[v3,3/7] net/virtio: add vectorized packed ring Rx function
diff mbox series

Message ID 20200408085313.4487-4-yong.liu@intel.com
State Superseded, archived
Delegated to: Maxime Coquelin
Headers show
Series
  • add packed ring vectorized datapath
Related show

Checks

Context Check Description
ci/Intel-compilation success Compilation OK
ci/checkpatch warning coding style issues

Commit Message

Liu, Yong April 8, 2020, 8:53 a.m. UTC
Optimize packed ring Rx datapath when AVX512 enabled and mergeable
buffer/Rx LRO offloading are not required. Solution of optimization
is pretty like vhost, is that split datapath into batch and single
functions. Batch function is further optimized by vector instructions.
Also pad desc extra structure to 16 bytes aligned, thus four elements
will be saved in one batch.

Signed-off-by: Marvin Liu <yong.liu@intel.com>

Patch
diff mbox series

diff --git a/drivers/net/virtio/Makefile b/drivers/net/virtio/Makefile
index efdcb0d93..7bdb87c49 100644
--- a/drivers/net/virtio/Makefile
+++ b/drivers/net/virtio/Makefile
@@ -37,6 +37,34 @@  else ifneq ($(filter y,$(CONFIG_RTE_ARCH_ARM) $(CONFIG_RTE_ARCH_ARM64)),)
 SRCS-$(CONFIG_RTE_LIBRTE_VIRTIO_PMD) += virtio_rxtx_simple_neon.c
 endif
 
+ifeq ($(RTE_TOOLCHAIN), gcc)
+ifeq ($(shell test $(GCC_VERSION) -ge 83 && echo 1), 1)
+CFLAGS += -DVIRTIO_GCC_UNROLL_PRAGMA
+endif
+endif
+
+ifeq ($(RTE_TOOLCHAIN), clang)
+ifeq ($(shell test $(CLANG_MAJOR_VERSION)$(CLANG_MINOR_VERSION) -ge 37 && echo 1), 1)
+CFLAGS += -DVIRTIO_CLANG_UNROLL_PRAGMA
+endif
+endif
+
+ifeq ($(RTE_TOOLCHAIN), icc)
+ifeq ($(shell test $(ICC_MAJOR_VERSION) -ge 16 && echo 1), 1)
+CFLAGS += -DVIRTIO_ICC_UNROLL_PRAGMA
+endif
+endif
+
+ifeq ($(findstring RTE_MACHINE_CPUFLAG_AVX512F,$(CFLAGS)),RTE_MACHINE_CPUFLAG_AVX512F)
+ifneq ($(FORCE_DISABLE_AVX512), y)
+CFLAGS += -DCC_AVX512_SUPPORT
+ifeq ($(shell test $(GCC_VERSION) -ge 100 && echo 1), 1)
+CFLAGS_virtio_rxtx_packed_avx.o += -Wno-zero-length-bounds
+endif
+SRCS-$(CONFIG_RTE_LIBRTE_VIRTIO_PMD) += virtio_rxtx_packed_avx.c
+endif
+endif
+
 ifeq ($(CONFIG_RTE_VIRTIO_USER),y)
 SRCS-$(CONFIG_RTE_LIBRTE_VIRTIO_PMD) += virtio_user/vhost_user.c
 SRCS-$(CONFIG_RTE_LIBRTE_VIRTIO_PMD) += virtio_user/vhost_kernel.c
diff --git a/drivers/net/virtio/meson.build b/drivers/net/virtio/meson.build
index 04c7fdf25..652ae39af 100644
--- a/drivers/net/virtio/meson.build
+++ b/drivers/net/virtio/meson.build
@@ -11,6 +11,17 @@  deps += ['kvargs', 'bus_pci']
 
 if arch_subdir == 'x86'
 	sources += files('virtio_rxtx_simple_sse.c')
+	if dpdk_conf.has('RTE_MACHINE_CPUFLAG_AVX512F')
+		cflags += ['-DCC_AVX512_SUPPORT']
+		if (toolchain == 'gcc' and cc.version().version_compare('>=8.3.0'))
+			cflags += '-DVHOST_GCC_UNROLL_PRAGMA'
+		elif (toolchain == 'clang' and cc.version().version_compare('>=3.7.0'))
+			cflags += '-DVHOST_CLANG_UNROLL_PRAGMA'
+		elif (toolchain == 'icc' and cc.version().version_compare('>=16.0.0'))
+			cflags += '-DVHOST_ICC_UNROLL_PRAGMA'
+		endif
+		sources += files('virtio_rxtx_packed_avx.c')
+	endif
 elif arch_subdir == 'ppc_64'
 	sources += files('virtio_rxtx_simple_altivec.c')
 elif arch_subdir == 'arm' and host_machine.cpu_family().startswith('aarch64')
diff --git a/drivers/net/virtio/virtio_ethdev.h b/drivers/net/virtio/virtio_ethdev.h
index cd8947656..10e39670e 100644
--- a/drivers/net/virtio/virtio_ethdev.h
+++ b/drivers/net/virtio/virtio_ethdev.h
@@ -104,6 +104,9 @@  uint16_t virtio_xmit_pkts_inorder(void *tx_queue, struct rte_mbuf **tx_pkts,
 uint16_t virtio_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
 		uint16_t nb_pkts);
 
