net/ice: add Tx AVX512 offload path

Add alternative Tx data path for AVX512 which can support partial
Tx offload features, including Tx checksum offload, vlan/QinQ
insertion offload.

Signed-off-by: Leyi Rong <leyi.rong@intel.com>
Signed-off-by: Wenzhuo Lu <wenzhuo.lu@intel.com>
Tested-by: Qin Sun <qinx.sun@intel.com>

drivers/net/ice/ice_rxtx.c

@@ -8,6 +8,7 @@
 
 #include "rte_pmd_ice.h"
 #include "ice_rxtx.h"
+#include "ice_rxtx_vec_common.h"
 
 #define ICE_TX_CKSUM_OFFLOAD_MASK (		 \
 		PKT_TX_IP_CKSUM |		 \
@@ -3267,12 +3268,14 @@ ice_set_tx_function(struct rte_eth_dev *dev)
 #ifdef RTE_ARCH_X86
 	struct ice_tx_queue *txq;
 	int i;
+	int tx_check_ret;
 	bool use_avx512 = false;
 	bool use_avx2 = false;
 
 	if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
-		if (!ice_tx_vec_dev_check(dev) &&
-				rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_128) {
+		tx_check_ret = ice_tx_vec_dev_check(dev);
+		if (tx_check_ret >= 0 &&
+		    rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_128) {
 			ad->tx_vec_allowed = true;
 			for (i = 0; i < dev->data->nb_tx_queues; i++) {
 				txq = dev->data->tx_queues[i];
@@ -3291,12 +3294,15 @@ ice_set_tx_function(struct rte_eth_dev *dev)
 			PMD_DRV_LOG(NOTICE,
 				"AVX512 is not supported in build env");
 #endif
-			if (!use_avx512 &&
+			if (!use_avx512 && tx_check_ret == ICE_VECTOR_PATH &&
 			(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;
 
+			if (!use_avx512 && tx_check_ret == ICE_VECTOR_OFFLOAD_PATH)
+				ad->tx_vec_allowed = false;
+
 		} else {
 			ad->tx_vec_allowed = false;
 		}
@@ -3305,9 +3311,18 @@ ice_set_tx_function(struct rte_eth_dev *dev)
 	if (ad->tx_vec_allowed) {
 		if (use_avx512) {
 #ifdef CC_AVX512_SUPPORT
-			PMD_DRV_LOG(NOTICE, "Using AVX512 Vector Tx (port %d).",
-				    dev->data->port_id);
-			dev->tx_pkt_burst = ice_xmit_pkts_vec_avx512;
+			if (tx_check_ret == ICE_VECTOR_OFFLOAD_PATH) {
+				PMD_DRV_LOG(NOTICE,
+					    "Using AVX512 OFFLOAD Vector Tx (port %d).",
+					    dev->data->port_id);
+				dev->tx_pkt_burst =
+					ice_xmit_pkts_vec_avx512_offload;
+			} else {
+				PMD_DRV_LOG(NOTICE,
+					    "Using AVX512 Vector Tx (port %d).",
+					    dev->data->port_id);
+				dev->tx_pkt_burst = ice_xmit_pkts_vec_avx512;
+			}
 #endif
 		} else {
 			PMD_DRV_LOG(DEBUG, "Using %sVector Tx (port %d).",
@@ -3343,6 +3358,7 @@ static const struct {
 #ifdef RTE_ARCH_X86
 #ifdef CC_AVX512_SUPPORT
 	{ ice_xmit_pkts_vec_avx512, "Vector AVX512" },
+	{ ice_xmit_pkts_vec_avx512_offload, "Offload Vector AVX512" },
 #endif
 	{ ice_xmit_pkts_vec_avx2, "Vector AVX2" },
 	{ ice_xmit_pkts_vec,      "Vector SSE" },

