drivers/net/virtio/virtio_rxtx_simple.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2015 Intel Corporation. All rights reserved.
   5  *   All rights reserved.
   6  *
   7  *   Redistribution and use in source and binary forms, with or without
   8  *   modification, are permitted provided that the following conditions
   9  *   are met:
  10  *
  11  *     * Redistributions of source code must retain the above copyright
  12  *       notice, this list of conditions and the following disclaimer.
  13  *     * Redistributions in binary form must reproduce the above copyright
  14  *       notice, this list of conditions and the following disclaimer in
  15  *       the documentation and/or other materials provided with the
  16  *       distribution.
  17  *     * Neither the name of Intel Corporation nor the names of its
  18  *       contributors may be used to endorse or promote products derived
  19  *       from this software without specific prior written permission.
  20  *
  21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  32  */
  33
  34 #include <stdint.h>
  35 #include <stdio.h>
  36 #include <stdlib.h>
  37 #include <string.h>
  38 #include <errno.h>
  39
  40 #include <rte_cycles.h>
  41 #include <rte_memory.h>
  42 #include <rte_memzone.h>
  43 #include <rte_branch_prediction.h>
  44 #include <rte_mempool.h>
  45 #include <rte_malloc.h>
  46 #include <rte_mbuf.h>
  47 #include <rte_ether.h>
  48 #include <rte_ethdev.h>
  49 #include <rte_prefetch.h>
  50 #include <rte_string_fns.h>
  51 #include <rte_errno.h>
  52 #include <rte_byteorder.h>
  53
  54 #include "virtio_logs.h"
  55 #include "virtio_ethdev.h"
  56 #include "virtqueue.h"
  57 #include "virtio_rxtx.h"
  58
  59 #define RTE_VIRTIO_VPMD_RX_BURST 32
  60 #define RTE_VIRTIO_DESC_PER_LOOP 8
  61 #define RTE_VIRTIO_VPMD_RX_REARM_THRESH RTE_VIRTIO_VPMD_RX_BURST
  62
  63 #ifndef __INTEL_COMPILER
  64 #pragma GCC diagnostic ignored "-Wcast-qual"
  65 #endif
  66
  67 int __attribute__((cold))
  68 virtqueue_enqueue_recv_refill_simple(struct virtqueue *vq,
  69         struct rte_mbuf *cookie)
  70 {
  71         struct vq_desc_extra *dxp;
  72         struct vring_desc *start_dp;
  73         uint16_t desc_idx;
  74
  75         desc_idx = vq->vq_avail_idx & (vq->vq_nentries - 1);
  76         dxp = &vq->vq_descx[desc_idx];
  77         dxp->cookie = (void *)cookie;
  78         vq->sw_ring[desc_idx] = cookie;
  79
  80         start_dp = vq->vq_ring.desc;
  81         start_dp[desc_idx].addr =
  82                 VIRTIO_MBUF_ADDR(cookie, vq) +
  83                 RTE_PKTMBUF_HEADROOM - vq->hw->vtnet_hdr_size;
  84         start_dp[desc_idx].len = cookie->buf_len -
  85                 RTE_PKTMBUF_HEADROOM + vq->hw->vtnet_hdr_size;
  86
  87         vq->vq_free_cnt--;
  88         vq->vq_avail_idx++;
  89
  90         return 0;
  91 }
  92
  93 static inline void
  94 virtio_rxq_rearm_vec(struct virtnet_rx *rxvq)
  95 {
  96         int i;
  97         uint16_t desc_idx;
  98         struct rte_mbuf **sw_ring;
  99         struct vring_desc *start_dp;
 100         int ret;
 101         struct virtqueue *vq = rxvq->vq;
 102
