drivers/net/virtio/virtio_rxtx_simple_sse.c

   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, nb_total;
  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         struct rte_mbuf **ref_rx_pkts;
  52         uint16_t nb_pkts_received = 0;
  53         __m128i shuf_msk1, shuf_msk2, len_adjust;
  54
  55         shuf_msk1 = _mm_set_epi8(
  56                 0xFF, 0xFF, 0xFF, 0xFF,
  57                 0xFF, 0xFF,             /* vlan tci */
  58                 5, 4,                   /* dat len */
  59                 0xFF, 0xFF, 5, 4,       /* pkt len */
  60                 0xFF, 0xFF, 0xFF, 0xFF  /* packet type */
  61
  62         );
  63
  64         shuf_msk2 = _mm_set_epi8(
  65                 0xFF, 0xFF, 0xFF, 0xFF,
  66                 0xFF, 0xFF,             /* vlan tci */
  67                 13, 12,                 /* dat len */
  68                 0xFF, 0xFF, 13, 12,     /* pkt len */
  69                 0xFF, 0xFF, 0xFF, 0xFF  /* packet type */
  70         );
  71
  72         /* Subtract the header length.
  73         *  In which case do we need the header length in used->len ?
  74         */
  75         len_adjust = _mm_set_epi16(
  76                 0, 0,
  77                 0,
  78                 (uint16_t)-vq->hw->vtnet_hdr_size,
  79                 0, (uint16_t)-vq->hw->vtnet_hdr_size,
  80                 0, 0);
  81
  82         if (unlikely(hw->started == 0))
  83                 return nb_pkts_received;
  84
  85         if (unlikely(nb_pkts < RTE_VIRTIO_DESC_PER_LOOP))
  86                 return 0;
  87
  88         nb_used = VIRTQUEUE_NUSED(vq);
  89
  90         rte_compiler_barrier();
  91
  92         if (unlikely(nb_used == 0))
  93                 return 0;
  94
  95         nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, RTE_VIRTIO_DESC_PER_LOOP);
  96         nb_used = RTE_MIN(nb_used, nb_pkts);
  97
  98         desc_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
  99         rused = &vq->vq_split.ring.used->ring[desc_idx];
 100         sw_ring  = &vq->sw_ring[desc_idx];
 101         sw_ring_end = &vq->sw_ring[vq->vq_nentries];
 102
 103         rte_prefetch0(rused);
 104
 105         if (vq->vq_free_cnt >= RTE_VIRTIO_VPMD_RX_REARM_THRESH) {
 106                 virtio_rxq_rearm_vec(rxvq);
 107                 if (unlikely(virtqueue_kick_prepare(vq)))
 108                         virtqueue_notify(vq);
 109         }
 110
 111         nb_total = nb_used;
 112         ref_rx_pkts = rx_pkts;
 113         for (nb_pkts_received = 0;
 114                 nb_pkts_received < nb_total;) {
 115                 __m128i desc[RTE_VIRTIO_DESC_PER_LOOP / 2];
 116                 __m128i mbp[RTE_VIRTIO_DESC_PER_LOOP / 2];
 117                 __m128i pkt_mb[RTE_VIRTIO_DESC_PER_LOOP];
 118
 119                 mbp[0] = _mm_loadu_si128((__m128i *)(sw_ring + 0));
 120                 desc[0] = _mm_loadu_si128((__m128i *)(rused + 0));
 121                 _mm_storeu_si128((__m128i *)&rx_pkts[0], mbp[0]);
 122
 123                 mbp[1] = _mm_loadu_si128((__m128i *)(sw_ring + 2));
 124                 desc[1] = _mm_loadu_si128((__m128i *)(rused + 2));
 125                 _mm_storeu_si128((__m128i *)&rx_pkts[2], mbp[1]);
 126
 127                 mbp[2] = _mm_loadu_si128((__m128i *)(sw_ring + 4));
 128                 desc[2] = _mm_loadu_si128((__m128i *)(rused + 4));
 129                 _mm_storeu_si128((__m128i *)&rx_pkts[4], mbp[2]);
 130
 131                 mbp[3] = _mm_loadu_si128((__m128i *)(sw_ring + 6));
 132                 desc[3] = _mm_loadu_si128((__m128i *)(rused + 6));
 133                 _mm_storeu_si128((__m128i *)&rx_pkts[6], mbp[3]);
