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 <tmmintrin.h>
  41
  42 #include <rte_cycles.h>
  43 #include <rte_memory.h>
  44 #include <rte_memzone.h>
  45 #include <rte_branch_prediction.h>
  46 #include <rte_mempool.h>
  47 #include <rte_malloc.h>
  48 #include <rte_mbuf.h>
  49 #include <rte_ether.h>
  50 #include <rte_ethdev.h>
  51 #include <rte_prefetch.h>
  52 #include <rte_string_fns.h>
  53 #include <rte_errno.h>
  54 #include <rte_byteorder.h>
  55
  56 #include "virtio_logs.h"
  57 #include "virtio_ethdev.h"
  58 #include "virtqueue.h"
  59 #include "virtio_rxtx.h"
  60
  61 #define RTE_VIRTIO_VPMD_RX_BURST 32
  62 #define RTE_VIRTIO_DESC_PER_LOOP 8
  63 #define RTE_VIRTIO_VPMD_RX_REARM_THRESH RTE_VIRTIO_VPMD_RX_BURST
  64
  65 #ifndef __INTEL_COMPILER
  66 #pragma GCC diagnostic ignored "-Wcast-qual"
  67 #endif
  68
  69 int __attribute__((cold))
  70 virtqueue_enqueue_recv_refill_simple(struct virtqueue *vq,
  71         struct rte_mbuf *cookie)
  72 {
  73         struct vq_desc_extra *dxp;
  74         struct vring_desc *start_dp;
  75         uint16_t desc_idx;
  76
  77         desc_idx = vq->vq_avail_idx & (vq->vq_nentries - 1);
  78         dxp = &vq->vq_descx[desc_idx];
  79         dxp->cookie = (void *)cookie;
  80         vq->sw_ring[desc_idx] = cookie;
  81
  82         start_dp = vq->vq_ring.desc;
  83         start_dp[desc_idx].addr = MBUF_DATA_DMA_ADDR(cookie, vq->offset) -
  84                                   vq->hw->vtnet_hdr_size;
  85         start_dp[desc_idx].len = cookie->buf_len -
  86                 RTE_PKTMBUF_HEADROOM + vq->hw->vtnet_hdr_size;
  87
  88         vq->vq_free_cnt--;
  89         vq->vq_avail_idx++;
  90
  91         return 0;
  92 }
  93
  94 static inline void
  95 virtio_rxq_rearm_vec(struct virtnet_rx *rxvq)
  96 {
  97         int i;
  98         uint16_t desc_idx;
  99         struct rte_mbuf **sw_ring;
 100         struct vring_desc *start_dp;
 101         int ret;
 102         struct virtqueue *vq = rxvq->vq;
 103
 104         desc_idx = vq->vq_avail_idx & (vq->vq_nentries - 1);
 105         sw_ring = &vq->sw_ring[desc_idx];
 106         start_dp = &vq->vq_ring.desc[desc_idx];
 107
 108         ret = rte_mempool_get_bulk(rxvq->mpool, (void **)sw_ring,
 109                 RTE_VIRTIO_VPMD_RX_REARM_THRESH);
 110         if (unlikely(ret)) {
 111                 rte_eth_devices[rxvq->port_id].data->rx_mbuf_alloc_failed +=
 112                         RTE_VIRTIO_VPMD_RX_REARM_THRESH;
 113                 return;
 114         }
 115
 116         for (i = 0; i < RTE_VIRTIO_VPMD_RX_REARM_THRESH; i++) {
 117                 uintptr_t p;
 118
 119                 p = (uintptr_t)&sw_ring[i]->rearm_data;
 120                 *(uint64_t *)p = rxvq->mbuf_initializer;
 121
 122                 start_dp[i].addr =
 123                         MBUF_DATA_DMA_ADDR(sw_ring[i], vq->offset) -
 124                         vq->hw->vtnet_hdr_size;
 125                 start_dp[i].len = sw_ring[i]->buf_len -
 126                         RTE_PKTMBUF_HEADROOM + vq->hw->vtnet_hdr_size;
 127         }
 128
 129         vq->vq_avail_idx += RTE_VIRTIO_VPMD_RX_REARM_THRESH;
 130         vq->vq_free_cnt -= RTE_VIRTIO_VPMD_RX_REARM_THRESH;
 131         vq_update_avail_idx(vq);
 132 }
 133
 134 /* virtio vPMD receive routine, only accept(nb_pkts >= RTE_VIRTIO_DESC_PER_LOOP)
 135  *
 136  * This routine is for non-mergeable RX, one desc for each guest buffer.
