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.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_VPMD_RX_BURST 32
  29 #define RTE_VIRTIO_DESC_PER_LOOP 8
  30 #define RTE_VIRTIO_VPMD_RX_REARM_THRESH RTE_VIRTIO_VPMD_RX_BURST
  31
  32 /* virtio vPMD receive routine, only accept(nb_pkts >= RTE_VIRTIO_DESC_PER_LOOP)
  33  *
  34  * This routine is for non-mergeable RX, one desc for each guest buffer.
  35  * This routine is based on the RX ring layout optimization. Each entry in the
  36  * avail ring points to the desc with the same index in the desc ring and this
  37  * will never be changed in the driver.
  38  *
  39  * - nb_pkts < RTE_VIRTIO_DESC_PER_LOOP, just return no packet
  40  */
  41 uint16_t
  42 virtio_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
  43         uint16_t nb_pkts)
  44 {
  45         struct virtnet_rx *rxvq = rx_queue;
  46         struct virtqueue *vq = rxvq->vq;
  47         struct virtio_hw *hw = vq->hw;
  48         uint16_t nb_used;
  49         uint16_t desc_idx;
  50         struct vring_used_elem *rused;
  51         struct rte_mbuf **sw_ring;
  52         struct rte_mbuf **sw_ring_end;
  53         uint16_t nb_pkts_received = 0;
  54         __m128i shuf_msk1, shuf_msk2, len_adjust;
  55
  56         shuf_msk1 = _mm_set_epi8(
  57                 0xFF, 0xFF, 0xFF, 0xFF,
  58                 0xFF, 0xFF,             /* vlan tci */
  59                 5, 4,                   /* dat len */
  60                 0xFF, 0xFF, 5, 4,       /* pkt len */
  61                 0xFF, 0xFF, 0xFF, 0xFF  /* packet type */
  62
  63         );
  64
  65         shuf_msk2 = _mm_set_epi8(
  66                 0xFF, 0xFF, 0xFF, 0xFF,
  67                 0xFF, 0xFF,             /* vlan tci */
  68                 13, 12,                 /* dat len */
  69                 0xFF, 0xFF, 13, 12,     /* pkt len */
  70                 0xFF, 0xFF, 0xFF, 0xFF  /* packet type */
  71         );
  72
  73         /* Subtract the header length.
  74         *  In which case do we need the header length in used->len ?
  75         */
  76         len_adjust = _mm_set_epi16(
  77                 0, 0,
  78                 0,
  79                 (uint16_t)-vq->hw->vtnet_hdr_size,
  80                 0, (uint16_t)-vq->hw->vtnet_hdr_size,
  81                 0, 0);
  82
  83         if (unlikely(hw->started == 0))
  84                 return nb_pkts_received;
  85
  86         if (unlikely(nb_pkts < RTE_VIRTIO_DESC_PER_LOOP))
  87                 return 0;
  88
  89         nb_used = VIRTQUEUE_NUSED(vq);
  90
  91         rte_compiler_barrier();
  92
  93         if (unlikely(nb_used == 0))
  94                 return 0;
  95
  96         nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, RTE_VIRTIO_DESC_PER_LOOP);
  97         nb_used = RTE_MIN(nb_used, nb_pkts);
  98
  99         desc_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
 100         rused = &vq->vq_ring.used->ring[desc_idx];
 101         sw_ring  = &vq->sw_ring[desc_idx];
 102         sw_ring_end = &vq->sw_ring[vq->vq_nentries];
 103
 104         rte_prefetch0(rused);
 105
 106         if (vq->vq_free_cnt >= RTE_VIRTIO_VPMD_RX_REARM_THRESH) {
 107                 virtio_rxq_rearm_vec(rxvq);
 108                 if (unlikely(virtqueue_kick_prepare(vq)))
 109                         virtqueue_notify(vq);
 110         }
 111
 112         for (nb_pkts_received = 0;
 113                 nb_pkts_received < nb_used;) {
 114                 __m128i desc[RTE_VIRTIO_DESC_PER_LOOP / 2];
 115                 __m128i mbp[RTE_VIRTIO_DESC_PER_LOOP / 2];
 116                 __m128i pkt_mb[RTE_VIRTIO_DESC_PER_LOOP];
 117
 118                 mbp[0] = _mm_loadu_si128((__m128i *)(sw_ring + 0));
 119                 desc[0] = _mm_loadu_si128((__m128i *)(rused + 0));
 120                 _mm_storeu_si128((__m128i *)&rx_pkts[0], mbp[0]);
 121
 122                 mbp[1] = _mm_loadu_si128((__m128i *)(sw_ring + 2));
 123                 desc[1] = _mm_loadu_si128((__m128i *)(rused + 2));
