lib/librte_vhost/virtio_net.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2016 Intel Corporation
   3  */
   4
   5 #include <stdint.h>
   6 #include <stdbool.h>
   7 #include <linux/virtio_net.h>
   8
   9 #include <rte_mbuf.h>
  10 #include <rte_memcpy.h>
  11 #include <rte_ether.h>
  12 #include <rte_ip.h>
  13 #include <rte_vhost.h>
  14 #include <rte_tcp.h>
  15 #include <rte_udp.h>
  16 #include <rte_sctp.h>
  17 #include <rte_arp.h>
  18 #include <rte_spinlock.h>
  19 #include <rte_malloc.h>
  20
  21 #include "iotlb.h"
  22 #include "vhost.h"
  23
  24 #define MAX_PKT_BURST 32
  25
  26 #define MAX_BATCH_LEN 256
  27
  28 static  __rte_always_inline bool
  29 rxvq_is_mergeable(struct virtio_net *dev)
  30 {
  31         return dev->features & (1ULL << VIRTIO_NET_F_MRG_RXBUF);
  32 }
  33
  34 static bool
  35 is_valid_virt_queue_idx(uint32_t idx, int is_tx, uint32_t nr_vring)
  36 {
  37         return (is_tx ^ (idx & 1)) == 0 && idx < nr_vring;
  38 }
  39
  40 static __rte_always_inline void *
  41 alloc_copy_ind_table(struct virtio_net *dev, struct vhost_virtqueue *vq,
  42                 uint64_t desc_addr, uint64_t desc_len)
  43 {
  44         void *idesc;
  45         uint64_t src, dst;
  46         uint64_t len, remain = desc_len;
  47
  48         idesc = rte_malloc(__func__, desc_len, 0);
  49         if (unlikely(!idesc))
  50                 return 0;
  51
  52         dst = (uint64_t)(uintptr_t)idesc;
  53
  54         while (remain) {
  55                 len = remain;
  56                 src = vhost_iova_to_vva(dev, vq, desc_addr, &len,
  57                                 VHOST_ACCESS_RO);
  58                 if (unlikely(!src || !len)) {
  59                         rte_free(idesc);
  60                         return 0;
  61                 }
  62
  63                 rte_memcpy((void *)(uintptr_t)dst, (void *)(uintptr_t)src, len);
  64
  65                 remain -= len;
  66                 dst += len;
  67                 desc_addr += len;
  68         }
  69
  70         return idesc;
  71 }
  72
  73 static __rte_always_inline void
  74 free_ind_table(void *idesc)
  75 {
  76         rte_free(idesc);
  77 }
  78
  79 static __rte_always_inline void
  80 do_flush_shadow_used_ring_split(struct virtio_net *dev,
  81                         struct vhost_virtqueue *vq,
  82                         uint16_t to, uint16_t from, uint16_t size)
  83 {
  84         rte_memcpy(&vq->used->ring[to],
  85                         &vq->shadow_used_split[from],
  86                         size * sizeof(struct vring_used_elem));
  87         vhost_log_cache_used_vring(dev, vq,
  88                         offsetof(struct vring_used, ring[to]),
  89                         size * sizeof(struct vring_used_elem));
  90 }
  91
  92 static __rte_always_inline void
  93 flush_shadow_used_ring_split(struct virtio_net *dev, struct vhost_virtqueue *vq)
  94 {
  95         uint16_t used_idx = vq->last_used_idx & (vq->size - 1);
  96
  97         if (used_idx + vq->shadow_used_idx <= vq->size) {
  98                 do_flush_shadow_used_ring_split(dev, vq, used_idx, 0,
  99                                           vq->shadow_used_idx);
 100         } else {
 101                 uint16_t size;
 102
 103                 /* update used ring interval [used_idx, vq->size] */
 104                 size = vq->size - used_idx;
 105                 do_flush_shadow_used_ring_split(dev, vq, used_idx, 0, size);
 106
 107                 /* update the left half used ring interval [0, left_size] */
 108                 do_flush_shadow_used_ring_split(dev, vq, 0, size,
 109                                           vq->shadow_used_idx - size);
 110         }
 111         vq->last_used_idx += vq->shadow_used_idx;
 112
 113         rte_smp_wmb();
 114
 115         vhost_log_cache_sync(dev, vq);
 116
 117         *(volatile uint16_t *)&vq->used->idx += vq->shadow_used_idx;
 118         vq->shadow_used_idx = 0;
 119         vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
 120                 sizeof(vq->used->idx));
 121 }
 122
 123 static __rte_always_inline void
 124 update_shadow_used_ring_split(struct vhost_virtqueue *vq,
 125                          uint16_t desc_idx, uint16_t len)
 126 {
 127         uint16_t i = vq->shadow_used_idx++;
 128
 129         vq->shadow_used_split[i].id  = desc_idx;
 130         vq->shadow_used_split[i].len = len;
 131 }
 132
 133 static __rte_unused __rte_always_inline void
 134 flush_shadow_used_ring_packed(struct virtio_net *dev,
 135                         struct vhost_virtqueue *vq)
 136 {
 137         int i;
 138         uint16_t used_idx = vq->last_used_idx;
 139
 140         /* Split loop in two to save memory barriers */
 141         for (i = 0; i < vq->shadow_used_idx; i++) {
 142                 vq->desc_packed[used_idx].id = vq->shadow_used_packed[i].id;
 143                 vq->desc_packed[used_idx].len = vq->shadow_used_packed[i].len;
 144
 145                 used_idx += vq->shadow_used_packed[i].count;
 146                 if (used_idx >= vq->size)
 147                         used_idx -= vq->size;
 148         }
 149
 150         rte_smp_wmb();
 151
 152         for (i = 0; i < vq->shadow_used_idx; i++) {
 153                 uint16_t flags;
 154
 155                 if (vq->shadow_used_packed[i].len)
 156                         flags = VRING_DESC_F_WRITE;
 157                 else
 158                         flags = 0;
 159
 160                 if (vq->used_wrap_counter) {
 161                         flags |= VRING_DESC_F_USED;
 162                         flags |= VRING_DESC_F_AVAIL;