 137  * This routine is based on the RX ring layout optimization. Each entry in the
 138  * avail ring points to the desc with the same index in the desc ring and this
 139  * will never be changed in the driver.
 140  *
 141  * - nb_pkts < RTE_VIRTIO_DESC_PER_LOOP, just return no packet
 142  */
 143 uint16_t
 144 virtio_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
 145         uint16_t nb_pkts)
 146 {
 147         struct virtnet_rx *rxvq = rx_queue;
 148         struct virtqueue *vq = rxvq->vq;
 149         uint16_t nb_used;
 150         uint16_t desc_idx;
 151         struct vring_used_elem *rused;
 152         struct rte_mbuf **sw_ring;
 153         struct rte_mbuf **sw_ring_end;
 154         uint16_t nb_pkts_received;
 155         __m128i shuf_msk1, shuf_msk2, len_adjust;
 156
 157         shuf_msk1 = _mm_set_epi8(
 158                 0xFF, 0xFF, 0xFF, 0xFF,
 159                 0xFF, 0xFF,             /* vlan tci */
 160                 5, 4,                   /* dat len */
 161                 0xFF, 0xFF, 5, 4,       /* pkt len */
 162                 0xFF, 0xFF, 0xFF, 0xFF  /* packet type */
 163
 164         );
 165
 166         shuf_msk2 = _mm_set_epi8(
 167                 0xFF, 0xFF, 0xFF, 0xFF,
 168                 0xFF, 0xFF,             /* vlan tci */
 169                 13, 12,                 /* dat len */
 170                 0xFF, 0xFF, 13, 12,     /* pkt len */
 171                 0xFF, 0xFF, 0xFF, 0xFF  /* packet type */
 172         );
 173
 174         /* Subtract the header length.
 175         *  In which case do we need the header length in used->len ?
 176         */
 177         len_adjust = _mm_set_epi16(
 178                 0, 0,
 179                 0,
 180                 (uint16_t)-vq->hw->vtnet_hdr_size,
 181                 0, (uint16_t)-vq->hw->vtnet_hdr_size,
 182                 0, 0);
 183
 184         if (unlikely(nb_pkts < RTE_VIRTIO_DESC_PER_LOOP))
 185                 return 0;
 186
 187         nb_used = *(volatile uint16_t *)&vq->vq_ring.used->idx -
 188                 vq->vq_used_cons_idx;
 189
 190         rte_compiler_barrier();
 191
 192         if (unlikely(nb_used == 0))
 193                 return 0;
 194
 195         nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, RTE_VIRTIO_DESC_PER_LOOP);
 196         nb_used = RTE_MIN(nb_used, nb_pkts);
 197
 198         desc_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
 199         rused = &vq->vq_ring.used->ring[desc_idx];
 200         sw_ring  = &vq->sw_ring[desc_idx];
 201         sw_ring_end = &vq->sw_ring[vq->vq_nentries];
 202
 203         _mm_prefetch((const void *)rused, _MM_HINT_T0);
 204
 205         if (vq->vq_free_cnt >= RTE_VIRTIO_VPMD_RX_REARM_THRESH) {
 206                 virtio_rxq_rearm_vec(rxvq);
 207                 if (unlikely(virtqueue_kick_prepare(vq)))
 208                         virtqueue_notify(vq);
 209         }
 210
 211         for (nb_pkts_received = 0;
 212                 nb_pkts_received < nb_used;) {
 213                 __m128i desc[RTE_VIRTIO_DESC_PER_LOOP / 2];
 214                 __m128i mbp[RTE_VIRTIO_DESC_PER_LOOP / 2];
 215                 __m128i pkt_mb[RTE_VIRTIO_DESC_PER_LOOP];
