lib/librte_vhost/virtio_net.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2016 Intel Corporation. All rights reserved.
   5  *   All rights reserved.
   6  *
   7  *   Redistribution and use in source and binary forms, with or without
   8  *   modification, are permitted provided that the following conditions
   9  *   are met:
  10  *
  11  *     * Redistributions of source code must retain the above copyright
  12  *       notice, this list of conditions and the following disclaimer.
  13  *     * Redistributions in binary form must reproduce the above copyright
  14  *       notice, this list of conditions and the following disclaimer in
  15  *       the documentation and/or other materials provided with the
  16  *       distribution.
  17  *     * Neither the name of Intel Corporation nor the names of its
  18  *       contributors may be used to endorse or promote products derived
  19  *       from this software without specific prior written permission.
  20  *
  21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  32  */
  33
  34 #include <stdint.h>
  35 #include <stdbool.h>
  36 #include <linux/virtio_net.h>
  37
  38 #include <rte_mbuf.h>
  39 #include <rte_memcpy.h>
  40 #include <rte_ether.h>
  41 #include <rte_ip.h>
  42 #include <rte_virtio_net.h>
  43 #include <rte_tcp.h>
  44 #include <rte_udp.h>
  45 #include <rte_sctp.h>
  46 #include <rte_arp.h>
  47
  48 #include "vhost.h"
  49
  50 #define MAX_PKT_BURST 32
  51 #define VHOST_LOG_PAGE  4096
  52
  53 static inline void __attribute__((always_inline))
  54 vhost_log_page(uint8_t *log_base, uint64_t page)
  55 {
  56         log_base[page / 8] |= 1 << (page % 8);
  57 }
  58
  59 static inline void __attribute__((always_inline))
  60 vhost_log_write(struct virtio_net *dev, uint64_t addr, uint64_t len)
  61 {
  62         uint64_t page;
  63
  64         if (likely(((dev->features & (1ULL << VHOST_F_LOG_ALL)) == 0) ||
  65                    !dev->log_base || !len))
  66                 return;
  67
  68         if (unlikely(dev->log_size <= ((addr + len - 1) / VHOST_LOG_PAGE / 8)))
  69                 return;
  70
  71         /* To make sure guest memory updates are committed before logging */
  72         rte_smp_wmb();
  73
  74         page = addr / VHOST_LOG_PAGE;
  75         while (page * VHOST_LOG_PAGE < addr + len) {
  76                 vhost_log_page((uint8_t *)(uintptr_t)dev->log_base, page);
  77                 page += 1;
  78         }
  79 }
  80
  81 static inline void __attribute__((always_inline))
  82 vhost_log_used_vring(struct virtio_net *dev, struct vhost_virtqueue *vq,
  83                      uint64_t offset, uint64_t len)
  84 {
  85         vhost_log_write(dev, vq->log_guest_addr + offset, len);
  86 }
  87
  88 static bool
  89 is_valid_virt_queue_idx(uint32_t idx, int is_tx, uint32_t qp_nb)
  90 {
  91         return (is_tx ^ (idx & 1)) == 0 && idx < qp_nb * VIRTIO_QNUM;
  92 }
  93
  94 static inline void __attribute__((always_inline))
  95 do_flush_shadow_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq,
  96                           uint16_t to, uint16_t from, uint16_t size)
  97 {
  98         rte_memcpy(&vq->used->ring[to],
  99                         &vq->shadow_used_ring[from],
 100                         size * sizeof(struct vring_used_elem));
 101         vhost_log_used_vring(dev, vq,
 102                         offsetof(struct vring_used, ring[to]),
 103                         size * sizeof(struct vring_used_elem));
 104 }
 105
 106 static inline void __attribute__((always_inline))
 107 flush_shadow_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq)
 108 {
 109         uint16_t used_idx = vq->last_used_idx & (vq->size - 1);
 110
 111         if (used_idx + vq->shadow_used_idx <= vq->size) {
 112                 do_flush_shadow_used_ring(dev, vq, used_idx, 0,
 113                                           vq->shadow_used_idx);
 114         } else {
 115                 uint16_t size;
 116
 117                 /* update used ring interval [used_idx, vq->size] */
 118                 size = vq->size - used_idx;
 119                 do_flush_shadow_used_ring(dev, vq, used_idx, 0, size);
 120
 121                 /* update the left half used ring interval [0, left_size] */
 122                 do_flush_shadow_used_ring(dev, vq, 0, size,
 123                                           vq->shadow_used_idx - size);
 124         }
 125         vq->last_used_idx += vq->shadow_used_idx;
 126
 127         rte_smp_wmb();
 128
 129         *(volatile uint16_t *)&vq->used->idx += vq->shadow_used_idx;
 130         vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
 131                 sizeof(vq->used->idx));
 132 }
 133
 134 static inline void __attribute__((always_inline))
 135 update_shadow_used_ring(struct vhost_virtqueue *vq,
 136                          uint16_t desc_idx, uint16_t len)
 137 {
 138         uint16_t i = vq->shadow_used_idx++;
 139
 140         vq->shadow_used_ring[i].id  = desc_idx;
 141         vq->shadow_used_ring[i].len = len;
 142 }
 143
 144 static void
 145 virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
 146 {
 147         if (m_buf->ol_flags & PKT_TX_L4_MASK) {
 148                 net_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 149                 net_hdr->csum_start = m_buf->l2_len + m_buf->l3_len;
 150
 151                 switch (m_buf->ol_flags & PKT_TX_L4_MASK) {
