[dpdk-dev] [PATCH v1 54/58] net/octeontx2: add Tx vector version

jerinj at marvell.com jerinj at marvell.com
Sun Jun 2 17:24:30 CEST 2019


From: Nithin Dabilpuram <ndabilpuram at marvell.com>

Add vector version of packet transmit function.

Signed-off-by: Nithin Dabilpuram <ndabilpuram at marvell.com>
Signed-off-by: Pavan Nikhilesh <pbhagavatula at marvell.com>
---
 drivers/net/octeontx2/otx2_tx.c | 883 +++++++++++++++++++++++++++++++-
 1 file changed, 882 insertions(+), 1 deletion(-)

diff --git a/drivers/net/octeontx2/otx2_tx.c b/drivers/net/octeontx2/otx2_tx.c
index 0ac5ea652..6bce55112 100644
--- a/drivers/net/octeontx2/otx2_tx.c
+++ b/drivers/net/octeontx2/otx2_tx.c
@@ -80,6 +80,859 @@ nix_xmit_pkts_mseg(void *tx_queue, struct rte_mbuf **tx_pkts,
 	return pkts;
 }
 
+#if defined(RTE_ARCH_ARM64)
+
+#define NIX_DESCS_PER_LOOP	4
+static __rte_always_inline uint16_t
+nix_xmit_pkts_vector(void *tx_queue, struct rte_mbuf **tx_pkts,
+		     uint16_t pkts, const uint16_t flags)
+{
+	uint64x2_t dataoff_iova0, dataoff_iova1, dataoff_iova2, dataoff_iova3;
+	uint64x2_t len_olflags0, len_olflags1, len_olflags2, len_olflags3;
+	uint64_t *mbuf0, *mbuf1, *mbuf2, *mbuf3;
+	uint64x2_t senddesc01_w0, senddesc23_w0;
+	uint64x2_t senddesc01_w1, senddesc23_w1;
+	uint64x2_t sgdesc01_w0, sgdesc23_w0;
+	uint64x2_t sgdesc01_w1, sgdesc23_w1;
+	struct otx2_eth_txq *txq = tx_queue;
+	uint64_t *lmt_addr = txq->lmt_addr;
+	rte_iova_t io_addr = txq->io_addr;
+	uint64x2_t ltypes01, ltypes23;
+	uint64x2_t xtmp128, ytmp128;
+	uint64x2_t xmask01, xmask23;
+	uint64x2_t mbuf01, mbuf23;
+	uint64x2_t cmd00, cmd01;
+	uint64x2_t cmd10, cmd11;
+	uint64x2_t cmd20, cmd21;
+	uint64x2_t cmd30, cmd31;
+	uint64_t lmt_status, i;
+
+	pkts = RTE_ALIGN_FLOOR(pkts, NIX_DESCS_PER_LOOP);
+
+	NIX_XMIT_FC_OR_RETURN(txq, pkts);
+
+	/* Reduce the cached count */
+	txq->fc_cache_pkts -= pkts;
+
+	/* Lets commit any changes in the packet */
+	rte_cio_wmb();
+
+	senddesc01_w0 = vld1q_dup_u64(&txq->cmd[0]);
+	senddesc23_w0 = senddesc01_w0;
+	senddesc01_w1 = vdupq_n_u64(0);
+	senddesc23_w1 = senddesc01_w1;
+	sgdesc01_w0 = vld1q_dup_u64(&txq->cmd[2]);
+	sgdesc23_w0 = sgdesc01_w0;
+
+	for (i = 0; i < pkts; i += NIX_DESCS_PER_LOOP) {
+		mbuf01 = vld1q_u64((uint64_t *)tx_pkts);
+		mbuf23 = vld1q_u64((uint64_t *)(tx_pkts + 2));
+
+		/* Clear lower 32bit of SEND_HDR_W0 and SEND_SG_W0 */
+		senddesc01_w0 = vbicq_u64(senddesc01_w0,
+					  vdupq_n_u64(0xFFFFFFFF));
+		sgdesc01_w0 = vbicq_u64(sgdesc01_w0,
+					vdupq_n_u64(0xFFFFFFFF));
+
+		senddesc23_w0 = senddesc01_w0;
+		sgdesc23_w0 = sgdesc01_w0;
+
+		tx_pkts = tx_pkts + NIX_DESCS_PER_LOOP;
+
+		/* Move mbufs to iova */
+		mbuf0 = (uint64_t *)vgetq_lane_u64(mbuf01, 0);
+		mbuf1 = (uint64_t *)vgetq_lane_u64(mbuf01, 1);
+		mbuf2 = (uint64_t *)vgetq_lane_u64(mbuf23, 0);
+		mbuf3 = (uint64_t *)vgetq_lane_u64(mbuf23, 1);
+
+		mbuf0 = (uint64_t *)((uintptr_t)mbuf0 +
+				     offsetof(struct rte_mbuf, buf_iova));
+		mbuf1 = (uint64_t *)((uintptr_t)mbuf1 +
+				     offsetof(struct rte_mbuf, buf_iova));
+		mbuf2 = (uint64_t *)((uintptr_t)mbuf2 +
+				     offsetof(struct rte_mbuf, buf_iova));
+		mbuf3 = (uint64_t *)((uintptr_t)mbuf3 +
+				     offsetof(struct rte_mbuf, buf_iova));
+		/*
+		 * Get mbuf's, olflags, iova, pktlen, dataoff
+		 * dataoff_iovaX.D[0] = iova,
+		 * dataoff_iovaX.D[1](15:0) = mbuf->dataoff
+		 * len_olflagsX.D[0] = ol_flags,
+		 * len_olflagsX.D[1](63:32) = mbuf->pkt_len
+		 */
+		dataoff_iova0  = vld1q_u64(mbuf0);
+		len_olflags0 = vld1q_u64(mbuf0 + 2);
+		dataoff_iova1  = vld1q_u64(mbuf1);
+		len_olflags1 = vld1q_u64(mbuf1 + 2);
+		dataoff_iova2  = vld1q_u64(mbuf2);
+		len_olflags2 = vld1q_u64(mbuf2 + 2);
+		dataoff_iova3  = vld1q_u64(mbuf3);
+		len_olflags3 = vld1q_u64(mbuf3 + 2);
+
+		if (flags & NIX_TX_OFFLOAD_MBUF_NOFF_F) {
+			struct rte_mbuf *mbuf;
+			/* Set don't free bit if reference count > 1 */
+			xmask01 = vdupq_n_u64(0);
+			xmask23 = xmask01;
+
+			mbuf = (struct rte_mbuf *)((uintptr_t)mbuf0 -
+				offsetof(struct rte_mbuf, buf_iova));
+
+			if (rte_pktmbuf_prefree_seg(mbuf) == NULL)
+				vsetq_lane_u64(0x80000, xmask01, 0);
+			else
+				__mempool_check_cookies(mbuf->pool,
+							(void **)&mbuf,
