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 fill_vec_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 296                          uint32_t avail_idx, uint16_t *vec_idx,
 297                          struct buf_vector *buf_vec, uint16_t *desc_chain_head,
 298                          uint16_t *desc_chain_len, uint8_t perm)
 299 {
 300         uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
 301         uint16_t vec_id = *vec_idx;
 302         uint32_t len    = 0;
 303         uint64_t dlen, desc_avail, desc_iova;
 304         struct vring_desc *descs = vq->desc;
 305         struct vring_desc *idesc = NULL;
 306
 307         *desc_chain_head = idx;
 308
 309         if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
 310                 dlen = vq->desc[idx].len;
 311                 descs = (struct vring_desc *)(uintptr_t)
 312                         vhost_iova_to_vva(dev, vq, vq->desc[idx].addr,
 313                                                 &dlen,
 314                                                 VHOST_ACCESS_RO);
 315                 if (unlikely(!descs))
 316                         return -1;
 317
 318                 if (unlikely(dlen < vq->desc[idx].len)) {
 319                         /*
 320                          * The indirect desc table is not contiguous
 321                          * in process VA space, we have to copy it.
 322                          */
 323                         idesc = alloc_copy_ind_table(dev, vq,
 324                                         vq->desc[idx].addr, vq->desc[idx].len);
 325                         if (unlikely(!idesc))
 326                                 return -1;
 327
 328                         descs = idesc;
 329                 }
 330
 331                 idx = 0;
 332         }
 333
 334         while (1) {
 335                 if (unlikely(idx >= vq->size)) {
 336                         free_ind_table(idesc);
 337                         return -1;
 338                 }
 339
 340
 341                 len += descs[idx].len;
 342                 desc_avail = descs[idx].len;
 343                 desc_iova = descs[idx].addr;
 344
 345                 while (desc_avail) {
 346                         uint64_t desc_addr;
 347                         uint64_t desc_chunck_len = desc_avail;
 348
 349                         if (unlikely(vec_id >= BUF_VECTOR_MAX)) {
 350                                 free_ind_table(idesc);
 351                                 return -1;
 352                         }
 353
 354                         desc_addr = vhost_iova_to_vva(dev, vq,
 355                                         desc_iova,
 356                                         &desc_chunck_len,
 357                                         perm);
 358                         if (unlikely(!desc_addr)) {
 359                                 free_ind_table(idesc);
 360                                 return -1;
 361                         }
 362
 363                         buf_vec[vec_id].buf_iova = desc_iova;
 364                         buf_vec[vec_id].buf_addr = desc_addr;
 365                         buf_vec[vec_id].buf_len  = desc_chunck_len;
 366                         buf_vec[vec_id].desc_idx = idx;
 367
 368                         desc_avail -= desc_chunck_len;
 369                         desc_iova += desc_chunck_len;
 370                         vec_id++;
 371                 }
 372
 373                 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
 374                         break;
 375
 376                 idx = descs[idx].next;
 377         }
 378
 379         *desc_chain_len = len;
 380         *vec_idx = vec_id;
 381
 382         if (unlikely(!!idesc))
 383                 free_ind_table(idesc);
 384
 385         return 0;
 386 }
 387
 388 /*
 389  * Returns -1 on fail, 0 on success
 390  */
 391 static inline int
 392 reserve_avail_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 393                                 uint32_t size, struct buf_vector *buf_vec,
 394                                 uint16_t *num_buffers, uint16_t avail_head,
 395                                 uint16_t *nr_vec)
 396 {
 397         uint16_t cur_idx;
 398         uint16_t vec_idx = 0;
 399         uint16_t max_tries, tries = 0;
 400
 401         uint16_t head_idx = 0;
 402         uint16_t len = 0;
 403
 404         *num_buffers = 0;
 405         cur_idx  = vq->last_avail_idx;
 406
 407         if (rxvq_is_mergeable(dev))
 408                 max_tries = vq->size;
 409         else
 410                 max_tries = 1;
 411
 412         while (size > 0) {
 413                 if (unlikely(cur_idx == avail_head))
 414                         return -1;
 415
 416                 if (unlikely(fill_vec_buf_split(dev, vq, cur_idx,
 417                                                 &vec_idx, buf_vec,
 418                                                 &head_idx, &len,
 419                                                 VHOST_ACCESS_RW) < 0))
 420                         return -1;
 421                 len = RTE_MIN(len, size);
 422                 update_shadow_used_ring_split(vq, head_idx, len);
 423                 size -= len;
 424
 425                 cur_idx++;
 426                 tries++;
 427                 *num_buffers += 1;
 428
 429                 /*
 430                  * if we tried all available ring items, and still
 431                  * can't get enough buf, it means something abnormal
 432                  * happened.