 152                 case PKT_TX_TCP_CKSUM:
 153                         net_hdr->csum_offset = (offsetof(struct tcp_hdr,
 154                                                 cksum));
 155                         break;
 156                 case PKT_TX_UDP_CKSUM:
 157                         net_hdr->csum_offset = (offsetof(struct udp_hdr,
 158                                                 dgram_cksum));
 159                         break;
 160                 case PKT_TX_SCTP_CKSUM:
 161                         net_hdr->csum_offset = (offsetof(struct sctp_hdr,
 162                                                 cksum));
 163                         break;
 164                 }
 165         }
 166
 167         if (m_buf->ol_flags & PKT_TX_TCP_SEG) {
 168                 if (m_buf->ol_flags & PKT_TX_IPV4)
 169                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 170                 else
 171                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 172                 net_hdr->gso_size = m_buf->tso_segsz;
 173                 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len
 174                                         + m_buf->l4_len;
 175         }
 176 }
 177
 178 static inline void
 179 copy_virtio_net_hdr(struct virtio_net *dev, uint64_t desc_addr,
 180                     struct virtio_net_hdr_mrg_rxbuf hdr)
 181 {
 182         if (dev->vhost_hlen == sizeof(struct virtio_net_hdr_mrg_rxbuf))
 183                 *(struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)desc_addr = hdr;
 184         else
 185                 *(struct virtio_net_hdr *)(uintptr_t)desc_addr = hdr.hdr;
 186 }
 187
 188 static inline int __attribute__((always_inline))
 189 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 190                   struct rte_mbuf *m, uint16_t desc_idx)
 191 {
 192         uint32_t desc_avail, desc_offset;
 193         uint32_t mbuf_avail, mbuf_offset;
 194         uint32_t cpy_len;
 195         struct vring_desc *desc;
 196         uint64_t desc_addr;
 197         struct virtio_net_hdr_mrg_rxbuf virtio_hdr = {{0, 0, 0, 0, 0, 0}, 0};
 198
 199         desc = &vq->desc[desc_idx];
 200         desc_addr = gpa_to_vva(dev, desc->addr);
 201         /*
 202          * Checking of 'desc_addr' placed outside of 'unlikely' macro to avoid
 203          * performance issue with some versions of gcc (4.8.4 and 5.3.0) which
 204          * otherwise stores offset on the stack instead of in a register.
 205          */
 206         if (unlikely(desc->len < dev->vhost_hlen) || !desc_addr)
 207                 return -1;
 208
 209         rte_prefetch0((void *)(uintptr_t)desc_addr);
 210
 211         virtio_enqueue_offload(m, &virtio_hdr.hdr);
 212         copy_virtio_net_hdr(dev, desc_addr, virtio_hdr);
 213         vhost_log_write(dev, desc->addr, dev->vhost_hlen);
 214         PRINT_PACKET(dev, (uintptr_t)desc_addr, dev->vhost_hlen, 0);
 215
 216         desc_offset = dev->vhost_hlen;
 217         desc_avail  = desc->len - dev->vhost_hlen;
 218
 219         mbuf_avail  = rte_pktmbuf_data_len(m);
 220         mbuf_offset = 0;
 221         while (mbuf_avail != 0 || m->next != NULL) {
 222                 /* done with current mbuf, fetch next */
 223                 if (mbuf_avail == 0) {
 224                         m = m->next;
 225
 226                         mbuf_offset = 0;
 227                         mbuf_avail  = rte_pktmbuf_data_len(m);
 228                 }
 229
 230                 /* done with current desc buf, fetch next */
 231                 if (desc_avail == 0) {
 232                         if ((desc->flags & VRING_DESC_F_NEXT) == 0) {
 233                                 /* Room in vring buffer is not enough */
 234                                 return -1;
 235                         }
 236                         if (unlikely(desc->next >= vq->size))
 237                                 return -1;
 238
 239                         desc = &vq->desc[desc->next];
 240                         desc_addr = gpa_to_vva(dev, desc->addr);
 241                         if (unlikely(!desc_addr))
 242                                 return -1;
 243
 244                         desc_offset = 0;
 245                         desc_avail  = desc->len;
 246                 }
 247
 248                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 249                 rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)),
 250                         rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 251                         cpy_len);
 252                 vhost_log_write(dev, desc->addr + desc_offset, cpy_len);
 253                 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
 254                              cpy_len, 0);
 255
 256                 mbuf_avail  -= cpy_len;
 257                 mbuf_offset += cpy_len;
 258                 desc_avail  -= cpy_len;
 259                 desc_offset += cpy_len;
 260         }
 261
 262         return 0;
 263 }
 264
 265 /**
 266  * This function adds buffers to the virtio devices RX virtqueue. Buffers can
 267  * be received from the physical port or from another virtio device. A packet
 268  * count is returned to indicate the number of packets that are succesfully
 269  * added to the RX queue. This function works when the mbuf is scattered, but
 270  * it doesn't support the mergeable feature.