+							1, 0);
+
+			mbuf = (struct rte_mbuf *)((uintptr_t)mbuf1 -
+				offsetof(struct rte_mbuf, buf_iova));
+			if (rte_pktmbuf_prefree_seg(mbuf) == NULL)
+				vsetq_lane_u64(0x80000, xmask01, 1);
+			else
+				__mempool_check_cookies(mbuf->pool,
+							(void **)&mbuf,
+							1, 0);
+
+			mbuf = (struct rte_mbuf *)((uintptr_t)mbuf2 -
+				offsetof(struct rte_mbuf, buf_iova));
+			if (rte_pktmbuf_prefree_seg(mbuf) == NULL)
+				vsetq_lane_u64(0x80000, xmask23, 0);
+			else
+				__mempool_check_cookies(mbuf->pool,
+							(void **)&mbuf,
+							1, 0);
+
+			mbuf = (struct rte_mbuf *)((uintptr_t)mbuf3 -
+				offsetof(struct rte_mbuf, buf_iova));
+			if (rte_pktmbuf_prefree_seg(mbuf) == NULL)
+				vsetq_lane_u64(0x80000, xmask23, 1);
+			else
+				__mempool_check_cookies(mbuf->pool,
+							(void **)&mbuf,
+							1, 0);
+			senddesc01_w0 = vorrq_u64(senddesc01_w0, xmask01);
+			senddesc23_w0 = vorrq_u64(senddesc23_w0, xmask23);
+		} else {
+			struct rte_mbuf *mbuf;
+			/* Mark mempool object as "put" since
+			 * it is freed by NIX
+			 */
+			mbuf = (struct rte_mbuf *)((uintptr_t)mbuf0 -
+				offsetof(struct rte_mbuf, buf_iova));
+			__mempool_check_cookies(mbuf->pool, (void **)&mbuf,
+						1, 0);
+
+			mbuf = (struct rte_mbuf *)((uintptr_t)mbuf1 -
+				offsetof(struct rte_mbuf, buf_iova));
+			__mempool_check_cookies(mbuf->pool, (void **)&mbuf,
+						1, 0);
+
+			mbuf = (struct rte_mbuf *)((uintptr_t)mbuf2 -
+				offsetof(struct rte_mbuf, buf_iova));
+			__mempool_check_cookies(mbuf->pool, (void **)&mbuf,
+						1, 0);
+
+			mbuf = (struct rte_mbuf *)((uintptr_t)mbuf3 -
+				offsetof(struct rte_mbuf, buf_iova));
+			__mempool_check_cookies(mbuf->pool, (void **)&mbuf,
+						1, 0);
+			RTE_SET_USED(mbuf);
+		}
+
+		/* Move mbufs to point pool */
+		mbuf0 = (uint64_t *)((uintptr_t)mbuf0 +
+			 offsetof(struct rte_mbuf, pool) -
+			 offsetof(struct rte_mbuf, buf_iova));
+		mbuf1 = (uint64_t *)((uintptr_t)mbuf1 +
+			 offsetof(struct rte_mbuf, pool) -
+			 offsetof(struct rte_mbuf, buf_iova));
+		mbuf2 = (uint64_t *)((uintptr_t)mbuf2 +
+			 offsetof(struct rte_mbuf, pool) -
+			 offsetof(struct rte_mbuf, buf_iova));
+		mbuf3 = (uint64_t *)((uintptr_t)mbuf3 +
+			 offsetof(struct rte_mbuf, pool) -
+			 offsetof(struct rte_mbuf, buf_iova));
+
+		if (flags &
+		    (NIX_TX_OFFLOAD_OL3_OL4_CSUM_F |
+		     NIX_TX_OFFLOAD_L3_L4_CSUM_F)) {
+			/* Get tx_offload for ol2, ol3, l2, l3 lengths */
+			/*
+			 * E(8):OL2_LEN(7):OL3_LEN(9):E(24):L3_LEN(9):L2_LEN(7)
+			 * E(8):OL2_LEN(7):OL3_LEN(9):E(24):L3_LEN(9):L2_LEN(7)
+			 */
+
+			asm volatile ("LD1 {%[a].D}[0],[%[in]]\n\t" :
+				      [a]"+w"(senddesc01_w1) :
+				      [in]"r"(mbuf0 + 2) : "memory");
+
+			asm volatile ("LD1 {%[a].D}[1],[%[in]]\n\t" :
+				      [a]"+w"(senddesc01_w1) :
+				      [in]"r"(mbuf1 + 2) : "memory");
+
+			asm volatile ("LD1 {%[b].D}[0],[%[in]]\n\t" :
+				      [b]"+w"(senddesc23_w1) :
+				      [in]"r"(mbuf2 + 2) : "memory");
+
+			asm volatile ("LD1 {%[b].D}[1],[%[in]]\n\t" :
+				      [b]"+w"(senddesc23_w1) :
+				      [in]"r"(mbuf3 + 2) : "memory");
+
+			/* Get pool pointer alone */
+			mbuf0 = (uint64_t *)*mbuf0;
+			mbuf1 = (uint64_t *)*mbuf1;
+			mbuf2 = (uint64_t *)*mbuf2;
+			mbuf3 = (uint64_t *)*mbuf3;
+		} else {
+			/* Get pool pointer alone */
+			mbuf0 = (uint64_t *)*mbuf0;
+			mbuf1 = (uint64_t *)*mbuf1;
+			mbuf2 = (uint64_t *)*mbuf2;
+			mbuf3 = (uint64_t *)*mbuf3;
+		}
+
+		const uint8x16_t shuf_mask2 = {
+			0x4, 0x5, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
+			0xc, 0xd, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
+		};
+		xtmp128 = vzip2q_u64(len_olflags0, len_olflags1);
+		ytmp128 = vzip2q_u64(len_olflags2, len_olflags3);
+
+		/* Clear dataoff_iovaX.D[1] bits other than dataoff(15:0) */
+		const uint64x2_t and_mask0 = {
+			0xFFFFFFFFFFFFFFFF,
+			0x000000000000FFFF,
+		};
+
+		dataoff_iova0 = vandq_u64(dataoff_iova0, and_mask0);
+		dataoff_iova1 = vandq_u64(dataoff_iova1, and_mask0);
+		dataoff_iova2 = vandq_u64(dataoff_iova2, and_mask0);
+		dataoff_iova3 = vandq_u64(dataoff_iova3, and_mask0);
+
+		/*
+		 * Pick only 16 bits of pktlen preset at bits 63:32
+		 * and place them at bits 15:0.