 163                 } else {
 164                         flags &= ~VRING_DESC_F_USED;
 165                         flags &= ~VRING_DESC_F_AVAIL;
 166                 }
 167
 168                 vq->desc_packed[vq->last_used_idx].flags = flags;
 169
 170                 vhost_log_cache_used_vring(dev, vq,
 171                                         vq->last_used_idx *
 172                                         sizeof(struct vring_packed_desc),
 173                                         sizeof(struct vring_packed_desc));
 174
 175                 vq->last_used_idx += vq->shadow_used_packed[i].count;
 176                 if (vq->last_used_idx >= vq->size) {
 177                         vq->used_wrap_counter ^= 1;
 178                         vq->last_used_idx -= vq->size;
 179                 }
 180         }
 181
 182         rte_smp_wmb();
 183         vq->shadow_used_idx = 0;
 184         vhost_log_cache_sync(dev, vq);
 185 }
 186
 187 static __rte_unused __rte_always_inline void
 188 update_shadow_used_ring_packed(struct vhost_virtqueue *vq,
 189                          uint16_t desc_idx, uint16_t len, uint16_t count)
 190 {
 191         uint16_t i = vq->shadow_used_idx++;
 192
 193         vq->shadow_used_packed[i].id  = desc_idx;
 194         vq->shadow_used_packed[i].len = len;
 195         vq->shadow_used_packed[i].count = count;
 196 }
 197
 198 static inline void
 199 do_data_copy_enqueue(struct virtio_net *dev, struct vhost_virtqueue *vq)
 200 {
 201         struct batch_copy_elem *elem = vq->batch_copy_elems;
 202         uint16_t count = vq->batch_copy_nb_elems;
 203         int i;
 204
 205         for (i = 0; i < count; i++) {
 206                 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
 207                 vhost_log_cache_write(dev, vq, elem[i].log_addr, elem[i].len);
 208                 PRINT_PACKET(dev, (uintptr_t)elem[i].dst, elem[i].len, 0);
 209         }
 210
 211         vq->batch_copy_nb_elems = 0;
 212 }
 213
 214 static inline void
 215 do_data_copy_dequeue(struct vhost_virtqueue *vq)
 216 {
 217         struct batch_copy_elem *elem = vq->batch_copy_elems;
 218         uint16_t count = vq->batch_copy_nb_elems;
 219         int i;
 220
 221         for (i = 0; i < count; i++)
 222                 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
 223
 224         vq->batch_copy_nb_elems = 0;
 225 }
 226
 227 /* avoid write operation when necessary, to lessen cache issues */
 228 #define ASSIGN_UNLESS_EQUAL(var, val) do {      \
 229         if ((var) != (val))                     \
 230                 (var) = (val);                  \
 231 } while (0)
 232
 233 static __rte_always_inline void
 234 virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
 235 {
 236         uint64_t csum_l4 = m_buf->ol_flags & PKT_TX_L4_MASK;
 237
 238         if (m_buf->ol_flags & PKT_TX_TCP_SEG)
 239                 csum_l4 |= PKT_TX_TCP_CKSUM;
 240
 241         if (csum_l4) {
 242                 net_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 243                 net_hdr->csum_start = m_buf->l2_len + m_buf->l3_len;
 244
 245                 switch (csum_l4) {
 246                 case PKT_TX_TCP_CKSUM:
 247                         net_hdr->csum_offset = (offsetof(struct tcp_hdr,
 248                                                 cksum));
 249                         break;
 250                 case PKT_TX_UDP_CKSUM:
 251                         net_hdr->csum_offset = (offsetof(struct udp_hdr,
 252                                                 dgram_cksum));
 253                         break;
 254                 case PKT_TX_SCTP_CKSUM:
 255                         net_hdr->csum_offset = (offsetof(struct sctp_hdr,
 256                                                 cksum));
 257                         break;
 258                 }
 259         } else {
 260                 ASSIGN_UNLESS_EQUAL(net_hdr->csum_start, 0);
 261                 ASSIGN_UNLESS_EQUAL(net_hdr->csum_offset, 0);
 262                 ASSIGN_UNLESS_EQUAL(net_hdr->flags, 0);
 263         }
 264
 265         /* IP cksum verification cannot be bypassed, then calculate here */
 266         if (m_buf->ol_flags & PKT_TX_IP_CKSUM) {
 267                 struct ipv4_hdr *ipv4_hdr;
 268
 269                 ipv4_hdr = rte_pktmbuf_mtod_offset(m_buf, struct ipv4_hdr *,
 270                                                    m_buf->l2_len);
 271                 ipv4_hdr->hdr_checksum = rte_ipv4_cksum(ipv4_hdr);
 272         }
 273
 274         if (m_buf->ol_flags & PKT_TX_TCP_SEG) {
 275                 if (m_buf->ol_flags & PKT_TX_IPV4)
 276                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 277                 else
 278                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 279                 net_hdr->gso_size = m_buf->tso_segsz;
 280                 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len
 281                                         + m_buf->l4_len;
 282         } else if (m_buf->ol_flags & PKT_TX_UDP_SEG) {
 283                 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
 284                 net_hdr->gso_size = m_buf->tso_segsz;
 285                 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len +
 286                         m_buf->l4_len;
 287         } else {
 288                 ASSIGN_UNLESS_EQUAL(net_hdr->gso_type, 0);
 289                 ASSIGN_UNLESS_EQUAL(net_hdr->gso_size, 0);
 290                 ASSIGN_UNLESS_EQUAL(net_hdr->hdr_len, 0);
 291         }
 292 }
 293
 294 static __rte_always_inline int