 216
 217                 mbp[0] = _mm_loadu_si128((__m128i *)(sw_ring + 0));
 218                 desc[0] = _mm_loadu_si128((__m128i *)(rused + 0));
 219                 _mm_storeu_si128((__m128i *)&rx_pkts[0], mbp[0]);
 220
 221                 mbp[1] = _mm_loadu_si128((__m128i *)(sw_ring + 2));
 222                 desc[1] = _mm_loadu_si128((__m128i *)(rused + 2));
 223                 _mm_storeu_si128((__m128i *)&rx_pkts[2], mbp[1]);
 224
 225                 mbp[2] = _mm_loadu_si128((__m128i *)(sw_ring + 4));
 226                 desc[2] = _mm_loadu_si128((__m128i *)(rused + 4));
 227                 _mm_storeu_si128((__m128i *)&rx_pkts[4], mbp[2]);
 228
 229                 mbp[3] = _mm_loadu_si128((__m128i *)(sw_ring + 6));
 230                 desc[3] = _mm_loadu_si128((__m128i *)(rused + 6));
 231                 _mm_storeu_si128((__m128i *)&rx_pkts[6], mbp[3]);
 232
 233                 pkt_mb[1] = _mm_shuffle_epi8(desc[0], shuf_msk2);
 234                 pkt_mb[0] = _mm_shuffle_epi8(desc[0], shuf_msk1);
 235                 pkt_mb[1] = _mm_add_epi16(pkt_mb[1], len_adjust);
 236                 pkt_mb[0] = _mm_add_epi16(pkt_mb[0], len_adjust);
 237                 _mm_storeu_si128((void *)&rx_pkts[1]->rx_descriptor_fields1,
 238                         pkt_mb[1]);
 239                 _mm_storeu_si128((void *)&rx_pkts[0]->rx_descriptor_fields1,
 240                         pkt_mb[0]);
 241
 242                 pkt_mb[3] = _mm_shuffle_epi8(desc[1], shuf_msk2);
 243                 pkt_mb[2] = _mm_shuffle_epi8(desc[1], shuf_msk1);
 244                 pkt_mb[3] = _mm_add_epi16(pkt_mb[3], len_adjust);
 245                 pkt_mb[2] = _mm_add_epi16(pkt_mb[2], len_adjust);
 246                 _mm_storeu_si128((void *)&rx_pkts[3]->rx_descriptor_fields1,
 247                         pkt_mb[3]);
 248                 _mm_storeu_si128((void *)&rx_pkts[2]->rx_descriptor_fields1,
 249                         pkt_mb[2]);
 250
 251                 pkt_mb[5] = _mm_shuffle_epi8(desc[2], shuf_msk2);
 252                 pkt_mb[4] = _mm_shuffle_epi8(desc[2], shuf_msk1);
 253                 pkt_mb[5] = _mm_add_epi16(pkt_mb[5], len_adjust);
 254                 pkt_mb[4] = _mm_add_epi16(pkt_mb[4], len_adjust);
 255                 _mm_storeu_si128((void *)&rx_pkts[5]->rx_descriptor_fields1,
 256                         pkt_mb[5]);
 257                 _mm_storeu_si128((void *)&rx_pkts[4]->rx_descriptor_fields1,
 258                         pkt_mb[4]);
 259
 260                 pkt_mb[7] = _mm_shuffle_epi8(desc[3], shuf_msk2);
 261                 pkt_mb[6] = _mm_shuffle_epi8(desc[3], shuf_msk1);
 262                 pkt_mb[7] = _mm_add_epi16(pkt_mb[7], len_adjust);
 263                 pkt_mb[6] = _mm_add_epi16(pkt_mb[6], len_adjust);
 264                 _mm_storeu_si128((void *)&rx_pkts[7]->rx_descriptor_fields1,
 265                         pkt_mb[7]);
 266                 _mm_storeu_si128((void *)&rx_pkts[6]->rx_descriptor_fields1,
 267                         pkt_mb[6]);
 268
 269                 if (unlikely(nb_used <= RTE_VIRTIO_DESC_PER_LOOP)) {
 270                         if (sw_ring + nb_used <= sw_ring_end)
