[dpdk-dev] [RFC 2/3] net/virtio: add vectorized packed ring Rx NEON path

[Joyce Kong <joyce.kong@arm.com>]

Optimize packed ring Rx batch path with NEON instructions.

Signed-off-by: Joyce Kong <joyce.kong@arm.com>
---
 drivers/net/virtio/meson.build               |   1 +
 drivers/net/virtio/virtio_rxtx.c             |   7 +-
 drivers/net/virtio/virtio_rxtx_packed.h      |  16 ++
 drivers/net/virtio/virtio_rxtx_packed_neon.c | 202 +++++++++++++++++++
 4 files changed, 224 insertions(+), 2 deletions(-)
 create mode 100644 drivers/net/virtio/virtio_rxtx_packed_neon.c

diff --git a/drivers/net/virtio/meson.build b/drivers/net/virtio/meson.build
index e1851b0a6..5af633686 100644
--- a/drivers/net/virtio/meson.build
+++ b/drivers/net/virtio/meson.build
@@ -34,6 +34,7 @@ elif arch_subdir == 'ppc'
 	sources += files('virtio_rxtx_simple_altivec.c')
 elif arch_subdir == 'arm' and host_machine.cpu_family().startswith('aarch64')
 	sources += files('virtio_rxtx_simple_neon.c')
+	sources += files('virtio_rxtx_packed_neon.c')
 endif
 
 if is_linux
diff --git a/drivers/net/virtio/virtio_rxtx.c b/drivers/net/virtio/virtio_rxtx.c
index f915b8a2c..1deb77569 100644
--- a/drivers/net/virtio/virtio_rxtx.c
+++ b/drivers/net/virtio/virtio_rxtx.c
@@ -2020,7 +2020,8 @@ virtio_xmit_pkts_inorder(void *tx_queue,
 	return nb_tx;
 }
 
-#ifndef CC_AVX512_SUPPORT
+#if !defined(CC_AVX512_SUPPORT) && !defined(RTE_ARCH_ARM) && \
+	!defined(RTE_ARCH_ARM64)
 uint16_t
 virtio_recv_pkts_packed_vec(void *rx_queue __rte_unused,
 			    struct rte_mbuf **rx_pkts __rte_unused,
@@ -2028,7 +2029,9 @@ virtio_recv_pkts_packed_vec(void *rx_queue __rte_unused,
 {
 	return 0;
 }
+#endif
 
+#if !defined(CC_AVX512_SUPPORT)
 uint16_t
 virtio_xmit_pkts_packed_vec(void *tx_queue __rte_unused,
 			    struct rte_mbuf **tx_pkts __rte_unused,
@@ -2036,4 +2039,4 @@ virtio_xmit_pkts_packed_vec(void *tx_queue __rte_unused,
 {
 	return 0;
 }
-#endif /* ifndef CC_AVX512_SUPPORT */
+#endif
diff --git a/drivers/net/virtio/virtio_rxtx_packed.h b/drivers/net/virtio/virtio_rxtx_packed.h
index b2447843b..fd2d6baa5 100644
--- a/drivers/net/virtio/virtio_rxtx_packed.h
+++ b/drivers/net/virtio/virtio_rxtx_packed.h
@@ -19,9 +19,16 @@
 #include "virtqueue.h"
 
 #define BYTE_SIZE 8
+
+#if defined(AVX512_SUPPORT)
 /* flag bits offset in packed ring desc higher 64bits */
 #define FLAGS_BITS_OFFSET ((offsetof(struct vring_packed_desc, flags) - \
 	offsetof(struct vring_packed_desc, len)) * BYTE_SIZE)
+#elif defined(RTE_ARCH_ARM) || defined(RTE_ARCH_ARM64)
+/* flag bits offset in packed ring desc 32bits */
+#define FLAGS_BITS_OFFSET ((offsetof(struct vring_packed_desc, flags) - \
+	offsetof(struct vring_packed_desc, id)) * BYTE_SIZE)
+#endif
 
 #define PACKED_FLAGS_MASK ((0ULL | VRING_PACKED_DESC_F_AVAIL_USED) << \
 	FLAGS_BITS_OFFSET)
@@ -44,8 +51,17 @@
 /* net hdr short size mask */
 #define NET_HDR_MASK 0x3F
 
+#if defined(RTE_ARCH_ARM) || defined(RTE_ARCH_ARM64)
+/* The cache line size on different aarh64 platforms are
+ * different, so put a four batch size here to match with
+ * the minimum cache line size.
+ */
+#define PACKED_BATCH_SIZE 4
+#else
 #define PACKED_BATCH_SIZE (RTE_CACHE_LINE_SIZE / \
 	sizeof(struct vring_packed_desc))
+#endif
+
 #define PACKED_BATCH_MASK (PACKED_BATCH_SIZE - 1)
 