 295 map_one_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 296                 struct buf_vector *buf_vec, uint16_t *vec_idx,
 297                 uint64_t desc_iova, uint64_t desc_len, uint8_t perm)
 298 {
 299         uint16_t vec_id = *vec_idx;
 300
 301         while (desc_len) {
 302                 uint64_t desc_addr;
 303                 uint64_t desc_chunck_len = desc_len;
 304
 305                 if (unlikely(vec_id >= BUF_VECTOR_MAX))
 306                         return -1;
 307
 308                 desc_addr = vhost_iova_to_vva(dev, vq,
 309                                 desc_iova,
 310                                 &desc_chunck_len,
 311                                 perm);
 312                 if (unlikely(!desc_addr))
 313                         return -1;
 314
 315                 buf_vec[vec_id].buf_iova = desc_iova;
 316                 buf_vec[vec_id].buf_addr = desc_addr;
 317                 buf_vec[vec_id].buf_len  = desc_chunck_len;
 318
 319                 desc_len -= desc_chunck_len;
 320                 desc_iova += desc_chunck_len;
 321                 vec_id++;
 322         }
 323         *vec_idx = vec_id;
 324
 325         return 0;
 326 }
 327
 328 static __rte_always_inline int
 329 fill_vec_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 330                          uint32_t avail_idx, uint16_t *vec_idx,
 331                          struct buf_vector *buf_vec, uint16_t *desc_chain_head,
 332                          uint16_t *desc_chain_len, uint8_t perm)
 333 {
 334         uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
 335         uint16_t vec_id = *vec_idx;
 336         uint32_t len    = 0;
 337         uint64_t dlen;
 338         struct vring_desc *descs = vq->desc;
 339         struct vring_desc *idesc = NULL;
 340
 341         *desc_chain_head = idx;
 342
 343         if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
 344                 dlen = vq->desc[idx].len;
 345                 descs = (struct vring_desc *)(uintptr_t)
 346                         vhost_iova_to_vva(dev, vq, vq->desc[idx].addr,
 347                                                 &dlen,
 348                                                 VHOST_ACCESS_RO);
 349                 if (unlikely(!descs))
 350                         return -1;
 351
 352                 if (unlikely(dlen < vq->desc[idx].len)) {
 353                         /*
 354                          * The indirect desc table is not contiguous
 355                          * in process VA space, we have to copy it.
 356                          */
 357                         idesc = alloc_copy_ind_table(dev, vq,
 358                                         vq->desc[idx].addr, vq->desc[idx].len);
 359                         if (unlikely(!idesc))
 360                                 return -1;
 361
 362                         descs = idesc;
 363                 }
 364
 365                 idx = 0;
 366         }
 367
 368         while (1) {
 369                 if (unlikely(idx >= vq->size)) {
 370                         free_ind_table(idesc);
 371                         return -1;
 372                 }
 373
 374                 len += descs[idx].len;
 375
 376                 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 377                                                 descs[idx].addr, descs[idx].len,
 378                                                 perm))) {
 379                         free_ind_table(idesc);
 380                         return -1;
 381                 }
 382
 383                 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
 384                         break;
 385
 386                 idx = descs[idx].next;
 387         }
 388
 389         *desc_chain_len = len;
 390         *vec_idx = vec_id;
 391
 392         if (unlikely(!!idesc))
 393                 free_ind_table(idesc);
 394
 395         return 0;
 396 }
 397
 398 /*
 399  * Returns -1 on fail, 0 on success
 400  */
 401 static inline int
 402 reserve_avail_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 403                                 uint32_t size, struct buf_vector *buf_vec,
 404                                 uint16_t *num_buffers, uint16_t avail_head,
 405                                 uint16_t *nr_vec)
 406 {
 407         uint16_t cur_idx;
 408         uint16_t vec_idx = 0;
 409         uint16_t max_tries, tries = 0;
 410
 411         uint16_t head_idx = 0;
 412         uint16_t len = 0;
 413
 414         *num_buffers = 0;
 415         cur_idx  = vq->last_avail_idx;
 416
 417         if (rxvq_is_mergeable(dev))
 418                 max_tries = vq->size;
 419         else
 420                 max_tries = 1;
 421
 422         while (size > 0) {
 423                 if (unlikely(cur_idx == avail_head))
 424                         return -1;
 425
 426                 if (unlikely(fill_vec_buf_split(dev, vq, cur_idx,
 427                                                 &vec_idx, buf_vec,
 428                                                 &head_idx, &len,
 429                                                 VHOST_ACCESS_RW) < 0))
 430                         return -1;
 431                 len = RTE_MIN(len, size);
 432                 update_shadow_used_ring_split(vq, head_idx, len);
 433                 size -= len;
 434
 435                 cur_idx++;
 436                 tries++;
 437                 *num_buffers += 1;
 438
 439                 /*
 440                  * if we tried all available ring items, and still
 441                  * can't get enough buf, it means something abnormal
 442                  * happened.