 271  */
 272 static inline uint32_t __attribute__((always_inline))
 273 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
 274               struct rte_mbuf **pkts, uint32_t count)
 275 {
 276         struct vhost_virtqueue *vq;
 277         uint16_t avail_idx, free_entries, start_idx;
 278         uint16_t desc_indexes[MAX_PKT_BURST];
 279         uint16_t used_idx;
 280         uint32_t i;
 281
 282         LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
 283         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->virt_qp_nb))) {
 284                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
 285                         dev->vid, __func__, queue_id);
 286                 return 0;
 287         }
 288
 289         vq = dev->virtqueue[queue_id];
 290         if (unlikely(vq->enabled == 0))
 291                 return 0;
 292
 293         avail_idx = *((volatile uint16_t *)&vq->avail->idx);
 294         start_idx = vq->last_used_idx;
 295         free_entries = avail_idx - start_idx;
 296         count = RTE_MIN(count, free_entries);
 297         count = RTE_MIN(count, (uint32_t)MAX_PKT_BURST);
 298         if (count == 0)
 299                 return 0;
 300
 301         LOG_DEBUG(VHOST_DATA, "(%d) start_idx %d | end_idx %d\n",
 302                 dev->vid, start_idx, start_idx + count);
 303
 304         /* Retrieve all of the desc indexes first to avoid caching issues. */
 305         rte_prefetch0(&vq->avail->ring[start_idx & (vq->size - 1)]);
 306         for (i = 0; i < count; i++) {
 307                 used_idx = (start_idx + i) & (vq->size - 1);
 308                 desc_indexes[i] = vq->avail->ring[used_idx];
 309                 vq->used->ring[used_idx].id = desc_indexes[i];
 310                 vq->used->ring[used_idx].len = pkts[i]->pkt_len +
 311                                                dev->vhost_hlen;
 312                 vhost_log_used_vring(dev, vq,
 313                         offsetof(struct vring_used, ring[used_idx]),
 314                         sizeof(vq->used->ring[used_idx]));
 315         }
 316
 317         rte_prefetch0(&vq->desc[desc_indexes[0]]);
 318         for (i = 0; i < count; i++) {
 319                 uint16_t desc_idx = desc_indexes[i];
 320                 int err;
 321
 322                 err = copy_mbuf_to_desc(dev, vq, pkts[i], desc_idx);
 323                 if (unlikely(err)) {
 324                         used_idx = (start_idx + i) & (vq->size - 1);
 325                         vq->used->ring[used_idx].len = dev->vhost_hlen;
 326                         vhost_log_used_vring(dev, vq,
 327                                 offsetof(struct vring_used, ring[used_idx]),
 328                                 sizeof(vq->used->ring[used_idx]));
 329                 }
 330
 331                 if (i + 1 < count)
 332                         rte_prefetch0(&vq->desc[desc_indexes[i+1]]);
 333         }
 334
 335         rte_smp_wmb();
 336
 337         *(volatile uint16_t *)&vq->used->idx += count;
 338         vq->last_used_idx += count;
 339         vhost_log_used_vring(dev, vq,
 340                 offsetof(struct vring_used, idx),
 341                 sizeof(vq->used->idx));
 342
 343         /* flush used->idx update before we read avail->flags. */
 344         rte_mb();
 345
 346         /* Kick the guest if necessary. */
 347         if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
 348                         && (vq->callfd >= 0))
 349                 eventfd_write(vq->callfd, (eventfd_t)1);
 350         return count;
 351 }
 352
 353 static inline int __attribute__((always_inline))
 354 fill_vec_buf(struct vhost_virtqueue *vq, uint32_t avail_idx,
 355              uint32_t *vec_idx, struct buf_vector *buf_vec,
 356              uint16_t *desc_chain_head, uint16_t *desc_chain_len)
 357 {
 358         uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
 359         uint32_t vec_id = *vec_idx;
 360         uint32_t len    = 0;
 361
 362         *desc_chain_head = idx;
 363         while (1) {
 364                 if (unlikely(vec_id >= BUF_VECTOR_MAX || idx >= vq->size))
 365                         return -1;
 366
 367                 len += vq->desc[idx].len;
 368                 buf_vec[vec_id].buf_addr = vq->desc[idx].addr;
 369                 buf_vec[vec_id].buf_len  = vq->desc[idx].len;
 370                 buf_vec[vec_id].desc_idx = idx;
 371                 vec_id++;
 372
 373                 if ((vq->desc[idx].flags & VRING_DESC_F_NEXT) == 0)
 374                         break;
 375
 376                 idx = vq->desc[idx].next;
 377         }
 378
 379         *desc_chain_len = len;
 380         *vec_idx = vec_id;
 381
 382         return 0;
 383 }
 384
 385 /*
 386  * Returns -1 on fail, 0 on success
 387  */
 388 static inline int
 389 reserve_avail_buf_mergeable(struct vhost_virtqueue *vq, uint32_t size,
 390                             struct buf_vector *buf_vec, uint16_t *num_buffers)
 391 {
 392         uint16_t cur_idx;
 393         uint16_t avail_idx;
 394         uint32_t vec_idx = 0;
 395         uint16_t tries = 0;
 396
 397         uint16_t head_idx = 0;
 398         uint16_t len = 0;
 399
 400         *num_buffers = 0;
 401         cur_idx  = vq->last_avail_idx;
 402
 403         while (size > 0) {
 404                 avail_idx = *((volatile uint16_t *)&vq->avail->idx);
 405                 if (unlikely(cur_idx == avail_idx))
 406                         return -1;
 407
 408                 if (unlikely(fill_vec_buf(vq, cur_idx, &vec_idx, buf_vec,
 409                                           &head_idx, &len) < 0))
 410                         return -1;
 411                 len = RTE_MIN(len, size);
 412                 update_shadow_used_ring(vq, head_idx, len);
 413                 size -= len;
 414
 415                 cur_idx++;
 416                 tries++;
 417                 *num_buffers += 1;
 418
 419                 /*
 420                  * if we tried all available ring items, and still
 421                  * can't get enough buf, it means something abnormal
 422                  * happened.
