[dpdk-dev] [PATCH v2 2/4] net/iavf: add offload path for Tx AVX512

[Wenzhuo Lu <wenzhuo.lu@intel.com>]

Add a specific path for TX AVX512.
In this path, support the HW offload features, like,
checksum insertion, VLAN insertion.
This path is chosen automatically according to the
configuration.

Signed-off-by: Wenzhuo Lu <wenzhuo.lu@intel.com>
---
 drivers/net/iavf/iavf_rxtx.c            |  50 ++++++----
 drivers/net/iavf/iavf_rxtx.h            |  13 ++-
 drivers/net/iavf/iavf_rxtx_vec_avx512.c | 165 ++++++++++++++++++++++++++++++++
 drivers/net/iavf/iavf_rxtx_vec_common.h |  98 +++++++++++++++++--
 4 files changed, 301 insertions(+), 25 deletions(-)

diff --git a/drivers/net/iavf/iavf_rxtx.c b/drivers/net/iavf/iavf_rxtx.c
index bf1064d..4744c35 100644
--- a/drivers/net/iavf/iavf_rxtx.c
+++ b/drivers/net/iavf/iavf_rxtx.c
@@ -135,7 +135,7 @@
 static inline bool
 check_tx_vec_allow(struct iavf_tx_queue *txq)
 {
-	if (!(txq->offloads & IAVF_NO_VECTOR_FLAGS) &&
+	if (!(txq->offloads & IAVF_TX_NO_VECTOR_FLAGS) &&
 	    txq->rs_thresh >= IAVF_VPMD_TX_MAX_BURST &&
 	    txq->rs_thresh <= IAVF_VPMD_TX_MAX_FREE_BUF) {
 		PMD_INIT_LOG(DEBUG, "Vector tx can be enabled on this txq.");
@@ -2473,17 +2473,23 @@
 #ifdef RTE_ARCH_X86
 	struct iavf_tx_queue *txq;
 	int i;
+	int check_ret;
+	bool use_sse = false;
 	bool use_avx2 = false;
-#ifdef CC_AVX512_SUPPORT
 	bool use_avx512 = false;
-#endif
 
-	if (!iavf_tx_vec_dev_check(dev) &&
-			rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_128) {
-		if ((rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2) == 1 ||
-		     rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1) &&
-				rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_256)
-			use_avx2 = true;
+	check_ret = iavf_tx_vec_dev_check(dev);
+
+	if (check_ret >= 0 &&
+	    rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_128) {
+		/* SSE and AVX2 not support offload path yet. */
+		if (check_ret == IAVF_VECTOR_PATH) {
+			use_sse = true;
+			if ((rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2) == 1 ||
+			     rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1) &&
+			    rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_256)
+				use_avx2 = true;
+		}
 #ifdef CC_AVX512_SUPPORT
 		if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1 &&
 		    rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512BW) == 1 &&
@@ -2491,15 +2497,24 @@
 			use_avx512 = true;
 #endif
 
-		PMD_DRV_LOG(DEBUG, "Using %sVector Tx (port %d).",
-			    use_avx2 ? "avx2 " : "",
-			    dev->data->port_id);
-		dev->tx_pkt_burst = use_avx2 ?
-				    iavf_xmit_pkts_vec_avx2 :
-				    iavf_xmit_pkts_vec;
+		if (!use_sse && !use_avx2 && !use_avx512)
+			goto normal;
+
+		if (!use_avx512) {
+			PMD_DRV_LOG(DEBUG, "Using %sVector Tx (port %d).",
+				    use_avx2 ? "avx2 " : "",
+				    dev->data->port_id);
+			dev->tx_pkt_burst = use_avx2 ?
+					    iavf_xmit_pkts_vec_avx2 :
+					    iavf_xmit_pkts_vec;
+		}
 #ifdef CC_AVX512_SUPPORT
-		if (use_avx512)
-			dev->tx_pkt_burst = iavf_xmit_pkts_vec_avx512;
+		if (use_avx512) {
+			if (check_ret == IAVF_VECTOR_PATH)
+				dev->tx_pkt_burst = iavf_xmit_pkts_vec_avx512;
+			else
+				dev->tx_pkt_burst = iavf_xmit_pkts_vec_avx512_offload;
+		}
 #endif
 		dev->tx_pkt_prepare = NULL;
 