 443                  */
 444                 if (unlikely(tries > max_tries))
 445                         return -1;
 446         }
 447
 448         *nr_vec = vec_idx;
 449
 450         return 0;
 451 }
 452
 453 static __rte_always_inline int
 454 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 455                             struct rte_mbuf *m, struct buf_vector *buf_vec,
 456                             uint16_t nr_vec, uint16_t num_buffers)
 457 {
 458         uint32_t vec_idx = 0;
 459         uint32_t mbuf_offset, mbuf_avail;
 460         uint32_t buf_offset, buf_avail;
 461         uint64_t buf_addr, buf_iova, buf_len;
 462         uint32_t cpy_len;
 463         uint64_t hdr_addr;
 464         struct rte_mbuf *hdr_mbuf;
 465         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 466         struct virtio_net_hdr_mrg_rxbuf tmp_hdr, *hdr = NULL;
 467         int error = 0;
 468
 469         if (unlikely(m == NULL)) {
 470                 error = -1;
 471                 goto out;
 472         }
 473
 474         buf_addr = buf_vec[vec_idx].buf_addr;
 475         buf_iova = buf_vec[vec_idx].buf_iova;
 476         buf_len = buf_vec[vec_idx].buf_len;
 477
 478         if (nr_vec > 1)
 479                 rte_prefetch0((void *)(uintptr_t)buf_vec[1].buf_addr);
 480
 481         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
 482                 error = -1;
 483                 goto out;
 484         }
 485
 486         hdr_mbuf = m;
 487         hdr_addr = buf_addr;
 488         if (unlikely(buf_len < dev->vhost_hlen))
 489                 hdr = &tmp_hdr;
 490         else
 491                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)hdr_addr;
 492
 493         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
 494                 dev->vid, num_buffers);
 495
 496         if (unlikely(buf_len < dev->vhost_hlen)) {
 497                 buf_offset = dev->vhost_hlen - buf_len;
 498                 vec_idx++;
 499                 buf_addr = buf_vec[vec_idx].buf_addr;
 500                 buf_iova = buf_vec[vec_idx].buf_iova;
 501                 buf_len = buf_vec[vec_idx].buf_len;
 502                 buf_avail = buf_len - buf_offset;
 503         } else {
 504                 buf_offset = dev->vhost_hlen;
 505                 buf_avail = buf_len - dev->vhost_hlen;
 506         }
 507
 508         mbuf_avail  = rte_pktmbuf_data_len(m);
 509         mbuf_offset = 0;
 510         while (mbuf_avail != 0 || m->next != NULL) {
 511                 /* done with current buf, get the next one */
 512                 if (buf_avail == 0) {
 513                         vec_idx++;
 514                         if (unlikely(vec_idx >= nr_vec)) {
 515                                 error = -1;
 516                                 goto out;
 517                         }
 518
 519                         buf_addr = buf_vec[vec_idx].buf_addr;
 520                         buf_iova = buf_vec[vec_idx].buf_iova;
 521                         buf_len = buf_vec[vec_idx].buf_len;
 522
 523                         /* Prefetch next buffer address. */
 524                         if (vec_idx + 1 < nr_vec)
 525                                 rte_prefetch0((void *)(uintptr_t)
 526                                                 buf_vec[vec_idx + 1].buf_addr);
 527                         buf_offset = 0;
 528                         buf_avail  = buf_len;
 529                 }
 530
 531                 /* done with current mbuf, get the next one */
 532                 if (mbuf_avail == 0) {
 533                         m = m->next;
 534
 535                         mbuf_offset = 0;
 536                         mbuf_avail  = rte_pktmbuf_data_len(m);
 537                 }
 538
 539                 if (hdr_addr) {
 540                         virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
 541                         if (rxvq_is_mergeable(dev))
 542                                 ASSIGN_UNLESS_EQUAL(hdr->num_buffers,
 543                                                 num_buffers);
 544
 545                         if (unlikely(hdr == &tmp_hdr)) {
 546                                 uint64_t len;
 547                                 uint64_t remain = dev->vhost_hlen;
 548                                 uint64_t src = (uint64_t)(uintptr_t)hdr, dst;
 549                                 uint64_t iova = buf_vec[0].buf_iova;
 550                                 uint16_t hdr_vec_idx = 0;
 551
 552                                 while (remain) {
 553                                         len = remain;
 554                                         dst = buf_vec[hdr_vec_idx].buf_addr;
 555                                         rte_memcpy((void *)(uintptr_t)dst,
 556                                                         (void *)(uintptr_t)src,
 557                                                         len);
 558
 559                                         PRINT_PACKET(dev, (uintptr_t)dst,
 560                                                         (uint32_t)len, 0);
 561                                         vhost_log_cache_write(dev, vq,
 562                                                         iova, len);
 563
 564                                         remain -= len;
 565                                         iova += len;
 566                                         src += len;
 567                                         hdr_vec_idx++;
 568                                 }
 569                         } else {
 570                                 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
 571                                                 dev->vhost_hlen, 0);
 572                                 vhost_log_cache_write(dev, vq,
 573                                                 buf_vec[0].buf_iova,
 574                                                 dev->vhost_hlen);
 575                         }
 576
 577                         hdr_addr = 0;
 578                 }
 579
 580                 cpy_len = RTE_MIN(buf_len, mbuf_avail);
 581
 582                 if (likely(cpy_len > MAX_BATCH_LEN ||
 583                                         vq->batch_copy_nb_elems >= vq->size)) {
 584                         rte_memcpy((void *)((uintptr_t)(buf_addr + buf_offset)),
 585                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 586                                 cpy_len);
 587                         vhost_log_cache_write(dev, vq, buf_iova + buf_offset,
 588                                         cpy_len);
 589                         PRINT_PACKET(dev, (uintptr_t)(buf_addr + buf_offset),
 590                                 cpy_len, 0);
 591                 } else {
 592                         batch_copy[vq->batch_copy_nb_elems].dst =
 593                                 (void *)((uintptr_t)(buf_addr + buf_offset));
 594                         batch_copy[vq->batch_copy_nb_elems].src =
 595                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
 596                         batch_copy[vq->batch_copy_nb_elems].log_addr =
 597                                 buf_iova + buf_offset;
 598                         batch_copy[vq->batch_copy_nb_elems].len = cpy_len;
 599                         vq->batch_copy_nb_elems++;