 423                  */
 424                 if (unlikely(tries >= vq->size))
 425                         return -1;
 426         }
 427
 428         return 0;
 429 }
 430
 431 static inline int __attribute__((always_inline))
 432 copy_mbuf_to_desc_mergeable(struct virtio_net *dev, struct rte_mbuf *m,
 433                             struct buf_vector *buf_vec, uint16_t num_buffers)
 434 {
 435         struct virtio_net_hdr_mrg_rxbuf virtio_hdr = {{0, 0, 0, 0, 0, 0}, 0};
 436         uint32_t vec_idx = 0;
 437         uint64_t desc_addr;
 438         uint32_t mbuf_offset, mbuf_avail;
 439         uint32_t desc_offset, desc_avail;
 440         uint32_t cpy_len;
 441         uint64_t hdr_addr, hdr_phys_addr;
 442         struct rte_mbuf *hdr_mbuf;
 443
 444         if (unlikely(m == NULL))
 445                 return -1;
 446
 447         desc_addr = gpa_to_vva(dev, buf_vec[vec_idx].buf_addr);
 448         if (buf_vec[vec_idx].buf_len < dev->vhost_hlen || !desc_addr)
 449                 return -1;
 450
 451         hdr_mbuf = m;
 452         hdr_addr = desc_addr;
 453         hdr_phys_addr = buf_vec[vec_idx].buf_addr;
 454         rte_prefetch0((void *)(uintptr_t)hdr_addr);
 455
 456         virtio_hdr.num_buffers = num_buffers;
 457         LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
 458                 dev->vid, num_buffers);
 459
 460         desc_avail  = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
 461         desc_offset = dev->vhost_hlen;
 462
 463         mbuf_avail  = rte_pktmbuf_data_len(m);
 464         mbuf_offset = 0;
 465         while (mbuf_avail != 0 || m->next != NULL) {
 466                 /* done with current desc buf, get the next one */
 467                 if (desc_avail == 0) {
 468                         vec_idx++;
 469                         desc_addr = gpa_to_vva(dev, buf_vec[vec_idx].buf_addr);
 470                         if (unlikely(!desc_addr))
 471                                 return -1;
 472
 473                         /* Prefetch buffer address. */
 474                         rte_prefetch0((void *)(uintptr_t)desc_addr);
 475                         desc_offset = 0;
 476                         desc_avail  = buf_vec[vec_idx].buf_len;
 477                 }
 478
 479                 /* done with current mbuf, get the next one */
 480                 if (mbuf_avail == 0) {
 481                         m = m->next;
 482
 483                         mbuf_offset = 0;
 484                         mbuf_avail  = rte_pktmbuf_data_len(m);
 485                 }
 486
 487                 if (hdr_addr) {
 488                         virtio_enqueue_offload(hdr_mbuf, &virtio_hdr.hdr);
 489                         copy_virtio_net_hdr(dev, hdr_addr, virtio_hdr);
 490                         vhost_log_write(dev, hdr_phys_addr, dev->vhost_hlen);
 491                         PRINT_PACKET(dev, (uintptr_t)hdr_addr,
 492                                      dev->vhost_hlen, 0);
 493
 494                         hdr_addr = 0;
 495                 }
 496
 497                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 498                 rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)),
 499                         rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 500                         cpy_len);
 501                 vhost_log_write(dev, buf_vec[vec_idx].buf_addr + desc_offset,
 502                         cpy_len);
 503                 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
 504                         cpy_len, 0);
 505
 506                 mbuf_avail  -= cpy_len;
 507                 mbuf_offset += cpy_len;
 508                 desc_avail  -= cpy_len;
 509                 desc_offset += cpy_len;
 510         }
 511
 512         return 0;
 513 }
 514
 515 static inline uint32_t __attribute__((always_inline))
 516 virtio_dev_merge_rx(struct virtio_net *dev, uint16_t queue_id,
 517         struct rte_mbuf **pkts, uint32_t count)
 518 {
 519         struct vhost_virtqueue *vq;
 520         uint32_t pkt_idx = 0;
 521         uint16_t num_buffers;
 522         struct buf_vector buf_vec[BUF_VECTOR_MAX];
 523
 524         LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
 525         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->virt_qp_nb))) {
 526                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
 527                         dev->vid, __func__, queue_id);
 528                 return 0;
 529         }
 530
 531         vq = dev->virtqueue[queue_id];
 532         if (unlikely(vq->enabled == 0))
 533                 return 0;
 534
 535         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
 536         if (count == 0)
 537                 return 0;
 538
 539         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
 540
 541         vq->shadow_used_idx = 0;
 542         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
 543                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
 544
 545                 if (unlikely(reserve_avail_buf_mergeable(vq, pkt_len, buf_vec,
 546                                                          &num_buffers) < 0)) {
 547                         LOG_DEBUG(VHOST_DATA,
 548                                 "(%d) failed to get enough desc from vring\n",
 549                                 dev->vid);
 550                         vq->shadow_used_idx -= num_buffers;
 551                         break;
 552                 }
 553
 554                 LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
 555                         dev->vid, vq->last_avail_idx,
 556                         vq->last_avail_idx + num_buffers);
 557
 558                 if (copy_mbuf_to_desc_mergeable(dev, pkts[pkt_idx],
 559                                                 buf_vec, num_buffers) < 0) {
 560                         vq->shadow_used_idx -= num_buffers;
 561                         break;
 562                 }
 563
 564                 vq->last_avail_idx += num_buffers;
 565         }
 566
 567         if (likely(vq->shadow_used_idx)) {
 568                 flush_shadow_used_ring(dev, vq);
