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(struct virtio_net *dev, struct vhost_virtqueue *vq,
 230                          uint32_t avail_idx, uint32_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         uint32_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(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         uint32_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(dev, vq, cur_idx, &vec_idx, buf_vec,
 351                                                 &head_idx, &len,
 352                                                 VHOST_ACCESS_RW) < 0))
 353                         return -1;
 354                 len = RTE_MIN(len, size);
 355                 update_shadow_used_ring(vq, head_idx, len);
 356                 size -= len;
 357
 358                 cur_idx++;
 359                 tries++;
 360                 *num_buffers += 1;
 361
 362                 /*
 363                  * if we tried all available ring items, and still
 364                  * can't get enough buf, it means something abnormal
 365                  * happened.
 366                  */
 367                 if (unlikely(tries > max_tries))
 368                         return -1;
 369         }
 370
 371         *nr_vec = vec_idx;
 372
 373         return 0;
 374 }
 375
 376 static __rte_always_inline int
 377 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 378                             struct rte_mbuf *m, struct buf_vector *buf_vec,
 379                             uint16_t nr_vec, uint16_t num_buffers)
 380 {
 381         uint32_t vec_idx = 0;
 382         uint32_t mbuf_offset, mbuf_avail;
 383         uint32_t buf_offset, buf_avail;
 384         uint64_t buf_addr, buf_iova, buf_len;
 385         uint32_t cpy_len;
 386         uint64_t hdr_addr;
 387         struct rte_mbuf *hdr_mbuf;
 388         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 389         struct virtio_net_hdr_mrg_rxbuf tmp_hdr, *hdr = NULL;
 390         int error = 0;
 391
 392         if (unlikely(m == NULL)) {
 393                 error = -1;
 394                 goto out;
 395         }
 396
 397         buf_addr = buf_vec[vec_idx].buf_addr;
 398         buf_iova = buf_vec[vec_idx].buf_iova;
 399         buf_len = buf_vec[vec_idx].buf_len;
 400
 401         if (nr_vec > 1)
 402                 rte_prefetch0((void *)(uintptr_t)buf_vec[1].buf_addr);
 403
 404         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
 405                 error = -1;
 406                 goto out;
 407         }
 408
 409         hdr_mbuf = m;
 410         hdr_addr = buf_addr;
 411         if (unlikely(buf_len < dev->vhost_hlen))
 412                 hdr = &tmp_hdr;
 413         else
 414                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)hdr_addr;
 415
 416         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
 417                 dev->vid, num_buffers);
 418
 419         if (unlikely(buf_len < dev->vhost_hlen)) {
 420                 buf_offset = dev->vhost_hlen - buf_len;
 421                 vec_idx++;
 422                 buf_addr = buf_vec[vec_idx].buf_addr;
 423                 buf_iova = buf_vec[vec_idx].buf_iova;
 424                 buf_len = buf_vec[vec_idx].buf_len;
 425                 buf_avail = buf_len - buf_offset;
 426         } else {
 427                 buf_offset = dev->vhost_hlen;
 428                 buf_avail = buf_len - dev->vhost_hlen;
 429         }
 430
 431         mbuf_avail  = rte_pktmbuf_data_len(m);
 432         mbuf_offset = 0;
 433         while (mbuf_avail != 0 || m->next != NULL) {
 434                 /* done with current buf, get the next one */
 435                 if (buf_avail == 0) {
 436                         vec_idx++;
 437                         if (unlikely(vec_idx >= nr_vec)) {
 438                                 error = -1;
 439                                 goto out;
 440                         }
 441
 442                         buf_addr = buf_vec[vec_idx].buf_addr;
 443                         buf_iova = buf_vec[vec_idx].buf_iova;
 444                         buf_len = buf_vec[vec_idx].buf_len;
 445
 446                         /* Prefetch next buffer address. */
 447                         if (vec_idx + 1 < nr_vec)
 448                                 rte_prefetch0((void *)(uintptr_t)
 449                                                 buf_vec[vec_idx + 1].buf_addr);
 450                         buf_offset = 0;
 451                         buf_avail  = buf_len;
 452                 }
 453
 454                 /* done with current mbuf, get the next one */
 455                 if (mbuf_avail == 0) {
 456                         m = m->next;
 457
 458                         mbuf_offset = 0;
 459                         mbuf_avail  = rte_pktmbuf_data_len(m);
