[dpdk-dev] [PATCH v2 50/57] net/octeontx2: add Rx vector version

[jerinj at marvell.com]

From: Jerin Jacob <jerinj at marvell.com>

Add vector version of packet Receive function.

Signed-off-by: Jerin Jacob <jerinj at marvell.com>
Signed-off-by: Nithin Dabilpuram <ndabilpuram at marvell.com>
---
 doc/guides/nics/octeontx2.rst     |   1 +
 drivers/net/octeontx2/Makefile    |   1 +
 drivers/net/octeontx2/meson.build |   2 +
 drivers/net/octeontx2/otx2_rx.c   | 259 +++++++++++++++++++++++++++++-
 4 files changed, 262 insertions(+), 1 deletion(-)

diff --git a/doc/guides/nics/octeontx2.rst b/doc/guides/nics/octeontx2.rst
index 3d9fc5f1d..9d6596ad8 100644
--- a/doc/guides/nics/octeontx2.rst
+++ b/doc/guides/nics/octeontx2.rst
@@ -30,6 +30,7 @@ Features of the OCTEON TX2 Ethdev PMD are:
 - Link state information
 - Link flow control
 - Scatter-Gather IO support
+- Vector Poll mode driver
 - Debug utilities - Context dump and error interrupt support
 - IEEE1588 timestamping
 
diff --git a/drivers/net/octeontx2/Makefile b/drivers/net/octeontx2/Makefile
index 3e25d2ad4..a5f125655 100644
--- a/drivers/net/octeontx2/Makefile
+++ b/drivers/net/octeontx2/Makefile
@@ -14,6 +14,7 @@ CFLAGS += -I$(RTE_SDK)/drivers/common/octeontx2
 CFLAGS += -I$(RTE_SDK)/drivers/mempool/octeontx2
 CFLAGS += -I$(RTE_SDK)/drivers/net/octeontx2
 CFLAGS += -O3
+CFLAGS += -flax-vector-conversions
 
 ifneq ($(CONFIG_RTE_ARCH_64),y)
 CFLAGS += -Wno-int-to-pointer-cast
diff --git a/drivers/net/octeontx2/meson.build b/drivers/net/octeontx2/meson.build
index 975b2e715..9d151f88d 100644
--- a/drivers/net/octeontx2/meson.build
+++ b/drivers/net/octeontx2/meson.build
@@ -24,6 +24,8 @@ sources = files('otx2_rx.c',
 
 deps += ['bus_pci', 'common_octeontx2', 'mempool_octeontx2']
 
+cflags += ['-flax-vector-conversions']
+
 extra_flags = []
 # This integrated controller runs only on a arm64 machine, remove 32bit warnings
 if not dpdk_conf.get('RTE_ARCH_64')
diff --git a/drivers/net/octeontx2/otx2_rx.c b/drivers/net/octeontx2/otx2_rx.c
index fca182785..deefe9588 100644
--- a/drivers/net/octeontx2/otx2_rx.c
+++ b/drivers/net/octeontx2/otx2_rx.c
@@ -84,6 +84,239 @@ nix_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
 	return nb_pkts;
 }
 
