net/mlx4: support hardware TSO
authorMoti Haimovsky <motih@mellanox.com>
Tue, 10 Jul 2018 10:45:54 +0000 (13:45 +0300)
committerShahaf Shuler <shahafs@mellanox.com>
Tue, 10 Jul 2018 12:02:57 +0000 (14:02 +0200)
Implement support for hardware TSO.

Signed-off-by: Moti Haimovsky <motih@mellanox.com>
Acked-by: Matan Azrad <matan@mellanox.com>
doc/guides/nics/features/mlx4.ini
doc/guides/nics/mlx4.rst
drivers/net/mlx4/Makefile
drivers/net/mlx4/mlx4.c
drivers/net/mlx4/mlx4.h
drivers/net/mlx4/mlx4_prm.h
drivers/net/mlx4/mlx4_rxtx.c
drivers/net/mlx4/mlx4_rxtx.h
drivers/net/mlx4/mlx4_txq.c

index f6efd21..98a3f61 100644 (file)
@@ -13,6 +13,7 @@ Queue start/stop     = Y
 MTU update           = Y
 Jumbo frame          = Y
 Scattered Rx         = Y
+TSO                  = Y
 Promiscuous mode     = Y
 Allmulticast mode    = Y
 Unicast MAC filter   = Y
index 491106a..7cfbda3 100644 (file)
@@ -142,6 +142,9 @@ Limitations
   The ability to enable/disable CRC stripping requires OFED version
   4.3-1.5.0.0 and above  or rdma-core version v18 and above.
 
+- TSO (Transmit Segmentation Offload) is supported in OFED version
+  4.4 and above.
+
 Prerequisites
 -------------
 
index 73f9d40..63bc003 100644 (file)
@@ -85,6 +85,11 @@ mlx4_autoconf.h.new: FORCE
 mlx4_autoconf.h.new: $(RTE_SDK)/buildtools/auto-config-h.sh
        $Q $(RM) -f -- '$@'
        $Q : > '$@'
+       $Q sh -- '$<' '$@' \
+               HAVE_IBV_MLX4_WQE_LSO_SEG \
+               infiniband/mlx4dv.h \
+               type 'struct mlx4_wqe_lso_seg' \
+               $(AUTOCONF_OUTPUT)
 
 # Create mlx4_autoconf.h or update it in case it differs from the new one.
 
index 06921c9..c67d2c0 100644 (file)
@@ -677,6 +677,15 @@ mlx4_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
                                        IBV_RAW_PACKET_CAP_SCATTER_FCS);
                DEBUG("FCS stripping toggling is %ssupported",
                      priv->hw_fcs_strip ? "" : "not ");
+               priv->tso =
+                       ((device_attr_ex.tso_caps.max_tso > 0) &&
+                        (device_attr_ex.tso_caps.supported_qpts &
+                         (1 << IBV_QPT_RAW_PACKET)));
+               if (priv->tso)
+                       priv->tso_max_payload_sz =
+                                       device_attr_ex.tso_caps.max_tso;
+               DEBUG("TSO is %ssupported",
+                     priv->tso ? "" : "not ");
                /* Configure the first MAC address by default. */
                err = mlx4_get_mac(priv, &mac.addr_bytes);
                if (err) {
index 300cb4d..89d8c38 100644 (file)
@@ -47,6 +47,9 @@
 /** Interrupt alarm timeout value in microseconds. */
 #define MLX4_INTR_ALARM_TIMEOUT 100000
 
+/* Maximum packet headers size (L2+L3+L4) for TSO. */
+#define MLX4_MAX_TSO_HEADER 192
+
 /** Port parameter. */
 #define MLX4_PMD_PORT_KVARG "port"
 