drivers/net/ice/ice_rxtx.h

@@ -261,6 +261,9 @@ uint16_t ice_recv_scattered_pkts_vec_avx512(void *rx_queue,
 					    uint16_t nb_pkts);
 uint16_t ice_xmit_pkts_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
 				  uint16_t nb_pkts);
+uint16_t ice_xmit_pkts_vec_avx512_offload(void *tx_queue,
+					  struct rte_mbuf **tx_pkts,
+					  uint16_t nb_pkts);
 int ice_fdir_programming(struct ice_pf *pf, struct ice_fltr_desc *fdir_desc);
 int ice_tx_done_cleanup(void *txq, uint32_t free_cnt);
 int ice_get_monitor_addr(void *rx_queue, struct rte_power_monitor_cond *pmc);

drivers/net/ice/ice_rxtx_vec_avx2.c

@@ -853,7 +853,7 @@ ice_xmit_fixed_burst_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts,
 	nb_pkts = RTE_MIN(nb_pkts, txq->tx_rs_thresh);
 
 	if (txq->nb_tx_free < txq->tx_free_thresh)
-		ice_tx_free_bufs(txq);
+		ice_tx_free_bufs_vec(txq);
 
 	nb_commit = nb_pkts = (uint16_t)RTE_MIN(txq->nb_tx_free, nb_pkts);
 	if (unlikely(nb_pkts == 0))

drivers/net/ice/ice_rxtx_vec_avx512.c

@@ -982,23 +982,26 @@ ice_tx_free_bufs_avx512(struct ice_tx_queue *txq)
 	return txq->tx_rs_thresh;
 }
 
-static inline void
+static __rte_always_inline void
 ice_vtx1(volatile struct ice_tx_desc *txdp,
-	 struct rte_mbuf *pkt, uint64_t flags)
+	 struct rte_mbuf *pkt, uint64_t flags, bool do_offload)
 {
 	uint64_t high_qw =
 		(ICE_TX_DESC_DTYPE_DATA |
 		 ((uint64_t)flags  << ICE_TXD_QW1_CMD_S) |
 		 ((uint64_t)pkt->data_len << ICE_TXD_QW1_TX_BUF_SZ_S));
 
+	if (do_offload)
+		ice_txd_enable_offload(pkt, &high_qw);
+
 	__m128i descriptor = _mm_set_epi64x(high_qw,
 				pkt->buf_iova + pkt->data_off);
 	_mm_store_si128((__m128i *)txdp, descriptor);
 }
 
-static inline void
-ice_vtx(volatile struct ice_tx_desc *txdp,
-	struct rte_mbuf **pkt, uint16_t nb_pkts,  uint64_t flags)
+static __rte_always_inline void
+ice_vtx(volatile struct ice_tx_desc *txdp, struct rte_mbuf **pkt,
+	uint16_t nb_pkts,  uint64_t flags, bool do_offload)
 {
 	const uint64_t hi_qw_tmpl = (ICE_TX_DESC_DTYPE_DATA |
 			((uint64_t)flags  << ICE_TXD_QW1_CMD_S));
@@ -1008,18 +1011,26 @@ ice_vtx(volatile struct ice_tx_desc *txdp,
 			hi_qw_tmpl |
 			((uint64_t)pkt[3]->data_len <<
 			 ICE_TXD_QW1_TX_BUF_SZ_S);
+		if (do_offload)
+			ice_txd_enable_offload(pkt[3], &hi_qw3);
 		uint64_t hi_qw2 =
 			hi_qw_tmpl |
 			((uint64_t)pkt[2]->data_len <<
 			 ICE_TXD_QW1_TX_BUF_SZ_S);
+		if (do_offload)
+			ice_txd_enable_offload(pkt[2], &hi_qw2);
 		uint64_t hi_qw1 =
 			hi_qw_tmpl |
 			((uint64_t)pkt[1]->data_len <<
 			 ICE_TXD_QW1_TX_BUF_SZ_S);
+		if (do_offload)
+			ice_txd_enable_offload(pkt[1], &hi_qw1);
 		uint64_t hi_qw0 =
 			hi_qw_tmpl |
 			((uint64_t)pkt[0]->data_len <<
 			 ICE_TXD_QW1_TX_BUF_SZ_S);
+		if (do_offload)
+			ice_txd_enable_offload(pkt[0], &hi_qw0);
 