 460                 }
 461
 462                 if (hdr_addr) {
 463                         virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
 464                         if (rxvq_is_mergeable(dev))
 465                                 ASSIGN_UNLESS_EQUAL(hdr->num_buffers,
 466                                                 num_buffers);
 467
 468                         if (unlikely(hdr == &tmp_hdr)) {
 469                                 uint64_t len;
 470                                 uint64_t remain = dev->vhost_hlen;
 471                                 uint64_t src = (uint64_t)(uintptr_t)hdr, dst;
 472                                 uint64_t iova = buf_vec[0].buf_iova;
 473                                 uint16_t hdr_vec_idx = 0;
 474
 475                                 while (remain) {
 476                                         len = remain;
 477                                         dst = buf_vec[hdr_vec_idx].buf_addr;
 478                                         rte_memcpy((void *)(uintptr_t)dst,
 479                                                         (void *)(uintptr_t)src,
 480                                                         len);
 481
 482                                         PRINT_PACKET(dev, (uintptr_t)dst,
 483                                                         (uint32_t)len, 0);
 484                                         vhost_log_cache_write(dev, vq,
 485                                                         iova, len);
 486
 487                                         remain -= len;
 488                                         iova += len;
 489                                         src += len;
 490                                         hdr_vec_idx++;
 491                                 }
 492                         } else {
 493                                 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
 494                                                 dev->vhost_hlen, 0);
 495                                 vhost_log_cache_write(dev, vq,
 496                                                 buf_vec[0].buf_iova,
 497                                                 dev->vhost_hlen);
 498                         }
 499
 500                         hdr_addr = 0;
 501                 }
 502
 503                 cpy_len = RTE_MIN(buf_len, mbuf_avail);
 504
 505                 if (likely(cpy_len > MAX_BATCH_LEN ||
 506                                         vq->batch_copy_nb_elems >= vq->size)) {
 507                         rte_memcpy((void *)((uintptr_t)(buf_addr + buf_offset)),
 508                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 509                                 cpy_len);
 510                         vhost_log_cache_write(dev, vq, buf_iova + buf_offset,
 511                                         cpy_len);
 512                         PRINT_PACKET(dev, (uintptr_t)(buf_addr + buf_offset),
 513                                 cpy_len, 0);
 514                 } else {
 515                         batch_copy[vq->batch_copy_nb_elems].dst =
 516                                 (void *)((uintptr_t)(buf_addr + buf_offset));
 517                         batch_copy[vq->batch_copy_nb_elems].src =
 518                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
 519                         batch_copy[vq->batch_copy_nb_elems].log_addr =
 520                                 buf_iova + buf_offset;
 521                         batch_copy[vq->batch_copy_nb_elems].len = cpy_len;
 522                         vq->batch_copy_nb_elems++;
 523                 }
 524
 525                 mbuf_avail  -= cpy_len;
 526                 mbuf_offset += cpy_len;
 527                 buf_avail  -= cpy_len;
 528                 buf_offset += cpy_len;
 529         }
 530
 531 out:
 532
 533         return error;
 534 }
 535
 536 static __rte_always_inline uint32_t
 537 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
 538         struct rte_mbuf **pkts, uint32_t count)
 539 {
 540         struct vhost_virtqueue *vq;
 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         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
 547         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
 548                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
 549                         dev->vid, __func__, queue_id);
 550                 return 0;
 551         }
 552
 553         vq = dev->virtqueue[queue_id];
 554
 555         rte_spinlock_lock(&vq->access_lock);
 556
 557         if (unlikely(vq->enabled == 0))
 558                 goto out_access_unlock;
 559
 560         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 561                 vhost_user_iotlb_rd_lock(vq);
 562
 563         if (unlikely(vq->access_ok == 0))