 569
 570                 /* flush used->idx update before we read avail->flags. */
 571                 rte_mb();
 572
 573                 /* Kick the guest if necessary. */
 574                 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
 575                                 && (vq->callfd >= 0))
 576                         eventfd_write(vq->callfd, (eventfd_t)1);
 577         }
 578
 579         return pkt_idx;
 580 }
 581
 582 uint16_t
 583 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
 584         struct rte_mbuf **pkts, uint16_t count)
 585 {
 586         struct virtio_net *dev = get_device(vid);
 587
 588         if (!dev)
 589                 return 0;
 590
 591         if (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF))
 592                 return virtio_dev_merge_rx(dev, queue_id, pkts, count);
 593         else
 594                 return virtio_dev_rx(dev, queue_id, pkts, count);
 595 }
 596
 597 static inline bool
 598 virtio_net_with_host_offload(struct virtio_net *dev)
 599 {
 600         if (dev->features &
 601                         (VIRTIO_NET_F_CSUM | VIRTIO_NET_F_HOST_ECN |
 602                          VIRTIO_NET_F_HOST_TSO4 | VIRTIO_NET_F_HOST_TSO6 |
 603                          VIRTIO_NET_F_HOST_UFO))
 604                 return true;
 605
 606         return false;
 607 }
 608
 609 static void
 610 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
 611 {
 612         struct ipv4_hdr *ipv4_hdr;
 613         struct ipv6_hdr *ipv6_hdr;
 614         void *l3_hdr = NULL;
 615         struct ether_hdr *eth_hdr;
 616         uint16_t ethertype;
 617
 618         eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
 619
 620         m->l2_len = sizeof(struct ether_hdr);
 621         ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
 622
 623         if (ethertype == ETHER_TYPE_VLAN) {
 624                 struct vlan_hdr *vlan_hdr = (struct vlan_hdr *)(eth_hdr + 1);
 625
 626                 m->l2_len += sizeof(struct vlan_hdr);
 627                 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
 628         }
 629
 630         l3_hdr = (char *)eth_hdr + m->l2_len;
 631
 632         switch (ethertype) {
 633         case ETHER_TYPE_IPv4:
 634                 ipv4_hdr = (struct ipv4_hdr *)l3_hdr;
 635                 *l4_proto = ipv4_hdr->next_proto_id;
 636                 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
 637                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 638                 m->ol_flags |= PKT_TX_IPV4;
 639                 break;
 640         case ETHER_TYPE_IPv6:
 641                 ipv6_hdr = (struct ipv6_hdr *)l3_hdr;
 642                 *l4_proto = ipv6_hdr->proto;
 643                 m->l3_len = sizeof(struct ipv6_hdr);
 644                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 645                 m->ol_flags |= PKT_TX_IPV6;
 646                 break;
 647         default:
 648                 m->l3_len = 0;
 649                 *l4_proto = 0;
 650                 break;
 651         }
 652 }
 653
 654 static inline void __attribute__((always_inline))
 655 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
 656 {
 657         uint16_t l4_proto = 0;
 658         void *l4_hdr = NULL;
 659         struct tcp_hdr *tcp_hdr = NULL;
 660
 661         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
 662                 return;
 663
 664         parse_ethernet(m, &l4_proto, &l4_hdr);
 665         if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 666                 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
 667                         switch (hdr->csum_offset) {
 668                         case (offsetof(struct tcp_hdr, cksum)):
 669                                 if (l4_proto == IPPROTO_TCP)
 670                                         m->ol_flags |= PKT_TX_TCP_CKSUM;
 671                                 break;
 672                         case (offsetof(struct udp_hdr, dgram_cksum)):
 673                                 if (l4_proto == IPPROTO_UDP)
 674                                         m->ol_flags |= PKT_TX_UDP_CKSUM;
 675                                 break;
 676                         case (offsetof(struct sctp_hdr, cksum)):
 677                                 if (l4_proto == IPPROTO_SCTP)
 678                                         m->ol_flags |= PKT_TX_SCTP_CKSUM;
 679                                 break;
 680                         default:
 681                                 break;
 682                         }
 683                 }
 684         }
 685
 686         if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 687                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 688                 case VIRTIO_NET_HDR_GSO_TCPV4:
 689                 case VIRTIO_NET_HDR_GSO_TCPV6:
 690                         tcp_hdr = (struct tcp_hdr *)l4_hdr;
 691                         m->ol_flags |= PKT_TX_TCP_SEG;
 692                         m->tso_segsz = hdr->gso_size;
 693                         m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
 694                         break;
 695                 default:
 696                         RTE_LOG(WARNING, VHOST_DATA,
 697                                 "unsupported gso type %u.\n", hdr->gso_type);
 698                         break;
 699                 }
 700         }
 701 }
 702
 703 #define RARP_PKT_SIZE   64
 704
 705 static int
 706 make_rarp_packet(struct rte_mbuf *rarp_mbuf, const struct ether_addr *mac)
 707 {
 708         struct ether_hdr *eth_hdr;
 709         struct arp_hdr  *rarp;
 710
 711         if (rarp_mbuf->buf_len < 64) {
 712                 RTE_LOG(WARNING, VHOST_DATA,
 713                         "failed to make RARP; mbuf size too small %u (< %d)\n",
 714                         rarp_mbuf->buf_len, RARP_PKT_SIZE);
 715                 return -1;
 716         }
 717
 718         /* Ethernet header. */
 719         eth_hdr = rte_pktmbuf_mtod_offset(rarp_mbuf, struct ether_hdr *, 0);
 720         memset(eth_hdr->d_addr.addr_bytes, 0xff, ETHER_ADDR_LEN);