 103         desc_idx = vq->vq_avail_idx & (vq->vq_nentries - 1);
 104         sw_ring = &vq->sw_ring[desc_idx];
 105         start_dp = &vq->vq_ring.desc[desc_idx];
 106
 107         ret = rte_mempool_get_bulk(rxvq->mpool, (void **)sw_ring,
 108                 RTE_VIRTIO_VPMD_RX_REARM_THRESH);
 109         if (unlikely(ret)) {
 110                 rte_eth_devices[rxvq->port_id].data->rx_mbuf_alloc_failed +=
 111                         RTE_VIRTIO_VPMD_RX_REARM_THRESH;
 112                 return;
 113         }
 114
 115         for (i = 0; i < RTE_VIRTIO_VPMD_RX_REARM_THRESH; i++) {
 116                 uintptr_t p;
 117
 118                 p = (uintptr_t)&sw_ring[i]->rearm_data;
 119                 *(uint64_t *)p = rxvq->mbuf_initializer;
 120
 121                 start_dp[i].addr =
 122                         VIRTIO_MBUF_ADDR(sw_ring[i], vq) +
 123                         RTE_PKTMBUF_HEADROOM - vq->hw->vtnet_hdr_size;
 124                 start_dp[i].len = sw_ring[i]->buf_len -
 125                         RTE_PKTMBUF_HEADROOM + vq->hw->vtnet_hdr_size;
 126         }
 127
 128         vq->vq_avail_idx += RTE_VIRTIO_VPMD_RX_REARM_THRESH;
 129         vq->vq_free_cnt -= RTE_VIRTIO_VPMD_RX_REARM_THRESH;
 130         vq_update_avail_idx(vq);
 131 }
 132
 133 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
 134
 135 #include <tmmintrin.h>
 136
 137 /* virtio vPMD receive routine, only accept(nb_pkts >= RTE_VIRTIO_DESC_PER_LOOP)
 138  *
 139  * This routine is for non-mergeable RX, one desc for each guest buffer.
 140  * This routine is based on the RX ring layout optimization. Each entry in the
 141  * avail ring points to the desc with the same index in the desc ring and this
 142  * will never be changed in the driver.
 143  *
 144  * - nb_pkts < RTE_VIRTIO_DESC_PER_LOOP, just return no packet
 145  */
 146 uint16_t
 147 virtio_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
 148         uint16_t nb_pkts)
 149 {
 150         struct virtnet_rx *rxvq = rx_queue;
 151         struct virtqueue *vq = rxvq->vq;
 152         uint16_t nb_used;
 153         uint16_t desc_idx;
 154         struct vring_used_elem *rused;
 155         struct rte_mbuf **sw_ring;
 156         struct rte_mbuf **sw_ring_end;
 157         uint16_t nb_pkts_received;
 158         __m128i shuf_msk1, shuf_msk2, len_adjust;
 159
 160         shuf_msk1 = _mm_set_epi8(
 161                 0xFF, 0xFF, 0xFF, 0xFF,
 162                 0xFF, 0xFF,             /* vlan tci */
 163                 5, 4,                   /* dat len */
 164                 0xFF, 0xFF, 5, 4,       /* pkt len */
 165                 0xFF, 0xFF, 0xFF, 0xFF  /* packet type */
 166
 167         );
 168
 169         shuf_msk2 = _mm_set_epi8(
 170                 0xFF, 0xFF, 0xFF, 0xFF,
 171                 0xFF, 0xFF,             /* vlan tci */
 172                 13, 12,                 /* dat len */
 173                 0xFF, 0xFF, 13, 12,     /* pkt len */
 174                 0xFF, 0xFF, 0xFF, 0xFF  /* packet type */
 175         );
 176
 177         /* Subtract the header length.
 178         *  In which case do we need the header length in used->len ?