+uint16_t virtio_recv_pkts_packed_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
+		uint16_t nb_pkts);
+
 int eth_virtio_dev_init(struct rte_eth_dev *eth_dev);
 
 void virtio_interrupt_handler(void *param);
diff --git a/drivers/net/virtio/virtio_rxtx.c b/drivers/net/virtio/virtio_rxtx.c
index 3a2dbc2e0..ac417232b 100644
--- a/drivers/net/virtio/virtio_rxtx.c
+++ b/drivers/net/virtio/virtio_rxtx.c
@@ -1245,7 +1245,6 @@  virtio_rx_offload(struct rte_mbuf *m, struct virtio_net_hdr *hdr)
 	return 0;
 }
 
-#define VIRTIO_MBUF_BURST_SZ 64
 #define DESC_PER_CACHELINE (RTE_CACHE_LINE_SIZE / sizeof(struct vring_desc))
 uint16_t
 virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
@@ -2328,3 +2327,11 @@  virtio_xmit_pkts_inorder(void *tx_queue,
 
 	return nb_tx;
 }
+
+__rte_weak uint16_t
+virtio_recv_pkts_packed_vec(void __rte_unused *rx_queue,
+			    struct rte_mbuf __rte_unused **rx_pkts,
+			    uint16_t __rte_unused nb_pkts)
+{
+	return 0;
+}
diff --git a/drivers/net/virtio/virtio_rxtx_packed_avx.c b/drivers/net/virtio/virtio_rxtx_packed_avx.c
new file mode 100644
index 000000000..f2976b98f
--- /dev/null
+++ b/drivers/net/virtio/virtio_rxtx_packed_avx.c
@@ -0,0 +1,358 @@ 
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2010-2020 Intel Corporation
+ */
+
+#include <stdint.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <errno.h>
+
+#include <rte_net.h>
+
+#include "virtio_logs.h"
+#include "virtio_ethdev.h"
+#include "virtio_pci.h"
+#include "virtqueue.h"
+
+#define PACKED_FLAGS_MASK (1ULL << 55 | 1ULL << 63)
+
+#define PACKED_BATCH_SIZE (RTE_CACHE_LINE_SIZE / \
+	sizeof(struct vring_packed_desc))
+#define PACKED_BATCH_MASK (PACKED_BATCH_SIZE - 1)
+
+#ifdef VIRTIO_GCC_UNROLL_PRAGMA
+#define virtio_for_each_try_unroll(iter, val, size) _Pragma("GCC unroll 4") \
+	for (iter = val; iter < size; iter++)
+#endif
+
+#ifdef VIRTIO_CLANG_UNROLL_PRAGMA
+#define virtio_for_each_try_unroll(iter, val, size) _Pragma("unroll 4") \
+	for (iter = val; iter < size; iter++)
+#endif
+
+#ifdef VIRTIO_ICC_UNROLL_PRAGMA
+#define virtio_for_each_try_unroll(iter, val, size) _Pragma("unroll (4)") \
+	for (iter = val; iter < size; iter++)
+#endif
+
+#ifndef virtio_for_each_try_unroll
+#define virtio_for_each_try_unroll(iter, val, num) \
+	for (iter = val; iter < num; iter++)
+#endif
+
+
+static inline void
+virtio_update_batch_stats(struct virtnet_stats *stats,
+			  uint16_t pkt_len1,
+			  uint16_t pkt_len2,
+			  uint16_t pkt_len3,
+			  uint16_t pkt_len4)
+{
+	stats->bytes += pkt_len1;
+	stats->bytes += pkt_len2;
+	stats->bytes += pkt_len3;
+	stats->bytes += pkt_len4;
+}
+/* Optionally fill offload information in structure */
+static inline int
+virtio_vec_rx_offload(struct rte_mbuf *m, struct virtio_net_hdr *hdr)
+{
+	struct rte_net_hdr_lens hdr_lens;
+	uint32_t hdrlen, ptype;
+	int l4_supported = 0;
+
+	/* nothing to do */
+	if (hdr->flags == 0)
+		return 0;
+
+	/* GSO not support in vec path, skip check */
+	m->ol_flags |= PKT_RX_IP_CKSUM_UNKNOWN;
+
+	ptype = rte_net_get_ptype(m, &hdr_lens, RTE_PTYPE_ALL_MASK);
+	m->packet_type = ptype;
+	if ((ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_TCP ||
+	    (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_UDP ||
+	    (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_SCTP)