+		 */
+		xtmp128 = vqtbl1q_u8(xtmp128, shuf_mask2);
+		ytmp128 = vqtbl1q_u8(ytmp128, shuf_mask2);
+
+		/* Add pairwise to get dataoff + iova in sgdesc_w1 */
+		sgdesc01_w1 = vpaddq_u64(dataoff_iova0, dataoff_iova1);
+		sgdesc23_w1 = vpaddq_u64(dataoff_iova2, dataoff_iova3);
+
+		/* Orr both sgdesc_w0 and senddesc_w0 with 16 bits of
+		 * pktlen at 15:0 position.
+		 */
+		sgdesc01_w0 = vorrq_u64(sgdesc01_w0, xtmp128);
+		sgdesc23_w0 = vorrq_u64(sgdesc23_w0, ytmp128);
+		senddesc01_w0 = vorrq_u64(senddesc01_w0, xtmp128);
+		senddesc23_w0 = vorrq_u64(senddesc23_w0, ytmp128);
+
+		if ((flags & NIX_TX_OFFLOAD_L3_L4_CSUM_F) &&
+		    !(flags & NIX_TX_OFFLOAD_OL3_OL4_CSUM_F)) {
+			/*
+			 * Lookup table to translate ol_flags to
+			 * il3/il4 types. But we still use ol3/ol4 types in
+			 * senddesc_w1 as only one header processing is enabled.
+			 */
+			const uint8x16_t tbl = {
+				/* [0-15] = il4type:il3type */
+				0x04, /* none (IPv6 assumed) */
+				0x14, /* PKT_TX_TCP_CKSUM (IPv6 assumed) */
+				0x24, /* PKT_TX_SCTP_CKSUM (IPv6 assumed) */
+				0x34, /* PKT_TX_UDP_CKSUM (IPv6 assumed) */
+				0x03, /* PKT_TX_IP_CKSUM */
+				0x13, /* PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM */
+				0x23, /* PKT_TX_IP_CKSUM | PKT_TX_SCTP_CKSUM */
+				0x33, /* PKT_TX_IP_CKSUM | PKT_TX_UDP_CKSUM */
+				0x02, /* PKT_TX_IPV4  */
+				0x12, /* PKT_TX_IPV4 | PKT_TX_TCP_CKSUM */
+				0x22, /* PKT_TX_IPV4 | PKT_TX_SCTP_CKSUM */
+				0x32, /* PKT_TX_IPV4 | PKT_TX_UDP_CKSUM */
+				0x03, /* PKT_TX_IPV4 | PKT_TX_IP_CKSUM */
+				0x13, /* PKT_TX_IPV4 | PKT_TX_IP_CKSUM |
+				       * PKT_TX_TCP_CKSUM
+				       */
+				0x23, /* PKT_TX_IPV4 | PKT_TX_IP_CKSUM |
+				       * PKT_TX_SCTP_CKSUM
+				       */
+				0x33, /* PKT_TX_IPV4 | PKT_TX_IP_CKSUM |
+				       * PKT_TX_UDP_CKSUM
+				       */
+			};
+
+			/* Extract olflags to translate to iltypes */
+			xtmp128 = vzip1q_u64(len_olflags0, len_olflags1);
+			ytmp128 = vzip1q_u64(len_olflags2, len_olflags3);
+
+			/*
+			 * E(47):L3_LEN(9):L2_LEN(7+z)
+			 * E(47):L3_LEN(9):L2_LEN(7+z)
+			 */
+			senddesc01_w1 = vshlq_n_u64(senddesc01_w1, 1);
+			senddesc23_w1 = vshlq_n_u64(senddesc23_w1, 1);
+
+			/* Move OLFLAGS bits 55:52 to 51:48
+			 * with zeros preprended on the byte and rest
+			 * don't care
+			 */
+			xtmp128 = vshrq_n_u8(xtmp128, 4);
+			ytmp128 = vshrq_n_u8(ytmp128, 4);
+			/*
+			 * E(48):L3_LEN(8):L2_LEN(z+7)
+			 * E(48):L3_LEN(8):L2_LEN(z+7)
+			 */
+			const int8x16_t tshft3 = {
+				-1, 0, 8, 8, 8,	8, 8, 8,
+				-1, 0, 8, 8, 8,	8, 8, 8,
+			};
+
+			senddesc01_w1 = vshlq_u8(senddesc01_w1, tshft3);
+			senddesc23_w1 = vshlq_u8(senddesc23_w1, tshft3);
+
+			/* Do the lookup */
+			ltypes01 = vqtbl1q_u8(tbl, xtmp128);
+			ltypes23 = vqtbl1q_u8(tbl, ytmp128);
+
+			/* Just use ld1q to retrieve aura
+			 * when we don't need tx_offload
+			 */
+			mbuf0 = (uint64_t *)((uintptr_t)mbuf0 +
+					offsetof(struct rte_mempool, pool_id));
+			mbuf1 = (uint64_t *)((uintptr_t)mbuf1 +
+					offsetof(struct rte_mempool, pool_id));
+			mbuf2 = (uint64_t *)((uintptr_t)mbuf2 +
+					offsetof(struct rte_mempool, pool_id));
+			mbuf3 = (uint64_t *)((uintptr_t)mbuf3 +
+					offsetof(struct rte_mempool, pool_id));
+
+			/* Pick only relevant fields i.e Bit 48:55 of iltype
+			 * and place it in ol3/ol4type of senddesc_w1
+			 */
+			const uint8x16_t shuf_mask0 = {
+				0xFF, 0xFF, 0xFF, 0xFF,	0x6, 0xFF, 0xFF, 0xFF,
+				0xFF, 0xFF, 0xFF, 0xFF, 0xE, 0xFF, 0xFF, 0xFF,
+			};
+
+			ltypes01 = vqtbl1q_u8(ltypes01, shuf_mask0);
+			ltypes23 = vqtbl1q_u8(ltypes23, shuf_mask0);
+
+			/* Prepare ol4ptr, ol3ptr from ol3len, ol2len.