 179         */
 180         len_adjust = _mm_set_epi16(
 181                 0, 0,
 182                 0,
 183                 (uint16_t)-vq->hw->vtnet_hdr_size,
 184                 0, (uint16_t)-vq->hw->vtnet_hdr_size,
 185                 0, 0);
 186
 187         if (unlikely(nb_pkts < RTE_VIRTIO_DESC_PER_LOOP))
 188                 return 0;
 189
 190         nb_used = VIRTQUEUE_NUSED(vq);
 191
 192         rte_compiler_barrier();
 193
 194         if (unlikely(nb_used == 0))
 195                 return 0;
 196
 197         nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, RTE_VIRTIO_DESC_PER_LOOP);
 198         nb_used = RTE_MIN(nb_used, nb_pkts);
 199
 200         desc_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
 201         rused = &vq->vq_ring.used->ring[desc_idx];
 202         sw_ring  = &vq->sw_ring[desc_idx];
 203         sw_ring_end = &vq->sw_ring[vq->vq_nentries];
 204
 205         rte_prefetch0(rused);
 206
 207         if (vq->vq_free_cnt >= RTE_VIRTIO_VPMD_RX_REARM_THRESH) {
 208                 virtio_rxq_rearm_vec(rxvq);
 209                 if (unlikely(virtqueue_kick_prepare(vq)))
 210                         virtqueue_notify(vq);
 211         }
 212
 213         for (nb_pkts_received = 0;
 214                 nb_pkts_received < nb_used;) {
 215                 __m128i desc[RTE_VIRTIO_DESC_PER_LOOP / 2];
 216                 __m128i mbp[RTE_VIRTIO_DESC_PER_LOOP / 2];
 217                 __m128i pkt_mb[RTE_VIRTIO_DESC_PER_LOOP];
 218
 219                 mbp[0] = _mm_loadu_si128((__m128i *)(sw_ring + 0));
 220                 desc[0] = _mm_loadu_si128((__m128i *)(rused + 0));
 221                 _mm_storeu_si128((__m128i *)&rx_pkts[0], mbp[0]);
 222
 223                 mbp[1] = _mm_loadu_si128((__m128i *)(sw_ring + 2));
 224                 desc[1] = _mm_loadu_si128((__m128i *)(rused + 2));
 225                 _mm_storeu_si128((__m128i *)&rx_pkts[2], mbp[1]);
 226
 227                 mbp[2] = _mm_loadu_si128((__m128i *)(sw_ring + 4));
 228                 desc[2] = _mm_loadu_si128((__m128i *)(rused + 4));
 229                 _mm_storeu_si128((__m128i *)&rx_pkts[4], mbp[2]);
 230
 231                 mbp[3] = _mm_loadu_si128((__m128i *)(sw_ring + 6));
 232                 desc[3] = _mm_loadu_si128((__m128i *)(rused + 6));
 233                 _mm_storeu_si128((__m128i *)&rx_pkts[6], mbp[3]);
 234
 235                 pkt_mb[1] = _mm_shuffle_epi8(desc[0], shuf_msk2);
 236                 pkt_mb[0] = _mm_shuffle_epi8(desc[0], shuf_msk1);
 237                 pkt_mb[1] = _mm_add_epi16(pkt_mb[1], len_adjust);
 238                 pkt_mb[0] = _mm_add_epi16(pkt_mb[0], len_adjust);
 239                 _mm_storeu_si128((void *)&rx_pkts[1]->rx_descriptor_fields1,
 240                         pkt_mb[1]);
 241                 _mm_storeu_si128((void *)&rx_pkts[0]->rx_descriptor_fields1,
 242                         pkt_mb[0]);
 243
 244                 pkt_mb[3] = _mm_shuffle_epi8(desc[1], shuf_msk2);
 245                 pkt_mb[2] = _mm_shuffle_epi8(desc[1], shuf_msk1);
 246                 pkt_mb[3] = _mm_add_epi16(pkt_mb[3], len_adjust);
 247                 pkt_mb[2] = _mm_add_epi16(pkt_mb[2], len_adjust);
 248                 _mm_storeu_si128((void *)&rx_pkts[3]->rx_descriptor_fields1,
 249                         pkt_mb[3]);
 250                 _mm_storeu_si128((void *)&rx_pkts[2]->rx_descriptor_fields1,
 251                         pkt_mb[2]);
 252
 253                 pkt_mb[5] = _mm_shuffle_epi8(desc[2], shuf_msk2);
 254                 pkt_mb[4] = _mm_shuffle_epi8(desc[2], shuf_msk1);
 255                 pkt_mb[5] = _mm_add_epi16(pkt_mb[5], len_adjust);
 256                 pkt_mb[4] = _mm_add_epi16(pkt_mb[4], len_adjust);
