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