 		__m512i desc0_3 =
 			_mm512_set_epi64
@@ -1036,7 +1047,7 @@ ice_vtx(volatile struct ice_tx_desc *txdp,
 
 	/* do any last ones */
 	while (nb_pkts) {
-		ice_vtx1(txdp, *pkt, flags);
+		ice_vtx1(txdp, *pkt, flags, do_offload);
 		txdp++, pkt++, nb_pkts--;
 	}
 }
@@ -1051,9 +1062,9 @@ ice_tx_backlog_entry_avx512(struct ice_vec_tx_entry *txep,
 		txep[i].mbuf = tx_pkts[i];
 }
 
-static inline uint16_t
+static __rte_always_inline uint16_t
 ice_xmit_fixed_burst_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
-				uint16_t nb_pkts)
+				uint16_t nb_pkts, bool do_offload)
 {
 	struct ice_tx_queue *txq = (struct ice_tx_queue *)tx_queue;
 	volatile struct ice_tx_desc *txdp;
@@ -1083,11 +1094,11 @@ ice_xmit_fixed_burst_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
 	if (nb_commit >= n) {
 		ice_tx_backlog_entry_avx512(txep, tx_pkts, n);
 
-		ice_vtx(txdp, tx_pkts, n - 1, flags);
+		ice_vtx(txdp, tx_pkts, n - 1, flags, do_offload);
 		tx_pkts += (n - 1);
 		txdp += (n - 1);
 
-		ice_vtx1(txdp, *tx_pkts++, rs);
+		ice_vtx1(txdp, *tx_pkts++, rs, do_offload);
 
 		nb_commit = (uint16_t)(nb_commit - n);
 
@@ -1101,7 +1112,7 @@ ice_xmit_fixed_burst_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
 
 	ice_tx_backlog_entry_avx512(txep, tx_pkts, nb_commit);
 
-	ice_vtx(txdp, tx_pkts, nb_commit, flags);
+	ice_vtx(txdp, tx_pkts, nb_commit, flags, do_offload);
 
 	tx_id = (uint16_t)(tx_id + nb_commit);
 	if (tx_id > txq->tx_next_rs) {
@@ -1131,7 +1142,30 @@ ice_xmit_pkts_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts,
 
 		num = (uint16_t)RTE_MIN(nb_pkts, txq->tx_rs_thresh);
 		ret = ice_xmit_fixed_burst_vec_avx512(tx_queue,
-						      &tx_pkts[nb_tx], num);
+				&tx_pkts[nb_tx], num, false);
+		nb_tx += ret;
+		nb_pkts -= ret;
+		if (ret < num)
+			break;
+	}
+
+	return nb_tx;
+}
+
+uint16_t
+ice_xmit_pkts_vec_avx512_offload(void *tx_queue, struct rte_mbuf **tx_pkts,
+				 uint16_t nb_pkts)
+{
+	uint16_t nb_tx = 0;
+	struct ice_tx_queue *txq = (struct ice_tx_queue *)tx_queue;
+
+	while (nb_pkts) {
+		uint16_t ret, num;
+
+		num = (uint16_t)RTE_MIN(nb_pkts, txq->tx_rs_thresh);
+		ret = ice_xmit_fixed_burst_vec_avx512(tx_queue,
+				&tx_pkts[nb_tx], num, true);
+
 		nb_tx += ret;
 		nb_pkts -= ret;
 		if (ret < num)

drivers/net/ice/ice_rxtx_vec_common.h

@@ -77,7 +77,7 @@ ice_rx_reassemble_packets(struct ice_rx_queue *rxq, struct rte_mbuf **rx_bufs,
 }
 
 static __rte_always_inline int
-ice_tx_free_bufs(struct ice_tx_queue *txq)
+ice_tx_free_bufs_vec(struct ice_tx_queue *txq)
 {
 	struct ice_tx_entry *txep;
 	uint32_t n;
@@ -197,7 +197,8 @@ _ice_tx_queue_release_mbufs_vec(struct ice_tx_queue *txq)
 #ifdef CC_AVX512_SUPPORT
 	struct rte_eth_dev *dev = txq->vsi->adapter->eth_dev;
 