+		l4_supported = 1;
+
+	if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
+		hdrlen = hdr_lens.l2_len + hdr_lens.l3_len + hdr_lens.l4_len;
+		if (hdr->csum_start <= hdrlen && l4_supported) {
+			m->ol_flags |= PKT_RX_L4_CKSUM_NONE;
+		} else {
+			/* Unknown proto or tunnel, do sw cksum. We can assume
+			 * the cksum field is in the first segment since the
+			 * buffers we provided to the host are large enough.
+			 * In case of SCTP, this will be wrong since it's a CRC
+			 * but there's nothing we can do.
+			 */
+			uint16_t csum = 0, off;
+
+			rte_raw_cksum_mbuf(m, hdr->csum_start,
+				rte_pktmbuf_pkt_len(m) - hdr->csum_start,
+				&csum);
+			if (likely(csum != 0xffff))
+				csum = ~csum;
+			off = hdr->csum_offset + hdr->csum_start;
+			if (rte_pktmbuf_data_len(m) >= off + 1)
+				*rte_pktmbuf_mtod_offset(m, uint16_t *,
+					off) = csum;
+		}
+	} else if (hdr->flags & VIRTIO_NET_HDR_F_DATA_VALID && l4_supported) {
+		m->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
+	}
+
+	return 0;
+}
+
+static 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 hdr_size = hw->vtnet_hdr_size;
+	uint64_t addrs[PACKED_BATCH_SIZE << 1];
+	uint16_t id = vq->vq_used_cons_idx;
+	uint8_t desc_stats;
+	uint16_t i;
+	void *desc_addr;
+
+	if (id & PACKED_BATCH_MASK)
+		return -1;
+
+	/* only care avail/used bits */
+	__m512i desc_flags = _mm512_maskz_set1_epi64(0xaa, PACKED_FLAGS_MASK);
+	desc_addr = &vq->vq_packed.ring.desc[id];
+
+	rte_smp_rmb();
+	__m512i packed_desc = _mm512_loadu_si512(desc_addr);
+	__m512i flags_mask  = _mm512_maskz_and_epi64(0xff, packed_desc,
+			desc_flags);
+
+	__m512i used_flags;
+	if (vq->vq_packed.used_wrap_counter)
+		used_flags = _mm512_maskz_set1_epi64(0xaa, PACKED_FLAGS_MASK);
+	else
+		used_flags = _mm512_setzero_si512();
+
+	/* Check all descs are used */
+	desc_stats = _mm512_cmp_epu64_mask(flags_mask, used_flags,
+			_MM_CMPINT_EQ);
+	if (desc_stats != 0xff)
+		return -1;
+
+	virtio_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
+		rx_pkts[i] = (struct rte_mbuf *)vq->vq_descx[id + i].cookie;
+		rte_packet_prefetch(rte_pktmbuf_mtod(rx_pkts[i], void *));
+
+		addrs[i << 1] = (uint64_t)rx_pkts[i]->rx_descriptor_fields1;
+		addrs[(i << 1) + 1] =
+			(uint64_t)rx_pkts[i]->rx_descriptor_fields1 + 8;
+	}
+
+	/* addresses of pkt_len and data_len */
+	__m512i vindex = _mm512_loadu_si512((void *)addrs);
+
+	/*
+	 * select 10b*4 load 32bit from packed_desc[95:64]
+	 * mmask  0110b*4 save 32bit into pkt_len and data_len
+	 */
+	__m512i value = _mm512_maskz_shuffle_epi32(0x6666, packed_desc, 0xAA);
+
+	/* mmask 0110b*4 reduce hdr_len from pkt_len and data_len */
+	__m512i mbuf_len_offset = _mm512_maskz_set1_epi32(0x6666,
+			(uint32_t)-hdr_size);
+
+	value = _mm512_add_epi32(value, mbuf_len_offset);
+	/* batch store into mbufs */
+	_mm512_i64scatter_epi64(0, vindex, value, 1);
+
+	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 - hdr_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;
+}
+
+static uint16_t
+virtqueue_dequeue_single_packed_vec(struct virtnet_rx *rxvq,
+				    struct rte_mbuf **rx_pkts)
+{
+	uint16_t used_idx, id;
+	uint32_t len;
+	struct virtqueue *vq = rxvq->vq;
+	struct virtio_hw *hw = vq->hw;
+	uint32_t hdr_size = hw->vtnet_hdr_size;
+	struct virtio_net_hdr *hdr;