 600                 }
 601
 602                 mbuf_avail  -= cpy_len;
 603                 mbuf_offset += cpy_len;
 604                 buf_avail  -= cpy_len;
 605                 buf_offset += cpy_len;
 606         }
 607
 608 out:
 609
 610         return error;
 611 }
 612
 613 static __rte_always_inline uint32_t
 614 virtio_dev_rx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 615         struct rte_mbuf **pkts, uint32_t count)
 616 {
 617         uint32_t pkt_idx = 0;
 618         uint16_t num_buffers;
 619         struct buf_vector buf_vec[BUF_VECTOR_MAX];
 620         uint16_t avail_head;
 621
 622         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
 623         avail_head = *((volatile uint16_t *)&vq->avail->idx);
 624
 625         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
 626                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
 627                 uint16_t nr_vec = 0;
 628
 629                 if (unlikely(reserve_avail_buf_split(dev, vq,
 630                                                 pkt_len, buf_vec, &num_buffers,
 631                                                 avail_head, &nr_vec) < 0)) {
 632                         VHOST_LOG_DEBUG(VHOST_DATA,
 633                                 "(%d) failed to get enough desc from vring\n",
 634                                 dev->vid);
 635                         vq->shadow_used_idx -= num_buffers;
 636                         break;
 637                 }
 638
 639                 rte_prefetch0((void *)(uintptr_t)buf_vec[0].buf_addr);
 640
 641                 VHOST_LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
 642                         dev->vid, vq->last_avail_idx,
 643                         vq->last_avail_idx + num_buffers);
 644
 645                 if (copy_mbuf_to_desc(dev, vq, pkts[pkt_idx],
 646                                                 buf_vec, nr_vec,
 647                                                 num_buffers) < 0) {
 648                         vq->shadow_used_idx -= num_buffers;
 649                         break;
 650                 }
 651
 652                 vq->last_avail_idx += num_buffers;
 653         }
 654
 655         do_data_copy_enqueue(dev, vq);
 656
 657         if (likely(vq->shadow_used_idx)) {
 658                 flush_shadow_used_ring_split(dev, vq);
 659                 vhost_vring_call(dev, vq);
 660         }
 661
 662         return pkt_idx;
 663 }
 664
 665 static __rte_always_inline uint32_t
 666 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
 667         struct rte_mbuf **pkts, uint32_t count)
 668 {
 669         struct vhost_virtqueue *vq;
 670
 671         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
 672         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
 673                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
 674                         dev->vid, __func__, queue_id);
 675                 return 0;
 676         }
 677
 678         vq = dev->virtqueue[queue_id];
 679
 680         rte_spinlock_lock(&vq->access_lock);
 681
 682         if (unlikely(vq->enabled == 0))
 683                 goto out_access_unlock;
 684
 685         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 686                 vhost_user_iotlb_rd_lock(vq);
 687
 688         if (unlikely(vq->access_ok == 0))
 689                 if (unlikely(vring_translate(dev, vq) < 0))
 690                         goto out;
 691
 692         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
 693         if (count == 0)
 694                 goto out;
 695
 696         count = virtio_dev_rx_split(dev, vq, pkts, count);
 697
 698 out:
 699         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 700                 vhost_user_iotlb_rd_unlock(vq);
 701
 702 out_access_unlock:
 703         rte_spinlock_unlock(&vq->access_lock);
 704
 705         return count;
 706 }
 707
 708 uint16_t
 709 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
 710         struct rte_mbuf **pkts, uint16_t count)
 711 {
 712         struct virtio_net *dev = get_device(vid);
 713
 714         if (!dev)
 715                 return 0;
 716
 717         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
 718                 RTE_LOG(ERR, VHOST_DATA,
 719                         "(%d) %s: built-in vhost net backend is disabled.\n",
 720                         dev->vid, __func__);
 721                 return 0;
 722         }
 723
 724         return virtio_dev_rx(dev, queue_id, pkts, count);
 725 }
 726
 727 static inline bool
 728 virtio_net_with_host_offload(struct virtio_net *dev)
 729 {
 730         if (dev->features &
 731                         ((1ULL << VIRTIO_NET_F_CSUM) |
 732                          (1ULL << VIRTIO_NET_F_HOST_ECN) |
 733                          (1ULL << VIRTIO_NET_F_HOST_TSO4) |
 734                          (1ULL << VIRTIO_NET_F_HOST_TSO6) |
 735                          (1ULL << VIRTIO_NET_F_HOST_UFO)))
 736                 return true;
 737
 738         return false;
 739 }
 740
 741 static void
 742 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
 743 {
 744         struct ipv4_hdr *ipv4_hdr;
 745         struct ipv6_hdr *ipv6_hdr;
 746         void *l3_hdr = NULL;
 747         struct ether_hdr *eth_hdr;
 748         uint16_t ethertype;
 749
 750         eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
 751
 752         m->l2_len = sizeof(struct ether_hdr);
 753         ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
 754
 755         if (ethertype == ETHER_TYPE_VLAN) {
 756                 struct vlan_hdr *vlan_hdr = (struct vlan_hdr *)(eth_hdr + 1);
 757
 758                 m->l2_len += sizeof(struct vlan_hdr);
 759                 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
 760         }
 761
 762         l3_hdr = (char *)eth_hdr + m->l2_len;
 763
 764         switch (ethertype) {
 765         case ETHER_TYPE_IPv4:
 766                 ipv4_hdr = l3_hdr;
 767                 *l4_proto = ipv4_hdr->next_proto_id;
 768                 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
 769                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 770                 m->ol_flags |= PKT_TX_IPV4;
 771                 break;
 772         case ETHER_TYPE_IPv6:
 773                 ipv6_hdr = l3_hdr;
 774                 *l4_proto = ipv6_hdr->proto;
 775                 m->l3_len = sizeof(struct ipv6_hdr);
 776                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 777                 m->ol_flags |= PKT_TX_IPV6;
 778                 break;
 779         default:
 780                 m->l3_len = 0;
 781                 *l4_proto = 0;
 782                 *l4_hdr = NULL;
 783                 break;
 784         }
 785 }
 786
 787 static __rte_always_inline void
 788 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
 789 {
 790         uint16_t l4_proto = 0;
 791         void *l4_hdr = NULL;
 792         struct tcp_hdr *tcp_hdr = NULL;
 793
 794         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
 795                 return;
 796
 797         parse_ethernet(m, &l4_proto, &l4_hdr);
 798         if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 799                 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