-	if (dev->tx_pkt_burst == ice_xmit_pkts_vec_avx512) {
+	if (dev->tx_pkt_burst == ice_xmit_pkts_vec_avx512 ||
+	    dev->tx_pkt_burst == ice_xmit_pkts_vec_avx512_offload) {
 		struct ice_vec_tx_entry *swr = (void *)txq->sw_ring;
 
 		if (txq->tx_tail < i) {
@@ -267,29 +268,39 @@ ice_rx_vec_queue_default(struct ice_rx_queue *rxq)
 	return 0;
 }
 
-#define ICE_NO_VECTOR_FLAGS (				 \
-		DEV_TX_OFFLOAD_MULTI_SEGS |		 \
-		DEV_TX_OFFLOAD_VLAN_INSERT |		 \
-		DEV_TX_OFFLOAD_IPV4_CKSUM |		 \
-		DEV_TX_OFFLOAD_SCTP_CKSUM |		 \
-		DEV_TX_OFFLOAD_UDP_CKSUM |		 \
-		DEV_TX_OFFLOAD_TCP_TSO |		 \
+#define ICE_TX_NO_VECTOR_FLAGS (			\
+		DEV_TX_OFFLOAD_MULTI_SEGS |		\
+		DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM |	\
+		DEV_TX_OFFLOAD_TCP_TSO)
+
+#define ICE_TX_VECTOR_OFFLOAD (				\
+		DEV_TX_OFFLOAD_VLAN_INSERT |		\
+		DEV_TX_OFFLOAD_QINQ_INSERT |		\
+		DEV_TX_OFFLOAD_IPV4_CKSUM |		\
+		DEV_TX_OFFLOAD_SCTP_CKSUM |		\
+		DEV_TX_OFFLOAD_UDP_CKSUM |		\
 		DEV_TX_OFFLOAD_TCP_CKSUM)
 
+#define ICE_VECTOR_PATH		0
+#define ICE_VECTOR_OFFLOAD_PATH	1
+
 static inline int
 ice_tx_vec_queue_default(struct ice_tx_queue *txq)
 {
 	if (!txq)
 		return -1;
 
-	if (txq->offloads & ICE_NO_VECTOR_FLAGS)
-		return -1;
-
 	if (txq->tx_rs_thresh < ICE_VPMD_TX_BURST ||
 	    txq->tx_rs_thresh > ICE_TX_MAX_FREE_BUF_SZ)
 		return -1;
 
-	return 0;
+	if (txq->offloads & ICE_TX_NO_VECTOR_FLAGS)
+		return -1;
+
+	if (txq->offloads & ICE_TX_VECTOR_OFFLOAD)
+		return ICE_VECTOR_OFFLOAD_PATH;
+
+	return ICE_VECTOR_PATH;
 }
 
 static inline int
@@ -312,14 +323,19 @@ ice_tx_vec_dev_check_default(struct rte_eth_dev *dev)
 {
 	int i;
 	struct ice_tx_queue *txq;
+	int ret = 0;
+	int result = 0;
 
 	for (i = 0; i < dev->data->nb_tx_queues; i++) {
 		txq = dev->data->tx_queues[i];
-		if (ice_tx_vec_queue_default(txq))
+		ret = ice_tx_vec_queue_default(txq);
+		if (ret < 0)
 			return -1;
+		if (ret == ICE_VECTOR_OFFLOAD_PATH)
+			result = ret;
 	}
 
-	return 0;
+	return result;
 }
 
 #ifdef CC_AVX2_SUPPORT
@@ -521,4 +537,64 @@ ice_rxq_rearm_common(struct ice_rx_queue *rxq, __rte_unused bool avx512)
 }
 #endif
 
