 1, skb, skb->len);
+       nicvf_sq_add_hdr_subdesc(nic, sq, qentry, subdesc_cnt - 1,
+                                skb, skb->len);
 
        /* Add SQ gather subdescs */
        qentry = nicvf_get_nxt_sqentry(sq, qentry);
index 3c1de97b1add9425f47477a9db04e740ea3d0c96..9e6d9876bfd0488361cc130b90fa7297fb9db0fb 100644 (file)
@@ -545,25 +545,28 @@ struct sq_hdr_subdesc {
        u64    subdesc_cnt:8;
        u64    csum_l4:2;
        u64    csum_l3:1;
-       u64    rsvd0:5;
+       u64    csum_inner_l4:2;
+       u64    csum_inner_l3:1;
+       u64    rsvd0:2;
        u64    l4_offset:8;
        u64    l3_offset:8;
        u64    rsvd1:4;
        u64    tot_len:20; /* W0 */
 
-       u64    tso_sdc_cont:8;
-       u64    tso_sdc_first:8;
-       u64    tso_l4_offset:8;
-       u64    tso_flags_last:12;
-       u64    tso_flags_first:12;
-       u64    rsvd2:2;
+       u64    rsvd2:24;
+       u64    inner_l4_offset:8;
+       u64    inner_l3_offset:8;
+       u64    tso_start:8;
+       u64    rsvd3:2;
        u64    tso_max_paysize:14; /* W1 */
 #elif defined(__LITTLE_ENDIAN_BITFIELD)
        u64    tot_len:20;
        u64    rsvd1:4;
        u64    l3_offset:8;
        u64    l4_offset:8;
-       u64    rsvd0:5;
+       u64    rsvd0:2;
+       u64    csum_inner_l3:1;
+       u64    csum_inner_l4:2;
        u64    csum_l3:1;
        u64    csum_l4:2;
        u64    subdesc_cnt:8;
@@ -574,12 +577,11 @@ struct sq_hdr_subdesc {
        u64    subdesc_type:4; /* W0 */
 
        u64    tso_max_paysize:14;
-       u64    rsvd2:2;
-       u64    tso_flags_first:12;
-       u64    tso_flags_last:12;
-       u64    tso_l4_offset:8;
-       u64    tso_sdc_first:8;
-       u64    tso_sdc_cont:8; /* W1 */
+       u64    rsvd3:2;
+       u64    tso_start:8;
+       u64    inner_l3_offset:8;
+       u64    inner_l4_offset:8;
+       u64    rsvd2:24; /* W1 */
 #endif
 };