 800                         switch (hdr->csum_offset) {
 801                         case (offsetof(struct tcp_hdr, cksum)):
 802                                 if (l4_proto == IPPROTO_TCP)
 803                                         m->ol_flags |= PKT_TX_TCP_CKSUM;
 804                                 break;
 805                         case (offsetof(struct udp_hdr, dgram_cksum)):
 806                                 if (l4_proto == IPPROTO_UDP)
 807                                         m->ol_flags |= PKT_TX_UDP_CKSUM;
 808                                 break;
 809                         case (offsetof(struct sctp_hdr, cksum)):
 810                                 if (l4_proto == IPPROTO_SCTP)
 811                                         m->ol_flags |= PKT_TX_SCTP_CKSUM;
 812                                 break;
 813                         default:
 814                                 break;
 815                         }
 816                 }
 817         }
 818
 819         if (l4_hdr && hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 820                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 821                 case VIRTIO_NET_HDR_GSO_TCPV4:
 822                 case VIRTIO_NET_HDR_GSO_TCPV6:
 823                         tcp_hdr = l4_hdr;
 824                         m->ol_flags |= PKT_TX_TCP_SEG;
 825                         m->tso_segsz = hdr->gso_size;
 826                         m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
 827                         break;
 828                 case VIRTIO_NET_HDR_GSO_UDP:
 829                         m->ol_flags |= PKT_TX_UDP_SEG;
 830                         m->tso_segsz = hdr->gso_size;
 831                         m->l4_len = sizeof(struct udp_hdr);
 832                         break;
 833                 default:
 834                         RTE_LOG(WARNING, VHOST_DATA,
 835                                 "unsupported gso type %u.\n", hdr->gso_type);
 836                         break;
 837                 }
 838         }
 839 }
 840
 841 static __rte_always_inline void
 842 put_zmbuf(struct zcopy_mbuf *zmbuf)
 843 {
 844         zmbuf->in_use = 0;
 845 }
 846
 847 static __rte_always_inline int
 848 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 849                   struct buf_vector *buf_vec, uint16_t nr_vec,
 850                   struct rte_mbuf *m, struct rte_mempool *mbuf_pool)
 851 {
 852         uint32_t buf_avail, buf_offset;
 853         uint64_t buf_addr, buf_iova, buf_len;
 854         uint32_t mbuf_avail, mbuf_offset;
 855         uint32_t cpy_len;
 856         struct rte_mbuf *cur = m, *prev = m;
 857         struct virtio_net_hdr tmp_hdr;
 858         struct virtio_net_hdr *hdr = NULL;
 859         /* A counter to avoid desc dead loop chain */
 860         uint16_t vec_idx = 0;
 861         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 862         int error = 0;
 863
 864         buf_addr = buf_vec[vec_idx].buf_addr;
 865         buf_iova = buf_vec[vec_idx].buf_iova;
 866         buf_len = buf_vec[vec_idx].buf_len;
 867
 868         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
 869                 error = -1;
 870                 goto out;
 871         }
 872
 873         if (likely(nr_vec > 1))
 874                 rte_prefetch0((void *)(uintptr_t)buf_vec[1].buf_addr);
 875
 876         if (virtio_net_with_host_offload(dev)) {
 877                 if (unlikely(buf_len < sizeof(struct virtio_net_hdr))) {
 878                         uint64_t len;
 879                         uint64_t remain = sizeof(struct virtio_net_hdr);
 880                         uint64_t src;
 881                         uint64_t dst = (uint64_t)(uintptr_t)&tmp_hdr;
 882                         uint16_t hdr_vec_idx = 0;
 883
 884                         /*
 885                          * No luck, the virtio-net header doesn't fit
 886                          * in a contiguous virtual area.
 887                          */
 888                         while (remain) {
 889                                 len = remain;
 890                                 src = buf_vec[hdr_vec_idx].buf_addr;
 891                                 rte_memcpy((void *)(uintptr_t)dst,
 892                                                    (void *)(uintptr_t)src, len);
 893
 894                                 remain -= len;
 895                                 dst += len;
 896                                 hdr_vec_idx++;
 897                         }
 898
 899                         hdr = &tmp_hdr;
 900                 } else {
 901                         hdr = (struct virtio_net_hdr *)((uintptr_t)buf_addr);
 902                         rte_prefetch0(hdr);
 903                 }
 904         }
 905
 906         /*
 907          * A virtio driver normally uses at least 2 desc buffers
 908          * for Tx: the first for storing the header, and others
 909          * for storing the data.
 910          */
 911         if (unlikely(buf_len < dev->vhost_hlen)) {
 912                 buf_offset = dev->vhost_hlen - buf_len;
 913                 vec_idx++;
 914                 buf_addr = buf_vec[vec_idx].buf_addr;
 915                 buf_iova = buf_vec[vec_idx].buf_iova;
 916                 buf_len = buf_vec[vec_idx].buf_len;
 917                 buf_avail  = buf_len - buf_offset;
 918         } else if (buf_len == dev->vhost_hlen) {
 919                 if (unlikely(++vec_idx >= nr_vec))
 920                         goto out;
 921                 buf_addr = buf_vec[vec_idx].buf_addr;
 922                 buf_iova = buf_vec[vec_idx].buf_iova;
 923                 buf_len = buf_vec[vec_idx].buf_len;
 924
 925                 buf_offset = 0;
 926                 buf_avail = buf_len;
 927         } else {
 928                 buf_offset = dev->vhost_hlen;
 929                 buf_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
 930         }
 931
 932         rte_prefetch0((void *)(uintptr_t)
 933                         (buf_addr + buf_offset));
 934
 935         PRINT_PACKET(dev,
 936                         (uintptr_t)(buf_addr + buf_offset),
 937                         (uint32_t)buf_avail, 0);
 938
 939         mbuf_offset = 0;
 940         mbuf_avail  = m->buf_len - RTE_PKTMBUF_HEADROOM;
 941         while (1) {
 942                 uint64_t hpa;
 943
 944                 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
 945
 946                 /*
 947                  * A desc buf might across two host physical pages that are
 948                  * not continuous. In such case (gpa_to_hpa returns 0), data
 949                  * will be copied even though zero copy is enabled.
 950                  */
 951                 if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
 952                                         buf_iova + buf_offset, cpy_len)))) {
 953                         cur->data_len = cpy_len;
 954                         cur->data_off = 0;
 955                         cur->buf_addr =
 956                                 (void *)(uintptr_t)(buf_addr + buf_offset);
 957                         cur->buf_iova = hpa;
 958
 959                         /*
 960                          * In zero copy mode, one mbuf can only reference data
 961                          * for one or partial of one desc buff.