@@ -90,6 +93,8 @@ struct priv {
        uint32_t hw_csum:1; /**< Checksum offload is supported. */
        uint32_t hw_csum_l2tun:1; /**< Checksum support for L2 tunnels. */
        uint32_t hw_fcs_strip:1; /**< FCS stripping toggling is supported. */
+       uint32_t tso:1; /**< Transmit segmentation offload is supported. */
+       uint32_t tso_max_payload_sz; /**< Max supported TSO payload size. */
        uint64_t hw_rss_sup; /**< Supported RSS hash fields (Verbs format). */
        struct rte_intr_handle intr_handle; /**< Port interrupt handle. */
        struct mlx4_drop *drop; /**< Shared resources for drop flow rules. */
index b771d8c..aef77ba 100644 (file)
@@ -19,6 +19,7 @@
 #ifdef PEDANTIC
 #pragma GCC diagnostic error "-Wpedantic"
 #endif
+#include "mlx4_autoconf.h"
 
 /* ConnectX-3 Tx queue basic block. */
 #define MLX4_TXBB_SHIFT 6
@@ -40,6 +41,7 @@
 /* Work queue element (WQE) flags. */
 #define MLX4_WQE_CTRL_IIP_HDR_CSUM (1 << 28)
 #define MLX4_WQE_CTRL_IL4_HDR_CSUM (1 << 27)
+#define MLX4_WQE_CTRL_RR (1 << 6)
 
 /* CQE checksum flags. */
 enum {
@@ -98,6 +100,19 @@ struct mlx4_cq {
        int arm_sn; /**< Rx event counter. */
 };
 
+#ifndef HAVE_IBV_MLX4_WQE_LSO_SEG
+/*
+ * WQE LSO segment structure.
+ * Defined here as backward compatibility for rdma-core v17 and below.
+ * Similar definition is found in infiniband/mlx4dv.h in rdma-core v18
+ * and above.
+ */
+struct mlx4_wqe_lso_seg {
+       rte_be32_t mss_hdr_size;
+       rte_be32_t header[];
+};
+#endif
+
 /**
  * Retrieve a CQE entry from a CQ.
  *
index 78b6dd5..8c88eff 100644 (file)
  * DWORD (32 byte) of a TXBB.
  */
 struct pv {
-       volatile struct mlx4_wqe_data_seg *dseg;
+       union {
+               volatile struct mlx4_wqe_data_seg *dseg;
+               volatile uint32_t *dst;
+       };
        uint32_t val;
 };
 
+/** A helper structure for TSO packet handling. */
+struct tso_info {
+       /** Pointer to the array of saved first DWORD (32 byte) of a TXBB. */
+       struct pv *pv;
+       /** Current entry in the pv array. */
+       int pv_counter;
+       /** Total size of the WQE including padding. */
+       uint32_t wqe_size;
+       /** Size of TSO header to prepend to each packet to send. */
+       uint16_t tso_header_size;
+       /** Total size of the TSO segment in the WQE. */
+       uint16_t wqe_tso_seg_size;
+       /** Raw WQE size in units of 16 Bytes and without padding. */
+       uint8_t fence_size;
+};
+
 /** A table to translate Rx completion flags to packet type. */
 uint32_t mlx4_ptype_table[0x100] __rte_cache_aligned = {
        /*
@@ -367,6 +386,342 @@ mlx4_fill_tx_data_seg(volatile struct mlx4_wqe_data_seg *dseg,
        dseg->byte_count = byte_count;
 }
 
