From fc3d66212fed67b499e80cae06106af64784bf54 Mon Sep 17 00:00:00 2001 From: Huawei Xie Date: Thu, 29 Oct 2015 22:53:26 +0800 Subject: [PATCH] virtio: add vector Rx With fixed avail ring, we don't need to get desc idx from avail ring. virtio driver only has to deal with desc ring. This patch uses vector instruction to accelerate processing desc ring. Signed-off-by: Huawei Xie Acked-by: Jianfeng Tan --- doc/guides/rel_notes/release_2_2.rst | 1 + drivers/net/virtio/virtio_ethdev.h | 2 + drivers/net/virtio/virtio_rxtx.c | 3 + drivers/net/virtio/virtio_rxtx.h | 2 + drivers/net/virtio/virtio_rxtx_simple.c | 228 ++++++++++++++++++++++++ drivers/net/virtio/virtqueue.h | 1 + 6 files changed, 237 insertions(+) diff --git a/doc/guides/rel_notes/release_2_2.rst b/doc/guides/rel_notes/release_2_2.rst index ba730f64e2..c523e576f8 100644 --- a/doc/guides/rel_notes/release_2_2.rst +++ b/doc/guides/rel_notes/release_2_2.rst @@ -58,6 +58,7 @@ New Features * **Enhanced support for virtio driver.** * Virtio ring layout optimization (fixed avail ring) + * Vector RX * **Added vhost-user multiple queue support.** diff --git a/drivers/net/virtio/virtio_ethdev.h b/drivers/net/virtio/virtio_ethdev.h index 9026d422e8..d7797ab9b0 100644 --- a/drivers/net/virtio/virtio_ethdev.h +++ b/drivers/net/virtio/virtio_ethdev.h @@ -108,6 +108,8 @@ uint16_t virtio_recv_mergeable_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts); +uint16_t virtio_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts, + uint16_t nb_pkts); /* * The VIRTIO_NET_F_GUEST_TSO[46] features permit the host to send us diff --git a/drivers/net/virtio/virtio_rxtx.c b/drivers/net/virtio/virtio_rxtx.c index 4fd75461f6..8db36907be 100644 --- a/drivers/net/virtio/virtio_rxtx.c +++ b/drivers/net/virtio/virtio_rxtx.c @@ -432,6 +432,9 @@ virtio_dev_rx_queue_setup(struct rte_eth_dev *dev, vq->mpool = mp; dev->data->rx_queues[queue_idx] = vq; + + virtio_rxq_vec_setup(vq); + return 0; } diff --git a/drivers/net/virtio/virtio_rxtx.h b/drivers/net/virtio/virtio_rxtx.h index 7d2d8fe96c..831e492672 100644 --- a/drivers/net/virtio/virtio_rxtx.h +++ b/drivers/net/virtio/virtio_rxtx.h @@ -33,5 +33,7 @@ #define RTE_PMD_VIRTIO_RX_MAX_BURST 64 +int virtio_rxq_vec_setup(struct virtqueue *rxq); + int virtqueue_enqueue_recv_refill_simple(struct virtqueue *vq, struct rte_mbuf *m); diff --git a/drivers/net/virtio/virtio_rxtx_simple.c b/drivers/net/virtio/virtio_rxtx_simple.c index cac5b9f17d..f6c00c118b 100644 --- a/drivers/net/virtio/virtio_rxtx_simple.c +++ b/drivers/net/virtio/virtio_rxtx_simple.c @@ -58,6 +58,14 @@ #include "virtqueue.h" #include "virtio_rxtx.h" +#define RTE_VIRTIO_VPMD_RX_BURST 32 +#define RTE_VIRTIO_DESC_PER_LOOP 8 +#define RTE_VIRTIO_VPMD_RX_REARM_THRESH RTE_VIRTIO_VPMD_RX_BURST + +#ifndef __INTEL_COMPILER +#pragma GCC diagnostic ignored "-Wcast-qual" +#endif + int __attribute__((cold)) virtqueue_enqueue_recv_refill_simple(struct virtqueue *vq, struct rte_mbuf *cookie) @@ -82,3 +90,223 @@ virtqueue_enqueue_recv_refill_simple(struct virtqueue *vq, return 0; } + +static inline void +virtio_rxq_rearm_vec(struct virtqueue *rxvq) +{ + int i; + uint16_t desc_idx; + struct rte_mbuf **sw_ring; + struct vring_desc *start_dp; + int ret; + + desc_idx = rxvq->vq_avail_idx & (rxvq->vq_nentries - 1); + sw_ring = &rxvq->sw_ring[desc_idx]; + start_dp = &rxvq->vq_ring.desc[desc_idx]; + + ret = rte_mempool_get_bulk(rxvq->mpool, (void **)sw_ring, + RTE_VIRTIO_VPMD_RX_REARM_THRESH); + if (unlikely(ret)) { + rte_eth_devices[rxvq->port_id].data->rx_mbuf_alloc_failed += + RTE_VIRTIO_VPMD_RX_REARM_THRESH; + return; + } + + for (i = 0; i < RTE_VIRTIO_VPMD_RX_REARM_THRESH; i++) { + uintptr_t p; + + p = (uintptr_t)&sw_ring[i]->rearm_data; + *(uint64_t *)p = rxvq->mbuf_initializer; + + start_dp[i].addr = + (uint64_t)((uintptr_t)sw_ring[i]->buf_physaddr + + RTE_PKTMBUF_HEADROOM - sizeof(struct virtio_net_hdr)); + start_dp[i].len = sw_ring[i]->buf_len - + RTE_PKTMBUF_HEADROOM + sizeof(struct virtio_net_hdr); + } + + rxvq->vq_avail_idx += RTE_VIRTIO_VPMD_RX_REARM_THRESH; + rxvq->vq_free_cnt -= RTE_VIRTIO_VPMD_RX_REARM_THRESH; + vq_update_avail_idx(rxvq); +} + +/* virtio vPMD receive routine, only accept(nb_pkts >= RTE_VIRTIO_DESC_PER_LOOP) + * + * This routine is for non-mergeable RX, one desc for each guest buffer. + * This routine is based on the RX ring layout optimization. Each entry in the + * avail ring points to the desc with the same index in the desc ring and this + * will never be changed in the driver. + * + * - nb_pkts < RTE_VIRTIO_DESC_PER_LOOP, just return no packet + */ +uint16_t +virtio_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts, + uint16_t nb_pkts) +{ + struct virtqueue *rxvq = rx_queue; + uint16_t nb_used; + uint16_t desc_idx; + struct vring_used_elem *rused; + struct rte_mbuf **sw_ring; + struct rte_mbuf **sw_ring_end; + uint16_t nb_pkts_received; + __m128i shuf_msk1, shuf_msk2, len_adjust; + + shuf_msk1 = _mm_set_epi8( + 0xFF, 0xFF, 0xFF, 0xFF, + 0xFF, 0xFF, /* vlan tci */ + 5, 4, /* dat len */ + 0xFF, 0xFF, 5, 4, /* pkt len */ + 0xFF, 0xFF, 0xFF, 0xFF /* packet type */ + + ); + + shuf_msk2 = _mm_set_epi8( + 0xFF, 0xFF, 0xFF, 0xFF, + 0xFF, 0xFF, /* vlan tci */ + 13, 12, /* dat len */ + 0xFF, 0xFF, 13, 12, /* pkt len */ + 0xFF, 0xFF, 0xFF, 0xFF /* packet type */ + ); + + /* Subtract the header length. + * In which case do we need the header length in used->len ? + */ + len_adjust = _mm_set_epi16( + 0, 0, + 0, + (uint16_t) -sizeof(struct virtio_net_hdr), + 0, (uint16_t) -sizeof(struct virtio_net_hdr), + 0, 0); + + if (unlikely(nb_pkts < RTE_VIRTIO_DESC_PER_LOOP)) + return 0; + + nb_used = *(volatile uint16_t *)&rxvq->vq_ring.used->idx - + rxvq->vq_used_cons_idx; + + rte_compiler_barrier(); + + if (unlikely(nb_used == 0)) + return 0; + + nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, RTE_VIRTIO_DESC_PER_LOOP); + nb_used = RTE_MIN(nb_used, nb_pkts); + + desc_idx = (uint16_t)(rxvq->vq_used_cons_idx & (rxvq->vq_nentries - 1)); + rused = &rxvq->vq_ring.used->ring[desc_idx]; + sw_ring = &rxvq->sw_ring[desc_idx]; + sw_ring_end = &rxvq->sw_ring[rxvq->vq_nentries]; + + _mm_prefetch((const void *)rused, _MM_HINT_T0); + + if (rxvq->vq_free_cnt >= RTE_VIRTIO_VPMD_RX_REARM_THRESH) { + virtio_rxq_rearm_vec(rxvq); + if (unlikely(virtqueue_kick_prepare(rxvq))) + virtqueue_notify(rxvq); + } + + for (nb_pkts_received = 0; + nb_pkts_received < nb_used;) { + __m128i desc[RTE_VIRTIO_DESC_PER_LOOP / 2]; + __m128i mbp[RTE_VIRTIO_DESC_PER_LOOP / 2]; + __m128i pkt_mb[RTE_VIRTIO_DESC_PER_LOOP]; + + mbp[0] = _mm_loadu_si128((__m128i *)(sw_ring + 0)); + desc[0] = _mm_loadu_si128((__m128i *)(rused + 0)); + _mm_storeu_si128((__m128i *)&rx_pkts[0], mbp[0]); + + mbp[1] = _mm_loadu_si128((__m128i *)(sw_ring + 2)); + desc[1] = _mm_loadu_si128((__m128i *)(rused + 2)); + _mm_storeu_si128((__m128i *)&rx_pkts[2], mbp[1]); + + mbp[2] = _mm_loadu_si128((__m128i *)(sw_ring + 4)); + desc[2] = _mm_loadu_si128((__m128i *)(rused + 4)); + _mm_storeu_si128((__m128i *)&rx_pkts[4], mbp[2]); + + mbp[3] = _mm_loadu_si128((__m128i *)(sw_ring + 6)); + desc[3] = _mm_loadu_si128((__m128i *)(rused + 6)); + _mm_storeu_si128((__m128i *)&rx_pkts[6], mbp[3]); + + pkt_mb[1] = _mm_shuffle_epi8(desc[0], shuf_msk2); + pkt_mb[0] = _mm_shuffle_epi8(desc[0], shuf_msk1); + pkt_mb[1] = _mm_add_epi16(pkt_mb[1], len_adjust); + pkt_mb[0] = _mm_add_epi16(pkt_mb[0], len_adjust); + _mm_storeu_si128((void *)&rx_pkts[1]->rx_descriptor_fields1, + pkt_mb[1]); + _mm_storeu_si128((void *)&rx_pkts[0]->rx_descriptor_fields1, + pkt_mb[0]); + + pkt_mb[3] = _mm_shuffle_epi8(desc[1], shuf_msk2); + pkt_mb[2] = _mm_shuffle_epi8(desc[1], shuf_msk1); + pkt_mb[3] = _mm_add_epi16(pkt_mb[3], len_adjust); + pkt_mb[2] = _mm_add_epi16(pkt_mb[2], len_adjust); + _mm_storeu_si128((void *)&rx_pkts[3]->rx_descriptor_fields1, + pkt_mb[3]); + _mm_storeu_si128((void *)&rx_pkts[2]->rx_descriptor_fields1, + pkt_mb[2]); + + pkt_mb[5] = _mm_shuffle_epi8(desc[2], shuf_msk2); + pkt_mb[4] = _mm_shuffle_epi8(desc[2], shuf_msk1); + pkt_mb[5] = _mm_add_epi16(pkt_mb[5], len_adjust); + pkt_mb[4] = _mm_add_epi16(pkt_mb[4], len_adjust); + _mm_storeu_si128((void *)&rx_pkts[5]->rx_descriptor_fields1, + pkt_mb[5]); + _mm_storeu_si128((void *)&rx_pkts[4]->rx_descriptor_fields1, + pkt_mb[4]); + + pkt_mb[7] = _mm_shuffle_epi8(desc[3], shuf_msk2); + pkt_mb[6] = _mm_shuffle_epi8(desc[3], shuf_msk1); + pkt_mb[7] = _mm_add_epi16(pkt_mb[7], len_adjust); + pkt_mb[6] = _mm_add_epi16(pkt_mb[6], len_adjust); + _mm_storeu_si128((void *)&rx_pkts[7]->rx_descriptor_fields1, + pkt_mb[7]); + _mm_storeu_si128((void *)&rx_pkts[6]->rx_descriptor_fields1, + pkt_mb[6]); + + if (unlikely(nb_used <= RTE_VIRTIO_DESC_PER_LOOP)) { + if (sw_ring + nb_used <= sw_ring_end) + nb_pkts_received += nb_used; + else + nb_pkts_received += sw_ring_end - sw_ring; + break; + } else { + if (unlikely(sw_ring + RTE_VIRTIO_DESC_PER_LOOP >= + sw_ring_end)) { + nb_pkts_received += sw_ring_end - sw_ring; + break; + } else { + nb_pkts_received += RTE_VIRTIO_DESC_PER_LOOP; + + rx_pkts += RTE_VIRTIO_DESC_PER_LOOP; + sw_ring += RTE_VIRTIO_DESC_PER_LOOP; + rused += RTE_VIRTIO_DESC_PER_LOOP; + nb_used -= RTE_VIRTIO_DESC_PER_LOOP; + } + } + } + + rxvq->vq_used_cons_idx += nb_pkts_received; + rxvq->vq_free_cnt += nb_pkts_received; + rxvq->packets += nb_pkts_received; + return nb_pkts_received; +} + +int __attribute__((cold)) +virtio_rxq_vec_setup(struct virtqueue *rxq) +{ + uintptr_t p; + struct rte_mbuf mb_def = { .buf_addr = 0 }; /* zeroed mbuf */ + + mb_def.nb_segs = 1; + mb_def.data_off = RTE_PKTMBUF_HEADROOM; + mb_def.port = rxq->port_id; + rte_mbuf_refcnt_set(&mb_def, 1); + + /* prevent compiler reordering: rearm_data covers previous fields */ + rte_compiler_barrier(); + p = (uintptr_t)&mb_def.rearm_data; + rxq->mbuf_initializer = *(uint64_t *)p; + + return 0; +} diff --git a/drivers/net/virtio/virtqueue.h b/drivers/net/virtio/virtqueue.h index 6a1ec48b86..98a77d5e03 100644 --- a/drivers/net/virtio/virtqueue.h +++ b/drivers/net/virtio/virtqueue.h @@ -188,6 +188,7 @@ struct virtqueue { */ uint16_t vq_used_cons_idx; uint16_t vq_avail_idx; + uint64_t mbuf_initializer; /**< value to init mbufs. */ phys_addr_t virtio_net_hdr_mem; /**< hdr for each xmit packet */ struct rte_mbuf **sw_ring; /**< RX software ring. */ -- 2.20.1