+#if defined(RTE_ARCH_ARM64)
+
+static __rte_always_inline uint64_t
+nix_vlan_update(const uint64_t w2, uint64_t ol_flags, uint8x16_t *f)
+{
+	if (w2 & BIT_ULL(21) /* vtag0_gone */) {
+		ol_flags |= PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED;
+		*f = vsetq_lane_u16((uint16_t)(w2 >> 32), *f, 5);
+	}
+
+	return ol_flags;
+}
+
+static __rte_always_inline uint64_t
+nix_qinq_update(const uint64_t w2, uint64_t ol_flags, struct rte_mbuf *mbuf)
+{
+	if (w2 & BIT_ULL(23) /* vtag1_gone */) {
+		ol_flags |= PKT_RX_QINQ | PKT_RX_QINQ_STRIPPED;
+		mbuf->vlan_tci_outer = (uint16_t)(w2 >> 48);
+	}
+
+	return ol_flags;
+}
+
+static __rte_always_inline uint16_t
+nix_recv_pkts_vector(void *rx_queue, struct rte_mbuf **rx_pkts,
+		     uint16_t pkts, const uint16_t flags)
+{
+	struct otx2_eth_rxq *rxq = rx_queue; uint16_t packets = 0;
+	uint64x2_t cq0_w8, cq1_w8, cq2_w8, cq3_w8, mbuf01, mbuf23;
+	const uint64_t mbuf_initializer = rxq->mbuf_initializer;
+	const uint64x2_t data_off = vdupq_n_u64(rxq->data_off);
+	uint64_t ol_flags0, ol_flags1, ol_flags2, ol_flags3;
+	uint64x2_t rearm0 = vdupq_n_u64(mbuf_initializer);
+	uint64x2_t rearm1 = vdupq_n_u64(mbuf_initializer);
+	uint64x2_t rearm2 = vdupq_n_u64(mbuf_initializer);
+	uint64x2_t rearm3 = vdupq_n_u64(mbuf_initializer);
+	struct rte_mbuf *mbuf0, *mbuf1, *mbuf2, *mbuf3;
+	const uint16_t *lookup_mem = rxq->lookup_mem;
+	const uint32_t qmask = rxq->qmask;
+	const uint64_t wdata = rxq->wdata;
+	const uintptr_t desc = rxq->desc;
+	uint8x16_t f0, f1, f2, f3;
+	uint32_t head = rxq->head;
+
+	pkts = nix_rx_nb_pkts(rxq, wdata, pkts, qmask);
+	/* Packets has to be floor-aligned to NIX_DESCS_PER_LOOP */
+	pkts = RTE_ALIGN_FLOOR(pkts, NIX_DESCS_PER_LOOP);
+
+	while (packets < pkts) {
+		/* Get the CQ pointers, since the ring size is multiple of
+		 * 4, We can avoid checking the wrap around of head
+		 * value after the each access unlike scalar version.
+		 */
+		const uintptr_t cq0 = desc + CQE_SZ(head);
+
+		/* Prefetch N desc ahead */
+		rte_prefetch_non_temporal((void *)(cq0 + CQE_SZ(8)));
+		rte_prefetch_non_temporal((void *)(cq0 + CQE_SZ(9)));
+		rte_prefetch_non_temporal((void *)(cq0 + CQE_SZ(10)));
+		rte_prefetch_non_temporal((void *)(cq0 + CQE_SZ(11)));
+
+		/* Get NIX_RX_SG_S for size and buffer pointer */
+		cq0_w8 = vld1q_u64((uint64_t *)(cq0 + CQE_SZ(0) + 64));
+		cq1_w8 = vld1q_u64((uint64_t *)(cq0 + CQE_SZ(1) + 64));
+		cq2_w8 = vld1q_u64((uint64_t *)(cq0 + CQE_SZ(2) + 64));
+		cq3_w8 = vld1q_u64((uint64_t *)(cq0 + CQE_SZ(3) + 64));
+
+		/* Extract mbuf from NIX_RX_SG_S */
+		mbuf01 = vzip2q_u64(cq0_w8, cq1_w8);
+		mbuf23 = vzip2q_u64(cq2_w8, cq3_w8);
+		mbuf01 = vqsubq_u64(mbuf01, data_off);
+		mbuf23 = vqsubq_u64(mbuf23, data_off);
+
+		/* Move mbufs to scalar registers for future use */
+		mbuf0 = (struct rte_mbuf *)vgetq_lane_u64(mbuf01, 0);