+			 * a [E(32):E(16):OL3(8):OL2(8)]
+			 * a = a + (a << 8)
+			 * a [E(32):E(16):(OL3+OL2):OL2]
+			 * => E(32):E(16)::OL4PTR(8):OL3PTR(8)
+			 */
+			senddesc01_w1 = vaddq_u8(senddesc01_w1,
+						 vshlq_n_u16(senddesc01_w1, 8));
+			senddesc23_w1 = vaddq_u8(senddesc23_w1,
+						 vshlq_n_u16(senddesc23_w1, 8));
+
+			/* Create first half of 4W cmd for 4 mbufs (sgdesc) */
+			cmd01 = vzip1q_u64(sgdesc01_w0, sgdesc01_w1);
+			cmd11 = vzip2q_u64(sgdesc01_w0, sgdesc01_w1);
+			cmd21 = vzip1q_u64(sgdesc23_w0, sgdesc23_w1);
+			cmd31 = vzip2q_u64(sgdesc23_w0, sgdesc23_w1);
+
+			xmask01 = vdupq_n_u64(0);
+			xmask23 = xmask01;
+			asm volatile ("LD1 {%[a].H}[0],[%[in]]\n\t" :
+				[a]"+w"(xmask01) : [in]"r"(mbuf0) : "memory");
+
+			asm volatile ("LD1 {%[a].H}[4],[%[in]]\n\t" :
+				 [a]"+w"(xmask01) : [in]"r"(mbuf1) : "memory");
+
+			asm volatile ("LD1 {%[b].H}[0],[%[in]]\n\t" :
+				 [b]"+w"(xmask23) : [in]"r"(mbuf2) : "memory");
+
+			asm volatile ("LD1 {%[b].H}[4],[%[in]]\n\t" :
+				 [b]"+w"(xmask23) : [in]"r"(mbuf3) : "memory");
+			xmask01 = vshlq_n_u64(xmask01, 20);
+			xmask23 = vshlq_n_u64(xmask23, 20);
+
+			senddesc01_w0 = vorrq_u64(senddesc01_w0, xmask01);
+			senddesc23_w0 = vorrq_u64(senddesc23_w0, xmask23);
+			/* Move ltypes to senddesc*_w1 */
+			senddesc01_w1 = vorrq_u64(senddesc01_w1, ltypes01);
+			senddesc23_w1 = vorrq_u64(senddesc23_w1, ltypes23);
+
+			/* Create first half of 4W cmd for 4 mbufs (sendhdr) */
+			cmd00 = vzip1q_u64(senddesc01_w0, senddesc01_w1);
+			cmd10 = vzip2q_u64(senddesc01_w0, senddesc01_w1);
+			cmd20 = vzip1q_u64(senddesc23_w0, senddesc23_w1);
+			cmd30 = vzip2q_u64(senddesc23_w0, senddesc23_w1);
+
+		} else if (!(flags & NIX_TX_OFFLOAD_L3_L4_CSUM_F) &&
+			   (flags & NIX_TX_OFFLOAD_OL3_OL4_CSUM_F)) {
+			/*
+			 * Lookup table to translate ol_flags to
+			 * ol3/ol4 types.
+			 */
+
+			const uint8x16_t tbl = {
+				/* [0-15] = ol4type:ol3type */
+				0x00, /* none */
+				0x03, /* OUTER_IP_CKSUM */
+				0x02, /* OUTER_IPV4 */
+				0x03, /* OUTER_IPV4 | OUTER_IP_CKSUM */
+				0x04, /* OUTER_IPV6 */
+				0x00, /* OUTER_IPV6 | OUTER_IP_CKSUM */
+				0x00, /* OUTER_IPV6 | OUTER_IPV4 */
+				0x00, /* OUTER_IPV6 | OUTER_IPV4 |
+				       * OUTER_IP_CKSUM
+				       */
+				0x00, /* OUTER_UDP_CKSUM */
+				0x33, /* OUTER_UDP_CKSUM | OUTER_IP_CKSUM */
+				0x32, /* OUTER_UDP_CKSUM | OUTER_IPV4 */
+				0x33, /* OUTER_UDP_CKSUM | OUTER_IPV4 |
+				       * OUTER_IP_CKSUM
+				       */
+				0x34, /* OUTER_UDP_CKSUM | OUTER_IPV6 */
+				0x00, /* OUTER_UDP_CKSUM | OUTER_IPV6 |
+				       * OUTER_IP_CKSUM
+				       */
+				0x00, /* OUTER_UDP_CKSUM | OUTER_IPV6 |
+				       * OUTER_IPV4
+				       */
+				0x00, /* OUTER_UDP_CKSUM | OUTER_IPV6 |
+				       * OUTER_IPV4 | OUTER_IP_CKSUM
+				       */
+			};
+
+			/* Extract olflags to translate to iltypes */
+			xtmp128 = vzip1q_u64(len_olflags0, len_olflags1);
+			ytmp128 = vzip1q_u64(len_olflags2, len_olflags3);
+
+			/*
+			 * E(47):OL3_LEN(9):OL2_LEN(7+z)
+			 * E(47):OL3_LEN(9):OL2_LEN(7+z)
+			 */
+			const uint8x16_t shuf_mask5 = {
+				0x6, 0x5, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
+				0xE, 0xD, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
+			};
+			senddesc01_w1 = vqtbl1q_u8(senddesc01_w1, shuf_mask5);
+			senddesc23_w1 = vqtbl1q_u8(senddesc23_w1, shuf_mask5);
+
+			/* Extract outer ol flags only */
+			const uint64x2_t o_cksum_mask = {
+				0x1C00020000000000,
+				0x1C00020000000000,
+			};
+
+			xtmp128 = vandq_u64(xtmp128, o_cksum_mask);
+			ytmp128 = vandq_u64(ytmp128, o_cksum_mask);