 433                  */
 434                 if (unlikely(tries > max_tries))
 435                         return -1;
 436         }
 437
 438         *nr_vec = vec_idx;
 439
 440         return 0;
 441 }
 442
 443 static __rte_always_inline int
 444 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 445                             struct rte_mbuf *m, struct buf_vector *buf_vec,
 446                             uint16_t nr_vec, uint16_t num_buffers)
 447 {
 448         uint32_t vec_idx = 0;
 449         uint32_t mbuf_offset, mbuf_avail;
 450         uint32_t buf_offset, buf_avail;
 451         uint64_t buf_addr, buf_iova, buf_len;
 452         uint32_t cpy_len;
 453         uint64_t hdr_addr;
 454         struct rte_mbuf *hdr_mbuf;
 455         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 456         struct virtio_net_hdr_mrg_rxbuf tmp_hdr, *hdr = NULL;
 457         int error = 0;
 458
 459         if (unlikely(m == NULL)) {
 460                 error = -1;
 461                 goto out;
 462         }
 463
 464         buf_addr = buf_vec[vec_idx].buf_addr;
 465         buf_iova = buf_vec[vec_idx].buf_iova;
 466         buf_len = buf_vec[vec_idx].buf_len;
 467
 468         if (nr_vec > 1)
 469                 rte_prefetch0((void *)(uintptr_t)buf_vec[1].buf_addr);
 470
 471         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
 472                 error = -1;
 473                 goto out;
 474         }
 475
 476         hdr_mbuf = m;
 477         hdr_addr = buf_addr;
 478         if (unlikely(buf_len < dev->vhost_hlen))
 479                 hdr = &tmp_hdr;
 480         else
 481                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)hdr_addr;
 482
 483         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
 484                 dev->vid, num_buffers);
 485
 486         if (unlikely(buf_len < dev->vhost_hlen)) {
 487                 buf_offset = dev->vhost_hlen - buf_len;
 488                 vec_idx++;
 489                 buf_addr = buf_vec[vec_idx].buf_addr;
 490                 buf_iova = buf_vec[vec_idx].buf_iova;
 491                 buf_len = buf_vec[vec_idx].buf_len;
 492                 buf_avail = buf_len - buf_offset;
 493         } else {
 494                 buf_offset = dev->vhost_hlen;
 495                 buf_avail = buf_len - dev->vhost_hlen;
 496         }
 497
 498         mbuf_avail  = rte_pktmbuf_data_len(m);
 499         mbuf_offset = 0;
 500         while (mbuf_avail != 0 || m->next != NULL) {
 501                 /* done with current buf, get the next one */
 502                 if (buf_avail == 0) {
 503                         vec_idx++;
 504                         if (unlikely(vec_idx >= nr_vec)) {
 505                                 error = -1;
 506                                 goto out;
 507                         }
 508
 509                         buf_addr = buf_vec[vec_idx].buf_addr;
 510                         buf_iova = buf_vec[vec_idx].buf_iova;
 511                         buf_len = buf_vec[vec_idx].buf_len;
 512
 513                         /* Prefetch next buffer address. */
 514                         if (vec_idx + 1 < nr_vec)
 515                                 rte_prefetch0((void *)(uintptr_t)
 516                                                 buf_vec[vec_idx + 1].buf_addr);
 517                         buf_offset = 0;
 518                         buf_avail  = buf_len;
 519                 }
 520
 521                 /* done with current mbuf, get the next one */
 522                 if (mbuf_avail == 0) {
 523                         m = m->next;
 524
 525                         mbuf_offset = 0;
 526                         mbuf_avail  = rte_pktmbuf_data_len(m);
 527                 }
 528
 529                 if (hdr_addr) {
 530                         virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
 531                         if (rxvq_is_mergeable(dev))
 532                                 ASSIGN_UNLESS_EQUAL(hdr->num_buffers,
 533                                                 num_buffers);
 534
 535                         if (unlikely(hdr == &tmp_hdr)) {
 536                                 uint64_t len;
 537                                 uint64_t remain = dev->vhost_hlen;
 538                                 uint64_t src = (uint64_t)(uintptr_t)hdr, dst;
 539                                 uint64_t iova = buf_vec[0].buf_iova;
 540                                 uint16_t hdr_vec_idx = 0;
 541
 542                                 while (remain) {
 543                                         len = remain;
 544                                         dst = buf_vec[hdr_vec_idx].buf_addr;
 545                                         rte_memcpy((void *)(uintptr_t)dst,
 546                                                         (void *)(uintptr_t)src,
 547                                                         len);
 548
 549                                         PRINT_PACKET(dev, (uintptr_t)dst,
 550                                                         (uint32_t)len, 0);
 551                                         vhost_log_cache_write(dev, vq,
 552                                                         iova, len);
 553
 554                                         remain -= len;
 555                                         iova += len;
