1 /* SPDX-License-Identifier: BSD-3-Clause 2 * Copyright(c) 2010-2015 Intel Corporation 3 */ 4 5 #include <stdint.h> 6 #include <stdio.h> 7 #include <stdlib.h> 8 #include <string.h> 9 #include <errno.h> 10 11 #include <tmmintrin.h> 12 13 #include <rte_byteorder.h> 14 #include <rte_branch_prediction.h> 15 #include <rte_cycles.h> 16 #include <rte_ether.h> 17 #include <rte_ethdev_driver.h> 18 #include <rte_errno.h> 19 #include <rte_memory.h> 20 #include <rte_mempool.h> 21 #include <rte_malloc.h> 22 #include <rte_mbuf.h> 23 #include <rte_prefetch.h> 24 #include <rte_string_fns.h> 25 26 #include "virtio_rxtx_simple.h" 27 28 #define RTE_VIRTIO_DESC_PER_LOOP 8 29 30 /* virtio vPMD receive routine, only accept(nb_pkts >= RTE_VIRTIO_DESC_PER_LOOP) 31 * 32 * This routine is for non-mergeable RX, one desc for each guest buffer. 33 * This routine is based on the RX ring layout optimization. Each entry in the 34 * avail ring points to the desc with the same index in the desc ring and this 35 * will never be changed in the driver. 36 * 37 * - nb_pkts < RTE_VIRTIO_DESC_PER_LOOP, just return no packet 38 */ 39 uint16_t 40 virtio_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts, 41 uint16_t nb_pkts) 42 { 43 struct virtnet_rx *rxvq = rx_queue; 44 struct virtqueue *vq = rxvq->vq; 45 struct virtio_hw *hw = vq->hw; 46 uint16_t nb_used; 47 uint16_t desc_idx; 48 struct vring_used_elem *rused; 49 struct rte_mbuf **sw_ring; 50 struct rte_mbuf **sw_ring_end; 51 uint16_t nb_pkts_received = 0; 52 __m128i shuf_msk1, shuf_msk2, len_adjust; 53 54 shuf_msk1 = _mm_set_epi8( 55 0xFF, 0xFF, 0xFF, 0xFF, 56 0xFF, 0xFF, /* vlan tci */ 57 5, 4, /* dat len */ 58 0xFF, 0xFF, 5, 4, /* pkt len */ 59 0xFF, 0xFF, 0xFF, 0xFF /* packet type */ 60 61 ); 62 63 shuf_msk2 = _mm_set_epi8( 64 0xFF, 0xFF, 0xFF, 0xFF, 65 0xFF, 0xFF, /* vlan tci */ 66 13, 12, /* dat len */ 67 0xFF, 0xFF, 13, 12, /* pkt len */ 68 0xFF, 0xFF, 0xFF, 0xFF /* packet type */ 69 ); 70 71 /* Subtract the header length. 72 * In which case do we need the header length in used->len ? 73 */ 74 len_adjust = _mm_set_epi16( 75 0, 0, 76 0, 77 (uint16_t)-vq->hw->vtnet_hdr_size, 78 0, (uint16_t)-vq->hw->vtnet_hdr_size, 79 0, 0); 80 81 if (unlikely(hw->started == 0)) 82 return nb_pkts_received; 83 84 if (unlikely(nb_pkts < RTE_VIRTIO_DESC_PER_LOOP)) 85 return 0; 86 87 nb_used = VIRTQUEUE_NUSED(vq); 88 89 rte_compiler_barrier(); 90 91 if (unlikely(nb_used == 0)) 92 return 0; 93 94 nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, RTE_VIRTIO_DESC_PER_LOOP); 95 nb_used = RTE_MIN(nb_used, nb_pkts); 96 97 desc_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1)); 98 rused = &vq->vq_ring.used->ring[desc_idx]; 99 sw_ring = &vq->sw_ring[desc_idx]; 100 sw_ring_end = &vq->sw_ring[vq->vq_nentries]; 101 102 rte_prefetch0(rused); 103 104 if (vq->vq_free_cnt >= RTE_VIRTIO_VPMD_RX_REARM_THRESH) { 105 virtio_rxq_rearm_vec(rxvq); 106 if (unlikely(virtqueue_kick_prepare(vq))) 107 virtqueue_notify(vq); 108 } 109 110 for (nb_pkts_received = 0; 111 nb_pkts_received < nb_used;) { 112 __m128i desc[RTE_VIRTIO_DESC_PER_LOOP / 2]; 113 __m128i mbp[RTE_VIRTIO_DESC_PER_LOOP / 2]; 114 __m128i pkt_mb[RTE_VIRTIO_DESC_PER_LOOP]; 115 116 mbp[0] = _mm_loadu_si128((__m128i *)(sw_ring + 0)); 117 desc[0] = _mm_loadu_si128((__m128i *)(rused + 0)); 118 _mm_storeu_si128((__m128i *)&rx_pkts[0], mbp[0]); 119 120 mbp[1] = _mm_loadu_si128((__m128i *)(sw_ring + 2)); 121 desc[1] = _mm_loadu_si128((__m128i *)(rused + 2)); 122 _mm_storeu_si128((__m128i *)&rx_pkts[2], mbp[1]); 123 124 mbp[2] = _mm_loadu_si128((__m128i *)(sw_ring + 4)); 125 desc[2] = _mm_loadu_si128((__m128i *)(rused + 4)); 126 _mm_storeu_si128((__m128i *)&rx_pkts[4], mbp[2]); 127 128 mbp[3] = _mm_loadu_si128((__m128i *)(sw_ring + 6)); 129 desc[3] = _mm_loadu_si128((__m128i *)(rused + 6)); 130 _mm_storeu_si128((__m128i *)&rx_pkts[6], mbp[3]); 131 132 pkt_mb[1] = _mm_shuffle_epi8(desc[0], shuf_msk2); 133 pkt_mb[0] = _mm_shuffle_epi8(desc[0], shuf_msk1); 134 pkt_mb[1] = _mm_add_epi16(pkt_mb[1], len_adjust); 135 pkt_mb[0] = _mm_add_epi16(pkt_mb[0], len_adjust); 136 _mm_storeu_si128((void *)&rx_pkts[1]->rx_descriptor_fields1, 137 pkt_mb[1]); 138 _mm_storeu_si128((void *)&rx_pkts[0]->rx_descriptor_fields1, 139 pkt_mb[0]); 140 141 pkt_mb[3] = _mm_shuffle_epi8(desc[1], shuf_msk2); 142 pkt_mb[2] = _mm_shuffle_epi8(desc[1], shuf_msk1); 143 pkt_mb[3] = _mm_add_epi16(pkt_mb[3], len_adjust); 144 pkt_mb[2] = _mm_add_epi16(pkt_mb[2], len_adjust); 145 _mm_storeu_si128((void *)&rx_pkts[3]->rx_descriptor_fields1, 146 pkt_mb[3]); 147 _mm_storeu_si128((void *)&rx_pkts[2]->rx_descriptor_fields1, 148 pkt_mb[2]); 149 150 pkt_mb[5] = _mm_shuffle_epi8(desc[2], shuf_msk2); 151 pkt_mb[4] = _mm_shuffle_epi8(desc[2], shuf_msk1); 152 pkt_mb[5] = _mm_add_epi16(pkt_mb[5], len_adjust); 153 pkt_mb[4] = _mm_add_epi16(pkt_mb[4], len_adjust); 154 _mm_storeu_si128((void *)&rx_pkts[5]->rx_descriptor_fields1, 155 pkt_mb[5]); 156 _mm_storeu_si128((void *)&rx_pkts[4]->rx_descriptor_fields1, 157 pkt_mb[4]); 158 159 pkt_mb[7] = _mm_shuffle_epi8(desc[3], shuf_msk2); 160 pkt_mb[6] = _mm_shuffle_epi8(desc[3], shuf_msk1); 161 pkt_mb[7] = _mm_add_epi16(pkt_mb[7], len_adjust); 162 pkt_mb[6] = _mm_add_epi16(pkt_mb[6], len_adjust); 163 _mm_storeu_si128((void *)&rx_pkts[7]->rx_descriptor_fields1, 164 pkt_mb[7]); 165 _mm_storeu_si128((void *)&rx_pkts[6]->rx_descriptor_fields1, 166 pkt_mb[6]); 167 168 if (unlikely(nb_used <= RTE_VIRTIO_DESC_PER_LOOP)) { 169 if (sw_ring + nb_used <= sw_ring_end) 170 nb_pkts_received += nb_used; 171 else 172 nb_pkts_received += sw_ring_end - sw_ring; 173 break; 174 } else { 175 if (unlikely(sw_ring + RTE_VIRTIO_DESC_PER_LOOP >= 176 sw_ring_end)) { 177 nb_pkts_received += sw_ring_end - sw_ring; 178 break; 179 } else { 180 nb_pkts_received += RTE_VIRTIO_DESC_PER_LOOP; 181 182 rx_pkts += RTE_VIRTIO_DESC_PER_LOOP; 183 sw_ring += RTE_VIRTIO_DESC_PER_LOOP; 184 rused += RTE_VIRTIO_DESC_PER_LOOP; 185 nb_used -= RTE_VIRTIO_DESC_PER_LOOP; 186 } 187 } 188 } 189 190 vq->vq_used_cons_idx += nb_pkts_received; 191 vq->vq_free_cnt += nb_pkts_received; 192 rxvq->stats.packets += nb_pkts_received; 193 return nb_pkts_received; 194 } 195