+
+			/* Extract OUTER_UDP_CKSUM bit 41 and
+			 * move it to bit 61
+			 */
+
+			xtmp128 = xtmp128 | vshlq_n_u64(xtmp128, 20);
+			ytmp128 = ytmp128 | vshlq_n_u64(ytmp128, 20);
+
+			/* Shift oltype by 2 to start nibble from BIT(56)
+			 * instead of BIT(58)
+			 */
+			xtmp128 = vshrq_n_u8(xtmp128, 2);
+			ytmp128 = vshrq_n_u8(ytmp128, 2);
+			/*
+			 * E(48):L3_LEN(8):L2_LEN(z+7)
+			 * E(48):L3_LEN(8):L2_LEN(z+7)
+			 */
+			const int8x16_t tshft3 = {
+				-1, 0, 8, 8, 8, 8, 8, 8,
+				-1, 0, 8, 8, 8, 8, 8, 8,
+			};
+
+			senddesc01_w1 = vshlq_u8(senddesc01_w1, tshft3);
+			senddesc23_w1 = vshlq_u8(senddesc23_w1, tshft3);
+
+			/* Do the lookup */
+			ltypes01 = vqtbl1q_u8(tbl, xtmp128);
+			ltypes23 = vqtbl1q_u8(tbl, ytmp128);
+
+			/* Just use ld1q to retrieve aura
+			 * when we don't need tx_offload
+			 */
+			mbuf0 = (uint64_t *)((uintptr_t)mbuf0 +
+					offsetof(struct rte_mempool, pool_id));
+			mbuf1 = (uint64_t *)((uintptr_t)mbuf1 +
+					offsetof(struct rte_mempool, pool_id));
+			mbuf2 = (uint64_t *)((uintptr_t)mbuf2 +
+					offsetof(struct rte_mempool, pool_id));
+			mbuf3 = (uint64_t *)((uintptr_t)mbuf3 +
+					offsetof(struct rte_mempool, pool_id));
+
+			/* Pick only relevant fields i.e Bit 56:63 of oltype
+			 * and place it in ol3/ol4type of senddesc_w1
+			 */
+			const uint8x16_t shuf_mask0 = {
+				0xFF, 0xFF, 0xFF, 0xFF,	0x7, 0xFF, 0xFF, 0xFF,
+				0xFF, 0xFF, 0xFF, 0xFF, 0xF, 0xFF, 0xFF, 0xFF,
+			};
+
+			ltypes01 = vqtbl1q_u8(ltypes01, shuf_mask0);
+			ltypes23 = vqtbl1q_u8(ltypes23, shuf_mask0);
+
+			/* Prepare ol4ptr, ol3ptr from ol3len, ol2len.
+			 * a [E(32):E(16):OL3(8):OL2(8)]
+			 * a = a + (a << 8)
+			 * a [E(32):E(16):(OL3+OL2):OL2]
+			 * => E(32):E(16)::OL4PTR(8):OL3PTR(8)
+			 */
+			senddesc01_w1 = vaddq_u8(senddesc01_w1,
+						 vshlq_n_u16(senddesc01_w1, 8));
+			senddesc23_w1 = vaddq_u8(senddesc23_w1,
+						 vshlq_n_u16(senddesc23_w1, 8));
+
+			/* Create second half of 4W cmd for 4 mbufs (sgdesc) */
+			cmd01 = vzip1q_u64(sgdesc01_w0, sgdesc01_w1);
+			cmd11 = vzip2q_u64(sgdesc01_w0, sgdesc01_w1);
+			cmd21 = vzip1q_u64(sgdesc23_w0, sgdesc23_w1);
+			cmd31 = vzip2q_u64(sgdesc23_w0, sgdesc23_w1);
+
+			xmask01 = vdupq_n_u64(0);
+			xmask23 = xmask01;
+			asm volatile ("LD1 {%[a].H}[0],[%[in]]\n\t" :
+				 [a]"+w"(xmask01) : [in]"r"(mbuf0) : "memory");
+
+			asm volatile ("LD1 {%[a].H}[4],[%[in]]\n\t" :
+				 [a]"+w"(xmask01) : [in]"r"(mbuf1) : "memory");
+
+			asm volatile ("LD1 {%[b].H}[0],[%[in]]\n\t" :
+				 [b]"+w"(xmask23) : [in]"r"(mbuf2) : "memory");
+
+			asm volatile ("LD1 {%[b].H}[4],[%[in]]\n\t" :
+				 [b]"+w"(xmask23) : [in]"r"(mbuf3) : "memory");
+			xmask01 = vshlq_n_u64(xmask01, 20);
+			xmask23 = vshlq_n_u64(xmask23, 20);
+
+			senddesc01_w0 = vorrq_u64(senddesc01_w0, xmask01);
+			senddesc23_w0 = vorrq_u64(senddesc23_w0, xmask23);
+			/* Move ltypes to senddesc*_w1 */
+			senddesc01_w1 = vorrq_u64(senddesc01_w1, ltypes01);
+			senddesc23_w1 = vorrq_u64(senddesc23_w1, ltypes23);
+
+			/* Create first half of 4W cmd for 4 mbufs (sendhdr) */
+			cmd00 = vzip1q_u64(senddesc01_w0, senddesc01_w1);
+			cmd10 = vzip2q_u64(senddesc01_w0, senddesc01_w1);
+			cmd20 = vzip1q_u64(senddesc23_w0, senddesc23_w1);
+			cmd30 = vzip2q_u64(senddesc23_w0, senddesc23_w1);
+
+		} else if ((flags & NIX_TX_OFFLOAD_L3_L4_CSUM_F) &&
+			   (flags & NIX_TX_OFFLOAD_OL3_OL4_CSUM_F)) {
+			/* Lookup table to translate ol_flags to
+			 * ol4type, ol3type, il4type, il3type of senddesc_w1
+			 */
+			const uint8x16x2_t tbl = {
+			{
+				{
+					/* [0-15] = il4type:il3type */