 556                                         src += len;
 557                                         hdr_vec_idx++;
 558                                 }
 559                         } else {
 560                                 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
 561                                                 dev->vhost_hlen, 0);
 562                                 vhost_log_cache_write(dev, vq,
 563                                                 buf_vec[0].buf_iova,
 564                                                 dev->vhost_hlen);
 565                         }
 566
 567                         hdr_addr = 0;
 568                 }
 569
 570                 cpy_len = RTE_MIN(buf_len, mbuf_avail);
 571
 572                 if (likely(cpy_len > MAX_BATCH_LEN ||
 573                                         vq->batch_copy_nb_elems >= vq->size)) {
 574                         rte_memcpy((void *)((uintptr_t)(buf_addr + buf_offset)),
 575                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 576                                 cpy_len);
 577                         vhost_log_cache_write(dev, vq, buf_iova + buf_offset,
 578                                         cpy_len);
 579                         PRINT_PACKET(dev, (uintptr_t)(buf_addr + buf_offset),
 580                                 cpy_len, 0);
 581                 } else {
 582                         batch_copy[vq->batch_copy_nb_elems].dst =
 583                                 (void *)((uintptr_t)(buf_addr + buf_offset));
 584                         batch_copy[vq->batch_copy_nb_elems].src =
 585                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
 586                         batch_copy[vq->batch_copy_nb_elems].log_addr =
 587                                 buf_iova + buf_offset;
 588                         batch_copy[vq->batch_copy_nb_elems].len = cpy_len;
 589                         vq->batch_copy_nb_elems++;
 590                 }
 591
 592                 mbuf_avail  -= cpy_len;
 593                 mbuf_offset += cpy_len;
 594                 buf_avail  -= cpy_len;
 595                 buf_offset += cpy_len;
 596         }
 597
 598 out:
 599
 600         return error;
 601 }
 602
 603 static __rte_always_inline uint32_t
 604 virtio_dev_rx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 605         struct rte_mbuf **pkts, uint32_t count)
 606 {
 607         uint32_t pkt_idx = 0;
 608         uint16_t num_buffers;
 609         struct buf_vector buf_vec[BUF_VECTOR_MAX];
 610         uint16_t avail_head;
 611
 612         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
 613         avail_head = *((volatile uint16_t *)&vq->avail->idx);
 614
 615         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
 616                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
 617                 uint16_t nr_vec = 0;
 618
 619                 if (unlikely(reserve_avail_buf_split(dev, vq,
 620                                                 pkt_len, buf_vec, &num_buffers,
 621                                                 avail_head, &nr_vec) < 0)) {
 622                         VHOST_LOG_DEBUG(VHOST_DATA,
 623                                 "(%d) failed to get enough desc from vring\n",
 624                                 dev->vid);
 625                         vq->shadow_used_idx -= num_buffers;
 626                         break;
 627                 }
 628
 629                 rte_prefetch0((void *)(uintptr_t)buf_vec[0].buf_addr);
 630
 631                 VHOST_LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
 632                         dev->vid, vq->last_avail_idx,
 633                         vq->last_avail_idx + num_buffers);
 634
 635                 if (copy_mbuf_to_desc(dev, vq, pkts[pkt_idx],
 636                                                 buf_vec, nr_vec,
 637                                                 num_buffers) < 0) {
 638                         vq->shadow_used_idx -= num_buffers;
 639                         break;
 640                 }
 641
 642                 vq->last_avail_idx += num_buffers;
 643         }
 644
 645         do_data_copy_enqueue(dev, vq);
 646
 647         if (likely(vq->shadow_used_idx)) {
 648                 flush_shadow_used_ring_split(dev, vq);
 649                 vhost_vring_call(dev, vq);
 650         }
 651
 652         return pkt_idx;
 653 }
 654
 655 static __rte_always_inline uint32_t
 656 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
 657         struct rte_mbuf **pkts, uint32_t count)
 658 {
 659         struct vhost_virtqueue *vq;
 660
 661         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
 662         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
 663                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
 664                         dev->vid, __func__, queue_id);
 665                 return 0;
 666         }
 667
 668         vq = dev->virtqueue[queue_id];
 669
 670         rte_spinlock_lock(&vq->access_lock);
 671
 672         if (unlikely(vq->enabled == 0))
 673                 goto out_access_unlock;
 674
 675         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 676                 vhost_user_iotlb_rd_lock(vq);
 677
 678         if (unlikely(vq->access_ok == 0))