 564                 if (unlikely(vring_translate(dev, vq) < 0))
 565                         goto out;
 566
 567         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
 568         if (count == 0)
 569                 goto out;
 570
 571         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
 572
 573         avail_head = *((volatile uint16_t *)&vq->avail->idx);
 574         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
 575                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
 576                 uint16_t nr_vec = 0;
 577
 578                 if (unlikely(reserve_avail_buf(dev, vq,
 579                                                 pkt_len, buf_vec, &num_buffers,
 580                                                 avail_head, &nr_vec) < 0)) {
 581                         VHOST_LOG_DEBUG(VHOST_DATA,
 582                                 "(%d) failed to get enough desc from vring\n",
 583                                 dev->vid);
 584                         vq->shadow_used_idx -= num_buffers;
 585                         break;
 586                 }
 587
 588                 rte_prefetch0((void *)(uintptr_t)buf_vec[0].buf_addr);
 589
 590                 VHOST_LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
 591                         dev->vid, vq->last_avail_idx,
 592                         vq->last_avail_idx + num_buffers);
 593
 594                 if (copy_mbuf_to_desc(dev, vq, pkts[pkt_idx],
 595                                                 buf_vec, nr_vec,
 596                                                 num_buffers) < 0) {
 597                         vq->shadow_used_idx -= num_buffers;
 598                         break;
 599                 }
 600
 601                 vq->last_avail_idx += num_buffers;
 602         }
 603
 604         do_data_copy_enqueue(dev, vq);
 605
 606         if (likely(vq->shadow_used_idx)) {
 607                 flush_shadow_used_ring(dev, vq);
 608                 vhost_vring_call(dev, vq);
 609         }
 610
 611 out:
 612         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 613                 vhost_user_iotlb_rd_unlock(vq);
 614
 615 out_access_unlock:
 616         rte_spinlock_unlock(&vq->access_lock);
 617
 618         return pkt_idx;
 619 }
 620
 621 uint16_t
 622 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
 623         struct rte_mbuf **pkts, uint16_t count)
 624 {
 625         struct virtio_net *dev = get_device(vid);
 626
 627         if (!dev)
 628                 return 0;
 629
 630         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
 631                 RTE_LOG(ERR, VHOST_DATA,
 632                         "(%d) %s: built-in vhost net backend is disabled.\n",
 633                         dev->vid, __func__);
 634                 return 0;
 635         }
 636
 637         return virtio_dev_rx(dev, queue_id, pkts, count);
 638 }
 639
 640 static inline bool
 641 virtio_net_with_host_offload(struct virtio_net *dev)
 642 {
 643         if (dev->features &
 644                         ((1ULL << VIRTIO_NET_F_CSUM) |
 645                          (1ULL << VIRTIO_NET_F_HOST_ECN) |
 646                          (1ULL << VIRTIO_NET_F_HOST_TSO4) |
 647                          (1ULL << VIRTIO_NET_F_HOST_TSO6) |
 648                          (1ULL << VIRTIO_NET_F_HOST_UFO)))
 649                 return true;
 650
 651         return false;
 652 }
 653
 654 static void
 655 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
 656 {
 657         struct ipv4_hdr *ipv4_hdr;
 658         struct ipv6_hdr *ipv6_hdr;
 659         void *l3_hdr = NULL;
 660         struct ether_hdr *eth_hdr;
 661         uint16_t ethertype;
 662
 663         eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
 664
 665         m->l2_len = sizeof(struct ether_hdr);
 666         ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
 667
 668         if (ethertype == ETHER_TYPE_VLAN) {
 669                 struct vlan_hdr *vlan_hdr = (struct vlan_hdr *)(eth_hdr + 1);
 670
 671                 m->l2_len += sizeof(struct vlan_hdr);
 672                 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
 673         }
 674
 675         l3_hdr = (char *)eth_hdr + m->l2_len;
 676
 677         switch (ethertype) {
 678         case ETHER_TYPE_IPv4:
 679                 ipv4_hdr = l3_hdr;
 680                 *l4_proto = ipv4_hdr->next_proto_id;
 681                 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
 682                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 683                 m->ol_flags |= PKT_TX_IPV4;
 684                 break;
 685         case ETHER_TYPE_IPv6:
 686                 ipv6_hdr = l3_hdr;
 687                 *l4_proto = ipv6_hdr->proto;
 688                 m->l3_len = sizeof(struct ipv6_hdr);