+					0x04, /* none (IPv6) */
+					0x14, /* PKT_TX_TCP_CKSUM (IPv6) */
+					0x24, /* PKT_TX_SCTP_CKSUM (IPv6) */
+					0x34, /* PKT_TX_UDP_CKSUM (IPv6) */
+					0x03, /* PKT_TX_IP_CKSUM */
+					0x13, /* PKT_TX_IP_CKSUM |
+					       * PKT_TX_TCP_CKSUM
+					       */
+					0x23, /* PKT_TX_IP_CKSUM |
+					       * PKT_TX_SCTP_CKSUM
+					       */
+					0x33, /* PKT_TX_IP_CKSUM |
+					       * PKT_TX_UDP_CKSUM
+					       */
+					0x02, /* PKT_TX_IPV4 */
+					0x12, /* PKT_TX_IPV4 |
+					       * PKT_TX_TCP_CKSUM
+					       */
+					0x22, /* PKT_TX_IPV4 |
+					       * PKT_TX_SCTP_CKSUM
+					       */
+					0x32, /* PKT_TX_IPV4 |
+					       * PKT_TX_UDP_CKSUM
+					       */
+					0x03, /* PKT_TX_IPV4 |
+					       * PKT_TX_IP_CKSUM
+					       */
+					0x13, /* PKT_TX_IPV4 | PKT_TX_IP_CKSUM |
+					       * PKT_TX_TCP_CKSUM
+					       */
+					0x23, /* PKT_TX_IPV4 | PKT_TX_IP_CKSUM |
+					       * PKT_TX_SCTP_CKSUM
+					       */
+					0x33, /* PKT_TX_IPV4 | PKT_TX_IP_CKSUM |
+					       * PKT_TX_UDP_CKSUM
+					       */
+				},
+
+				{
+					/* [16-31] = ol4type:ol3type */
+					0x00, /* none */
+					0x03, /* OUTER_IP_CKSUM */
+					0x02, /* OUTER_IPV4 */
+					0x03, /* OUTER_IPV4 | OUTER_IP_CKSUM */
+					0x04, /* OUTER_IPV6 */
+					0x00, /* OUTER_IPV6 | OUTER_IP_CKSUM */
+					0x00, /* OUTER_IPV6 | OUTER_IPV4 */
+					0x00, /* OUTER_IPV6 | OUTER_IPV4 |
+					       * OUTER_IP_CKSUM
+					       */
+					0x00, /* OUTER_UDP_CKSUM */
+					0x33, /* OUTER_UDP_CKSUM |
+					       * OUTER_IP_CKSUM
+					       */
+					0x32, /* OUTER_UDP_CKSUM |
+					       * OUTER_IPV4
+					       */
+					0x33, /* OUTER_UDP_CKSUM |
+					       * OUTER_IPV4 | OUTER_IP_CKSUM
+					       */
+					0x34, /* OUTER_UDP_CKSUM |
+					       * OUTER_IPV6
+					       */
+					0x00, /* OUTER_UDP_CKSUM | OUTER_IPV6 |
+					       * OUTER_IP_CKSUM
+					       */
+					0x00, /* OUTER_UDP_CKSUM | OUTER_IPV6 |
+					       * OUTER_IPV4
+					       */
+					0x00, /* OUTER_UDP_CKSUM | OUTER_IPV6 |
+					       * OUTER_IPV4 | OUTER_IP_CKSUM
+					       */
+				},
+			}
+			};
+
+			/* Extract olflags to translate to oltype & iltype */
+			xtmp128 = vzip1q_u64(len_olflags0, len_olflags1);
+			ytmp128 = vzip1q_u64(len_olflags2, len_olflags3);
+
+			/*
+			 * E(8):OL2_LN(7):OL3_LN(9):E(23):L3_LN(9):L2_LN(7+z)
+			 * E(8):OL2_LN(7):OL3_LN(9):E(23):L3_LN(9):L2_LN(7+z)
+			 */
+			const uint32x4_t tshft_4 = {
+				1, 0,
+				1, 0,
+			};
+			senddesc01_w1 = vshlq_u32(senddesc01_w1, tshft_4);
+			senddesc23_w1 = vshlq_u32(senddesc23_w1, tshft_4);
+
+			/*
+			 * E(32):L3_LEN(8):L2_LEN(7+Z):OL3_LEN(8):OL2_LEN(7+Z)
+			 * E(32):L3_LEN(8):L2_LEN(7+Z):OL3_LEN(8):OL2_LEN(7+Z)
+			 */
+			const uint8x16_t shuf_mask5 = {
+				0x6, 0x5, 0x0, 0x1, 0xFF, 0xFF, 0xFF, 0xFF,
+				0xE, 0xD, 0x8, 0x9, 0xFF, 0xFF,	0xFF, 0xFF,
+			};
+			senddesc01_w1 = vqtbl1q_u8(senddesc01_w1, shuf_mask5);
+			senddesc23_w1 = vqtbl1q_u8(senddesc23_w1, shuf_mask5);
+
+			/* Extract outer and inner header ol_flags */
+			const uint64x2_t oi_cksum_mask = {
+				0x1CF0020000000000,
+				0x1CF0020000000000,
+			};
+
+			xtmp128 = vandq_u64(xtmp128, oi_cksum_mask);
+			ytmp128 = vandq_u64(ytmp128, oi_cksum_mask);
+
+			/* Extract OUTER_UDP_CKSUM bit 41 and
+			 * move it to bit 61
+			 */
+
+			xtmp128 = xtmp128 | vshlq_n_u64(xtmp128, 20);
+			ytmp128 = ytmp128 | vshlq_n_u64(ytmp128, 20);
+
+			/* Shift right oltype by 2 and iltype by 4
+			 * to start oltype nibble from BIT(58)
+			 * instead of BIT(56) and iltype nibble from BIT(48)
+			 * instead of BIT(52).