 679                 if (unlikely(vring_translate(dev, vq) < 0))
 680                         goto out;
 681
 682         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
 683         if (count == 0)
 684                 goto out;
 685
 686         count = virtio_dev_rx_split(dev, vq, pkts, count);
 687
 688 out:
 689         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 690                 vhost_user_iotlb_rd_unlock(vq);
 691
 692 out_access_unlock:
 693         rte_spinlock_unlock(&vq->access_lock);
 694
 695         return count;
 696 }
 697
 698 uint16_t
 699 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
 700         struct rte_mbuf **pkts, uint16_t count)
 701 {
 702         struct virtio_net *dev = get_device(vid);
 703
 704         if (!dev)
 705                 return 0;
 706
 707         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
 708                 RTE_LOG(ERR, VHOST_DATA,
 709                         "(%d) %s: built-in vhost net backend is disabled.\n",
 710                         dev->vid, __func__);
 711                 return 0;
 712         }
 713
 714         return virtio_dev_rx(dev, queue_id, pkts, count);
 715 }
 716
 717 static inline bool
 718 virtio_net_with_host_offload(struct virtio_net *dev)
 719 {
 720         if (dev->features &
 721                         ((1ULL << VIRTIO_NET_F_CSUM) |
 722                          (1ULL << VIRTIO_NET_F_HOST_ECN) |
 723                          (1ULL << VIRTIO_NET_F_HOST_TSO4) |
 724                          (1ULL << VIRTIO_NET_F_HOST_TSO6) |
 725                          (1ULL << VIRTIO_NET_F_HOST_UFO)))
 726                 return true;
 727
 728         return false;
 729 }
 730
 731 static void
 732 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
 733 {
 734         struct ipv4_hdr *ipv4_hdr;
 735         struct ipv6_hdr *ipv6_hdr;
 736         void *l3_hdr = NULL;
 737         struct ether_hdr *eth_hdr;
 738         uint16_t ethertype;
 739
 740         eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
 741
 742         m->l2_len = sizeof(struct ether_hdr);
 743         ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
 744
 745         if (ethertype == ETHER_TYPE_VLAN) {
 746                 struct vlan_hdr *vlan_hdr = (struct vlan_hdr *)(eth_hdr + 1);
 747
 748                 m->l2_len += sizeof(struct vlan_hdr);
 749                 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
 750         }
 751
 752         l3_hdr = (char *)eth_hdr + m->l2_len;
 753
 754         switch (ethertype) {
 755         case ETHER_TYPE_IPv4:
 756                 ipv4_hdr = l3_hdr;
 757                 *l4_proto = ipv4_hdr->next_proto_id;
 758                 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
 759                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 760                 m->ol_flags |= PKT_TX_IPV4;
 761                 break;
 762         case ETHER_TYPE_IPv6:
 763                 ipv6_hdr = l3_hdr;
 764                 *l4_proto = ipv6_hdr->proto;
 765                 m->l3_len = sizeof(struct ipv6_hdr);
 766                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 767                 m->ol_flags |= PKT_TX_IPV6;
 768                 break;
 769         default:
 770                 m->l3_len = 0;
 771                 *l4_proto = 0;
 772                 *l4_hdr = NULL;
 773                 break;
 774         }
 775 }
 776
 777 static __rte_always_inline void
 778 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
 779 {
 780         uint16_t l4_proto = 0;
 781         void *l4_hdr = NULL;
 782         struct tcp_hdr *tcp_hdr = NULL;
 783
 784         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
 785                 return;
 786
 787         parse_ethernet(m, &l4_proto, &l4_hdr);
 788         if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 789                 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
 790                         switch (hdr->csum_offset) {
 791                         case (offsetof(struct tcp_hdr, cksum)):
 792                                 if (l4_proto == IPPROTO_TCP)
 793                                         m->ol_flags |= PKT_TX_TCP_CKSUM;
 794                                 break;
 795                         case (offsetof(struct udp_hdr, dgram_cksum)):
 796                                 if (l4_proto == IPPROTO_UDP)
 797                                         m->ol_flags |= PKT_TX_UDP_CKSUM;
 798                                 break;
 799                         case (offsetof(struct sctp_hdr, cksum)):
 800                                 if (l4_proto == IPPROTO_SCTP)
 801                                         m->ol_flags |= PKT_TX_SCTP_CKSUM;
 802                                 break;
 803                         default:
 804                                 break;
 805                         }
 806                 }
 807         }
 808
 809         if (l4_hdr && hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 810                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 811                 case VIRTIO_NET_HDR_GSO_TCPV4:
 812                 case VIRTIO_NET_HDR_GSO_TCPV6:
 813                         tcp_hdr = l4_hdr;
 814                         m->ol_flags |= PKT_TX_TCP_SEG;
