[dpdk-dev] [PATCH v2] net/mlx4: support hardware TSO

[Moti Haimovsky <motih@mellanox.com>]

Implement support for hardware TSO.

Signed-off-by: Moti Haimovsky <motih@mellanox.com>
---
v2:
* Fixed coding style warning.
in reply to
1530184583-30166-1-git-send-email-motih@mellanox.com

v1:
* Fixed coding style warnings.
in reply to
1530181779-19716-1-git-send-email-motih@mellanox.com
---

 doc/guides/nics/features/mlx4.ini |   1 +
 doc/guides/nics/mlx4.rst          |   3 +
 drivers/net/mlx4/mlx4.c           |  16 ++
 drivers/net/mlx4/mlx4.h           |   5 +
 drivers/net/mlx4/mlx4_prm.h       |  12 ++
 drivers/net/mlx4/mlx4_rxtx.c      | 372 +++++++++++++++++++++++++++++++++++++-
 drivers/net/mlx4/mlx4_rxtx.h      |   2 +-
 drivers/net/mlx4/mlx4_txq.c       |   8 +-
 8 files changed, 415 insertions(+), 4 deletions(-)

diff --git a/doc/guides/nics/features/mlx4.ini b/doc/guides/nics/features/mlx4.ini
index f6efd21..98a3f61 100644
--- a/doc/guides/nics/features/mlx4.ini
+++ b/doc/guides/nics/features/mlx4.ini
@@ -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
diff --git a/doc/guides/nics/mlx4.rst b/doc/guides/nics/mlx4.rst
index 491106a..12adaeb 100644
--- a/doc/guides/nics/mlx4.rst
+++ b/doc/guides/nics/mlx4.rst
@@ -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 or in rdma-core version v18 and above.
+
 Prerequisites
 -------------
 
diff --git a/drivers/net/mlx4/mlx4.c b/drivers/net/mlx4/mlx4.c
index d151a90..61b7844 100644
--- a/drivers/net/mlx4/mlx4.c
+++ b/drivers/net/mlx4/mlx4.c
@@ -519,6 +519,8 @@ struct mlx4_conf {
 		.ports.present = 0,
 	};
 	unsigned int vf;
+	struct rte_mbuf mbuf;
+	uint64_t size_test = UINT_MAX;
 	int i;
 
 	(void)pci_drv;
@@ -677,6 +679,20 @@ struct mlx4_conf {
 					IBV_RAW_PACKET_CAP_SCATTER_FCS);
 		DEBUG("FCS stripping toggling is %ssupported",
 		      priv->hw_fcs_strip ? "" : "not ");
+		/*
+		 * No TSO SIZE is defined in DPDK, need to figure it out
+		 * in order to see if we can support it.
+		 */
+		mbuf.tso_segsz = size_test;
+		priv->tso =
+			((device_attr_ex.tso_caps.max_tso >= mbuf.tso_segsz) &&
+			 (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) {
diff --git a/drivers/net/mlx4/mlx4.h b/drivers/net/mlx4/mlx4.h
index 300cb4d..742d741 100644
--- a/drivers/net/mlx4/mlx4.h
+++ b/drivers/net/mlx4/mlx4.h
@@ -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  // TODO: find the real value
+
 /** 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 TSO payload size being supported */
 	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. */
diff --git a/drivers/net/mlx4/mlx4_prm.h b/drivers/net/mlx4/mlx4_prm.h
index e15a3c1..915796b 100644
--- a/drivers/net/mlx4/mlx4_prm.h
+++ b/drivers/net/mlx4/mlx4_prm.h
@@ -40,6 +40,7 @@
 /* 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 {
@@ -97,6 +98,17 @@ struct mlx4_cq {
 	int arm_sn; /**< Rx event counter. */
 };
 
+/*
+ * 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_ {
+	__be32 mss_hdr_size;
+	__be32 header[0];
+};
+
 /**
  * Retrieve a CQE entry from a CQ.
  *
diff --git a/drivers/net/mlx4/mlx4_rxtx.c b/drivers/net/mlx4/mlx4_rxtx.c
index a92da66..992d193 100644
--- a/drivers/net/mlx4/mlx4_rxtx.c
+++ b/drivers/net/mlx4/mlx4_rxtx.c
@@ -38,10 +38,25 @@
  * 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 struct for TSO packet handling. */
+struct tso_info {
+	/* 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_sz;
+	/* Total size of the TSO entry 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 = {
 	/*
@@ -377,6 +392,349 @@ struct pv {
 }
 