+		mbuf1 = (struct rte_mbuf *)vgetq_lane_u64(mbuf01, 1);
+		mbuf2 = (struct rte_mbuf *)vgetq_lane_u64(mbuf23, 0);
+		mbuf3 = (struct rte_mbuf *)vgetq_lane_u64(mbuf23, 1);
+
+		/* Mask to get packet len from NIX_RX_SG_S */
+		const uint8x16_t shuf_msk = {
+			0xFF, 0xFF,   /* pkt_type set as unknown */
+			0xFF, 0xFF,   /* pkt_type set as unknown */
+			0, 1,         /* octet 1~0, low 16 bits pkt_len */
+			0xFF, 0xFF,   /* skip high 16 bits pkt_len, zero out */
+			0, 1,         /* octet 1~0, 16 bits data_len */
+			0xFF, 0xFF,
+			0xFF, 0xFF, 0xFF, 0xFF
+			};
+
+		/* Form the rx_descriptor_fields1 with pkt_len and data_len */
+		f0 = vqtbl1q_u8(cq0_w8, shuf_msk);
+		f1 = vqtbl1q_u8(cq1_w8, shuf_msk);
+		f2 = vqtbl1q_u8(cq2_w8, shuf_msk);
+		f3 = vqtbl1q_u8(cq3_w8, shuf_msk);
+
+		/* Load CQE word0 and word 1 */
+		uint64x2_t cq0_w0 = vld1q_u64((uint64_t *)(cq0 + CQE_SZ(0)));
+		uint64x2_t cq1_w0 = vld1q_u64((uint64_t *)(cq0 + CQE_SZ(1)));
+		uint64x2_t cq2_w0 = vld1q_u64((uint64_t *)(cq0 + CQE_SZ(2)));
+		uint64x2_t cq3_w0 = vld1q_u64((uint64_t *)(cq0 + CQE_SZ(3)));
+
+		if (flags & NIX_RX_OFFLOAD_RSS_F) {
+			/* Fill rss in the rx_descriptor_fields1 */
+			f0 = vsetq_lane_u32(vgetq_lane_u32(cq0_w0, 0), f0, 3);
+			f1 = vsetq_lane_u32(vgetq_lane_u32(cq1_w0, 0), f1, 3);
+			f2 = vsetq_lane_u32(vgetq_lane_u32(cq2_w0, 0), f2, 3);
+			f3 = vsetq_lane_u32(vgetq_lane_u32(cq3_w0, 0), f3, 3);
+			ol_flags0 = PKT_RX_RSS_HASH;
+			ol_flags1 = PKT_RX_RSS_HASH;
+			ol_flags2 = PKT_RX_RSS_HASH;
+			ol_flags3 = PKT_RX_RSS_HASH;
+		} else {
+			ol_flags0 = 0; ol_flags1 = 0;
+			ol_flags2 = 0; ol_flags3 = 0;
+		}
+
+		if (flags & NIX_RX_OFFLOAD_PTYPE_F) {
+			/* Fill packet_type in the rx_descriptor_fields1 */
+			f0 = vsetq_lane_u32(nix_ptype_get(lookup_mem,
+					    vgetq_lane_u64(cq0_w0, 1)), f0, 0);
+			f1 = vsetq_lane_u32(nix_ptype_get(lookup_mem,
+					    vgetq_lane_u64(cq1_w0, 1)), f1, 0);
+			f2 = vsetq_lane_u32(nix_ptype_get(lookup_mem,
+					    vgetq_lane_u64(cq2_w0, 1)), f2, 0);
+			f3 = vsetq_lane_u32(nix_ptype_get(lookup_mem,
+					    vgetq_lane_u64(cq3_w0, 1)), f3, 0);
+		}
+
+		if (flags & NIX_RX_OFFLOAD_CHECKSUM_F) {
+			ol_flags0 |= nix_rx_olflags_get(lookup_mem,
+						vgetq_lane_u64(cq0_w0, 1));
+			ol_flags1 |= nix_rx_olflags_get(lookup_mem,
+						vgetq_lane_u64(cq1_w0, 1));
+			ol_flags2 |= nix_rx_olflags_get(lookup_mem,
+						vgetq_lane_u64(cq2_w0, 1));
+			ol_flags3 |= nix_rx_olflags_get(lookup_mem,
+						vgetq_lane_u64(cq3_w0, 1));
+		}
+
+		if (flags & NIX_RX_OFFLOAD_VLAN_STRIP_F) {
+			uint64_t cq0_w2 = *(uint64_t *)(cq0 + CQE_SZ(0) + 16);
+			uint64_t cq1_w2 = *(uint64_t *)(cq0 + CQE_SZ(1) + 16);
+			uint64_t cq2_w2 = *(uint64_t *)(cq0 + CQE_SZ(2) + 16);
+			uint64_t cq3_w2 = *(uint64_t *)(cq0 + CQE_SZ(3) + 16);
+
+			ol_flags0 = nix_vlan_update(cq0_w2, ol_flags0, &f0);