 815                         m->tso_segsz = hdr->gso_size;
 816                         m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
 817                         break;
 818                 case VIRTIO_NET_HDR_GSO_UDP:
 819                         m->ol_flags |= PKT_TX_UDP_SEG;
 820                         m->tso_segsz = hdr->gso_size;
 821                         m->l4_len = sizeof(struct udp_hdr);
 822                         break;
 823                 default:
 824                         RTE_LOG(WARNING, VHOST_DATA,
 825                                 "unsupported gso type %u.\n", hdr->gso_type);
 826                         break;
 827                 }
 828         }
 829 }
 830
 831 static __rte_always_inline void
 832 put_zmbuf(struct zcopy_mbuf *zmbuf)
 833 {
 834         zmbuf->in_use = 0;
 835 }
 836
 837 static __rte_always_inline int
 838 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 839                   struct buf_vector *buf_vec, uint16_t nr_vec,
 840                   struct rte_mbuf *m, struct rte_mempool *mbuf_pool)
 841 {
 842         uint32_t buf_avail, buf_offset;
 843         uint64_t buf_addr, buf_iova, buf_len;
 844         uint32_t mbuf_avail, mbuf_offset;
 845         uint32_t cpy_len;
 846         struct rte_mbuf *cur = m, *prev = m;
 847         struct virtio_net_hdr tmp_hdr;
 848         struct virtio_net_hdr *hdr = NULL;
 849         /* A counter to avoid desc dead loop chain */
 850         uint16_t vec_idx = 0;
 851         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 852         int error = 0;
 853
 854         buf_addr = buf_vec[vec_idx].buf_addr;
 855         buf_iova = buf_vec[vec_idx].buf_iova;
 856         buf_len = buf_vec[vec_idx].buf_len;
 857
 858         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
 859                 error = -1;
 860                 goto out;
 861         }
 862
 863         if (likely(nr_vec > 1))
 864                 rte_prefetch0((void *)(uintptr_t)buf_vec[1].buf_addr);
 865
 866         if (virtio_net_with_host_offload(dev)) {
 867                 if (unlikely(buf_len < sizeof(struct virtio_net_hdr))) {
 868                         uint64_t len;
 869                         uint64_t remain = sizeof(struct virtio_net_hdr);
 870                         uint64_t src;
 871                         uint64_t dst = (uint64_t)(uintptr_t)&tmp_hdr;
 872                         uint16_t hdr_vec_idx = 0;
 873
 874                         /*
 875                          * No luck, the virtio-net header doesn't fit
 876                          * in a contiguous virtual area.
 877                          */
 878                         while (remain) {
 879                                 len = remain;
 880                                 src = buf_vec[hdr_vec_idx].buf_addr;
 881                                 rte_memcpy((void *)(uintptr_t)dst,
 882                                                    (void *)(uintptr_t)src, len);
 883
 884                                 remain -= len;
 885                                 dst += len;
 886                                 hdr_vec_idx++;
 887                         }
 888
 889                         hdr = &tmp_hdr;
 890                 } else {
 891                         hdr = (struct virtio_net_hdr *)((uintptr_t)buf_addr);
 892                         rte_prefetch0(hdr);
 893                 }
 894         }
 895
 896         /*
 897          * A virtio driver normally uses at least 2 desc buffers
 898          * for Tx: the first for storing the header, and others
 899          * for storing the data.
 900          */
 901         if (unlikely(buf_len < dev->vhost_hlen)) {
 902                 buf_offset = dev->vhost_hlen - buf_len;
 903                 vec_idx++;
 904                 buf_addr = buf_vec[vec_idx].buf_addr;
 905                 buf_iova = buf_vec[vec_idx].buf_iova;
 906                 buf_len = buf_vec[vec_idx].buf_len;
 907                 buf_avail  = buf_len - buf_offset;
 908         } else if (buf_len == dev->vhost_hlen) {
 909                 if (unlikely(++vec_idx >= nr_vec))
 910                         goto out;
 911                 buf_addr = buf_vec[vec_idx].buf_addr;
 912                 buf_iova = buf_vec[vec_idx].buf_iova;
 913                 buf_len = buf_vec[vec_idx].buf_len;
 914
 915                 buf_offset = 0;
 916                 buf_avail = buf_len;
 917         } else {
 918                 buf_offset = dev->vhost_hlen;
 919                 buf_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
 920         }
 921
 922         rte_prefetch0((void *)(uintptr_t)
 923                         (buf_addr + buf_offset));
 924
 925         PRINT_PACKET(dev,
 926                         (uintptr_t)(buf_addr + buf_offset),
 927                         (uint32_t)buf_avail, 0);
 928
 929         mbuf_offset = 0;
 930         mbuf_avail  = m->buf_len - RTE_PKTMBUF_HEADROOM;
 931         while (1) {
 932                 uint64_t hpa;
 933
 934                 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
 935
 936                 /*
 937                  * A desc buf might across two host physical pages that are
 938                  * not continuous. In such case (gpa_to_hpa returns 0), data
 939                  * will be copied even though zero copy is enabled.