 721         ether_addr_copy(mac, &eth_hdr->s_addr);
 722         eth_hdr->ether_type = htons(ETHER_TYPE_RARP);
 723
 724         /* RARP header. */
 725         rarp = (struct arp_hdr *)(eth_hdr + 1);
 726         rarp->arp_hrd = htons(ARP_HRD_ETHER);
 727         rarp->arp_pro = htons(ETHER_TYPE_IPv4);
 728         rarp->arp_hln = ETHER_ADDR_LEN;
 729         rarp->arp_pln = 4;
 730         rarp->arp_op  = htons(ARP_OP_REVREQUEST);
 731
 732         ether_addr_copy(mac, &rarp->arp_data.arp_sha);
 733         ether_addr_copy(mac, &rarp->arp_data.arp_tha);
 734         memset(&rarp->arp_data.arp_sip, 0x00, 4);
 735         memset(&rarp->arp_data.arp_tip, 0x00, 4);
 736
 737         rarp_mbuf->pkt_len  = rarp_mbuf->data_len = RARP_PKT_SIZE;
 738
 739         return 0;
 740 }
 741
 742 static inline void __attribute__((always_inline))
 743 put_zmbuf(struct zcopy_mbuf *zmbuf)
 744 {
 745         zmbuf->in_use = 0;
 746 }
 747
 748 static inline int __attribute__((always_inline))
 749 copy_desc_to_mbuf(struct virtio_net *dev, struct vring_desc *descs,
 750                   uint16_t max_desc, struct rte_mbuf *m, uint16_t desc_idx,
 751                   struct rte_mempool *mbuf_pool)
 752 {
 753         struct vring_desc *desc;
 754         uint64_t desc_addr;
 755         uint32_t desc_avail, desc_offset;
 756         uint32_t mbuf_avail, mbuf_offset;
 757         uint32_t cpy_len;
 758         struct rte_mbuf *cur = m, *prev = m;
 759         struct virtio_net_hdr *hdr = NULL;
 760         /* A counter to avoid desc dead loop chain */
 761         uint32_t nr_desc = 1;
 762
 763         desc = &descs[desc_idx];
 764         if (unlikely((desc->len < dev->vhost_hlen)) ||
 765                         (desc->flags & VRING_DESC_F_INDIRECT))
 766                 return -1;
 767
 768         desc_addr = gpa_to_vva(dev, desc->addr);
 769         if (unlikely(!desc_addr))
 770                 return -1;
 771
 772         if (virtio_net_with_host_offload(dev)) {
 773                 hdr = (struct virtio_net_hdr *)((uintptr_t)desc_addr);
 774                 rte_prefetch0(hdr);
 775         }
 776
 777         /*
 778          * A virtio driver normally uses at least 2 desc buffers
 779          * for Tx: the first for storing the header, and others
 780          * for storing the data.
 781          */
 782         if (likely((desc->len == dev->vhost_hlen) &&
 783                    (desc->flags & VRING_DESC_F_NEXT) != 0)) {
 784                 desc = &descs[desc->next];
 785                 if (unlikely(desc->flags & VRING_DESC_F_INDIRECT))
 786                         return -1;
 787
 788                 desc_addr = gpa_to_vva(dev, desc->addr);
 789                 if (unlikely(!desc_addr))
 790                         return -1;
 791
 792                 desc_offset = 0;
 793                 desc_avail  = desc->len;
 794                 nr_desc    += 1;
 795         } else {
 796                 desc_avail  = desc->len - dev->vhost_hlen;
 797                 desc_offset = dev->vhost_hlen;
 798         }
 799
 800         rte_prefetch0((void *)(uintptr_t)(desc_addr + desc_offset));
 801
 802         PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset), desc_avail, 0);
 803
 804         mbuf_offset = 0;
 805         mbuf_avail  = m->buf_len - RTE_PKTMBUF_HEADROOM;
 806         while (1) {
 807                 uint64_t hpa;
 808
 809                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 810
 811                 /*
 812                  * A desc buf might across two host physical pages that are
 813                  * not continuous. In such case (gpa_to_hpa returns 0), data
 814                  * will be copied even though zero copy is enabled.
 815                  */
 816                 if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
 817                                         desc->addr + desc_offset, cpy_len)))) {
 818                         cur->data_len = cpy_len;
 819                         cur->data_off = 0;
 820                         cur->buf_addr = (void *)(uintptr_t)desc_addr;
 821                         cur->buf_physaddr = hpa;
 822
 823                         /*
 824                          * In zero copy mode, one mbuf can only reference data
 825                          * for one or partial of one desc buff.
 826                          */
 827                         mbuf_avail = cpy_len;
 828                 } else {
 829                         rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
 830                                                            mbuf_offset),
 831                                 (void *)((uintptr_t)(desc_addr + desc_offset)),
 832                                 cpy_len);
 833                 }
 834
 835                 mbuf_avail  -= cpy_len;
 836                 mbuf_offset += cpy_len;
 837                 desc_avail  -= cpy_len;
 838                 desc_offset += cpy_len;
 839
 840                 /* This desc reaches to its end, get the next one */
 841                 if (desc_avail == 0) {
 842                         if ((desc->flags & VRING_DESC_F_NEXT) == 0)
 843                                 break;
 844
 845                         if (unlikely(desc->next >= max_desc ||
 846                                      ++nr_desc > max_desc))
 847                                 return -1;
 848                         desc = &descs[desc->next];
 849                         if (unlikely(desc->flags & VRING_DESC_F_INDIRECT))
 850                                 return -1;
 851
 852                         desc_addr = gpa_to_vva(dev, desc->addr);
 853                         if (unlikely(!desc_addr))
 854                                 return -1;
 855
 856                         rte_prefetch0((void *)(uintptr_t)desc_addr);
 857
 858                         desc_offset = 0;
 859                         desc_avail  = desc->len;
 860
 861                         PRINT_PACKET(dev, (uintptr_t)desc_addr, desc->len, 0);
 862                 }
 863
 864                 /*
 865                  * This mbuf reaches to its end, get a new one
 866                  * to hold more data.