 962                          */
 963                         mbuf_avail = cpy_len;
 964                 } else {
 965                         if (likely(cpy_len > MAX_BATCH_LEN ||
 966                                    vq->batch_copy_nb_elems >= vq->size ||
 967                                    (hdr && cur == m))) {
 968                                 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
 969                                                                    mbuf_offset),
 970                                            (void *)((uintptr_t)(buf_addr +
 971                                                            buf_offset)),
 972                                            cpy_len);
 973                         } else {
 974                                 batch_copy[vq->batch_copy_nb_elems].dst =
 975                                         rte_pktmbuf_mtod_offset(cur, void *,
 976                                                                 mbuf_offset);
 977                                 batch_copy[vq->batch_copy_nb_elems].src =
 978                                         (void *)((uintptr_t)(buf_addr +
 979                                                                 buf_offset));
 980                                 batch_copy[vq->batch_copy_nb_elems].len =
 981                                         cpy_len;
 982                                 vq->batch_copy_nb_elems++;
 983                         }
 984                 }
 985
 986                 mbuf_avail  -= cpy_len;
 987                 mbuf_offset += cpy_len;
 988                 buf_avail -= cpy_len;
 989                 buf_offset += cpy_len;
 990
 991                 /* This buf reaches to its end, get the next one */
 992                 if (buf_avail == 0) {
 993                         if (++vec_idx >= nr_vec)
 994                                 break;
 995
 996                         buf_addr = buf_vec[vec_idx].buf_addr;
 997                         buf_iova = buf_vec[vec_idx].buf_iova;
 998                         buf_len = buf_vec[vec_idx].buf_len;
 999
1000                         /*
1001                          * Prefecth desc n + 1 buffer while
1002                          * desc n buffer is processed.
1003                          */
1004                         if (vec_idx + 1 < nr_vec)
1005                                 rte_prefetch0((void *)(uintptr_t)
1006                                                 buf_vec[vec_idx + 1].buf_addr);
1007
1008                         buf_offset = 0;
1009                         buf_avail  = buf_len;
1010
1011                         PRINT_PACKET(dev, (uintptr_t)buf_addr,
1012                                         (uint32_t)buf_avail, 0);
1013                 }
1014
1015                 /*
1016                  * This mbuf reaches to its end, get a new one
1017                  * to hold more data.
1018                  */
1019                 if (mbuf_avail == 0) {
1020                         cur = rte_pktmbuf_alloc(mbuf_pool);
1021                         if (unlikely(cur == NULL)) {
1022                                 RTE_LOG(ERR, VHOST_DATA, "Failed to "
1023                                         "allocate memory for mbuf.\n");
1024                                 error = -1;
1025                                 goto out;
1026                         }
1027                         if (unlikely(dev->dequeue_zero_copy))
1028                                 rte_mbuf_refcnt_update(cur, 1);
1029
1030                         prev->next = cur;
1031                         prev->data_len = mbuf_offset;
1032                         m->nb_segs += 1;
1033                         m->pkt_len += mbuf_offset;
1034                         prev = cur;
1035
1036                         mbuf_offset = 0;
1037                         mbuf_avail  = cur->buf_len - RTE_PKTMBUF_HEADROOM;
1038                 }
1039         }
1040
1041         prev->data_len = mbuf_offset;
1042         m->pkt_len    += mbuf_offset;
1043
1044         if (hdr)
1045                 vhost_dequeue_offload(hdr, m);
1046
1047 out:
1048
1049         return error;
1050 }
1051
1052 static __rte_always_inline struct zcopy_mbuf *
1053 get_zmbuf(struct vhost_virtqueue *vq)
1054 {
1055         uint16_t i;
1056         uint16_t last;
1057         int tries = 0;
1058
1059         /* search [last_zmbuf_idx, zmbuf_size) */
1060         i = vq->last_zmbuf_idx;
1061         last = vq->zmbuf_size;
1062
1063 again:
1064         for (; i < last; i++) {
1065                 if (vq->zmbufs[i].in_use == 0) {
1066                         vq->last_zmbuf_idx = i + 1;
1067                         vq->zmbufs[i].in_use = 1;
1068                         return &vq->zmbufs[i];
1069                 }
1070         }
1071
1072         tries++;
1073         if (tries == 1) {
1074                 /* search [0, last_zmbuf_idx) */
1075                 i = 0;
1076                 last = vq->last_zmbuf_idx;
1077                 goto again;
1078         }
1079
1080         return NULL;
1081 }
1082
1083 static __rte_always_inline bool
1084 mbuf_is_consumed(struct rte_mbuf *m)
1085 {
1086         while (m) {
1087                 if (rte_mbuf_refcnt_read(m) > 1)
1088                         return false;
1089                 m = m->next;
1090         }
1091
1092         return true;
1093 }
1094
1095 static __rte_always_inline void
1096 restore_mbuf(struct rte_mbuf *m)
1097 {
1098         uint32_t mbuf_size, priv_size;
1099
1100         while (m) {
1101                 priv_size = rte_pktmbuf_priv_size(m->pool);
1102                 mbuf_size = sizeof(struct rte_mbuf) + priv_size;
1103                 /* start of buffer is after mbuf structure and priv data */
1104
1105                 m->buf_addr = (char *)m + mbuf_size;
1106                 m->buf_iova = rte_mempool_virt2iova(m) + mbuf_size;
1107                 m = m->next;
1108         }
1109 }
1110
1111 static __rte_always_inline uint16_t
1112 virtio_dev_tx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
1113         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1114 {
1115         uint16_t i;
1116         uint16_t free_entries;
1117
1118         if (unlikely(dev->dequeue_zero_copy)) {
1119                 struct zcopy_mbuf *zmbuf, *next;
1120                 int nr_updated = 0;
1121
1122                 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1123                      zmbuf != NULL; zmbuf = next) {
1124                         next = TAILQ_NEXT(zmbuf, next);
1125
1126                         if (mbuf_is_consumed(zmbuf->mbuf)) {
1127                                 update_shadow_used_ring_split(vq,
1128                                                 zmbuf->desc_idx, 0);
1129                                 nr_updated += 1;
1130
1131                                 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1132                                 restore_mbuf(zmbuf->mbuf);
1133                                 rte_pktmbuf_free(zmbuf->mbuf);
1134                                 put_zmbuf(zmbuf);
1135                                 vq->nr_zmbuf -= 1;
1136                         }
1137                 }
1138
