drivers/net/virtio/virtio_rxtx_simple_altivec.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2015 Intel Corporation
   3  * Copyright(C) 2019 IBM Corporation
   4  */
   5
   6 #include <stdint.h>
   7 #include <stdio.h>
   8 #include <stdlib.h>
   9 #include <string.h>
  10 #include <errno.h>
  11
  12 #include <altivec.h>
  13
  14 #include <rte_byteorder.h>
  15 #include <rte_branch_prediction.h>
  16 #include <rte_cycles.h>
  17 #include <rte_ether.h>
  18 #include <rte_ethdev_driver.h>
  19 #include <rte_errno.h>
  20 #include <rte_memory.h>
  21 #include <rte_mempool.h>
  22 #include <rte_malloc.h>
  23 #include <rte_mbuf.h>
  24 #include <rte_prefetch.h>
  25 #include <rte_string_fns.h>
  26
  27 #include "virtio_rxtx_simple.h"
  28
  29 #define RTE_VIRTIO_DESC_PER_LOOP 8
  30
  31 /* virtio vPMD receive routine, only accept(nb_pkts >= RTE_VIRTIO_DESC_PER_LOOP)
  32  *
  33  * This routine is for non-mergeable RX, one desc for each guest buffer.
  34  * This routine is based on the RX ring layout optimization. Each entry in the
  35  * avail ring points to the desc with the same index in the desc ring and this
  36  * will never be changed in the driver.
  37  *
  38  * - nb_pkts < RTE_VIRTIO_DESC_PER_LOOP, just return no packet
  39  */
  40 uint16_t
  41 virtio_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
  42         uint16_t nb_pkts)
  43 {
  44         struct virtnet_rx *rxvq = rx_queue;
  45         struct virtqueue *vq = rxvq->vq;
  46         struct virtio_hw *hw = vq->hw;
  47         uint16_t nb_used, nb_total;
  48         uint16_t desc_idx;
  49         struct vring_used_elem *rused;
  50         struct rte_mbuf **sw_ring;
  51         struct rte_mbuf **sw_ring_end;
  52         struct rte_mbuf **ref_rx_pkts;
  53         uint16_t nb_pkts_received = 0;
  54         const vector unsigned char zero = {0};
  55
  56         const vector unsigned char shuf_msk1 = {
  57                 0xFF, 0xFF, 0xFF, 0xFF, /* packet type */
  58                 4, 5, 0xFF, 0xFF, /* vlan tci */
  59                 4, 5,                   /* dat len */
  60                 0xFF, 0xFF,             /* vlan tci */
  61                 0xFF, 0xFF, 0xFF, 0xFF
  62         };
  63
  64         const vector unsigned char shuf_msk2 = {
  65                 0xFF, 0xFF, 0xFF, 0xFF, /* packet type */
  66                 12, 13, 0xFF, 0xFF,     /* pkt len */
  67                 12, 13,                 /* dat len */
  68                 0xFF, 0xFF,             /* vlan tci */
  69                 0xFF, 0xFF, 0xFF, 0xFF
  70         };
  71
  72         /*
  73          * Subtract the header length.
  74          *  In which case do we need the header length in used->len ?
  75          */
  76         const vector unsigned short len_adjust = {
  77                 0, 0,
  78                 (uint16_t)-vq->hw->vtnet_hdr_size, 0,
  79                 (uint16_t)-vq->hw->vtnet_hdr_size, 0,
  80                 0, 0
  81         };
  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_split.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         nb_total = nb_used;
 113         ref_rx_pkts = rx_pkts;
 114         for (nb_pkts_received = 0;
 115                 nb_pkts_received < nb_total;) {
 116                 vector unsigned char desc[RTE_VIRTIO_DESC_PER_LOOP / 2];
 117                 vector unsigned char mbp[RTE_VIRTIO_DESC_PER_LOOP / 2];
 118                 vector unsigned char pkt_mb[RTE_VIRTIO_DESC_PER_LOOP];
 119
 120                 mbp[0] = vec_vsx_ld(0, (unsigned char const *)(sw_ring + 0));
 121                 desc[0] = vec_vsx_ld(0, (unsigned char const *)(rused + 0));
 122                 *(vector unsigned char *)&rx_pkts[0] = mbp[0];
 123
 124                 mbp[1] = vec_vsx_ld(0, (unsigned char const *)(sw_ring + 2));
 125                 desc[1] = vec_vsx_ld(0, (unsigned char const *)(rused + 2));
 126                 *(vector unsigned char *)&rx_pkts[2] = mbp[1];
 127
 128                 mbp[2] = vec_vsx_ld(0, (unsigned char const *)(sw_ring + 4));
 129                 desc[2] = vec_vsx_ld(0, (unsigned char const *)(rused + 4));
 130                 *(vector unsigned char *)&rx_pkts[4] = mbp[2];
 131
 132                 mbp[3] = vec_vsx_ld(0, (unsigned char const *)(sw_ring + 6));
