On 4/24/20 11:24 AM, Marvin Liu wrote:
> Optimize packed ring Rx path with SIMD instructions. Solution of
> optimization is pretty like vhost, is that split path into batch and
> single functions. Batch function is further optimized by AVX512
> 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....@intel.com>
>
> diff --git a/drivers/net/virtio/Makefile b/drivers/net/virtio/Makefile
> index c9edb84ee..102b1deab 100644
> --- a/drivers/net/virtio/Makefile
> +++ b/drivers/net/virtio/Makefile
> @@ -36,6 +36,41 @@ else ifneq ($(filter y,$(CONFIG_RTE_ARCH_ARM)
> $(CONFIG_RTE_ARCH_ARM64)),)
> SRCS-$(CONFIG_RTE_LIBRTE_VIRTIO_PMD) += virtio_rxtx_simple_neon.c
> endif
>
> +ifneq ($(FORCE_DISABLE_AVX512), y)
> + CC_AVX512_SUPPORT=\
> + $(shell $(CC) -march=native -dM -E - </dev/null 2>&1 | \
> + sed '/./{H;$$!d} ; x ; /AVX512F/!d; /AVX512BW/!d; /AVX512VL/!d' | \
> + grep -q AVX512 && echo 1)
> +endif
> +
> +ifeq ($(CC_AVX512_SUPPORT), 1)
> +CFLAGS += -DCC_AVX512_SUPPORT
> +SRCS-$(CONFIG_RTE_LIBRTE_VIRTIO_PMD) += virtio_rxtx_packed_avx.c
> +
> +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 $(ICC_MAJOR_VERSION) -ge 16 && echo 1), 1)
> +CFLAGS += -DVIRTIO_ICC_UNROLL_PRAGMA
> +endif
> +endif
> +
> +CFLAGS_virtio_rxtx_packed_avx.o += -mavx512f -mavx512bw -mavx512vl
> +ifeq ($(shell test $(GCC_VERSION) -ge 100 && echo 1), 1)
> +CFLAGS_virtio_rxtx_packed_avx.o += -Wno-zero-length-bounds
> +endif
> +endif
> +
> ifeq ($(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 15150eea1..8e68c3039 100644
> --- a/drivers/net/virtio/meson.build
> +++ b/drivers/net/virtio/meson.build
> @@ -9,6 +9,20 @@ sources += files('virtio_ethdev.c',
> deps += ['kvargs', 'bus_pci']
>
> if arch_subdir == 'x86'
> + if '-mno-avx512f' not in machine_args
> + if cc.has_argument('-mavx512f') and
> cc.has_argument('-mavx512vl') and cc.has_argument('-mavx512bw')
> + cflags += ['-mavx512f', '-mavx512bw', '-mavx512vl']
> + 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
> + endif
> sources += files('virtio_rxtx_simple_sse.c')
> elif arch_subdir == 'ppc'
> sources += files('virtio_rxtx_simple_altivec.c')
> diff --git a/drivers/net/virtio/virtio_ethdev.h
> b/drivers/net/virtio/virtio_ethdev.h
> index febaf17a8..5c112cac7 100644
> --- a/drivers/net/virtio/virtio_ethdev.h
> +++ b/drivers/net/virtio/virtio_ethdev.h
> @@ -105,6 +105,9 @@ 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 84f4cf946..c9b6e7844 100644
> --- a/drivers/net/virtio/virtio_rxtx.c
> +++ b/drivers/net/virtio/virtio_rxtx.c
> @@ -2329,3 +2329,11 @@ virtio_xmit_pkts_inorder(void *tx_queue,
>
> return nb_tx;
> }
> +
> +__rte_weak uint16_t
> +virtio_recv_pkts_packed_vec(void *rx_queue __rte_unused,
> + struct rte_mbuf **rx_pkts __rte_unused,
> + uint16_t nb_pkts __rte_unused)
> +{
> + 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..8a7b459eb
> --- /dev/null
> +++ b/drivers/net/virtio/virtio_rxtx_packed_avx.c
> @@ -0,0 +1,374 @@
> +/* 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 BYTE_SIZE 8
> +/* 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)
> +
> +#define PACKED_FLAGS_MASK ((0ULL | VRING_PACKED_DESC_F_AVAIL_USED) << \
> + FLAGS_BITS_OFFSET)
> +
> +#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;
IIUC, the only difference with the non-vectorized version is the GSO
support removed here.
gso_type being in the same cacheline as flags in virtio_net_hdr, I don't
think checking the performance gain is worth the added maintainance
effort due to code duplication.
Please prove I'm wrong, otherwise please move virtio_rx_offload() in a
header and use it here. Alternative if it really imapcts performance is
to put all the shared code in a dedicated function that can be re-used
by both implementations.
> +
> + /* 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;
> +}
Otherwise, the patch looks okay to me.
Thanks,
Maxime