 940                  */
 941                 if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
 942                                         buf_iova + buf_offset, cpy_len)))) {
 943                         cur->data_len = cpy_len;
 944                         cur->data_off = 0;
 945                         cur->buf_addr =
 946                                 (void *)(uintptr_t)(buf_addr + buf_offset);
 947                         cur->buf_iova = hpa;
 948
 949                         /*
 950                          * In zero copy mode, one mbuf can only reference data
 951                          * for one or partial of one desc buff.
 952                          */
 953                         mbuf_avail = cpy_len;
 954                 } else {
 955                         if (likely(cpy_len > MAX_BATCH_LEN ||
 956                                    vq->batch_copy_nb_elems >= vq->size ||
 957                                    (hdr && cur == m))) {
 958                                 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
 959                                                                    mbuf_offset),
 960                                            (void *)((uintptr_t)(buf_addr +
 961                                                            buf_offset)),
 962                                            cpy_len);
 963                         } else {
 964                                 batch_copy[vq->batch_copy_nb_elems].dst =
 965                                         rte_pktmbuf_mtod_offset(cur, void *,
 966                                                                 mbuf_offset);
 967                                 batch_copy[vq->batch_copy_nb_elems].src =
 968                                         (void *)((uintptr_t)(buf_addr +
 969                                                                 buf_offset));
 970                                 batch_copy[vq->batch_copy_nb_elems].len =
 971                                         cpy_len;
 972                                 vq->batch_copy_nb_elems++;
 973                         }
 974                 }
 975
 976                 mbuf_avail  -= cpy_len;
 977                 mbuf_offset += cpy_len;
 978                 buf_avail -= cpy_len;
 979                 buf_offset += cpy_len;
 980
 981                 /* This buf reaches to its end, get the next one */
 982                 if (buf_avail == 0) {
 983                         if (++vec_idx >= nr_vec)
 984                                 break;
 985
 986                         buf_addr = buf_vec[vec_idx].buf_addr;
 987                         buf_iova = buf_vec[vec_idx].buf_iova;
 988                         buf_len = buf_vec[vec_idx].buf_len;
 989
 990                         /*
 991                          * Prefecth desc n + 1 buffer while
 992                          * desc n buffer is processed.
 993                          */
 994                         if (vec_idx + 1 < nr_vec)
 995                                 rte_prefetch0((void *)(uintptr_t)
 996                                                 buf_vec[vec_idx + 1].buf_addr);
 997
 998                         buf_offset = 0;
 999                         buf_avail  = buf_len;
1000
1001                         PRINT_PACKET(dev, (uintptr_t)buf_addr,
1002                                         (uint32_t)buf_avail, 0);
1003                 }
1004
1005                 /*
1006                  * This mbuf reaches to its end, get a new one
1007                  * to hold more data.