@@ -2521,6 +2536,7 @@
 	}
 #endif
 
+normal:
 	PMD_DRV_LOG(DEBUG, "Using Basic Tx callback (port=%d).",
 		    dev->data->port_id);
 	dev->tx_pkt_burst = iavf_xmit_pkts;
diff --git a/drivers/net/iavf/iavf_rxtx.h b/drivers/net/iavf/iavf_rxtx.h
index 06ff528..da39f78 100644
--- a/drivers/net/iavf/iavf_rxtx.h
+++ b/drivers/net/iavf/iavf_rxtx.h
@@ -23,14 +23,20 @@
 #define IAVF_VPMD_DESCS_PER_LOOP  4
 #define IAVF_VPMD_TX_MAX_FREE_BUF 64
 
-#define IAVF_NO_VECTOR_FLAGS (				 \
+#define IAVF_TX_NO_VECTOR_FLAGS (				 \
 		DEV_TX_OFFLOAD_MULTI_SEGS |		 \
+		DEV_TX_OFFLOAD_TCP_TSO)
+
+#define IAVF_TX_VECTOR_OFFLOAD (				 \
 		DEV_TX_OFFLOAD_VLAN_INSERT |		 \
+		DEV_TX_OFFLOAD_QINQ_INSERT |		 \
 		DEV_TX_OFFLOAD_SCTP_CKSUM |		 \
 		DEV_TX_OFFLOAD_UDP_CKSUM |		 \
-		DEV_TX_OFFLOAD_TCP_TSO |		 \
 		DEV_TX_OFFLOAD_TCP_CKSUM)
 
+#define IAVF_VECTOR_PATH 0
+#define IAVF_VECTOR_OFFLOAD_PATH 1
+
 #define DEFAULT_TX_RS_THRESH     32
 #define DEFAULT_TX_FREE_THRESH   32
 
@@ -487,6 +493,9 @@ uint16_t iavf_recv_scattered_pkts_vec_avx512_flex_rxd(void *rx_queue,
 						      uint16_t nb_pkts);
 uint16_t iavf_xmit_pkts_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
 				   uint16_t nb_pkts);
+uint16_t iavf_xmit_pkts_vec_avx512_offload(void *tx_queue,
+					   struct rte_mbuf **tx_pkts,
+					   uint16_t nb_pkts);
 int iavf_txq_vec_setup_avx512(struct iavf_tx_queue *txq);
 