 134
 135                 pkt_mb[1] = _mm_shuffle_epi8(desc[0], shuf_msk2);
 136                 pkt_mb[0] = _mm_shuffle_epi8(desc[0], shuf_msk1);
 137                 pkt_mb[1] = _mm_add_epi16(pkt_mb[1], len_adjust);
 138                 pkt_mb[0] = _mm_add_epi16(pkt_mb[0], len_adjust);
 139                 _mm_storeu_si128((void *)&rx_pkts[1]->rx_descriptor_fields1,
 140                         pkt_mb[1]);
 141                 _mm_storeu_si128((void *)&rx_pkts[0]->rx_descriptor_fields1,
 142                         pkt_mb[0]);
 143
 144                 pkt_mb[3] = _mm_shuffle_epi8(desc[1], shuf_msk2);
 145                 pkt_mb[2] = _mm_shuffle_epi8(desc[1], shuf_msk1);
 146                 pkt_mb[3] = _mm_add_epi16(pkt_mb[3], len_adjust);
 147                 pkt_mb[2] = _mm_add_epi16(pkt_mb[2], len_adjust);
 148                 _mm_storeu_si128((void *)&rx_pkts[3]->rx_descriptor_fields1,
 149                         pkt_mb[3]);
 150                 _mm_storeu_si128((void *)&rx_pkts[2]->rx_descriptor_fields1,
 151                         pkt_mb[2]);
 152
 153                 pkt_mb[5] = _mm_shuffle_epi8(desc[2], shuf_msk2);
 154                 pkt_mb[4] = _mm_shuffle_epi8(desc[2], shuf_msk1);
 155                 pkt_mb[5] = _mm_add_epi16(pkt_mb[5], len_adjust);
 156                 pkt_mb[4] = _mm_add_epi16(pkt_mb[4], len_adjust);
 157                 _mm_storeu_si128((void *)&rx_pkts[5]->rx_descriptor_fields1,
 158                         pkt_mb[5]);
 159                 _mm_storeu_si128((void *)&rx_pkts[4]->rx_descriptor_fields1,
 160                         pkt_mb[4]);
 161
 162                 pkt_mb[7] = _mm_shuffle_epi8(desc[3], shuf_msk2);
 163                 pkt_mb[6] = _mm_shuffle_epi8(desc[3], shuf_msk1);
 164                 pkt_mb[7] = _mm_add_epi16(pkt_mb[7], len_adjust);
 165                 pkt_mb[6] = _mm_add_epi16(pkt_mb[6], len_adjust);
 166                 _mm_storeu_si128((void *)&rx_pkts[7]->rx_descriptor_fields1,
 167                         pkt_mb[7]);
 168                 _mm_storeu_si128((void *)&rx_pkts[6]->rx_descriptor_fields1,
 169                         pkt_mb[6]);
 170
 171                 if (unlikely(nb_used <= RTE_VIRTIO_DESC_PER_LOOP)) {
 172                         if (sw_ring + nb_used <= sw_ring_end)
 173                                 nb_pkts_received += nb_used;
 174                         else
 175                                 nb_pkts_received += sw_ring_end - sw_ring;
 176                         break;
 177                 } else {
 178                         if (unlikely(sw_ring + RTE_VIRTIO_DESC_PER_LOOP >=
 179                                 sw_ring_end)) {
 180                                 nb_pkts_received += sw_ring_end - sw_ring;
 181                                 break;
 182                         } else {
 183                                 nb_pkts_received += RTE_VIRTIO_DESC_PER_LOOP;
 184
 185                                 rx_pkts += RTE_VIRTIO_DESC_PER_LOOP;
 186                                 sw_ring += RTE_VIRTIO_DESC_PER_LOOP;
 187                                 rused   += RTE_VIRTIO_DESC_PER_LOOP;
 188                                 nb_used -= RTE_VIRTIO_DESC_PER_LOOP;
 189                         }
 190                 }
 191         }
 192
 193         vq->vq_used_cons_idx += nb_pkts_received;
 194         vq->vq_free_cnt += nb_pkts_received;
 195         rxvq->stats.packets += nb_pkts_received;
 196         for (nb_used = 0; nb_used < nb_pkts_received; nb_used++)
 197                 virtio_update_packet_stats(&rxvq->stats, ref_rx_pkts[nb_used]);
 198
 199         return nb_pkts_received;
 200 }