 257                 _mm_storeu_si128((void *)&rx_pkts[5]->rx_descriptor_fields1,
 258                         pkt_mb[5]);
 259                 _mm_storeu_si128((void *)&rx_pkts[4]->rx_descriptor_fields1,
 260                         pkt_mb[4]);
 261
 262                 pkt_mb[7] = _mm_shuffle_epi8(desc[3], shuf_msk2);
 263                 pkt_mb[6] = _mm_shuffle_epi8(desc[3], shuf_msk1);
 264                 pkt_mb[7] = _mm_add_epi16(pkt_mb[7], len_adjust);
 265                 pkt_mb[6] = _mm_add_epi16(pkt_mb[6], len_adjust);
 266                 _mm_storeu_si128((void *)&rx_pkts[7]->rx_descriptor_fields1,
 267                         pkt_mb[7]);
 268                 _mm_storeu_si128((void *)&rx_pkts[6]->rx_descriptor_fields1,
 269                         pkt_mb[6]);
 270
 271                 if (unlikely(nb_used <= RTE_VIRTIO_DESC_PER_LOOP)) {
 272                         if (sw_ring + nb_used <= sw_ring_end)
 273                                 nb_pkts_received += nb_used;
 274                         else
 275                                 nb_pkts_received += sw_ring_end - sw_ring;
 276                         break;
 277                 } else {
 278                         if (unlikely(sw_ring + RTE_VIRTIO_DESC_PER_LOOP >=
 279                                 sw_ring_end)) {
 280                                 nb_pkts_received += sw_ring_end - sw_ring;
 281                                 break;
 282                         } else {
 283                                 nb_pkts_received += RTE_VIRTIO_DESC_PER_LOOP;
 284
 285                                 rx_pkts += RTE_VIRTIO_DESC_PER_LOOP;
 286                                 sw_ring += RTE_VIRTIO_DESC_PER_LOOP;
 287                                 rused   += RTE_VIRTIO_DESC_PER_LOOP;
 288                                 nb_used -= RTE_VIRTIO_DESC_PER_LOOP;
 289                         }
 290                 }
 291         }
 292
 293         vq->vq_used_cons_idx += nb_pkts_received;
 294         vq->vq_free_cnt += nb_pkts_received;
 295         rxvq->stats.packets += nb_pkts_received;
 296         return nb_pkts_received;
 297 }
 298
 299 #endif
 300
 301 #define VIRTIO_TX_FREE_THRESH 32
 302 #define VIRTIO_TX_MAX_FREE_BUF_SZ 32
 303 #define VIRTIO_TX_FREE_NR 32
 304 /* TODO: vq->tx_free_cnt could mean num of free slots so we could avoid shift */
 305 static inline void
 306 virtio_xmit_cleanup(struct virtqueue *vq)
 307 {
 308         uint16_t i, desc_idx;
 309         uint32_t nb_free = 0;
 310         struct rte_mbuf *m, *free[VIRTIO_TX_MAX_FREE_BUF_SZ];
 311
 312         desc_idx = (uint16_t)(vq->vq_used_cons_idx &
 313                    ((vq->vq_nentries >> 1) - 1));
 314         m = (struct rte_mbuf *)vq->vq_descx[desc_idx++].cookie;
 315         m = __rte_pktmbuf_prefree_seg(m);
 316         if (likely(m != NULL)) {
 317                 free[0] = m;
 318                 nb_free = 1;
 319                 for (i = 1; i < VIRTIO_TX_FREE_NR; i++) {
 320                         m = (struct rte_mbuf *)vq->vq_descx[desc_idx++].cookie;
 321                         m = __rte_pktmbuf_prefree_seg(m);
 322                         if (likely(m != NULL)) {
 323                                 if (likely(m->pool == free[0]->pool))
 324                                         free[nb_free++] = m;
 325                                 else {
 326                                         rte_mempool_put_bulk(free[0]->pool,
 327                                                 (void **)free,
 328                                                 RTE_MIN(RTE_DIM(free),
 329                                                         nb_free));
 330                                         free[0] = m;
 331                                         nb_free = 1;
 332                                 }