 124                 _mm_storeu_si128((__m128i *)&rx_pkts[2], mbp[1]);
 125
 126                 mbp[2] = _mm_loadu_si128((__m128i *)(sw_ring + 4));
 127                 desc[2] = _mm_loadu_si128((__m128i *)(rused + 4));
 128                 _mm_storeu_si128((__m128i *)&rx_pkts[4], mbp[2]);
 129
 130                 mbp[3] = _mm_loadu_si128((__m128i *)(sw_ring + 6));
 131                 desc[3] = _mm_loadu_si128((__m128i *)(rused + 6));
 132                 _mm_storeu_si128((__m128i *)&rx_pkts[6], mbp[3]);
 133
 134                 pkt_mb[1] = _mm_shuffle_epi8(desc[0], shuf_msk2);
 135                 pkt_mb[0] = _mm_shuffle_epi8(desc[0], shuf_msk1);
 136                 pkt_mb[1] = _mm_add_epi16(pkt_mb[1], len_adjust);
 137                 pkt_mb[0] = _mm_add_epi16(pkt_mb[0], len_adjust);
 138                 _mm_storeu_si128((void *)&rx_pkts[1]->rx_descriptor_fields1,
 139                         pkt_mb[1]);
 140                 _mm_storeu_si128((void *)&rx_pkts[0]->rx_descriptor_fields1,
 141                         pkt_mb[0]);
 142
 143                 pkt_mb[3] = _mm_shuffle_epi8(desc[1], shuf_msk2);
 144                 pkt_mb[2] = _mm_shuffle_epi8(desc[1], shuf_msk1);
 145                 pkt_mb[3] = _mm_add_epi16(pkt_mb[3], len_adjust);
 146                 pkt_mb[2] = _mm_add_epi16(pkt_mb[2], len_adjust);
 147                 _mm_storeu_si128((void *)&rx_pkts[3]->rx_descriptor_fields1,
 148                         pkt_mb[3]);
 149                 _mm_storeu_si128((void *)&rx_pkts[2]->rx_descriptor_fields1,
 150                         pkt_mb[2]);
 151
 152                 pkt_mb[5] = _mm_shuffle_epi8(desc[2], shuf_msk2);
 153                 pkt_mb[4] = _mm_shuffle_epi8(desc[2], shuf_msk1);
 154                 pkt_mb[5] = _mm_add_epi16(pkt_mb[5], len_adjust);
 155                 pkt_mb[4] = _mm_add_epi16(pkt_mb[4], len_adjust);
 156                 _mm_storeu_si128((void *)&rx_pkts[5]->rx_descriptor_fields1,
 157                         pkt_mb[5]);
 158                 _mm_storeu_si128((void *)&rx_pkts[4]->rx_descriptor_fields1,
 159                         pkt_mb[4]);
 160
 161                 pkt_mb[7] = _mm_shuffle_epi8(desc[3], shuf_msk2);
 162                 pkt_mb[6] = _mm_shuffle_epi8(desc[3], shuf_msk1);
 163                 pkt_mb[7] = _mm_add_epi16(pkt_mb[7], len_adjust);
 164                 pkt_mb[6] = _mm_add_epi16(pkt_mb[6], len_adjust);
 165                 _mm_storeu_si128((void *)&rx_pkts[7]->rx_descriptor_fields1,
 166                         pkt_mb[7]);
 167                 _mm_storeu_si128((void *)&rx_pkts[6]->rx_descriptor_fields1,
 168                         pkt_mb[6]);
 169
 170                 if (unlikely(nb_used <= RTE_VIRTIO_DESC_PER_LOOP)) {
 171                         if (sw_ring + nb_used <= sw_ring_end)
 172                                 nb_pkts_received += nb_used;
 173                         else
 174                                 nb_pkts_received += sw_ring_end - sw_ring;
 175                         break;
 176                 } else {
 177                         if (unlikely(sw_ring + RTE_VIRTIO_DESC_PER_LOOP >=
 178                                 sw_ring_end)) {
 179                                 nb_pkts_received += sw_ring_end - sw_ring;
 180                                 break;
 181                         } else {
 182                                 nb_pkts_received += RTE_VIRTIO_DESC_PER_LOOP;
 183
 184                                 rx_pkts += RTE_VIRTIO_DESC_PER_LOOP;
 185                                 sw_ring += RTE_VIRTIO_DESC_PER_LOOP;
 186                                 rused   += RTE_VIRTIO_DESC_PER_LOOP;
 187                                 nb_used -= RTE_VIRTIO_DESC_PER_LOOP;
 188                         }
 189                 }
 190         }
 191
 192         vq->vq_used_cons_idx += nb_pkts_received;
 193         vq->vq_free_cnt += nb_pkts_received;
 194         rxvq->stats.packets += nb_pkts_received;
 195         return nb_pkts_received;
 196 }