 867                  */
 868                 if (mbuf_avail == 0) {
 869                         cur = rte_pktmbuf_alloc(mbuf_pool);
 870                         if (unlikely(cur == NULL)) {
 871                                 RTE_LOG(ERR, VHOST_DATA, "Failed to "
 872                                         "allocate memory for mbuf.\n");
 873                                 return -1;
 874                         }
 875
 876                         prev->next = cur;
 877                         prev->data_len = mbuf_offset;
 878                         m->nb_segs += 1;
 879                         m->pkt_len += mbuf_offset;
 880                         prev = cur;
 881
 882                         mbuf_offset = 0;
 883                         mbuf_avail  = cur->buf_len - RTE_PKTMBUF_HEADROOM;
 884                 }
 885         }
 886
 887         prev->data_len = mbuf_offset;
 888         m->pkt_len    += mbuf_offset;
 889
 890         if (hdr)
 891                 vhost_dequeue_offload(hdr, m);
 892
 893         return 0;
 894 }
 895
 896 static inline void __attribute__((always_inline))
 897 update_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq,
 898                  uint32_t used_idx, uint32_t desc_idx)
 899 {
 900         vq->used->ring[used_idx].id  = desc_idx;
 901         vq->used->ring[used_idx].len = 0;
 902         vhost_log_used_vring(dev, vq,
 903                         offsetof(struct vring_used, ring[used_idx]),
 904                         sizeof(vq->used->ring[used_idx]));
 905 }
 906
 907 static inline void __attribute__((always_inline))
 908 update_used_idx(struct virtio_net *dev, struct vhost_virtqueue *vq,
 909                 uint32_t count)
 910 {
 911         if (unlikely(count == 0))
 912                 return;
 913
 914         rte_smp_wmb();
 915         rte_smp_rmb();
 916
 917         vq->used->idx += count;
 918         vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
 919                         sizeof(vq->used->idx));
 920
 921         /* Kick guest if required. */
 922         if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
 923                         && (vq->callfd >= 0))
 924                 eventfd_write(vq->callfd, (eventfd_t)1);
 925 }
 926
 927 static inline struct zcopy_mbuf *__attribute__((always_inline))
 928 get_zmbuf(struct vhost_virtqueue *vq)
 929 {
 930         uint16_t i;
 931         uint16_t last;
 932         int tries = 0;
 933
 934         /* search [last_zmbuf_idx, zmbuf_size) */
 935         i = vq->last_zmbuf_idx;
 936         last = vq->zmbuf_size;
 937
 938 again:
 939         for (; i < last; i++) {
 940                 if (vq->zmbufs[i].in_use == 0) {
 941                         vq->last_zmbuf_idx = i + 1;
 942                         vq->zmbufs[i].in_use = 1;
 943                         return &vq->zmbufs[i];
 944                 }
 945         }
 946
 947         tries++;
 948         if (tries == 1) {
 949                 /* search [0, last_zmbuf_idx) */
 950                 i = 0;
 951                 last = vq->last_zmbuf_idx;
 952                 goto again;
 953         }
 954
 955         return NULL;
 956 }
 957
 958 static inline bool __attribute__((always_inline))
 959 mbuf_is_consumed(struct rte_mbuf *m)
 960 {
 961         while (m) {
 962                 if (rte_mbuf_refcnt_read(m) > 1)
 963                         return false;
 964                 m = m->next;
 965         }
 966
 967         return true;
 968 }
 969
 970 uint16_t
 971 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
 972         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
 973 {
 974         struct virtio_net *dev;
 975         struct rte_mbuf *rarp_mbuf = NULL;
 976         struct vhost_virtqueue *vq;
 977         uint32_t desc_indexes[MAX_PKT_BURST];
 978         uint32_t used_idx;
 979         uint32_t i = 0;
 980         uint16_t free_entries;
 981         uint16_t avail_idx;
 982
 983         dev = get_device(vid);
 984         if (!dev)
 985                 return 0;
 986
 987         if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->virt_qp_nb))) {
 988                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
 989                         dev->vid, __func__, queue_id);
 990                 return 0;
 991         }
 992
 993         vq = dev->virtqueue[queue_id];
 994         if (unlikely(vq->enabled == 0))
 995                 return 0;
 996
 997         if (unlikely(dev->dequeue_zero_copy)) {
 998                 struct zcopy_mbuf *zmbuf, *next;
 999                 int nr_updated = 0;
1000
1001                 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1002                      zmbuf != NULL; zmbuf = next) {
1003                         next = TAILQ_NEXT(zmbuf, next);
1004
1005                         if (mbuf_is_consumed(zmbuf->mbuf)) {
1006                                 used_idx = vq->last_used_idx++ & (vq->size - 1);
1007                                 update_used_ring(dev, vq, used_idx,
1008                                                  zmbuf->desc_idx);
1009                                 nr_updated += 1;
1010
1011                                 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1012                                 rte_pktmbuf_free(zmbuf->mbuf);
1013                                 put_zmbuf(zmbuf);
1014                                 vq->nr_zmbuf -= 1;
1015                         }
1016                 }
1017
1018                 update_used_idx(dev, vq, nr_updated);
1019         }
1020
1021         /*
1022          * Construct a RARP broadcast packet, and inject it to the "pkts"
1023          * array, to looks like that guest actually send such packet.