 #ifdef VIRTIO_GCC_UNROLL_PRAGMA
diff --git a/drivers/net/virtio/virtio_rxtx_packed_neon.c b/drivers/net/virtio/virtio_rxtx_packed_neon.c
new file mode 100644
index 000000000..182afe5c6
--- /dev/null
+++ b/drivers/net/virtio/virtio_rxtx_packed_neon.c
@@ -0,0 +1,202 @@
+#include <stdlib.h>
+#include <stdint.h>
+#include <stdio.h>
+#include <string.h>
+#include <errno.h>
+
+#include <rte_net.h>
+#include <rte_vect.h>
+
+#include "virtio_logs.h"
+#include "virtio_ethdev.h"
+#include "virtio_pci.h"
+#include "virtio_rxtx_packed.h"
+#include "virtqueue.h"
+
+static inline uint16_t
+virtqueue_dequeue_batch_packed_vec(struct virtnet_rx *rxvq,
+				   struct rte_mbuf **rx_pkts)
+{
+	struct virtqueue *vq = rxvq->vq;
+	struct virtio_hw *hw = vq->hw;
+	uint16_t head_size = hw->vtnet_hdr_size;
+	uint16_t id = vq->vq_used_cons_idx;
+	struct vring_packed_desc *p_desc;
+	uint16_t i;
+
+	if (id & PACKED_BATCH_MASK)
+		return -1;
+
+	if (unlikely((id + PACKED_BATCH_SIZE) > vq->vq_nentries))
+		return -1;
+
+	/* Map packed descriptor to mbuf fields. */
+	uint8x16_t shuf_msk1 = {
+		0xFF, 0xFF, 0xFF, 0xFF, /* pkt_type set as unknown */
+		0, 1,			/* octet 1~0, low 16 bits pkt_len */
+		0xFF, 0xFF,		/* skip high 16 bits of pkt_len, zero out */
+		0, 1,			/* octet 1~0, 16 bits data_len */
+		0xFF, 0xFF,		/* vlan tci set as unknown */
+		0xFF, 0xFF, 0xFF, 0xFF
+	};
+
+	uint8x16_t shuf_msk2 = {
+		0xFF, 0xFF, 0xFF, 0xFF, /* pkt_type set as unknown */
+		8, 9,			/* octet 9~8, low 16 bits pkt_len */
+		0xFF, 0xFF,		/* skip high 16 bits of pkt_len, zero out */
+		8, 9,			/* octet 9~8, 16 bits data_len */
+		0xFF, 0xFF,		/* vlan tci set as unknown */
+		0xFF, 0xFF, 0xFF, 0xFF
+	};
+
+	/* Subtract the header length. */
+	uint16x8_t len_adjust = {
+		0, 0,		/* ignore pkt_type field */
+		head_size,	/* sub head_size on pkt_len */
+		0,		/* ignore high 16 bits of pkt_len */
+		head_size,	/* sub head_size on data_len */
+		0, 0, 0		/* ignore non-length fields */
+	};
+
+	uint64x2_t desc[PACKED_BATCH_SIZE / 2];
+	uint64x2x2_t mbp[PACKED_BATCH_SIZE / 2];
+	uint64x2_t pkt_mb[PACKED_BATCH_SIZE];
+
+	p_desc = &vq->vq_packed.ring.desc[id];
+	/* Load packed descriptor 0,1. */
+	desc[0] = vld2q_u64((uint64_t *)(p_desc)).val[1];
+	/* Load packed descriptor 2,3. */
+	desc[1] = vld2q_u64((uint64_t *)(p_desc + 2)).val[1];
+
+	/* Only care avail/used bits. */
+	uint32x4_t v_mask = vdupq_n_u32(PACKED_FLAGS_MASK);
+	uint32x4_t v_desc = vuzp2q_u32(vreinterpretq_u32_u64(desc[0]),
+				vreinterpretq_u32_u64(desc[1]));
+	uint32x4_t v_flag = vandq_u32(v_desc, v_mask);
+
+	uint32x4_t v_used_flag = vdupq_n_u32(0);
+	if (vq->vq_packed.used_wrap_counter)
+		v_used_flag = vdupq_n_u32(PACKED_FLAGS_MASK);
+
+	poly128_t desc_stats = vreinterpretq_p128_u32(vceqq_u32(v_flag,
+					v_used_flag));
+
+	/* Check all descs are used. */
+	if (!desc_stats)
+		return -1;
+
+	/* Load 2 mbuf pointers per time. */
+	mbp[0] = vld2q_u64((uint64_t *)&vq->vq_descx[id]);
+	vst1q_u64((uint64_t *)&rx_pkts[0], mbp[0].val[0]);
+
+	mbp[1] = vld2q_u64((uint64_t *)&vq->vq_descx[id + 2]);
+	vst1q_u64((uint64_t *)&rx_pkts[2], mbp[1].val[0]);
+
+	/**
+	 *  Update data length and packet length for descriptor.
+	 *  structure of pkt_mb:
+	 *  --------------------------------------------------------------------