+			 */
+			const int8x16_t tshft5 = {
+				8, 8, 8, 8, 8, 8, -4, -2,
+				8, 8, 8, 8, 8, 8, -4, -2,
+			};
+
+			xtmp128 = vshlq_u8(xtmp128, tshft5);
+			ytmp128 = vshlq_u8(ytmp128, tshft5);
+			/*
+			 * E(32):L3_LEN(8):L2_LEN(8):OL3_LEN(8):OL2_LEN(8)
+			 * E(32):L3_LEN(8):L2_LEN(8):OL3_LEN(8):OL2_LEN(8)
+			 */
+			const int8x16_t tshft3 = {
+				-1, 0, -1, 0, 0, 0, 0, 0,
+				-1, 0, -1, 0, 0, 0, 0, 0,
+			};
+
+			senddesc01_w1 = vshlq_u8(senddesc01_w1, tshft3);
+			senddesc23_w1 = vshlq_u8(senddesc23_w1, tshft3);
+
+			/* Mark Bit(4) of oltype */
+			const uint64x2_t oi_cksum_mask2 = {
+				0x1000000000000000,
+				0x1000000000000000,
+			};
+
+			xtmp128 = vorrq_u64(xtmp128, oi_cksum_mask2);
+			ytmp128 = vorrq_u64(ytmp128, oi_cksum_mask2);
+
+			/* Do the lookup */
+			ltypes01 = vqtbl2q_u8(tbl, xtmp128);
+			ltypes23 = vqtbl2q_u8(tbl, ytmp128);
+
+			/* Just use ld1q to retrieve aura
+			 * when we don't need tx_offload
+			 */
+			mbuf0 = (uint64_t *)((uintptr_t)mbuf0 +
+					offsetof(struct rte_mempool, pool_id));
+			mbuf1 = (uint64_t *)((uintptr_t)mbuf1 +
+					offsetof(struct rte_mempool, pool_id));
+			mbuf2 = (uint64_t *)((uintptr_t)mbuf2 +
+					offsetof(struct rte_mempool, pool_id));
+			mbuf3 = (uint64_t *)((uintptr_t)mbuf3 +
+					offsetof(struct rte_mempool, pool_id));
+
+			/* Pick only relevant fields i.e Bit 48:55 of iltype and
+			 * Bit 56:63 of oltype and place it in corresponding
+			 * place in senddesc_w1.
+			 */
+			const uint8x16_t shuf_mask0 = {
+				0xFF, 0xFF, 0xFF, 0xFF, 0x7, 0x6, 0xFF, 0xFF,
+				0xFF, 0xFF, 0xFF, 0xFF,	0xF, 0xE, 0xFF, 0xFF,
+			};
+
+			ltypes01 = vqtbl1q_u8(ltypes01, shuf_mask0);
+			ltypes23 = vqtbl1q_u8(ltypes23, shuf_mask0);
+
+			/* Prepare l4ptr, l3ptr, ol4ptr, ol3ptr from
+			 * l3len, l2len, ol3len, ol2len.
+			 * a [E(32):L3(8):L2(8):OL3(8):OL2(8)]
+			 * a = a + (a << 8)
+			 * a [E:(L3+L2):(L2+OL3):(OL3+OL2):OL2]
+			 * a = a + (a << 16)
+			 * a [E:(L3+L2+OL3+OL2):(L2+OL3+OL2):(OL3+OL2):OL2]
+			 * => E(32):IL4PTR(8):IL3PTR(8):OL4PTR(8):OL3PTR(8)
+			 */
+			senddesc01_w1 = vaddq_u8(senddesc01_w1,
+						 vshlq_n_u32(senddesc01_w1, 8));
+			senddesc23_w1 = vaddq_u8(senddesc23_w1,
+						 vshlq_n_u32(senddesc23_w1, 8));
+
+			/* Create second half of 4W cmd for 4 mbufs (sgdesc) */
+			cmd01 = vzip1q_u64(sgdesc01_w0, sgdesc01_w1);
+			cmd11 = vzip2q_u64(sgdesc01_w0, sgdesc01_w1);
+			cmd21 = vzip1q_u64(sgdesc23_w0, sgdesc23_w1);
+			cmd31 = vzip2q_u64(sgdesc23_w0, sgdesc23_w1);
+
+			/* Continue preparing l4ptr, l3ptr, ol4ptr, ol3ptr */
+			senddesc01_w1 = vaddq_u8(senddesc01_w1,
+						vshlq_n_u32(senddesc01_w1, 16));
+			senddesc23_w1 = vaddq_u8(senddesc23_w1,
+						vshlq_n_u32(senddesc23_w1, 16));
+
+			xmask01 = vdupq_n_u64(0);
+			xmask23 = xmask01;
+			asm volatile ("LD1 {%[a].H}[0],[%[in]]\n\t" :
+				 [a]"+w"(xmask01) : [in]"r"(mbuf0) : "memory");
+
+			asm volatile ("LD1 {%[a].H}[4],[%[in]]\n\t" :
+				 [a]"+w"(xmask01) : [in]"r"(mbuf1) : "memory");
+
+			asm volatile ("LD1 {%[b].H}[0],[%[in]]\n\t" :
+				 [b]"+w"(xmask23) : [in]"r"(mbuf2) : "memory");
+
+			asm volatile ("LD1 {%[b].H}[4],[%[in]]\n\t" :
+				 [b]"+w"(xmask23) : [in]"r"(mbuf3) : "memory");
+			xmask01 = vshlq_n_u64(xmask01, 20);
+			xmask23 = vshlq_n_u64(xmask23, 20);
+
+			senddesc01_w0 = vorrq_u64(senddesc01_w0, xmask01);
+			senddesc23_w0 = vorrq_u64(senddesc23_w0, xmask23);
+			/* Move ltypes to senddesc*_w1 */
+			senddesc01_w1 = vorrq_u64(senddesc01_w1, ltypes01);
+			senddesc23_w1 = vorrq_u64(senddesc23_w1, ltypes23);
+
+			/* Create first half of 4W cmd for 4 mbufs (sendhdr) */
+			cmd00 = vzip1q_u64(senddesc01_w0, senddesc01_w1);
+			cmd10 = vzip2q_u64(senddesc01_w0, senddesc01_w1);
+			cmd20 = vzip1q_u64(senddesc23_w0, senddesc23_w1);
+			cmd30 = vzip2q_u64(senddesc23_w0, senddesc23_w1);
+		} else {