1024          *
1025          * Check user_send_rarp() for more information.
1026          */
1027         if (unlikely(rte_atomic16_cmpset((volatile uint16_t *)
1028                                          &dev->broadcast_rarp.cnt, 1, 0))) {
1029                 rarp_mbuf = rte_pktmbuf_alloc(mbuf_pool);
1030                 if (rarp_mbuf == NULL) {
1031                         RTE_LOG(ERR, VHOST_DATA,
1032                                 "Failed to allocate memory for mbuf.\n");
1033                         return 0;
1034                 }
1035
1036                 if (make_rarp_packet(rarp_mbuf, &dev->mac)) {
1037                         rte_pktmbuf_free(rarp_mbuf);
1038                         rarp_mbuf = NULL;
1039                 } else {
1040                         count -= 1;
1041                 }
1042         }
1043
1044         free_entries = *((volatile uint16_t *)&vq->avail->idx) -
1045                         vq->last_avail_idx;
1046         if (free_entries == 0)
1047                 goto out;
1048
1049         LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1050
1051         /* Prefetch available and used ring */
1052         avail_idx = vq->last_avail_idx & (vq->size - 1);
1053         used_idx  = vq->last_used_idx  & (vq->size - 1);
1054         rte_prefetch0(&vq->avail->ring[avail_idx]);
1055         rte_prefetch0(&vq->used->ring[used_idx]);
1056
1057         count = RTE_MIN(count, MAX_PKT_BURST);
1058         count = RTE_MIN(count, free_entries);
1059         LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1060                         dev->vid, count);
1061
1062         /* Retrieve all of the head indexes first to avoid caching issues. */
1063         for (i = 0; i < count; i++) {
1064                 avail_idx = (vq->last_avail_idx + i) & (vq->size - 1);
1065                 used_idx  = (vq->last_used_idx  + i) & (vq->size - 1);
1066                 desc_indexes[i] = vq->avail->ring[avail_idx];
1067
1068                 if (likely(dev->dequeue_zero_copy == 0))
1069                         update_used_ring(dev, vq, used_idx, desc_indexes[i]);
1070         }
1071
1072         /* Prefetch descriptor index. */
1073         rte_prefetch0(&vq->desc[desc_indexes[0]]);
1074         for (i = 0; i < count; i++) {
1075                 struct vring_desc *desc;
1076                 uint16_t sz, idx;
1077                 int err;
1078
1079                 if (likely(i + 1 < count))
1080                         rte_prefetch0(&vq->desc[desc_indexes[i + 1]]);
1081
1082                 if (vq->desc[desc_indexes[i]].flags & VRING_DESC_F_INDIRECT) {
1083                         desc = (struct vring_desc *)(uintptr_t)gpa_to_vva(dev,
1084                                         vq->desc[desc_indexes[i]].addr);
1085                         if (unlikely(!desc))
1086                                 break;
1087
1088                         rte_prefetch0(desc);
1089                         sz = vq->desc[desc_indexes[i]].len / sizeof(*desc);
1090                         idx = 0;
1091                 } else {
1092                         desc = vq->desc;
1093                         sz = vq->size;
1094                         idx = desc_indexes[i];
1095                 }
1096
1097                 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
1098                 if (unlikely(pkts[i] == NULL)) {
1099                         RTE_LOG(ERR, VHOST_DATA,
1100                                 "Failed to allocate memory for mbuf.\n");
1101                         break;
1102                 }
1103
1104                 err = copy_desc_to_mbuf(dev, desc, sz, pkts[i], idx, mbuf_pool);
1105                 if (unlikely(err)) {
1106                         rte_pktmbuf_free(pkts[i]);
1107                         break;
1108                 }
1109
1110                 if (unlikely(dev->dequeue_zero_copy)) {
1111                         struct zcopy_mbuf *zmbuf;
1112
1113                         zmbuf = get_zmbuf(vq);
1114                         if (!zmbuf) {
1115                                 rte_pktmbuf_free(pkts[i]);
1116                                 break;
1117                         }
1118                         zmbuf->mbuf = pkts[i];
1119                         zmbuf->desc_idx = desc_indexes[i];
1120
1121                         /*
1122                          * Pin lock the mbuf; we will check later to see
1123                          * whether the mbuf is freed (when we are the last
1124                          * user) or not. If that's the case, we then could
1125                          * update the used ring safely.
1126                          */
1127                         rte_mbuf_refcnt_update(pkts[i], 1);
1128
1129                         vq->nr_zmbuf += 1;
1130                         TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1131                 }
1132         }
1133         vq->last_avail_idx += i;
1134
1135         if (likely(dev->dequeue_zero_copy == 0)) {
1136                 vq->last_used_idx += i;
1137                 update_used_idx(dev, vq, i);
1138         }
1139
1140 out:
1141         if (unlikely(rarp_mbuf != NULL)) {
1142                 /*
1143                  * Inject it to the head of "pkts" array, so that switch's mac
1144                  * learning table will get updated first.
1145                  */
1146                 memmove(&pkts[1], pkts, i * sizeof(struct rte_mbuf *));
1147                 pkts[0] = rarp_mbuf;
1148                 i += 1;
1149         }
1150
1151         return i;
1152 }