 689                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 690                 m->ol_flags |= PKT_TX_IPV6;
 691                 break;
 692         default:
 693                 m->l3_len = 0;
 694                 *l4_proto = 0;
 695                 *l4_hdr = NULL;
 696                 break;
 697         }
 698 }
 699
 700 static __rte_always_inline void
 701 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
 702 {
 703         uint16_t l4_proto = 0;
 704         void *l4_hdr = NULL;
 705         struct tcp_hdr *tcp_hdr = NULL;
 706
 707         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
 708                 return;
 709
 710         parse_ethernet(m, &l4_proto, &l4_hdr);
 711         if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 712                 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
 713                         switch (hdr->csum_offset) {
 714                         case (offsetof(struct tcp_hdr, cksum)):
 715                                 if (l4_proto == IPPROTO_TCP)
 716                                         m->ol_flags |= PKT_TX_TCP_CKSUM;
 717                                 break;
 718                         case (offsetof(struct udp_hdr, dgram_cksum)):
 719                                 if (l4_proto == IPPROTO_UDP)
 720                                         m->ol_flags |= PKT_TX_UDP_CKSUM;
 721                                 break;
 722                         case (offsetof(struct sctp_hdr, cksum)):
 723                                 if (l4_proto == IPPROTO_SCTP)
 724                                         m->ol_flags |= PKT_TX_SCTP_CKSUM;
 725                                 break;
 726                         default:
 727                                 break;
 728                         }
 729                 }
 730         }
 731
 732         if (l4_hdr && hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 733                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 734                 case VIRTIO_NET_HDR_GSO_TCPV4:
 735                 case VIRTIO_NET_HDR_GSO_TCPV6:
 736                         tcp_hdr = l4_hdr;
 737                         m->ol_flags |= PKT_TX_TCP_SEG;
 738                         m->tso_segsz = hdr->gso_size;
 739                         m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
 740                         break;
 741                 case VIRTIO_NET_HDR_GSO_UDP:
 742                         m->ol_flags |= PKT_TX_UDP_SEG;
 743                         m->tso_segsz = hdr->gso_size;
 744                         m->l4_len = sizeof(struct udp_hdr);
 745                         break;
 746                 default:
 747                         RTE_LOG(WARNING, VHOST_DATA,
 748                                 "unsupported gso type %u.\n", hdr->gso_type);
 749                         break;
 750                 }
 751         }
 752 }
 753
 754 static __rte_always_inline void
 755 put_zmbuf(struct zcopy_mbuf *zmbuf)
 756 {
 757         zmbuf->in_use = 0;
 758 }
 759
 760 static __rte_always_inline int
 761 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 762                   struct buf_vector *buf_vec, uint16_t nr_vec,
 763                   struct rte_mbuf *m, struct rte_mempool *mbuf_pool)
 764 {
 765         uint32_t buf_avail, buf_offset;
 766         uint64_t buf_addr, buf_iova, buf_len;
 767         uint32_t mbuf_avail, mbuf_offset;
 768         uint32_t cpy_len;
 769         struct rte_mbuf *cur = m, *prev = m;
 770         struct virtio_net_hdr tmp_hdr;
 771         struct virtio_net_hdr *hdr = NULL;
 772         /* A counter to avoid desc dead loop chain */
 773         uint16_t vec_idx = 0;
 774         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 775         int error = 0;
 776
 777         buf_addr = buf_vec[vec_idx].buf_addr;
 778         buf_iova = buf_vec[vec_idx].buf_iova;
 779         buf_len = buf_vec[vec_idx].buf_len;
 780
 781         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
 782                 error = -1;
 783                 goto out;
 784         }
 785
 786         if (likely(nr_vec > 1))
 787                 rte_prefetch0((void *)(uintptr_t)buf_vec[1].buf_addr);
 788
 789         if (virtio_net_with_host_offload(dev)) {
 790                 if (unlikely(buf_len < sizeof(struct virtio_net_hdr))) {
 791                         uint64_t len;
 792                         uint64_t remain = sizeof(struct virtio_net_hdr);
 793                         uint64_t src;
 794                         uint64_t dst = (uint64_t)(uintptr_t)&tmp_hdr;
 795                         uint16_t hdr_vec_idx = 0;
 796
 797                         /*
 798                          * No luck, the virtio-net header doesn't fit
 799                          * in a contiguous virtual area.