 333                         }
 334                 }
 335                 rte_mempool_put_bulk(free[0]->pool, (void **)free,
 336                         RTE_MIN(RTE_DIM(free), nb_free));
 337         } else {
 338                 for (i = 1; i < VIRTIO_TX_FREE_NR; i++) {
 339                         m = (struct rte_mbuf *)vq->vq_descx[desc_idx++].cookie;
 340                         m = __rte_pktmbuf_prefree_seg(m);
 341                         if (m != NULL)
 342                                 rte_mempool_put(m->pool, m);
 343                 }
 344         }
 345
 346         vq->vq_used_cons_idx += VIRTIO_TX_FREE_NR;
 347         vq->vq_free_cnt += (VIRTIO_TX_FREE_NR << 1);
 348 }
 349
 350 uint16_t
 351 virtio_xmit_pkts_simple(void *tx_queue, struct rte_mbuf **tx_pkts,
 352         uint16_t nb_pkts)
 353 {
 354         struct virtnet_tx *txvq = tx_queue;
 355         struct virtqueue *vq = txvq->vq;
 356         uint16_t nb_used;
 357         uint16_t desc_idx;
 358         struct vring_desc *start_dp;
 359         uint16_t nb_tail, nb_commit;
 360         int i;
 361         uint16_t desc_idx_max = (vq->vq_nentries >> 1) - 1;
 362
 363         nb_used = VIRTQUEUE_NUSED(vq);
 364         rte_compiler_barrier();
 365
 366         if (nb_used >= VIRTIO_TX_FREE_THRESH)
 367                 virtio_xmit_cleanup(vq);
 368
 369         nb_commit = nb_pkts = RTE_MIN((vq->vq_free_cnt >> 1), nb_pkts);
 370         desc_idx = (uint16_t)(vq->vq_avail_idx & desc_idx_max);
 371         start_dp = vq->vq_ring.desc;
 372         nb_tail = (uint16_t) (desc_idx_max + 1 - desc_idx);
 373
 374         if (nb_commit >= nb_tail) {
 375                 for (i = 0; i < nb_tail; i++)
 376                         vq->vq_descx[desc_idx + i].cookie = tx_pkts[i];
 377                 for (i = 0; i < nb_tail; i++) {
 378                         start_dp[desc_idx].addr =
 379                                 VIRTIO_MBUF_DATA_DMA_ADDR(*tx_pkts, vq);
 380                         start_dp[desc_idx].len = (*tx_pkts)->pkt_len;
 381                         tx_pkts++;
 382                         desc_idx++;
 383                 }
 384                 nb_commit -= nb_tail;
 385                 desc_idx = 0;
 386         }
 387         for (i = 0; i < nb_commit; i++)
 388                 vq->vq_descx[desc_idx + i].cookie = tx_pkts[i];
 389         for (i = 0; i < nb_commit; i++) {
 390                 start_dp[desc_idx].addr =
 391                         VIRTIO_MBUF_DATA_DMA_ADDR(*tx_pkts, vq);
 392                 start_dp[desc_idx].len = (*tx_pkts)->pkt_len;
 393                 tx_pkts++;
 394                 desc_idx++;
 395         }
 396
 397         rte_compiler_barrier();
 398
 399         vq->vq_free_cnt -= (uint16_t)(nb_pkts << 1);
 400         vq->vq_avail_idx += nb_pkts;
 401         vq->vq_ring.avail->idx = vq->vq_avail_idx;
 402         txvq->stats.packets += nb_pkts;
 403
 404         if (likely(nb_pkts)) {
 405                 if (unlikely(virtqueue_kick_prepare(vq)))
 406                         virtqueue_notify(vq);
 407         }
 408
 409         return nb_pkts;
 410 }
 411
 412 int __attribute__((cold))
 413 virtio_rxq_vec_setup(struct virtnet_rx *rxq)
 414 {
 415         uintptr_t p;
 416         struct rte_mbuf mb_def = { .buf_addr = 0 }; /* zeroed mbuf */
 417
 418         mb_def.nb_segs = 1;
 419         mb_def.data_off = RTE_PKTMBUF_HEADROOM;
 420         mb_def.port = rxq->port_id;
 421         rte_mbuf_refcnt_set(&mb_def, 1);
 422
 423         /* prevent compiler reordering: rearm_data covers previous fields */
 424         rte_compiler_barrier();
 425         p = (uintptr_t)&mb_def.rearm_data;
 426         rxq->mbuf_initializer = *(uint64_t *)p;
 427
 428         return 0;
 429 }