1008                  */
1009                 if (mbuf_avail == 0) {
1010                         cur = rte_pktmbuf_alloc(mbuf_pool);
1011                         if (unlikely(cur == NULL)) {
1012                                 RTE_LOG(ERR, VHOST_DATA, "Failed to "
1013                                         "allocate memory for mbuf.\n");
1014                                 error = -1;
1015                                 goto out;
1016                         }
1017                         if (unlikely(dev->dequeue_zero_copy))
1018                                 rte_mbuf_refcnt_update(cur, 1);
1019
1020                         prev->next = cur;
1021                         prev->data_len = mbuf_offset;
1022                         m->nb_segs += 1;
1023                         m->pkt_len += mbuf_offset;
1024                         prev = cur;
1025
1026                         mbuf_offset = 0;
1027                         mbuf_avail  = cur->buf_len - RTE_PKTMBUF_HEADROOM;
1028                 }
1029         }
1030
1031         prev->data_len = mbuf_offset;
1032         m->pkt_len    += mbuf_offset;
1033
1034         if (hdr)
1035                 vhost_dequeue_offload(hdr, m);
1036
1037 out:
1038
1039         return error;
1040 }
1041
1042 static __rte_always_inline struct zcopy_mbuf *
1043 get_zmbuf(struct vhost_virtqueue *vq)
1044 {
1045         uint16_t i;
1046         uint16_t last;
1047         int tries = 0;
1048
1049         /* search [last_zmbuf_idx, zmbuf_size) */
1050         i = vq->last_zmbuf_idx;
1051         last = vq->zmbuf_size;
1052
1053 again:
1054         for (; i < last; i++) {
1055                 if (vq->zmbufs[i].in_use == 0) {
1056                         vq->last_zmbuf_idx = i + 1;
1057                         vq->zmbufs[i].in_use = 1;
1058                         return &vq->zmbufs[i];
1059                 }
1060         }
1061
1062         tries++;
1063         if (tries == 1) {
1064                 /* search [0, last_zmbuf_idx) */
1065                 i = 0;
1066                 last = vq->last_zmbuf_idx;
1067                 goto again;
1068         }
1069
1070         return NULL;
1071 }
1072
1073 static __rte_always_inline bool
1074 mbuf_is_consumed(struct rte_mbuf *m)
1075 {
1076         while (m) {
1077                 if (rte_mbuf_refcnt_read(m) > 1)
1078                         return false;
1079                 m = m->next;
1080         }
1081
1082         return true;
1083 }
1084
1085 static __rte_always_inline void
1086 restore_mbuf(struct rte_mbuf *m)
1087 {
1088         uint32_t mbuf_size, priv_size;
1089
1090         while (m) {
1091                 priv_size = rte_pktmbuf_priv_size(m->pool);
1092                 mbuf_size = sizeof(struct rte_mbuf) + priv_size;
1093                 /* start of buffer is after mbuf structure and priv data */
1094
1095                 m->buf_addr = (char *)m + mbuf_size;
1096                 m->buf_iova = rte_mempool_virt2iova(m) + mbuf_size;
1097                 m = m->next;
1098         }
1099 }
1100
1101 static __rte_always_inline uint16_t
1102 virtio_dev_tx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
1103         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1104 {
1105         uint16_t i;
1106         uint16_t free_entries;
1107
1108         if (unlikely(dev->dequeue_zero_copy)) {
1109                 struct zcopy_mbuf *zmbuf, *next;
1110                 int nr_updated = 0;
1111
1112                 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1113                      zmbuf != NULL; zmbuf = next) {
1114                         next = TAILQ_NEXT(zmbuf, next);
1115
1116                         if (mbuf_is_consumed(zmbuf->mbuf)) {
1117                                 update_shadow_used_ring_split(vq,
1118                                                 zmbuf->desc_idx, 0);
1119                                 nr_updated += 1;
1120
1121                                 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1122                                 restore_mbuf(zmbuf->mbuf);
1123                                 rte_pktmbuf_free(zmbuf->mbuf);
1124                                 put_zmbuf(zmbuf);
1125                                 vq->nr_zmbuf -= 1;
1126                         }
1127                 }
1128
1129                 flush_shadow_used_ring_split(dev, vq);
1130                 vhost_vring_call(dev, vq);
1131         }
1132
1133         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
1134
1135         free_entries = *((volatile uint16_t *)&vq->avail->idx) -
1136                         vq->last_avail_idx;
1137         if (free_entries == 0)
1138                 return 0;
1139
1140         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1141
1142         count = RTE_MIN(count, MAX_PKT_BURST);
1143         count = RTE_MIN(count, free_entries);
1144         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1145                         dev->vid, count);
1146
1147         for (i = 0; i < count; i++) {
1148                 struct buf_vector buf_vec[BUF_VECTOR_MAX];
1149                 uint16_t head_idx, dummy_len;
1150                 uint16_t nr_vec = 0;
1151                 int err;
1152
1153                 if (unlikely(fill_vec_buf_split(dev, vq,
1154                                                 vq->last_avail_idx + i,
1155                                                 &nr_vec, buf_vec,
1156                                                 &head_idx, &dummy_len,
1157                                                 VHOST_ACCESS_RO) < 0))
1158                         break;
1159
1160                 if (likely(dev->dequeue_zero_copy == 0))
1161                         update_shadow_used_ring_split(vq, head_idx, 0);
1162
1163                 rte_prefetch0((void *)(uintptr_t)buf_vec[0].buf_addr);
1164
1165                 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
1166                 if (unlikely(pkts[i] == NULL)) {
1167                         RTE_LOG(ERR, VHOST_DATA,
1168                                 "Failed to allocate memory for mbuf.\n");
1169                         break;
1170                 }
1171
1172                 err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts[i],
1173                                 mbuf_pool);
1174                 if (unlikely(err)) {
1175                         rte_pktmbuf_free(pkts[i]);
1176                         break;
1177                 }
1178
1179                 if (unlikely(dev->dequeue_zero_copy)) {
1180                         struct zcopy_mbuf *zmbuf;
1181
1182                         zmbuf = get_zmbuf(vq);
1183                         if (!zmbuf) {
1184                                 rte_pktmbuf_free(pkts[i]);
1185                                 break;
1186                         }
1187                         zmbuf->mbuf = pkts[i];
1188                         zmbuf->desc_idx = head_idx;
1189
1190                         /*
1191                          * Pin lock the mbuf; we will check later to see
1192                          * whether the mbuf is freed (when we are the last
1193                          * user) or not. If that's the case, we then could
1194                          * update the used ring safely.