+/**
+ * Obtain and calculate TSO information needed for assembling a TSO WQE.
+ *
+ * @param buf
+ *   Pointer to the first packet mbuf.
+ * @param txq
+ *   Pointer to Tx queue structure.
+ * @param tinfo
+ *   Pointer to a structure to fill the info with.
+ *
+ * @return
+ *   0 on success, negative value upon error.
+ */
+static inline int
+mlx4_tx_burst_tso_get_params(struct rte_mbuf *buf,
+                            struct txq *txq,
+                            struct tso_info *tinfo)
+{
+       struct mlx4_sq *sq = &txq->msq;
+       const uint8_t tunneled = txq->priv->hw_csum_l2tun &&
+                                (buf->ol_flags & PKT_TX_TUNNEL_MASK);
+
+       tinfo->tso_header_size = buf->l2_len + buf->l3_len + buf->l4_len;
+       if (tunneled)
+               tinfo->tso_header_size +=
+                               buf->outer_l2_len + buf->outer_l3_len;
+       if (unlikely(buf->tso_segsz == 0 ||
+                    tinfo->tso_header_size == 0 ||
+                    tinfo->tso_header_size > MLX4_MAX_TSO_HEADER ||
+                    tinfo->tso_header_size > buf->data_len))
+               return -EINVAL;
+       /*
+        * Calculate the WQE TSO segment size
+        * Note:
+        * 1. An LSO segment must be padded such that the subsequent data
+        *    segment is 16-byte aligned.
+        * 2. The start address of the TSO segment is always 16 Bytes aligned.
+        */
+       tinfo->wqe_tso_seg_size = RTE_ALIGN(sizeof(struct mlx4_wqe_lso_seg) +
+                                           tinfo->tso_header_size,
+                                           sizeof(struct mlx4_wqe_data_seg));
+       tinfo->fence_size = ((sizeof(struct mlx4_wqe_ctrl_seg) +
+                            tinfo->wqe_tso_seg_size) >> MLX4_SEG_SHIFT) +
+                            buf->nb_segs;
+       tinfo->wqe_size =
+               RTE_ALIGN((uint32_t)(tinfo->fence_size << MLX4_SEG_SHIFT),
+                         MLX4_TXBB_SIZE);
+       /* Validate WQE size and WQE space in the send queue. */
+       if (sq->remain_size < tinfo->wqe_size ||
+           tinfo->wqe_size > MLX4_MAX_WQE_SIZE)
+               return -ENOMEM;
+       /* Init pv. */
+       tinfo->pv = (struct pv *)txq->bounce_buf;
+       tinfo->pv_counter = 0;
+       return 0;
+}
+
+/**
+ * Fill the TSO WQE data segments with info on buffers to transmit .
+ *
+ * @param buf
+ *   Pointer to the first packet mbuf.
+ * @param txq
+ *   Pointer to Tx queue structure.
+ * @param tinfo
+ *   Pointer to TSO info to use.
+ * @param dseg
+ *   Pointer to the first data segment in the TSO WQE.
+ * @param ctrl
+ *   Pointer to the control segment in the TSO WQE.
+ *
+ * @return
+ *   0 on success, negative value upon error.
+ */
+static inline volatile struct mlx4_wqe_ctrl_seg *
+mlx4_tx_burst_fill_tso_dsegs(struct rte_mbuf *buf,
+                            struct txq *txq,
+                            struct tso_info *tinfo,
+                            volatile struct mlx4_wqe_data_seg *dseg,
+                            volatile struct mlx4_wqe_ctrl_seg *ctrl)
+{
+       uint32_t lkey;
+       int nb_segs = buf->nb_segs;
+       int nb_segs_txbb;
+       struct mlx4_sq *sq = &txq->msq;
+       struct rte_mbuf *sbuf = buf;
+       struct pv *pv = tinfo->pv;
+       int *pv_counter = &tinfo->pv_counter;
+       volatile struct mlx4_wqe_ctrl_seg *ctrl_next =
+                       (volatile struct mlx4_wqe_ctrl_seg *)
+                               ((volatile uint8_t *)ctrl + tinfo->wqe_size);
+       uint16_t data_len = sbuf->data_len - tinfo->tso_header_size;
+       uintptr_t data_addr = rte_pktmbuf_mtod_offset(sbuf, uintptr_t,
+                                                     tinfo->tso_header_size);
+
+       do {
+               /* how many dseg entries do we have in the current TXBB ? */
+               nb_segs_txbb = (MLX4_TXBB_SIZE -
+                               ((uintptr_t)dseg & (MLX4_TXBB_SIZE - 1))) >>
+                              MLX4_SEG_SHIFT;
+               switch (nb_segs_txbb) {
+#ifndef NDEBUG
+               default:
+                       /* Should never happen. */
+                       rte_panic("%p: Invalid number of SGEs(%d) for a TXBB",
+                       (void *)txq, nb_segs_txbb);
+                       /* rte_panic never returns. */
+                       break;
+#endif /* NDEBUG */
+               case 4:
+                       /* Memory region key for this memory pool. */
+                       lkey = mlx4_tx_mb2mr(txq, sbuf);
+                       if (unlikely(lkey == (uint32_t)-1))
+                               goto err;
+                       dseg->addr = rte_cpu_to_be_64(data_addr);
+                       dseg->lkey = lkey;
+                       /*
+                        * This data segment starts at the beginning of a new
+                        * TXBB, so we need to postpone its byte_count writing