 800                          */
 801                         while (remain) {
 802                                 len = remain;
 803                                 src = buf_vec[hdr_vec_idx].buf_addr;
 804                                 rte_memcpy((void *)(uintptr_t)dst,
 805                                                    (void *)(uintptr_t)src, len);
 806
 807                                 remain -= len;
 808                                 dst += len;
 809                                 hdr_vec_idx++;
 810                         }
 811
 812                         hdr = &tmp_hdr;
 813                 } else {
 814                         hdr = (struct virtio_net_hdr *)((uintptr_t)buf_addr);
 815                         rte_prefetch0(hdr);
 816                 }
 817         }
 818
 819         /*
 820          * A virtio driver normally uses at least 2 desc buffers
 821          * for Tx: the first for storing the header, and others
 822          * for storing the data.
 823          */
 824         if (unlikely(buf_len < dev->vhost_hlen)) {
 825                 buf_offset = dev->vhost_hlen - buf_len;
 826                 vec_idx++;
 827                 buf_addr = buf_vec[vec_idx].buf_addr;
 828                 buf_iova = buf_vec[vec_idx].buf_iova;
 829                 buf_len = buf_vec[vec_idx].buf_len;
 830                 buf_avail  = buf_len - buf_offset;
 831         } else if (buf_len == dev->vhost_hlen) {
 832                 if (unlikely(++vec_idx >= nr_vec))
 833                         goto out;
 834                 buf_addr = buf_vec[vec_idx].buf_addr;
 835                 buf_iova = buf_vec[vec_idx].buf_iova;
 836                 buf_len = buf_vec[vec_idx].buf_len;
 837
 838                 buf_offset = 0;
 839                 buf_avail = buf_len;
 840         } else {
 841                 buf_offset = dev->vhost_hlen;
 842                 buf_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
 843         }
 844
 845         rte_prefetch0((void *)(uintptr_t)
 846                         (buf_addr + buf_offset));
 847
 848         PRINT_PACKET(dev,
 849                         (uintptr_t)(buf_addr + buf_offset),
 850                         (uint32_t)buf_avail, 0);
 851
 852         mbuf_offset = 0;
 853         mbuf_avail  = m->buf_len - RTE_PKTMBUF_HEADROOM;
 854         while (1) {
 855                 uint64_t hpa;
 856
 857                 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
 858
 859                 /*
 860                  * A desc buf might across two host physical pages that are
 861                  * not continuous. In such case (gpa_to_hpa returns 0), data
 862                  * will be copied even though zero copy is enabled.
 863                  */
 864                 if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
 865                                         buf_iova + buf_offset, cpy_len)))) {
 866                         cur->data_len = cpy_len;
 867                         cur->data_off = 0;
 868                         cur->buf_addr =
 869                                 (void *)(uintptr_t)(buf_addr + buf_offset);
 870                         cur->buf_iova = hpa;
 871
 872                         /*
 873                          * In zero copy mode, one mbuf can only reference data
 874                          * for one or partial of one desc buff.