+static inline void
+ice_txd_enable_offload(struct rte_mbuf *tx_pkt,
+		       uint64_t *txd_hi)
+{
+	uint64_t ol_flags = tx_pkt->ol_flags;
+	uint32_t td_cmd = 0;
+	uint32_t td_offset = 0;
+
+	/* Tx Checksum Offload */
+	/* SET MACLEN */
+	td_offset |= (tx_pkt->l2_len >> 1) <<
+			ICE_TX_DESC_LEN_MACLEN_S;
+
+	/* Enable L3 checksum offload */
+	if (ol_flags & PKT_TX_IP_CKSUM) {
+		td_cmd |= ICE_TX_DESC_CMD_IIPT_IPV4_CSUM;
+		td_offset |= (tx_pkt->l3_len >> 2) <<
+			ICE_TX_DESC_LEN_IPLEN_S;
+	} else if (ol_flags & PKT_TX_IPV4) {
+		td_cmd |= ICE_TX_DESC_CMD_IIPT_IPV4;
+		td_offset |= (tx_pkt->l3_len >> 2) <<
+			ICE_TX_DESC_LEN_IPLEN_S;
+	} else if (ol_flags & PKT_TX_IPV6) {
+		td_cmd |= ICE_TX_DESC_CMD_IIPT_IPV6;
+		td_offset |= (tx_pkt->l3_len >> 2) <<
+			ICE_TX_DESC_LEN_IPLEN_S;
+	}
+
+	/* Enable L4 checksum offloads */
+	switch (ol_flags & PKT_TX_L4_MASK) {
+	case PKT_TX_TCP_CKSUM:
+		td_cmd |= ICE_TX_DESC_CMD_L4T_EOFT_TCP;
+		td_offset |= (sizeof(struct rte_tcp_hdr) >> 2) <<
+			ICE_TX_DESC_LEN_L4_LEN_S;
+		break;
+	case PKT_TX_SCTP_CKSUM:
+		td_cmd |= ICE_TX_DESC_CMD_L4T_EOFT_SCTP;
+		td_offset |= (sizeof(struct rte_sctp_hdr) >> 2) <<
+			ICE_TX_DESC_LEN_L4_LEN_S;
+		break;
+	case PKT_TX_UDP_CKSUM:
+		td_cmd |= ICE_TX_DESC_CMD_L4T_EOFT_UDP;
+		td_offset |= (sizeof(struct rte_udp_hdr) >> 2) <<
+			ICE_TX_DESC_LEN_L4_LEN_S;
+		break;
+	default:
+		break;
+	}
+
+	*txd_hi |= ((uint64_t)td_offset) << ICE_TXD_QW1_OFFSET_S;
+
+	/* Tx VLAN/QINQ insertion Offload */
+	if (ol_flags & (PKT_TX_VLAN | PKT_TX_QINQ)) {
+		td_cmd |= ICE_TX_DESC_CMD_IL2TAG1;
+		*txd_hi |= ((uint64_t)tx_pkt->vlan_tci <<
+				ICE_TXD_QW1_L2TAG1_S);
+	}
+
+	*txd_hi |= ((uint64_t)td_cmd) << ICE_TXD_QW1_CMD_S;
+}
 #endif

drivers/net/ice/ice_rxtx_vec_sse.c

@@ -702,7 +702,7 @@ ice_xmit_fixed_burst_vec(void *tx_queue, struct rte_mbuf **tx_pkts,
 	nb_pkts = RTE_MIN(nb_pkts, txq->tx_rs_thresh);
 
 	if (txq->nb_tx_free < txq->tx_free_thresh)
-		ice_tx_free_bufs(txq);
+		ice_tx_free_bufs_vec(txq);
 
 	nb_pkts = (uint16_t)RTE_MIN(txq->nb_tx_free, nb_pkts);
 	nb_commit = nb_pkts;