+                        * for later.
+                        */
+                       pv[*pv_counter].dseg = dseg;
+                       /*
+                        * Zero length segment is treated as inline segment
+                        * with zero data.
+                        */
+                       pv[(*pv_counter)++].val =
+                               rte_cpu_to_be_32(data_len ?
+                                                data_len :
+                                                0x80000000);
+                       if (--nb_segs == 0)
+                               return ctrl_next;
+                       /* Prepare next buf info */
+                       sbuf = sbuf->next;
+                       dseg++;
+                       data_len = sbuf->data_len;
+                       data_addr = rte_pktmbuf_mtod(sbuf, uintptr_t);
+                       /* fallthrough */
+               case 3:
+                       lkey = mlx4_tx_mb2mr(txq, sbuf);
+                       if (unlikely(lkey == (uint32_t)-1))
+                               goto err;
+                       mlx4_fill_tx_data_seg(dseg, lkey, data_addr,
+                                       rte_cpu_to_be_32(data_len ?
+                                                        data_len :
+                                                        0x80000000));
+                       if (--nb_segs == 0)
+                               return ctrl_next;
+                       /* Prepare next buf info */
+                       sbuf = sbuf->next;
+                       dseg++;
+                       data_len = sbuf->data_len;
+                       data_addr = rte_pktmbuf_mtod(sbuf, uintptr_t);
+                       /* fallthrough */
+               case 2:
+                       lkey = mlx4_tx_mb2mr(txq, sbuf);
+                       if (unlikely(lkey == (uint32_t)-1))
+                               goto err;
+                       mlx4_fill_tx_data_seg(dseg, lkey, data_addr,
+                                       rte_cpu_to_be_32(data_len ?
+                                                        data_len :
+                                                        0x80000000));
+                       if (--nb_segs == 0)
+                               return ctrl_next;
+                       /* Prepare next buf info */
+                       sbuf = sbuf->next;
+                       dseg++;
+                       data_len = sbuf->data_len;
+                       data_addr = rte_pktmbuf_mtod(sbuf, uintptr_t);
+                       /* fallthrough */
+               case 1:
+                       lkey = mlx4_tx_mb2mr(txq, sbuf);
+                       if (unlikely(lkey == (uint32_t)-1))
+                               goto err;
+                       mlx4_fill_tx_data_seg(dseg, lkey, data_addr,
+                                       rte_cpu_to_be_32(data_len ?
+                                                        data_len :
+                                                        0x80000000));
+                       if (--nb_segs == 0)
+                               return ctrl_next;
+                       /* Prepare next buf info */
+                       sbuf = sbuf->next;
+                       dseg++;
+                       data_len = sbuf->data_len;
+                       data_addr = rte_pktmbuf_mtod(sbuf, uintptr_t);
+                       /* fallthrough */
+               }
+               /* Wrap dseg if it points at the end of the queue. */
+               if ((volatile uint8_t *)dseg >= sq->eob)
+                       dseg = (volatile struct mlx4_wqe_data_seg *)
+                                       ((volatile uint8_t *)dseg - sq->size);
+       } while (true);
+err:
+       return NULL;
+}
+
+/**
+ * Fill the packet's l2, l3 and l4 headers to the WQE.
+ *
+ * This will be used as the header for each TSO segment that is transmitted.
+ *
+ * @param buf
+ *   Pointer to the first packet mbuf.
+ * @param txq
+ *   Pointer to Tx queue structure.
+ * @param tinfo
+ *   Pointer to TSO info to use.
+ * @param ctrl
+ *   Pointer to the control segment in the TSO WQE.
+ *
+ * @return
+ *   0 on success, negative value upon error.
+ */
+static inline volatile struct mlx4_wqe_data_seg *
+mlx4_tx_burst_fill_tso_hdr(struct rte_mbuf *buf,
+                          struct txq *txq,
+                          struct tso_info *tinfo,
+                          volatile struct mlx4_wqe_ctrl_seg *ctrl)
+{
+       volatile struct mlx4_wqe_lso_seg *tseg =
+               (volatile struct mlx4_wqe_lso_seg *)(ctrl + 1);
+       struct mlx4_sq *sq = &txq->msq;
+       struct pv *pv = tinfo->pv;
+       int *pv_counter = &tinfo->pv_counter;
+       int remain_size = tinfo->tso_header_size;
+       char *from = rte_pktmbuf_mtod(buf, char *);