 875                          */
 876                         mbuf_avail = cpy_len;
 877                 } else {
 878                         if (likely(cpy_len > MAX_BATCH_LEN ||
 879                                    vq->batch_copy_nb_elems >= vq->size ||
 880                                    (hdr && cur == m))) {
 881                                 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
 882                                                                    mbuf_offset),
 883                                            (void *)((uintptr_t)(buf_addr +
 884                                                            buf_offset)),
 885                                            cpy_len);
 886                         } else {
 887                                 batch_copy[vq->batch_copy_nb_elems].dst =
 888                                         rte_pktmbuf_mtod_offset(cur, void *,
 889                                                                 mbuf_offset);
 890                                 batch_copy[vq->batch_copy_nb_elems].src =
 891                                         (void *)((uintptr_t)(buf_addr +
 892                                                                 buf_offset));
 893                                 batch_copy[vq->batch_copy_nb_elems].len =
 894                                         cpy_len;
 895                                 vq->batch_copy_nb_elems++;
 896                         }
 897                 }
 898
 899                 mbuf_avail  -= cpy_len;
 900                 mbuf_offset += cpy_len;
 901                 buf_avail -= cpy_len;
 902                 buf_offset += cpy_len;
 903
 904                 /* This buf reaches to its end, get the next one */
 905                 if (buf_avail == 0) {
 906                         if (++vec_idx >= nr_vec)
 907                                 break;
 908
 909                         buf_addr = buf_vec[vec_idx].buf_addr;
 910                         buf_iova = buf_vec[vec_idx].buf_iova;
 911                         buf_len = buf_vec[vec_idx].buf_len;
 912
 913                         /*
 914                          * Prefecth desc n + 1 buffer while
 915                          * desc n buffer is processed.
 916                          */
 917                         if (vec_idx + 1 < nr_vec)
 918                                 rte_prefetch0((void *)(uintptr_t)
 919                                                 buf_vec[vec_idx + 1].buf_addr);
 920
 921                         buf_offset = 0;
 922                         buf_avail  = buf_len;
 923
 924                         PRINT_PACKET(dev, (uintptr_t)buf_addr,
 925                                         (uint32_t)buf_avail, 0);
 926                 }
 927
 928                 /*
 929                  * This mbuf reaches to its end, get a new one
 930                  * to hold more data.
 931                  */
 932                 if (mbuf_avail == 0) {
 933                         cur = rte_pktmbuf_alloc(mbuf_pool);
 934                         if (unlikely(cur == NULL)) {
 935                                 RTE_LOG(ERR, VHOST_DATA, "Failed to "
 936                                         "allocate memory for mbuf.\n");
 937                                 error = -1;
 938                                 goto out;
 939                         }
 940                         if (unlikely(dev->dequeue_zero_copy))
 941                                 rte_mbuf_refcnt_update(cur, 1);
 942
 943                         prev->next = cur;
 944                         prev->data_len = mbuf_offset;
 945                         m->nb_segs += 1;
 946                         m->pkt_len += mbuf_offset;
 947                         prev = cur;
 948
 949                         mbuf_offset = 0;
 950                         mbuf_avail  = cur->buf_len - RTE_PKTMBUF_HEADROOM;
 951                 }
 952         }
 953
 954         prev->data_len = mbuf_offset;
 955         m->pkt_len    += mbuf_offset;
 956
 957         if (hdr)
 958                 vhost_dequeue_offload(hdr, m);
 959
 960 out:
 961
 962         return error;
 963 }
 964
 965 static __rte_always_inline struct zcopy_mbuf *
 966 get_zmbuf(struct vhost_virtqueue *vq)
 967 {
 968         uint16_t i;
 969         uint16_t last;
 970         int tries = 0;
 971
 972         /* search [last_zmbuf_idx, zmbuf_size) */
 973         i = vq->last_zmbuf_idx;
 974         last = vq->zmbuf_size;
 975
 976 again:
 977         for (; i < last; i++) {
 978                 if (vq->zmbufs[i].in_use == 0) {
 979                         vq->last_zmbuf_idx = i + 1;
 980                         vq->zmbufs[i].in_use = 1;