 /**
+ * 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_sz = buf->l2_len + buf->l3_len + buf->l4_len;
+	if (tunneled)
+		tinfo->tso_header_sz += buf->outer_l2_len + buf->outer_l3_len;
+	if (unlikely(buf->tso_segsz == 0 || tinfo->tso_header_sz == 0)) {
+		DEBUG("%p: Invalid TSO parameters", (void *)txq);
+		return -EINVAL;
+	}
+	/* First segment must contain all TSO headers. */
+	if (unlikely(tinfo->tso_header_sz > MLX4_MAX_TSO_HEADER) ||
+		     tinfo->tso_header_sz > buf->data_len) {
+		DEBUG("%p: Invalid TSO header length", (void *)txq);
+		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_sz,
+					    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;
+	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 pv
+ *   Pointer to a stash area for saving the first 32bit word of each TXBB
+ *   used for the TSO WQE.
+ * @param pv_counter
+ *   Current location in the stash.
+ *
+ * @return
+ *   0 on success, negative value upon error.
+ */
+static inline int
+mlx4_tx_burst_fill_tso_segs(struct rte_mbuf *buf,
+			    struct txq *txq,
+			    const struct tso_info *tinfo,
+			    volatile struct mlx4_wqe_data_seg *dseg,
+			    struct pv *pv, int *pv_counter)
+{
+	uint32_t lkey;
+	int nb_segs = buf->nb_segs;
+	int nb_segs_txbb;
+	struct mlx4_sq *sq = &txq->msq;
+	struct rte_mbuf *sbuf = buf;
+	uint16_t sb_of = tinfo->tso_header_sz;
+	uint16_t data_len;
+
+	while (nb_segs > 0) {
+		/* 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;
+		/* how many dseg entries do we have in the current TXBB ? */
+		nb_segs_txbb =
+			(MLX4_TXBB_SIZE / sizeof(struct mlx4_wqe_data_seg)) -
+			((uintptr_t)dseg & (MLX4_TXBB_SIZE - 1)) /
+			sizeof(struct mlx4_wqe_data_seg);
+		switch (nb_segs_txbb) {
+		case 4:
+			/* Memory region key for this memory pool. */
+			lkey = mlx4_tx_mb2mr(txq, sbuf);
+			if (unlikely(lkey == (uint32_t)-1))
+				goto lkey_err;
+			dseg->addr =
+			    rte_cpu_to_be_64(rte_pktmbuf_mtod_offset(sbuf,
+								     uintptr_t,
+								     sb_of));
+			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.
+			 */
+			data_len = sbuf->data_len - sb_of;
+			pv[(*pv_counter)++].val =
+				rte_cpu_to_be_32(data_len ?
+						 data_len :
+						 0x80000000);
+			sb_of = 0;
+			sbuf = sbuf->next;
+			dseg++;
+			if (--nb_segs == 0)
+				break;
+			/* fallthrough */
+		case 3:
+			lkey = mlx4_tx_mb2mr(txq, sbuf);
+			if (unlikely(lkey == (uint32_t)-1))
+				goto lkey_err;
+			data_len = sbuf->data_len - sb_of;
+			mlx4_fill_tx_data_seg(dseg,
+					lkey,
+					rte_pktmbuf_mtod_offset(sbuf,
+								uintptr_t,
+								sb_of),
+					rte_cpu_to_be_32(data_len ?
+							 data_len :
+							 0x80000000));
+			sb_of = 0;
+			sbuf = sbuf->next;
+			dseg++;
+			if (--nb_segs == 0)
+				break;
+			/* fallthrough */
+		case 2:
+			lkey = mlx4_tx_mb2mr(txq, sbuf);
+			if (unlikely(lkey == (uint32_t)-1))
+				goto lkey_err;
+			data_len = sbuf->data_len - sb_of;
+			mlx4_fill_tx_data_seg(dseg,
+					lkey,
+					rte_pktmbuf_mtod_offset(sbuf,
+								uintptr_t,
+								sb_of),
+					rte_cpu_to_be_32(data_len ?
+							 data_len :
+							 0x80000000));
+			sb_of = 0;
+			sbuf = sbuf->next;
+			dseg++;
+			if (--nb_segs == 0)
+				break;
+			/* fallthrough */
+		case 1:
+			lkey = mlx4_tx_mb2mr(txq, sbuf);
+			if (unlikely(lkey == (uint32_t)-1))
+				goto lkey_err;
+			data_len = sbuf->data_len - sb_of;
+			mlx4_fill_tx_data_seg(dseg,
+					lkey,
+					rte_pktmbuf_mtod_offset(sbuf,
+								uintptr_t,
+								sb_of),
+					rte_cpu_to_be_32(data_len ?
+							 data_len :
+							 0x80000000));
+			sb_of = 0;
+			sbuf = sbuf->next;
+			dseg++;
+			--nb_segs;
+			break;
+		default:
+			/* Should never happen */
+			ERROR("%p: invalid number of txbb data segments %d",
+			      (void *)txq, nb_segs_txbb);
+			return -EINVAL;
+		}
+	}
+	return 0;
+lkey_err:
+	DEBUG("%p: unable to get MP <-> MR association",