+       uint16_t txbb_avail_space;
+       /* Union to overcome volatile constraints when copying TSO header. */
+       union {
+               volatile uint8_t *vto;
+               uint8_t *to;
+       } thdr = { .vto = (volatile uint8_t *)tseg->header, };
+
+       /*
+        * TSO data always starts at offset 20 from the beginning of the TXBB
+        * (16 byte ctrl + 4byte TSO desc). Since each TXBB is 64Byte aligned
+        * we can write the first 44 TSO header bytes without worry for TxQ
+        * wrapping or overwriting the first TXBB 32bit word.
+        */
+       txbb_avail_space = MLX4_TXBB_SIZE -
+                          (sizeof(struct mlx4_wqe_ctrl_seg) +
+                           sizeof(struct mlx4_wqe_lso_seg));
+       while (remain_size >= (int)(txbb_avail_space + sizeof(uint32_t))) {
+               /* Copy to end of txbb. */
+               rte_memcpy(thdr.to, from, txbb_avail_space);
+               from += txbb_avail_space;
+               thdr.to += txbb_avail_space;
+               /* New TXBB, Check for TxQ wrap. */
+               if (thdr.to >= sq->eob)
+                       thdr.vto = sq->buf;
+               /* New TXBB, stash the first 32bits for later use. */
+               pv[*pv_counter].dst = (volatile uint32_t *)thdr.to;
+               pv[(*pv_counter)++].val = *(uint32_t *)from,
+               from += sizeof(uint32_t);
+               thdr.to += sizeof(uint32_t);
+               remain_size -= txbb_avail_space + sizeof(uint32_t);
+               /* Avail space in new TXBB is TXBB size - 4 */
+               txbb_avail_space = MLX4_TXBB_SIZE - sizeof(uint32_t);
+       }
+       if (remain_size > txbb_avail_space) {
+               rte_memcpy(thdr.to, from, txbb_avail_space);
+               from += txbb_avail_space;
+               thdr.to += txbb_avail_space;
+               remain_size -= txbb_avail_space;
+               /* New TXBB, Check for TxQ wrap. */
+               if (thdr.to >= sq->eob)
+                       thdr.vto = sq->buf;
+               pv[*pv_counter].dst = (volatile uint32_t *)thdr.to;
+               rte_memcpy(&pv[*pv_counter].val, from, remain_size);
+               (*pv_counter)++;
+       } else if (remain_size) {
+               rte_memcpy(thdr.to, from, remain_size);
+       }
+       tseg->mss_hdr_size = rte_cpu_to_be_32((buf->tso_segsz << 16) |
+                                             tinfo->tso_header_size);
+       /* Calculate data segment location */
+       return (volatile struct mlx4_wqe_data_seg *)
+                               ((uintptr_t)tseg + tinfo->wqe_tso_seg_size);
+}
+
+/**
+ * Write data segments and header for TSO uni/multi segment packet.
+ *
+ * @param buf
+ *   Pointer to the first packet mbuf.
+ * @param txq
+ *   Pointer to Tx queue structure.
+ * @param ctrl
+ *   Pointer to the WQE control segment.
+ *
+ * @return
+ *   Pointer to the next WQE control segment on success, NULL otherwise.
+ */
+static volatile struct mlx4_wqe_ctrl_seg *
+mlx4_tx_burst_tso(struct rte_mbuf *buf, struct txq *txq,
+                 volatile struct mlx4_wqe_ctrl_seg *ctrl)
+{
+       volatile struct mlx4_wqe_data_seg *dseg;
+       volatile struct mlx4_wqe_ctrl_seg *ctrl_next;
+       struct mlx4_sq *sq = &txq->msq;
+       struct tso_info tinfo;
+       struct pv *pv;
+       int pv_counter;
+       int ret;
+
+       ret = mlx4_tx_burst_tso_get_params(buf, txq, &tinfo);
+       if (unlikely(ret))
+               goto error;
+       dseg = mlx4_tx_burst_fill_tso_hdr(buf, txq, &tinfo, ctrl);
+       if (unlikely(dseg == NULL))
+               goto error;
+       if ((uintptr_t)dseg >= (uintptr_t)sq->eob)
+               dseg = (volatile struct mlx4_wqe_data_seg *)
+                                       ((uintptr_t)dseg - sq->size);
+       ctrl_next = mlx4_tx_burst_fill_tso_dsegs(buf, txq, &tinfo, dseg, ctrl);
+       if (unlikely(ctrl_next == NULL))
+               goto error;
+       /* Write the first DWORD of each TXBB save earlier. */
+       if (likely(tinfo.pv_counter)) {
+               pv = tinfo.pv;
+               pv_counter = tinfo.pv_counter;
+               /* Need a barrier here before writing the first TXBB word. */
+               rte_io_wmb();
+               do {
+                       --pv_counter;
+                       *pv[pv_counter].dst = pv[pv_counter].val;
+               } while (pv_counter > 0);
+       }
+       ctrl->fence_size = tinfo.fence_size;
+       sq->remain_size -= tinfo.wqe_size;
+       return ctrl_next;
+error:
+       txq->stats.odropped++;
+       return NULL;
+}
+
 /**
  * Write data segments of multi-segment packet.
  *
@@ -560,6 +915,7 @@ mlx4_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
                        uint16_t flags16[2];
                } srcrb;
                uint32_t lkey;
+               bool tso = txq->priv->tso && (buf->ol_flags & PKT_TX_TCP_SEG);
 