 uint8_t iavf_proto_xtr_type_to_rxdid(uint8_t xtr_type);
diff --git a/drivers/net/iavf/iavf_rxtx_vec_avx512.c b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
index 5cb4c7c..2891a1a 100644
--- a/drivers/net/iavf/iavf_rxtx_vec_avx512.c
+++ b/drivers/net/iavf/iavf_rxtx_vec_avx512.c
@@ -1706,3 +1706,168 @@
 	txq->ops = &avx512_vec_txq_ops;
 	return 0;
 }
+
+static inline void
+iavf_vtx1_offload(volatile struct iavf_tx_desc *txdp,
+		  struct rte_mbuf *pkt, uint64_t flags)
+{
+	uint64_t high_qw =
+		(IAVF_TX_DESC_DTYPE_DATA |
+		 ((uint64_t)flags  << IAVF_TXD_QW1_CMD_SHIFT) |
+		 ((uint64_t)pkt->data_len << IAVF_TXD_QW1_TX_BUF_SZ_SHIFT));
+	iavf_txd_enable_offload(pkt, &high_qw);
+
+	__m128i descriptor = _mm_set_epi64x(high_qw,
+					    pkt->buf_iova + pkt->data_off);
+	_mm_storeu_si128((__m128i *)txdp, descriptor);
+}
+
+static inline void
+iavf_vtx_offload(volatile struct iavf_tx_desc *txdp,
+		 struct rte_mbuf **pkt, uint16_t nb_pkts,  uint64_t flags)
+{
+	const uint64_t hi_qw_tmpl = (IAVF_TX_DESC_DTYPE_DATA |
+			((uint64_t)flags  << IAVF_TXD_QW1_CMD_SHIFT));
+
+	/* if unaligned on 32-bit boundary, do one to align */
+	if (((uintptr_t)txdp & 0x1F) != 0 && nb_pkts != 0) {
+		iavf_vtx1_offload(txdp, *pkt, flags);
+		nb_pkts--, txdp++, pkt++;
+	}
+
+	/* do 4 at a time while possible, in bursts */
+	for (; nb_pkts > 3; txdp += 4, pkt += 4, nb_pkts -= 4) {
+		uint64_t hi_qw3 =
+			hi_qw_tmpl |
+			((uint64_t)pkt[3]->data_len <<
+			 IAVF_TXD_QW1_TX_BUF_SZ_SHIFT);
+		iavf_txd_enable_offload(pkt[3], &hi_qw3);
+		uint64_t hi_qw2 =
+			hi_qw_tmpl |
+			((uint64_t)pkt[2]->data_len <<
+			 IAVF_TXD_QW1_TX_BUF_SZ_SHIFT);
+		iavf_txd_enable_offload(pkt[2], &hi_qw2);
+		uint64_t hi_qw1 =
+			hi_qw_tmpl |
+			((uint64_t)pkt[1]->data_len <<
+			 IAVF_TXD_QW1_TX_BUF_SZ_SHIFT);
+		iavf_txd_enable_offload(pkt[1], &hi_qw1);
+		uint64_t hi_qw0 =
+			hi_qw_tmpl |
+			((uint64_t)pkt[0]->data_len <<
+			 IAVF_TXD_QW1_TX_BUF_SZ_SHIFT);
+		iavf_txd_enable_offload(pkt[0], &hi_qw0);
+
+		__m512i desc0_3 =
+			_mm512_set_epi64
+				(hi_qw3,
+				 pkt[3]->buf_iova + pkt[3]->data_off,
+				 hi_qw2,
+				 pkt[2]->buf_iova + pkt[2]->data_off,
+				 hi_qw1,
+				 pkt[1]->buf_iova + pkt[1]->data_off,
+				 hi_qw0,
+				 pkt[0]->buf_iova + pkt[0]->data_off);
+		_mm512_storeu_si512((void *)txdp, desc0_3);
+	}
+
+	/* do any last ones */
+	while (nb_pkts) {
+		iavf_vtx1_offload(txdp, *pkt, flags);
+		txdp++, pkt++, nb_pkts--;
+	}
+}
+
+static inline uint16_t
+iavf_xmit_fixed_burst_vec_avx512_offload(void *tx_queue,
+					 struct rte_mbuf **tx_pkts,