 271                                 nb_pkts_received += nb_used;
 272                         else
 273                                 nb_pkts_received += sw_ring_end - sw_ring;
 274                         break;
 275                 } else {
 276                         if (unlikely(sw_ring + RTE_VIRTIO_DESC_PER_LOOP >=
 277                                 sw_ring_end)) {
 278                                 nb_pkts_received += sw_ring_end - sw_ring;
 279                                 break;
 280                         } else {
 281                                 nb_pkts_received += RTE_VIRTIO_DESC_PER_LOOP;
 282
 283                                 rx_pkts += RTE_VIRTIO_DESC_PER_LOOP;
 284                                 sw_ring += RTE_VIRTIO_DESC_PER_LOOP;
 285                                 rused   += RTE_VIRTIO_DESC_PER_LOOP;
 286                                 nb_used -= RTE_VIRTIO_DESC_PER_LOOP;
 287                         }
 288                 }
 289         }
 290
 291         vq->vq_used_cons_idx += nb_pkts_received;
 292         vq->vq_free_cnt += nb_pkts_received;
 293         rxvq->stats.packets += nb_pkts_received;
 294         return nb_pkts_received;
 295 }
 296
 297 #define VIRTIO_TX_FREE_THRESH 32
 298 #define VIRTIO_TX_MAX_FREE_BUF_SZ 32
 299 #define VIRTIO_TX_FREE_NR 32
 300 /* TODO: vq->tx_free_cnt could mean num of free slots so we could avoid shift */
 301 static inline void
 302 virtio_xmit_cleanup(struct virtqueue *vq)
 303 {
 304         uint16_t i, desc_idx;
 305         int nb_free = 0;
 306         struct rte_mbuf *m, *free[VIRTIO_TX_MAX_FREE_BUF_SZ];
 307
 308         desc_idx = (uint16_t)(vq->vq_used_cons_idx &
 309                    ((vq->vq_nentries >> 1) - 1));
 310         m = (struct rte_mbuf *)vq->vq_descx[desc_idx++].cookie;
 311         m = __rte_pktmbuf_prefree_seg(m);
 312         if (likely(m != NULL)) {
 313                 free[0] = m;
 314                 nb_free = 1;
 315                 for (i = 1; i < VIRTIO_TX_FREE_NR; i++) {
 316                         m = (struct rte_mbuf *)vq->vq_descx[desc_idx++].cookie;
 317                         m = __rte_pktmbuf_prefree_seg(m);
 318                         if (likely(m != NULL)) {
 319                                 if (likely(m->pool == free[0]->pool))
 320                                         free[nb_free++] = m;
 321                                 else {
 322                                         rte_mempool_put_bulk(free[0]->pool,
 323                                                 (void **)free, nb_free);
 324                                         free[0] = m;
 325                                         nb_free = 1;
 326                                 }
 327                         }
 328                 }
 329                 rte_mempool_put_bulk(free[0]->pool, (void **)free, nb_free);
 330         } else {
 331                 for (i = 1; i < VIRTIO_TX_FREE_NR; i++) {
 332                         m = (struct rte_mbuf *)vq->vq_descx[desc_idx++].cookie;
 333                         m = __rte_pktmbuf_prefree_seg(m);
 334                         if (m != NULL)
 335                                 rte_mempool_put(m->pool, m);
 336                 }