1195                          */
1196                         rte_mbuf_refcnt_update(pkts[i], 1);
1197
1198                         vq->nr_zmbuf += 1;
1199                         TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1200                 }
1201         }
1202         vq->last_avail_idx += i;
1203
1204         if (likely(dev->dequeue_zero_copy == 0)) {
1205                 do_data_copy_dequeue(vq);
1206                 if (unlikely(i < count))
1207                         vq->shadow_used_idx = i;
1208                 flush_shadow_used_ring_split(dev, vq);
1209                 vhost_vring_call(dev, vq);
1210         }
1211
1212         return i;
1213 }
1214
1215 uint16_t
1216 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
1217         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1218 {
1219         struct virtio_net *dev;
1220         struct rte_mbuf *rarp_mbuf = NULL;
1221         struct vhost_virtqueue *vq;
1222
1223         dev = get_device(vid);
1224         if (!dev)
1225                 return 0;
1226
1227         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
1228                 RTE_LOG(ERR, VHOST_DATA,
1229                         "(%d) %s: built-in vhost net backend is disabled.\n",
1230                         dev->vid, __func__);
1231                 return 0;
1232         }
1233
1234         if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
1235                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
1236                         dev->vid, __func__, queue_id);
1237                 return 0;
1238         }
1239
1240         vq = dev->virtqueue[queue_id];
1241
1242         if (unlikely(rte_spinlock_trylock(&vq->access_lock) == 0))
1243                 return 0;
1244
1245         if (unlikely(vq->enabled == 0))
1246                 goto out_access_unlock;
1247
1248         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1249                 vhost_user_iotlb_rd_lock(vq);
1250
1251         if (unlikely(vq->access_ok == 0))
1252                 if (unlikely(vring_translate(dev, vq) < 0))
1253                         goto out;
1254
1255         /*
1256          * Construct a RARP broadcast packet, and inject it to the "pkts"
1257          * array, to looks like that guest actually send such packet.
1258          *
1259          * Check user_send_rarp() for more information.
1260          *
1261          * broadcast_rarp shares a cacheline in the virtio_net structure
1262          * with some fields that are accessed during enqueue and
1263          * rte_atomic16_cmpset() causes a write if using cmpxchg. This could
1264          * result in false sharing between enqueue and dequeue.
1265          *
1266          * Prevent unnecessary false sharing by reading broadcast_rarp first
1267          * and only performing cmpset if the read indicates it is likely to
1268          * be set.
1269          */
1270         if (unlikely(rte_atomic16_read(&dev->broadcast_rarp) &&
1271                         rte_atomic16_cmpset((volatile uint16_t *)
1272                                 &dev->broadcast_rarp.cnt, 1, 0))) {
1273
1274                 rarp_mbuf = rte_net_make_rarp_packet(mbuf_pool, &dev->mac);
1275                 if (rarp_mbuf == NULL) {
1276                         RTE_LOG(ERR, VHOST_DATA,
1277                                 "Failed to make RARP packet.\n");
1278                         return 0;
1279                 }
1280                 count -= 1;
1281         }
1282
1283         count = virtio_dev_tx_split(dev, vq, mbuf_pool, pkts, count);
1284
1285 out:
1286         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1287                 vhost_user_iotlb_rd_unlock(vq);
1288
1289 out_access_unlock:
1290         rte_spinlock_unlock(&vq->access_lock);
1291
1292         if (unlikely(rarp_mbuf != NULL)) {
1293                 /*
1294                  * Inject it to the head of "pkts" array, so that switch's mac
1295                  * learning table will get updated first.
1296                  */
1297                 memmove(&pkts[1], pkts, count * sizeof(struct rte_mbuf *));
1298                 pkts[0] = rarp_mbuf;
1299                 count += 1;
1300         }
1301
1302         return count;
1303 }