 133                 desc[3] = vec_vsx_ld(0, (unsigned char const *)(rused + 6));
 134                 *(vector unsigned char *)&rx_pkts[6] = mbp[3];
 135
 136                 pkt_mb[0] = vec_perm(desc[0], zero, shuf_msk1);
 137                 pkt_mb[1] = vec_perm(desc[0], zero, shuf_msk2);
 138                 pkt_mb[0] = (vector unsigned char)
 139                         ((vector unsigned short)pkt_mb[0] + len_adjust);
 140                 pkt_mb[1] = (vector unsigned char)
 141                         ((vector unsigned short)pkt_mb[1] + len_adjust);
 142                 *(vector unsigned char *)&rx_pkts[0]->rx_descriptor_fields1 =
 143                         pkt_mb[0];
 144                 *(vector unsigned char *)&rx_pkts[1]->rx_descriptor_fields1 =
 145                         pkt_mb[1];
 146
 147                 pkt_mb[2] = vec_perm(desc[1], zero, shuf_msk1);
 148                 pkt_mb[3] = vec_perm(desc[1], zero, shuf_msk2);
 149                 pkt_mb[2] = (vector unsigned char)
 150                         ((vector unsigned short)pkt_mb[2] + len_adjust);
 151                 pkt_mb[3] = (vector unsigned char)
 152                         ((vector unsigned short)pkt_mb[3] + len_adjust);
 153                 *(vector unsigned char *)&rx_pkts[2]->rx_descriptor_fields1 =
 154                         pkt_mb[2];
 155                 *(vector unsigned char *)&rx_pkts[3]->rx_descriptor_fields1 =
 156                         pkt_mb[3];
 157
 158                 pkt_mb[4] = vec_perm(desc[2], zero, shuf_msk1);
 159                 pkt_mb[5] = vec_perm(desc[2], zero, shuf_msk2);
 160                 pkt_mb[4] = (vector unsigned char)
 161                         ((vector unsigned short)pkt_mb[4] + len_adjust);
 162                 pkt_mb[5] = (vector unsigned char)
 163                         ((vector unsigned short)pkt_mb[5] + len_adjust);
 164                 *(vector unsigned char *)&rx_pkts[4]->rx_descriptor_fields1 =
 165                         pkt_mb[4];
 166                 *(vector unsigned char *)&rx_pkts[5]->rx_descriptor_fields1 =
 167                         pkt_mb[5];
 168
 169                 pkt_mb[6] = vec_perm(desc[3], zero, shuf_msk1);
 170                 pkt_mb[7] = vec_perm(desc[3], zero, shuf_msk2);
 171                 pkt_mb[6] = (vector unsigned char)
 172                         ((vector unsigned short)pkt_mb[6] + len_adjust);
 173                 pkt_mb[7] = (vector unsigned char)
 174                         ((vector unsigned short)pkt_mb[7] + len_adjust);
 175                 *(vector unsigned char *)&rx_pkts[6]->rx_descriptor_fields1 =
 176                         pkt_mb[6];
 177                 *(vector unsigned char *)&rx_pkts[7]->rx_descriptor_fields1 =
 178                         pkt_mb[7];
 179
 180                 if (unlikely(nb_used <= RTE_VIRTIO_DESC_PER_LOOP)) {
 181                         if (sw_ring + nb_used <= sw_ring_end)
 182                                 nb_pkts_received += nb_used;
 183                         else
 184                                 nb_pkts_received += sw_ring_end - sw_ring;
 185                         break;
 186                 } else {
 187                         if (unlikely(sw_ring + RTE_VIRTIO_DESC_PER_LOOP >=
 188                                 sw_ring_end)) {
 189                                 nb_pkts_received += sw_ring_end - sw_ring;
 190                                 break;
 191                         } else {
 192                                 nb_pkts_received += RTE_VIRTIO_DESC_PER_LOOP;
 193
 194                                 rx_pkts += RTE_VIRTIO_DESC_PER_LOOP;
 195                                 sw_ring += RTE_VIRTIO_DESC_PER_LOOP;
 196                                 rused   += RTE_VIRTIO_DESC_PER_LOOP;
 197                                 nb_used -= RTE_VIRTIO_DESC_PER_LOOP;
 198                         }
 199                 }
 200         }
 201
 202         vq->vq_used_cons_idx += nb_pkts_received;
 203         vq->vq_free_cnt += nb_pkts_received;
 204         rxvq->stats.packets += nb_pkts_received;
 205         for (nb_used = 0; nb_used < nb_pkts_received; nb_used++)
 206                 virtio_update_packet_stats(&rxvq->stats, ref_rx_pkts[nb_used]);
 207
 208         return nb_pkts_received;
 209 }