+					 uint16_t nb_pkts)
+{
+	struct iavf_tx_queue *txq = (struct iavf_tx_queue *)tx_queue;
+	volatile struct iavf_tx_desc *txdp;
+	struct iavf_tx_vec_entry *txep;
+	uint16_t n, nb_commit, tx_id;
+	/* bit2 is reserved and must be set to 1 according to Spec */
+	uint64_t flags = IAVF_TX_DESC_CMD_EOP | IAVF_TX_DESC_CMD_ICRC;
+	uint64_t rs = IAVF_TX_DESC_CMD_RS | flags;
+
+	/* cross rx_thresh boundary is not allowed */
+	nb_pkts = RTE_MIN(nb_pkts, txq->rs_thresh);
+
+	if (txq->nb_free < txq->free_thresh)
+		iavf_tx_free_bufs_avx512(txq);
+
+	nb_commit = nb_pkts = (uint16_t)RTE_MIN(txq->nb_free, nb_pkts);
+	if (unlikely(nb_pkts == 0))
+		return 0;
+
+	tx_id = txq->tx_tail;
+	txdp = &txq->tx_ring[tx_id];
+	txep = (void *)txq->sw_ring;
+	txep += tx_id;
+
+	txq->nb_free = (uint16_t)(txq->nb_free - nb_pkts);
+
+	n = (uint16_t)(txq->nb_tx_desc - tx_id);
+	if (nb_commit >= n) {
+		tx_backlog_entry_avx512(txep, tx_pkts, n);
+
+		iavf_vtx_offload(txdp, tx_pkts, n - 1, flags);
+		tx_pkts += (n - 1);
+		txdp += (n - 1);
+
+		iavf_vtx1_offload(txdp, *tx_pkts++, rs);
+
+		nb_commit = (uint16_t)(nb_commit - n);
+
+		tx_id = 0;
+		txq->next_rs = (uint16_t)(txq->rs_thresh - 1);
+
+		/* avoid reach the end of ring */
+		txdp = &txq->tx_ring[tx_id];
+		txep = (void *)txq->sw_ring;
+		txep += tx_id;
+	}
+
+	tx_backlog_entry_avx512(txep, tx_pkts, nb_commit);
+
+	iavf_vtx_offload(txdp, tx_pkts, nb_commit, flags);
+
+	tx_id = (uint16_t)(tx_id + nb_commit);
+	if (tx_id > txq->next_rs) {
+		txq->tx_ring[txq->next_rs].cmd_type_offset_bsz |=
+			rte_cpu_to_le_64(((uint64_t)IAVF_TX_DESC_CMD_RS) <<
+					 IAVF_TXD_QW1_CMD_SHIFT);
+		txq->next_rs =
+			(uint16_t)(txq->next_rs + txq->rs_thresh);
+	}
+
+	txq->tx_tail = tx_id;
+
+	IAVF_PCI_REG_WRITE(txq->qtx_tail, txq->tx_tail);
+
+	return nb_pkts;
+}
+
+uint16_t
+iavf_xmit_pkts_vec_avx512_offload(void *tx_queue, struct rte_mbuf **tx_pkts,
+				  uint16_t nb_pkts)
+{
+	uint16_t nb_tx = 0;
+	struct iavf_tx_queue *txq = (struct iavf_tx_queue *)tx_queue;
+
+	while (nb_pkts) {
+		uint16_t ret, num;
+
+		num = (uint16_t)RTE_MIN(nb_pkts, txq->rs_thresh);
+		ret = iavf_xmit_fixed_burst_vec_avx512_offload(tx_queue,
+							       &tx_pkts[nb_tx],
+							       num);
+		nb_tx += ret;
+		nb_pkts -= ret;
+		if (ret < num)
+			break;
+	}
+
+	return nb_tx;
+}
diff --git a/drivers/net/iavf/iavf_rxtx_vec_common.h b/drivers/net/iavf/iavf_rxtx_vec_common.h
index 46a1873..54227b6 100644
--- a/drivers/net/iavf/iavf_rxtx_vec_common.h
+++ b/drivers/net/iavf/iavf_rxtx_vec_common.h
@@ -236,14 +236,17 @@
 	if (!txq)
 		return -1;
 