 337         }
 338
 339         vq->vq_used_cons_idx += VIRTIO_TX_FREE_NR;
 340         vq->vq_free_cnt += (VIRTIO_TX_FREE_NR << 1);
 341 }
 342
 343 uint16_t
 344 virtio_xmit_pkts_simple(void *tx_queue, struct rte_mbuf **tx_pkts,
 345         uint16_t nb_pkts)
 346 {
 347         struct virtnet_tx *txvq = tx_queue;
 348         struct virtqueue *vq = txvq->vq;
 349         uint16_t nb_used;
 350         uint16_t desc_idx;
 351         struct vring_desc *start_dp;
 352         uint16_t nb_tail, nb_commit;
 353         int i;
 354         uint16_t desc_idx_max = (vq->vq_nentries >> 1) - 1;
 355
 356         nb_used = VIRTQUEUE_NUSED(vq);
 357         rte_compiler_barrier();
 358
 359         if (nb_used >= VIRTIO_TX_FREE_THRESH)
 360                 virtio_xmit_cleanup(vq);
 361
 362         nb_commit = nb_pkts = RTE_MIN((vq->vq_free_cnt >> 1), nb_pkts);
 363         desc_idx = (uint16_t)(vq->vq_avail_idx & desc_idx_max);
 364         start_dp = vq->vq_ring.desc;
 365         nb_tail = (uint16_t) (desc_idx_max + 1 - desc_idx);
 366
 367         if (nb_commit >= nb_tail) {
 368                 for (i = 0; i < nb_tail; i++)
 369                         vq->vq_descx[desc_idx + i].cookie = tx_pkts[i];
 370                 for (i = 0; i < nb_tail; i++) {
 371                         start_dp[desc_idx].addr =
 372                                 MBUF_DATA_DMA_ADDR(*tx_pkts, vq->offset);
 373                         start_dp[desc_idx].len = (*tx_pkts)->pkt_len;
 374                         tx_pkts++;
 375                         desc_idx++;
 376                 }
 377                 nb_commit -= nb_tail;
 378                 desc_idx = 0;
 379         }
 380         for (i = 0; i < nb_commit; i++)
 381                 vq->vq_descx[desc_idx + i].cookie = tx_pkts[i];
 382         for (i = 0; i < nb_commit; i++) {
 383                 start_dp[desc_idx].addr =
 384                         MBUF_DATA_DMA_ADDR(*tx_pkts, vq->offset);
 385                 start_dp[desc_idx].len = (*tx_pkts)->pkt_len;
 386                 tx_pkts++;
 387                 desc_idx++;
 388         }
 389
 390         rte_compiler_barrier();
 391
 392         vq->vq_free_cnt -= (uint16_t)(nb_pkts << 1);
 393         vq->vq_avail_idx += nb_pkts;
 394         vq->vq_ring.avail->idx = vq->vq_avail_idx;
 395         txvq->stats.packets += nb_pkts;
 396
 397         if (likely(nb_pkts)) {
 398                 if (unlikely(virtqueue_kick_prepare(vq)))
 399                         virtqueue_notify(vq);
 400         }
 401
 402         return nb_pkts;
 403 }
 404
 405 int __attribute__((cold))
 406 virtio_rxq_vec_setup(struct virtnet_rx *rxq)
 407 {
 408         uintptr_t p;
 409         struct rte_mbuf mb_def = { .buf_addr = 0 }; /* zeroed mbuf */
 410
 411         mb_def.nb_segs = 1;
 412         mb_def.data_off = RTE_PKTMBUF_HEADROOM;
 413         mb_def.port = rxq->port_id;
 414         rte_mbuf_refcnt_set(&mb_def, 1);
 415
 416         /* prevent compiler reordering: rearm_data covers previous fields */
 417         rte_compiler_barrier();
 418         p = (uintptr_t)&mb_def.rearm_data;
 419         rxq->mbuf_initializer = *(uint64_t *)p;
 420
 421         return 0;
 422 }