+			ol_flags1 = nix_vlan_update(cq1_w2, ol_flags1, &f1);
+			ol_flags2 = nix_vlan_update(cq2_w2, ol_flags2, &f2);
+			ol_flags3 = nix_vlan_update(cq3_w2, ol_flags3, &f3);
+
+			ol_flags0 = nix_qinq_update(cq0_w2, ol_flags0, mbuf0);
+			ol_flags1 = nix_qinq_update(cq1_w2, ol_flags1, mbuf1);
+			ol_flags2 = nix_qinq_update(cq2_w2, ol_flags2, mbuf2);
+			ol_flags3 = nix_qinq_update(cq3_w2, ol_flags3, mbuf3);
+		}
+
+		if (flags & NIX_RX_OFFLOAD_MARK_UPDATE_F) {
+			ol_flags0 = nix_update_match_id(*(uint16_t *)
+				    (cq0 + CQE_SZ(0) + 38), ol_flags0, mbuf0);
+			ol_flags1 = nix_update_match_id(*(uint16_t *)
+				    (cq0 + CQE_SZ(1) + 38), ol_flags1, mbuf1);
+			ol_flags2 = nix_update_match_id(*(uint16_t *)
+				    (cq0 + CQE_SZ(2) + 38), ol_flags2, mbuf2);
+			ol_flags3 = nix_update_match_id(*(uint16_t *)
+				    (cq0 + CQE_SZ(3) + 38), ol_flags3, mbuf3);
+		}
+
+		/* Form rearm_data with ol_flags */
+		rearm0 = vsetq_lane_u64(ol_flags0, rearm0, 1);
+		rearm1 = vsetq_lane_u64(ol_flags1, rearm1, 1);
+		rearm2 = vsetq_lane_u64(ol_flags2, rearm2, 1);
+		rearm3 = vsetq_lane_u64(ol_flags3, rearm3, 1);
+
+		/* Update rx_descriptor_fields1 */
+		vst1q_u64((uint64_t *)mbuf0->rx_descriptor_fields1, f0);
+		vst1q_u64((uint64_t *)mbuf1->rx_descriptor_fields1, f1);
+		vst1q_u64((uint64_t *)mbuf2->rx_descriptor_fields1, f2);
+		vst1q_u64((uint64_t *)mbuf3->rx_descriptor_fields1, f3);
+
+		/* Update rearm_data */
+		vst1q_u64((uint64_t *)mbuf0->rearm_data, rearm0);
+		vst1q_u64((uint64_t *)mbuf1->rearm_data, rearm1);
+		vst1q_u64((uint64_t *)mbuf2->rearm_data, rearm2);
+		vst1q_u64((uint64_t *)mbuf3->rearm_data, rearm3);
+
+		/* Store the mbufs to rx_pkts */
+		vst1q_u64((uint64_t *)&rx_pkts[packets], mbuf01);
+		vst1q_u64((uint64_t *)&rx_pkts[packets + 2], mbuf23);
+
+		/* Prefetch mbufs */
+		otx2_prefetch_store_keep(mbuf0);
+		otx2_prefetch_store_keep(mbuf1);
+		otx2_prefetch_store_keep(mbuf2);
+		otx2_prefetch_store_keep(mbuf3);
+
+		/* Mark mempool obj as "get" as it is alloc'ed by NIX */
+		__mempool_check_cookies(mbuf0->pool, (void **)&mbuf0, 1, 1);
+		__mempool_check_cookies(mbuf1->pool, (void **)&mbuf1, 1, 1);
+		__mempool_check_cookies(mbuf2->pool, (void **)&mbuf2, 1, 1);
+		__mempool_check_cookies(mbuf3->pool, (void **)&mbuf3, 1, 1);
+
+		/* Advance head pointer and packets */
+		head += NIX_DESCS_PER_LOOP; head &= qmask;
+		packets += NIX_DESCS_PER_LOOP;
+	}
+
+	rxq->head = head;
+	rxq->available -= packets;
+
+	rte_cio_wmb();
+	/* Free all the CQs that we've processed */
+	otx2_write64((rxq->wdata | packets), rxq->cq_door);
+
+	return packets;
+}
+
+#else
+
+static inline uint16_t
+nix_recv_pkts_vector(void *rx_queue, struct rte_mbuf **rx_pkts,
+		     uint16_t pkts, const uint16_t flags)
+{
+	RTE_SET_USED(rx_queue);
+	RTE_SET_USED(rx_pkts);
+	RTE_SET_USED(pkts);
+	RTE_SET_USED(flags);
+
+	return 0;
+}
+
+#endif
 