+			/* Just use ld1q to retrieve aura
+			 * when we don't need tx_offload
+			 */
+			mbuf0 = (uint64_t *)((uintptr_t)mbuf0 +
+					offsetof(struct rte_mempool, pool_id));
+			mbuf1 = (uint64_t *)((uintptr_t)mbuf1 +
+					offsetof(struct rte_mempool, pool_id));
+			mbuf2 = (uint64_t *)((uintptr_t)mbuf2 +
+					offsetof(struct rte_mempool, pool_id));
+			mbuf3 = (uint64_t *)((uintptr_t)mbuf3 +
+					offsetof(struct rte_mempool, pool_id));
+			xmask01 = vdupq_n_u64(0);
+			xmask23 = xmask01;
+			asm volatile ("LD1 {%[a].H}[0],[%[in]]\n\t" :
+				 [a]"+w"(xmask01) : [in]"r"(mbuf0) : "memory");
+
+			asm volatile ("LD1 {%[a].H}[4],[%[in]]\n\t" :
+				 [a]"+w"(xmask01) : [in]"r"(mbuf1) : "memory");
+
+			asm volatile ("LD1 {%[b].H}[0],[%[in]]\n\t" :
+				 [b]"+w"(xmask23) : [in]"r"(mbuf2) : "memory");
+
+			asm volatile ("LD1 {%[b].H}[4],[%[in]]\n\t" :
+				 [b]"+w"(xmask23) : [in]"r"(mbuf3) : "memory");
+			xmask01 = vshlq_n_u64(xmask01, 20);
+			xmask23 = vshlq_n_u64(xmask23, 20);
+
+			senddesc01_w0 = vorrq_u64(senddesc01_w0, xmask01);
+			senddesc23_w0 = vorrq_u64(senddesc23_w0, xmask23);
+
+			/* Create 4W cmd for 4 mbufs (sendhdr, sgdesc) */
+			cmd00 = vzip1q_u64(senddesc01_w0, senddesc01_w1);
+			cmd01 = vzip1q_u64(sgdesc01_w0, sgdesc01_w1);
+			cmd10 = vzip2q_u64(senddesc01_w0, senddesc01_w1);
+			cmd11 = vzip2q_u64(sgdesc01_w0, sgdesc01_w1);
+			cmd20 = vzip1q_u64(senddesc23_w0, senddesc23_w1);
+			cmd21 = vzip1q_u64(sgdesc23_w0, sgdesc23_w1);
+			cmd30 = vzip2q_u64(senddesc23_w0, senddesc23_w1);
+			cmd31 = vzip2q_u64(sgdesc23_w0, sgdesc23_w1);
+		}
+
+		do {
+			vst1q_u64(lmt_addr, cmd00);
+			vst1q_u64(lmt_addr + 2, cmd01);
+			vst1q_u64(lmt_addr + 4, cmd10);
+			vst1q_u64(lmt_addr + 6, cmd11);
+			vst1q_u64(lmt_addr + 8, cmd20);
+			vst1q_u64(lmt_addr + 10, cmd21);
+			vst1q_u64(lmt_addr + 12, cmd30);
+			vst1q_u64(lmt_addr + 14, cmd31);
+			lmt_status = otx2_lmt_submit(io_addr);
+
+		} while (lmt_status == 0);
+	}
+
+	return pkts;
+}
+
+#else
+static __rte_always_inline uint16_t
+nix_xmit_pkts_vector(void *tx_queue, struct rte_mbuf **tx_pkts,
+		     uint16_t pkts, const uint16_t flags)
+{
+	RTE_SET_USED(tx_queue);
+	RTE_SET_USED(tx_pkts);
+	RTE_SET_USED(pkts);
+	RTE_SET_USED(flags);
+	return 0;
+}
+#endif
+
 #define T(name, f4, f3, f2, f1, f0, sz, flags)				\
 static uint16_t __rte_noinline	__hot					\
 otx2_nix_xmit_pkts_ ## name(void *tx_queue,				\
@@ -107,6 +960,21 @@ otx2_nix_xmit_pkts_mseg_ ## name(void *tx_queue,			\
 NIX_TX_FASTPATH_MODES
 #undef T
 
+#define T(name, f4, f3, f2, f1, f0, sz, flags)				\
+static uint16_t __rte_noinline	__hot					\
+otx2_nix_xmit_pkts_vec_ ## name(void *tx_queue,				\
+			struct rte_mbuf **tx_pkts, uint16_t pkts)	\
+{									\
+	/* VLAN and TSTMP is not supported by vec */			\
+	if ((flags) & NIX_TX_OFFLOAD_VLAN_QINQ_F ||			\
+	    (flags) & NIX_TX_OFFLOAD_TSTAMP_F)				\
+		return 0;						\
+	return nix_xmit_pkts_vector(tx_queue, tx_pkts, pkts, (flags));	\
+}
+
+NIX_TX_FASTPATH_MODES
+#undef T
+
 static inline void
 pick_tx_func(struct rte_eth_dev *eth_dev,
 	     const eth_tx_burst_t tx_burst[2][2][2][2][2])
@@ -143,7 +1011,20 @@ NIX_TX_FASTPATH_MODES
 #undef T
 	};
 
-	pick_tx_func(eth_dev, nix_eth_tx_burst);
+	const eth_tx_burst_t nix_eth_tx_vec_burst[2][2][2][2][2] = {
+#define T(name, f4, f3, f2, f1, f0, sz, flags)				\
+	[f4][f3][f2][f1][f0] =  otx2_nix_xmit_pkts_vec_ ## name,
+
+NIX_TX_FASTPATH_MODES
+#undef T
+	};
+
+	if (dev->scalar_ena ||
+	    (dev->tx_offload_flags &
+	     (NIX_TX_OFFLOAD_VLAN_QINQ_F | NIX_TX_OFFLOAD_TSTAMP_F)))
+		pick_tx_func(eth_dev, nix_eth_tx_burst);
+	else
+		pick_tx_func(eth_dev, nix_eth_tx_vec_burst);
 
 	if (dev->tx_offloads & DEV_TX_OFFLOAD_MULTI_SEGS)
 		pick_tx_func(eth_dev, nix_eth_tx_burst_mseg);
-- 
2.21.0



More information about the dev mailing list