+	struct vring_packed_desc *desc;
+	struct rte_mbuf *cookie;
+
+	desc = vq->vq_packed.ring.desc;
+	used_idx = vq->vq_used_cons_idx;
+	if (!desc_is_used(&desc[used_idx], vq))
+		return -1;
+
+	len = desc[used_idx].len;
+	id = desc[used_idx].id;
+	cookie = (struct rte_mbuf *)vq->vq_descx[id].cookie;
+	if (unlikely(cookie == NULL)) {
+		PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u",
+				vq->vq_used_cons_idx);
+		return -1;
+	}
+	rte_prefetch0(cookie);
+	rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
+
+	cookie->data_off = RTE_PKTMBUF_HEADROOM;
+	cookie->ol_flags = 0;
+	cookie->pkt_len = (uint32_t)(len - hdr_size);
+	cookie->data_len = (uint32_t)(len - hdr_size);
+
+	hdr = (struct virtio_net_hdr *)((char *)cookie->buf_addr +
+					RTE_PKTMBUF_HEADROOM - hdr_size);
+	if (hw->has_rx_offload)
+		virtio_vec_rx_offload(cookie, hdr);
+
+	*rx_pkts = cookie;
+
+	rxvq->stats.bytes += cookie->pkt_len;
+
+	vq->vq_free_cnt++;
+	vq->vq_used_cons_idx++;
+	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;
+}
+
+static inline void
+virtio_recv_refill_packed_vec(struct virtnet_rx *rxvq,
+			      struct rte_mbuf **cookie,
+			      uint16_t num)
+{
+	struct virtqueue *vq = rxvq->vq;
+	struct vring_packed_desc *start_dp = vq->vq_packed.ring.desc;
+	uint16_t flags = vq->vq_packed.cached_flags;
+	struct virtio_hw *hw = vq->hw;
+	struct vq_desc_extra *dxp;
+	uint16_t idx, i;
+	uint16_t total_num = 0;
+	uint16_t head_idx = vq->vq_avail_idx;
+	uint16_t head_flag = vq->vq_packed.cached_flags;
+	uint64_t addr;
+
+	do {
+		idx = vq->vq_avail_idx;
+		virtio_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
+			dxp = &vq->vq_descx[idx + i];
+			dxp->cookie = (void *)cookie[total_num + i];
+
+			addr = VIRTIO_MBUF_ADDR(cookie[total_num + i], vq) +
+				RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
+			start_dp[idx + i].addr = addr;
+			start_dp[idx + i].len = cookie[total_num + i]->buf_len
+				- RTE_PKTMBUF_HEADROOM + hw->vtnet_hdr_size;
+			if (total_num || i) {
+				virtqueue_store_flags_packed(&start_dp[idx + i],
+						flags, hw->weak_barriers);
+			}
+		}
+
+		vq->vq_avail_idx += PACKED_BATCH_SIZE;
+		if (vq->vq_avail_idx >= vq->vq_nentries) {
+			vq->vq_avail_idx -= vq->vq_nentries;
+			vq->vq_packed.cached_flags ^=
+				VRING_PACKED_DESC_F_AVAIL_USED;
+			flags = vq->vq_packed.cached_flags;
+		}
+		total_num += PACKED_BATCH_SIZE;
+	} while (total_num < num);
+
+	virtqueue_store_flags_packed(&start_dp[head_idx], head_flag,
+				hw->weak_barriers);
+	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
+}
+
+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;
+}
diff --git a/drivers/net/virtio/virtqueue.h b/drivers/net/virtio/virtqueue.h
index 6301c56b2..43e305ecc 100644
--- a/drivers/net/virtio/virtqueue.h
+++ b/drivers/net/virtio/virtqueue.h
@@ -20,6 +20,7 @@  struct rte_mbuf;
 
 #define DEFAULT_RX_FREE_THRESH 32
 
+#define VIRTIO_MBUF_BURST_SZ 64
 /*
  * Per virtio_ring.h in Linux.
  *     For virtio_pci on SMP, we don't need to order with respect to MMIO
@@ -236,7 +237,8 @@  struct vq_desc_extra {
 	void *cookie;
 	uint16_t ndescs;
 	uint16_t next;
-};
+	uint8_t padding[4];
+} __rte_packed __rte_aligned(16);
 
 struct virtqueue {
 	struct virtio_hw  *hw; /**< virtio_hw structure pointer. */