 981                         return &vq->zmbufs[i];
 982                 }
 983         }
 984
 985         tries++;
 986         if (tries == 1) {
 987                 /* search [0, last_zmbuf_idx) */
 988                 i = 0;
 989                 last = vq->last_zmbuf_idx;
 990                 goto again;
 991         }
 992
 993         return NULL;
 994 }
 995
 996 static __rte_always_inline bool
 997 mbuf_is_consumed(struct rte_mbuf *m)
 998 {
 999         while (m) {
1000                 if (rte_mbuf_refcnt_read(m) > 1)
1001                         return false;
1002                 m = m->next;
1003         }
1004
1005         return true;
1006 }
1007
1008 static __rte_always_inline void
1009 restore_mbuf(struct rte_mbuf *m)
1010 {
1011         uint32_t mbuf_size, priv_size;
1012
1013         while (m) {
1014                 priv_size = rte_pktmbuf_priv_size(m->pool);
1015                 mbuf_size = sizeof(struct rte_mbuf) + priv_size;
1016                 /* start of buffer is after mbuf structure and priv data */
1017
1018                 m->buf_addr = (char *)m + mbuf_size;
1019                 m->buf_iova = rte_mempool_virt2iova(m) + mbuf_size;
1020                 m = m->next;
1021         }
1022 }
1023
1024 uint16_t
1025 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
1026         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1027 {
1028         struct virtio_net *dev;
1029         struct rte_mbuf *rarp_mbuf = NULL;
1030         struct vhost_virtqueue *vq;
1031         uint32_t i = 0;
1032         uint16_t free_entries;
1033
1034         dev = get_device(vid);
1035         if (!dev)
1036                 return 0;
1037
1038         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
1039                 RTE_LOG(ERR, VHOST_DATA,
1040                         "(%d) %s: built-in vhost net backend is disabled.\n",
1041                         dev->vid, __func__);
1042                 return 0;
1043         }
1044
1045         if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
1046                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
1047                         dev->vid, __func__, queue_id);
1048                 return 0;
1049         }
1050
1051         vq = dev->virtqueue[queue_id];
1052
1053         if (unlikely(rte_spinlock_trylock(&vq->access_lock) == 0))
1054                 return 0;
1055
1056         if (unlikely(vq->enabled == 0))
1057                 goto out_access_unlock;
1058
1059         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1060                 vhost_user_iotlb_rd_lock(vq);
1061
1062         if (unlikely(vq->access_ok == 0))
1063                 if (unlikely(vring_translate(dev, vq) < 0))
1064                         goto out;
1065
1066         if (unlikely(dev->dequeue_zero_copy)) {
1067                 struct zcopy_mbuf *zmbuf, *next;
1068                 int nr_updated = 0;
1069
1070                 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1071                      zmbuf != NULL; zmbuf = next) {
1072                         next = TAILQ_NEXT(zmbuf, next);
1073
1074                         if (mbuf_is_consumed(zmbuf->mbuf)) {
1075                                 update_shadow_used_ring(vq, zmbuf->desc_idx, 0);
1076                                 nr_updated += 1;
1077
1078                                 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1079                                 restore_mbuf(zmbuf->mbuf);
1080                                 rte_pktmbuf_free(zmbuf->mbuf);
1081                                 put_zmbuf(zmbuf);
1082                                 vq->nr_zmbuf -= 1;
1083                         }
1084                 }
1085
1086                 flush_shadow_used_ring(dev, vq);
1087                 vhost_vring_call(dev, vq);
1088         }
1089
1090         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
1091
1092         /*
1093          * Construct a RARP broadcast packet, and inject it to the "pkts"
1094          * array, to looks like that guest actually send such packet.
1095          *
1096          * Check user_send_rarp() for more information.
1097          *
1098          * broadcast_rarp shares a cacheline in the virtio_net structure
1099          * with some fields that are accessed during enqueue and
1100          * rte_atomic16_cmpset() causes a write if using cmpxchg. This could
1101          * result in false sharing between enqueue and dequeue.
1102          *
1103          * Prevent unnecessary false sharing by reading broadcast_rarp first
1104          * and only performing cmpset if the read indicates it is likely to
1105          * be set.