                /* Clean up old buffer. */
                if (likely(elt->buf != NULL)) {
@@ -578,7 +934,16 @@ mlx4_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
                        } while (tmp != NULL);
                }
                RTE_MBUF_PREFETCH_TO_FREE(elt_next->buf);
-               if (buf->nb_segs == 1) {
+               if (tso) {
+                       /* Change opcode to TSO */
+                       owner_opcode &= ~MLX4_OPCODE_CONFIG_CMD;
+                       owner_opcode |= MLX4_OPCODE_LSO | MLX4_WQE_CTRL_RR;
+                       ctrl_next = mlx4_tx_burst_tso(buf, txq, ctrl);
+                       if (!ctrl_next) {
+                               elt->buf = NULL;
+                               break;
+                       }
+               } else if (buf->nb_segs == 1) {
                        /* Validate WQE space in the send queue. */
                        if (sq->remain_size < MLX4_TXBB_SIZE) {
                                elt->buf = NULL;
index 4c025e3..ffa8abf 100644 (file)
@@ -90,7 +90,7 @@ struct mlx4_txq_stats {
        unsigned int idx; /**< Mapping index. */
        uint64_t opackets; /**< Total of successfully sent packets. */
        uint64_t obytes; /**< Total of successfully sent bytes. */
-       uint64_t odropped; /**< Total of packets not sent when Tx ring full. */
+       uint64_t odropped; /**< Total number of packets failed to transmit. */
 };
 
 /** Tx queue descriptor. */
index 6edaadb..9aa7440 100644 (file)
@@ -116,8 +116,14 @@ mlx4_get_tx_port_offloads(struct priv *priv)
                             DEV_TX_OFFLOAD_UDP_CKSUM |
                             DEV_TX_OFFLOAD_TCP_CKSUM);
        }
-       if (priv->hw_csum_l2tun)
+       if (priv->tso)
+               offloads |= DEV_TX_OFFLOAD_TCP_TSO;
+       if (priv->hw_csum_l2tun) {
                offloads |= DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM;
+               if (priv->tso)
+                       offloads |= (DEV_TX_OFFLOAD_VXLAN_TNL_TSO |
+                                    DEV_TX_OFFLOAD_GRE_TNL_TSO);
+       }
        return offloads;
 }