+	      (void *)txq);
+	return -EFAULT;
+}
+
+/**
+ * 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 tseg
+ *   Pointer to the TSO header field in the TSO WQE.
+ * @param pv
+ *   Pointer to a stash area for saving the first 32bit word of each TXBB
+ *   used for the TSO WQE.
+ * @param pv_counter
+ *   Current location in the stash.
+ *
+ * @return
+ *   0 on success, negative value upon error.
+ */
+static inline int
+mlx4_tx_burst_fill_tso_hdr(struct rte_mbuf *buf,
+			   struct txq *txq,
+			   const struct tso_info *tinfo,
+			   volatile struct mlx4_wqe_lso_seg_ *tseg,
+			    struct pv *pv, int *pv_counter)
+{
+	struct mlx4_sq *sq = &txq->msq;
+	int remain_sz = tinfo->tso_header_sz;
+	char *from = rte_pktmbuf_mtod(buf, char *);
+	uint16_t txbb_avail_space;
+	int copy_sz;
+	/* 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_));
+	copy_sz = RTE_MIN(txbb_avail_space, remain_sz);
+	rte_memcpy(thdr.to, from, copy_sz);
+	remain_sz -= copy_sz;
+	while (remain_sz > 0) {
+		from += copy_sz;
+		thdr.to += copy_sz;
+		/* Start of TXBB need to 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.vto;
+		pv[(*pv_counter)++].val = *((uint32_t *)from);
+		from += sizeof(uint32_t);
+		thdr.to += sizeof(uint32_t);
+		remain_sz -= sizeof(uint32_t);
+		if (remain_sz <= 0)
+			break;
+		/* Now copy the rest */
+		txbb_avail_space = MLX4_TXBB_SIZE - sizeof(uint32_t);
+		copy_sz = RTE_MIN(txbb_avail_space, remain_sz);
+		rte_memcpy(thdr.to, from, copy_sz);
+		remain_sz -= copy_sz;
+	}
+	/* TODO: handle PID and IPID ? */
+	tseg->mss_hdr_size = rte_cpu_to_be_32((buf->tso_segsz << 16) |
+					      tinfo->tso_header_sz);
+	return 0;
+}
+
+/**
+ * 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_lso_seg_ *tseg =
+		(volatile struct mlx4_wqe_lso_seg_ *)(ctrl + 1);
+	struct mlx4_sq *sq = &txq->msq;
+	struct tso_info tinfo;
+	struct pv *pv = (struct pv *)txq->bounce_buf;
+	int pv_counter = 0;
+	int ret;
+
+	ret = mlx4_tx_burst_tso_get_params(buf, txq, &tinfo);
+	if (ret)
+		goto error;
+	ret = mlx4_tx_burst_fill_tso_hdr(buf, txq, &tinfo,
+					 tseg, pv, &pv_counter);
+	if (ret)
+		goto error;
+	/* Calculate data segment location */
+	dseg = (volatile struct mlx4_wqe_data_seg *)
+				((uintptr_t)tseg + tinfo.wqe_tso_seg_size);
+	if ((uintptr_t)dseg >= (uintptr_t)sq->eob)
+		dseg = (volatile struct mlx4_wqe_data_seg *)
+					((uintptr_t)dseg - sq->size);
+	ret = mlx4_tx_burst_fill_tso_segs(buf, txq, &tinfo,
+					  dseg, pv, &pv_counter);
+	if (ret)
+		goto error;
+	/* Write the first DWORD of each TXBB save earlier. */
+	if (pv_counter) {
+		/* Need a barrier here before writing the first TXBB word. */
+		rte_io_wmb();
+		for (--pv_counter; pv_counter  >= 0; pv_counter--)
+			*pv[pv_counter].dst = pv[pv_counter].val;
+	}
+	ctrl->fence_size = tinfo.fence_size;
+	sq->remain_size -= tinfo.wqe_size;
+	/* Align next WQE address to the next TXBB. */
+	return (volatile struct mlx4_wqe_ctrl_seg *)
+		((volatile uint8_t *)ctrl + tinfo.wqe_size);
+error:
+	txq->stats.odropped++;
+	rte_errno = ret;
+	return NULL;
+}
+
+/**
  * Write data segments of multi-segment packet.
  *
  * @param buf
@@ -569,6 +927,7 @@ struct pv {
 			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)) {
@@ -587,7 +946,16 @@ struct pv {
 			} 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;
diff --git a/drivers/net/mlx4/mlx4_rxtx.h b/drivers/net/mlx4/mlx4_rxtx.h
index 4c025e3..ffa8abf 100644
--- a/drivers/net/mlx4/mlx4_rxtx.h
+++ b/drivers/net/mlx4/mlx4_rxtx.h
@@ -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. */
diff --git a/drivers/net/mlx4/mlx4_txq.c b/drivers/net/mlx4/mlx4_txq.c
index 6edaadb..9aa7440 100644
--- a/drivers/net/mlx4/mlx4_txq.c
+++ b/drivers/net/mlx4/mlx4_txq.c
@@ -116,8 +116,14 @@
 			     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;
 }
 
-- 
1.8.3.1