1106          */
1107
1108         if (unlikely(rte_atomic16_read(&dev->broadcast_rarp) &&
1109                         rte_atomic16_cmpset((volatile uint16_t *)
1110                                 &dev->broadcast_rarp.cnt, 1, 0))) {
1111
1112                 rarp_mbuf = rte_net_make_rarp_packet(mbuf_pool, &dev->mac);
1113                 if (rarp_mbuf == NULL) {
1114                         RTE_LOG(ERR, VHOST_DATA,
1115                                 "Failed to make RARP packet.\n");
1116                         return 0;
1117                 }
1118                 count -= 1;
1119         }
1120
1121         free_entries = *((volatile uint16_t *)&vq->avail->idx) -
1122                         vq->last_avail_idx;
1123         if (free_entries == 0)
1124                 goto out;
1125
1126         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1127
1128         count = RTE_MIN(count, MAX_PKT_BURST);
1129         count = RTE_MIN(count, free_entries);
1130         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1131                         dev->vid, count);
1132
1133         for (i = 0; i < count; i++) {
1134                 struct buf_vector buf_vec[BUF_VECTOR_MAX];
1135                 uint16_t head_idx, dummy_len;
1136                 uint32_t nr_vec = 0;
1137                 int err;
1138
1139                 if (unlikely(fill_vec_buf(dev, vq,
1140                                                 vq->last_avail_idx + i,
1141                                                 &nr_vec, buf_vec,
1142                                                 &head_idx, &dummy_len,
1143                                                 VHOST_ACCESS_RO) < 0))
1144                         break;
1145
1146                 if (likely(dev->dequeue_zero_copy == 0))
1147                         update_shadow_used_ring(vq, head_idx, 0);
1148
1149                 rte_prefetch0((void *)(uintptr_t)buf_vec[0].buf_addr);
1150
1151                 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
1152                 if (unlikely(pkts[i] == NULL)) {
1153                         RTE_LOG(ERR, VHOST_DATA,
1154                                 "Failed to allocate memory for mbuf.\n");
1155                         break;
1156                 }
1157
1158                 err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts[i],
1159                                 mbuf_pool);
1160                 if (unlikely(err)) {
1161                         rte_pktmbuf_free(pkts[i]);
1162                         break;
1163                 }
1164
1165                 if (unlikely(dev->dequeue_zero_copy)) {
1166                         struct zcopy_mbuf *zmbuf;
1167
1168                         zmbuf = get_zmbuf(vq);
1169                         if (!zmbuf) {
1170                                 rte_pktmbuf_free(pkts[i]);
1171                                 break;
1172                         }
1173                         zmbuf->mbuf = pkts[i];
1174                         zmbuf->desc_idx = head_idx;
1175
1176                         /*
1177                          * Pin lock the mbuf; we will check later to see
1178                          * whether the mbuf is freed (when we are the last
1179                          * user) or not. If that's the case, we then could
1180                          * update the used ring safely.
1181                          */
1182                         rte_mbuf_refcnt_update(pkts[i], 1);
1183
1184                         vq->nr_zmbuf += 1;
1185                         TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1186                 }
1187         }
1188         vq->last_avail_idx += i;
1189
1190         if (likely(dev->dequeue_zero_copy == 0)) {
1191                 do_data_copy_dequeue(vq);
1192                 if (unlikely(i < count))
1193                         vq->shadow_used_idx = i;
1194                 flush_shadow_used_ring(dev, vq);
1195                 vhost_vring_call(dev, vq);
1196         }
1197
1198 out:
1199         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1200                 vhost_user_iotlb_rd_unlock(vq);
1201
1202 out_access_unlock:
1203         rte_spinlock_unlock(&vq->access_lock);
1204
1205         if (unlikely(rarp_mbuf != NULL)) {
1206                 /*
1207                  * Inject it to the head of "pkts" array, so that switch's mac
1208                  * learning table will get updated first.
1209                  */
1210                 memmove(&pkts[1], pkts, i * sizeof(struct rte_mbuf *));
1211                 pkts[0] = rarp_mbuf;
1212                 i += 1;
1213         }
1214
1215         return i;
1216 }