1139                 flush_shadow_used_ring_split(dev, vq);
1140                 vhost_vring_call(dev, vq);
1141         }
1142
1143         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
1144
1145         free_entries = *((volatile uint16_t *)&vq->avail->idx) -
1146                         vq->last_avail_idx;
1147         if (free_entries == 0)
1148                 return 0;
1149
1150         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1151
1152         count = RTE_MIN(count, MAX_PKT_BURST);
1153         count = RTE_MIN(count, free_entries);
1154         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1155                         dev->vid, count);
1156
1157         for (i = 0; i < count; i++) {
1158                 struct buf_vector buf_vec[BUF_VECTOR_MAX];
1159                 uint16_t head_idx, dummy_len;
1160                 uint16_t nr_vec = 0;
1161                 int err;
1162
1163                 if (unlikely(fill_vec_buf_split(dev, vq,
1164                                                 vq->last_avail_idx + i,
1165                                                 &nr_vec, buf_vec,
1166                                                 &head_idx, &dummy_len,
1167                                                 VHOST_ACCESS_RO) < 0))
1168                         break;
1169
1170                 if (likely(dev->dequeue_zero_copy == 0))
1171                         update_shadow_used_ring_split(vq, head_idx, 0);
1172
1173                 rte_prefetch0((void *)(uintptr_t)buf_vec[0].buf_addr);
1174
1175                 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
1176                 if (unlikely(pkts[i] == NULL)) {
1177                         RTE_LOG(ERR, VHOST_DATA,
1178                                 "Failed to allocate memory for mbuf.\n");
1179                         break;
1180                 }
1181
1182                 err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts[i],
1183                                 mbuf_pool);
1184                 if (unlikely(err)) {
1185                         rte_pktmbuf_free(pkts[i]);
1186                         break;
1187                 }
1188
1189                 if (unlikely(dev->dequeue_zero_copy)) {
1190                         struct zcopy_mbuf *zmbuf;
1191
1192                         zmbuf = get_zmbuf(vq);
1193                         if (!zmbuf) {
1194                                 rte_pktmbuf_free(pkts[i]);
1195                                 break;
1196                         }
1197                         zmbuf->mbuf = pkts[i];
1198                         zmbuf->desc_idx = head_idx;
1199
1200                         /*
1201                          * Pin lock the mbuf; we will check later to see
1202                          * whether the mbuf is freed (when we are the last
1203                          * user) or not. If that's the case, we then could
1204                          * update the used ring safely.
1205                          */
1206                         rte_mbuf_refcnt_update(pkts[i], 1);
1207
1208                         vq->nr_zmbuf += 1;
1209                         TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1210                 }
1211         }
1212         vq->last_avail_idx += i;
1213
1214         if (likely(dev->dequeue_zero_copy == 0)) {
1215                 do_data_copy_dequeue(vq);
1216                 if (unlikely(i < count))
1217                         vq->shadow_used_idx = i;
1218                 flush_shadow_used_ring_split(dev, vq);
1219                 vhost_vring_call(dev, vq);
1220         }
1221
1222         return i;
1223 }
1224
1225 uint16_t
1226 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
1227         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1228 {
1229         struct virtio_net *dev;
1230         struct rte_mbuf *rarp_mbuf = NULL;
1231         struct vhost_virtqueue *vq;
1232
1233         dev = get_device(vid);
1234         if (!dev)
1235                 return 0;
1236
1237         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
1238                 RTE_LOG(ERR, VHOST_DATA,
1239                         "(%d) %s: built-in vhost net backend is disabled.\n",
1240                         dev->vid, __func__);
1241                 return 0;
1242         }
1243
1244         if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
1245                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
1246                         dev->vid, __func__, queue_id);
1247                 return 0;
1248         }
1249
1250         vq = dev->virtqueue[queue_id];
1251
1252         if (unlikely(rte_spinlock_trylock(&vq->access_lock) == 0))
1253                 return 0;
1254
1255         if (unlikely(vq->enabled == 0))
1256                 goto out_access_unlock;
1257
1258         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1259                 vhost_user_iotlb_rd_lock(vq);
1260
1261         if (unlikely(vq->access_ok == 0))
1262                 if (unlikely(vring_translate(dev, vq) < 0))
1263                         goto out;
1264
1265         /*
1266          * Construct a RARP broadcast packet, and inject it to the "pkts"
1267          * array, to looks like that guest actually send such packet.
1268          *
1269          * Check user_send_rarp() for more information.
1270          *
1271          * broadcast_rarp shares a cacheline in the virtio_net structure
1272          * with some fields that are accessed during enqueue and
1273          * rte_atomic16_cmpset() causes a write if using cmpxchg. This could
1274          * result in false sharing between enqueue and dequeue.
1275          *
1276          * Prevent unnecessary false sharing by reading broadcast_rarp first
1277          * and only performing cmpset if the read indicates it is likely to
1278          * be set.
1279          */
1280         if (unlikely(rte_atomic16_read(&dev->broadcast_rarp) &&
1281                         rte_atomic16_cmpset((volatile uint16_t *)
1282                                 &dev->broadcast_rarp.cnt, 1, 0))) {
1283
1284                 rarp_mbuf = rte_net_make_rarp_packet(mbuf_pool, &dev->mac);
1285                 if (rarp_mbuf == NULL) {
1286                         RTE_LOG(ERR, VHOST_DATA,
1287                                 "Failed to make RARP packet.\n");
1288                         return 0;
1289                 }
1290                 count -= 1;
1291         }
1292
1293         count = virtio_dev_tx_split(dev, vq, mbuf_pool, pkts, count);
1294
1295 out:
1296         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1297                 vhost_user_iotlb_rd_unlock(vq);
1298
1299 out_access_unlock:
1300         rte_spinlock_unlock(&vq->access_lock);
1301
1302         if (unlikely(rarp_mbuf != NULL)) {
1303                 /*
1304                  * Inject it to the head of "pkts" array, so that switch's mac
1305                  * learning table will get updated first.
1306                  */
1307                 memmove(&pkts[1], pkts, count * sizeof(struct rte_mbuf *));
1308                 pkts[0] = rarp_mbuf;
1309                 count += 1;
1310         }
1311
1312         return count;
1313 }