+	 *  |4 octet pkt_type|4 octet pkt_len|2 octet data_len|2 octet vlan_tci|
+	 *  --------------------------------------------------------------------
+	 */
+	pkt_mb[0] = vreinterpretq_u64_u8(vqtbl1q_u8(vreinterpretq_u8_u64(desc[0]), shuf_msk1));
+	pkt_mb[1] = vreinterpretq_u64_u8(vqtbl1q_u8(vreinterpretq_u8_u64(desc[0]), shuf_msk2));
+	pkt_mb[2] = vreinterpretq_u64_u8(vqtbl1q_u8(vreinterpretq_u8_u64(desc[1]), shuf_msk1));
+	pkt_mb[3] = vreinterpretq_u64_u8(vqtbl1q_u8(vreinterpretq_u8_u64(desc[1]), shuf_msk2));
+
+	pkt_mb[0] = vreinterpretq_u64_u16(vsubq_u16(vreinterpretq_u16_u64(pkt_mb[0]), len_adjust));
+	pkt_mb[1] = vreinterpretq_u64_u16(vsubq_u16(vreinterpretq_u16_u64(pkt_mb[1]), len_adjust));
+	pkt_mb[2] = vreinterpretq_u64_u16(vsubq_u16(vreinterpretq_u16_u64(pkt_mb[2]), len_adjust));
+	pkt_mb[3] = vreinterpretq_u64_u16(vsubq_u16(vreinterpretq_u16_u64(pkt_mb[3]), len_adjust));
+
+	vst1q_u64((void *)&rx_pkts[0]->rx_descriptor_fields1, pkt_mb[0]);
+	vst1q_u64((void *)&rx_pkts[1]->rx_descriptor_fields1, pkt_mb[1]);
+	vst1q_u64((void *)&rx_pkts[2]->rx_descriptor_fields1, pkt_mb[2]);
+	vst1q_u64((void *)&rx_pkts[3]->rx_descriptor_fields1, pkt_mb[3]);
+
+	if (hw->has_rx_offload) {
+		virtio_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
+			char *addr = (char *)rx_pkts[i]->buf_addr +
+				RTE_PKTMBUF_HEADROOM - head_size;
+			virtio_vec_rx_offload(rx_pkts[i],
+					(struct virtio_net_hdr *)addr);
+		}
+	}
+
+	virtio_update_batch_stats(&rxvq->stats, rx_pkts[0]->pkt_len,
+			rx_pkts[1]->pkt_len, rx_pkts[2]->pkt_len,
+			rx_pkts[3]->pkt_len);
+
+	vq->vq_free_cnt += PACKED_BATCH_SIZE;
+
+	vq->vq_used_cons_idx += PACKED_BATCH_SIZE;
+	if (vq->vq_used_cons_idx >= vq->vq_nentries) {
+		vq->vq_used_cons_idx -= vq->vq_nentries;
+		vq->vq_packed.used_wrap_counter ^= 1;
+	}
+
+	return 0;
+}
+
+uint16_t
+virtio_recv_pkts_packed_vec(void *rx_queue,
+			    struct rte_mbuf **rx_pkts,
+			    uint16_t nb_pkts)
+{
+	struct virtnet_rx *rxvq = rx_queue;
+	struct virtqueue *vq = rxvq->vq;
+	struct virtio_hw *hw = vq->hw;
+	uint16_t num, nb_rx = 0;
+	uint32_t nb_enqueued = 0;
+	uint16_t free_cnt = vq->vq_free_thresh;
+
+	if (unlikely(hw->started == 0))
+		return nb_rx;
+
+	num = RTE_MIN(VIRTIO_MBUF_BURST_SZ, nb_pkts);
+	if (likely(num > PACKED_BATCH_SIZE))
+		num = num - ((vq->vq_used_cons_idx + num) % PACKED_BATCH_SIZE);
+
+	while (num) {
+		if (!virtqueue_dequeue_batch_packed_vec(rxvq,
+					&rx_pkts[nb_rx])) {
+			nb_rx += PACKED_BATCH_SIZE;
+			num -= PACKED_BATCH_SIZE;
+			continue;
+		}
+		if (!virtqueue_dequeue_single_packed_vec(rxvq,
+					&rx_pkts[nb_rx])) {
+			nb_rx++;
+			num--;
+			continue;
+		}
+		break;
+	};
+
+	PMD_RX_LOG(DEBUG, "dequeue:%d", num);
+
+	rxvq->stats.packets += nb_rx;
+
+	if (likely(vq->vq_free_cnt >= free_cnt)) {
+		struct rte_mbuf *new_pkts[free_cnt];
+		if (likely(rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts,
+						free_cnt) == 0)) {
+			virtio_recv_refill_packed_vec(rxvq, new_pkts,
+					free_cnt);
+			nb_enqueued += free_cnt;
+		} else {
+			struct rte_eth_dev *dev =
+				&rte_eth_devices[rxvq->port_id];
+			dev->data->rx_mbuf_alloc_failed += free_cnt;
+		}
+	}
+
+	if (likely(nb_enqueued)) {
+		if (unlikely(virtqueue_kick_prepare_packed(vq))) {
+			virtqueue_notify(vq);
+			PMD_RX_LOG(DEBUG, "Notified");
+		}
+	}
+
+	return nb_rx;
+}
-- 
2.28.0