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