 #define R(name, f5, f4, f3, f2, f1, f0, flags)				\
 static uint16_t __rte_noinline	__hot					       \
@@ -100,6 +333,16 @@ otx2_nix_recv_pkts_mseg_ ## name(void *rx_queue,			       \
 	return nix_recv_pkts(rx_queue, rx_pkts, pkts,			       \
 			     (flags) | NIX_RX_MULTI_SEG_F);		       \
 }									       \
+									       \
+static uint16_t __rte_noinline	__hot					       \
+otx2_nix_recv_pkts_vec_ ## name(void *rx_queue,				       \
+			struct rte_mbuf **rx_pkts, uint16_t pkts)	       \
+{									       \
+	/* TSTMP is not supported by vector */				       \
+	if ((flags) & NIX_RX_OFFLOAD_TSTAMP_F)				       \
+		return 0;						       \
+	return nix_recv_pkts_vector(rx_queue, rx_pkts, pkts, (flags));	       \
+}									       \
 
 NIX_RX_FASTPATH_MODES
 #undef R
@@ -141,7 +384,21 @@ NIX_RX_FASTPATH_MODES
 #undef R
 	};
 
-	pick_rx_func(eth_dev, nix_eth_rx_burst);
+	const eth_rx_burst_t nix_eth_rx_vec_burst[2][2][2][2][2][2] = {
+#define R(name, f5, f4, f3, f2, f1, f0, flags)				\
+	[f5][f4][f3][f2][f1][f0] =  otx2_nix_recv_pkts_vec_ ## name,
+
+NIX_RX_FASTPATH_MODES
+#undef R
+	};
+
+	/* For PTP enabled, scalar rx function should be chosen as most of the
+	 * PTP apps are implemented to rx burst 1 pkt.
+	 */
+	if (dev->scalar_ena || dev->rx_offloads & DEV_RX_OFFLOAD_TIMESTAMP)
+		pick_rx_func(eth_dev, nix_eth_rx_burst);
+	else
+		pick_rx_func(eth_dev, nix_eth_rx_vec_burst);
 
 	if (dev->rx_offloads & DEV_RX_OFFLOAD_SCATTER)
 		pick_rx_func(eth_dev, nix_eth_rx_burst_mseg);
-- 
2.21.0