-	if (txq->offloads & IAVF_NO_VECTOR_FLAGS)
-		return -1;
-
 	if (txq->rs_thresh < IAVF_VPMD_TX_MAX_BURST ||
 	    txq->rs_thresh > IAVF_VPMD_TX_MAX_FREE_BUF)
 		return -1;
 
-	return 0;
+	if (txq->offloads & IAVF_TX_NO_VECTOR_FLAGS)
+		return -1;
+
+	if (txq->offloads & IAVF_TX_VECTOR_OFFLOAD)
+		return IAVF_VECTOR_OFFLOAD_PATH;
+
+	return IAVF_VECTOR_PATH;
 }
 
 static inline int
@@ -266,14 +269,97 @@
 {
 	int i;
 	struct iavf_tx_queue *txq;
+	int ret;
+	int result = 0;
 
 	for (i = 0; i < dev->data->nb_tx_queues; i++) {
 		txq = dev->data->tx_queues[i];
-		if (iavf_tx_vec_queue_default(txq))
+		ret = iavf_tx_vec_queue_default(txq);
+
+		if (ret < 0)
 			return -1;
+		else if (ret > result)
+			result = ret;
 	}
 
-	return 0;
+	return result;
+}
+
+/******************************************************************************
+ * If user knows a specific offload is not enabled by APP,
+ * the macro can be commented to save the effort of fast path.
+ * Currently below 2 features are supported in TX path,
+ * 1, checksum offload
+ * 2, VLAN/QINQ insertion
+ ******************************************************************************/
+#define IAVF_TX_CSUM_OFFLOAD
+#define IAVF_TX_VLAN_QINQ_OFFLOAD
+
+static inline void
+iavf_txd_enable_offload(__rte_unused struct rte_mbuf *tx_pkt,
+			uint64_t *txd_hi)
+{
+#if defined(IAVF_TX_CSUM_OFFLOAD) || defined(IAVF_TX_VLAN_QINQ_OFFLOAD)
+	uint64_t ol_flags = tx_pkt->ol_flags;
+#endif
+	uint32_t td_cmd = 0;
+#ifdef IAVF_TX_CSUM_OFFLOAD
+	uint32_t td_offset = 0;
+#endif
+
+#ifdef IAVF_TX_CSUM_OFFLOAD
+	/* Set MACLEN */
+	td_offset |= (tx_pkt->l2_len >> 1) <<
+		     IAVF_TX_DESC_LENGTH_MACLEN_SHIFT;
+
+	/* Enable L3 checksum offloads */
+	if (ol_flags & PKT_TX_IP_CKSUM) {
+		td_cmd |= IAVF_TX_DESC_CMD_IIPT_IPV4_CSUM;
+		td_offset |= (tx_pkt->l3_len >> 2) <<
+			     IAVF_TX_DESC_LENGTH_IPLEN_SHIFT;
+	} else if (ol_flags & PKT_TX_IPV4) {
+		td_cmd |= IAVF_TX_DESC_CMD_IIPT_IPV4;
+		td_offset |= (tx_pkt->l3_len >> 2) <<
+			     IAVF_TX_DESC_LENGTH_IPLEN_SHIFT;
+	} else if (ol_flags & PKT_TX_IPV6) {
+		td_cmd |= IAVF_TX_DESC_CMD_IIPT_IPV6;
+		td_offset |= (tx_pkt->l3_len >> 2) <<
+			     IAVF_TX_DESC_LENGTH_IPLEN_SHIFT;
+	}
+
+	/* Enable L4 checksum offloads */
+	switch (ol_flags & PKT_TX_L4_MASK) {
+	case PKT_TX_TCP_CKSUM:
+		td_cmd |= IAVF_TX_DESC_CMD_L4T_EOFT_TCP;
+		td_offset |= (sizeof(struct rte_tcp_hdr) >> 2) <<
+			     IAVF_TX_DESC_LENGTH_L4_FC_LEN_SHIFT;
+		break;
+	case PKT_TX_SCTP_CKSUM:
+		td_cmd |= IAVF_TX_DESC_CMD_L4T_EOFT_SCTP;
+		td_offset |= (sizeof(struct rte_sctp_hdr) >> 2) <<
+			     IAVF_TX_DESC_LENGTH_L4_FC_LEN_SHIFT;
+		break;
+	case PKT_TX_UDP_CKSUM:
+		td_cmd |= IAVF_TX_DESC_CMD_L4T_EOFT_UDP;
+		td_offset |= (sizeof(struct rte_udp_hdr) >> 2) <<
+			     IAVF_TX_DESC_LENGTH_L4_FC_LEN_SHIFT;
+		break;
+	default:
+		break;
+	}
+
+	*txd_hi |= ((uint64_t)td_offset) << IAVF_TXD_QW1_OFFSET_SHIFT;
+#endif
+
+#ifdef IAVF_TX_VLAN_QINQ_OFFLOAD
+	if (ol_flags & (PKT_TX_VLAN | PKT_TX_QINQ)) {
+		td_cmd |= IAVF_TX_DESC_CMD_IL2TAG1;
+		*txd_hi |= ((uint64_t)tx_pkt->vlan_tci <<
+			    IAVF_TXD_QW1_L2TAG1_SHIFT);
+	}
+#endif
+
+	*txd_hi |= ((uint64_t)td_cmd) << IAVF_TXD_QW1_CMD_SHIFT;
 }
 
 #endif
-- 
1.9.3