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 nr_vring)
  90 {
  91         return (is_tx ^ (idx & 1)) == 0 && idx < nr_vring;
  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 vring_desc *descs,
 190                   struct rte_mbuf *m, uint16_t desc_idx, uint32_t size)
 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         /* A counter to avoid desc dead loop chain */
 199         uint16_t nr_desc = 1;
 200
 201         desc = &descs[desc_idx];
 202         desc_addr = rte_vhost_gpa_to_vva(dev->mem, desc->addr);
 203         /*
 204          * Checking of 'desc_addr' placed outside of 'unlikely' macro to avoid
 205          * performance issue with some versions of gcc (4.8.4 and 5.3.0) which
 206          * otherwise stores offset on the stack instead of in a register.
 207          */
 208         if (unlikely(desc->len < dev->vhost_hlen) || !desc_addr)
 209                 return -1;
 210
 211         rte_prefetch0((void *)(uintptr_t)desc_addr);
 212
 213         virtio_enqueue_offload(m, &virtio_hdr.hdr);
 214         copy_virtio_net_hdr(dev, desc_addr, virtio_hdr);
 215         vhost_log_write(dev, desc->addr, dev->vhost_hlen);
 216         PRINT_PACKET(dev, (uintptr_t)desc_addr, dev->vhost_hlen, 0);
 217
 218         desc_offset = dev->vhost_hlen;
 219         desc_avail  = desc->len - dev->vhost_hlen;
 220
 221         mbuf_avail  = rte_pktmbuf_data_len(m);
 222         mbuf_offset = 0;
 223         while (mbuf_avail != 0 || m->next != NULL) {
 224                 /* done with current mbuf, fetch next */
 225                 if (mbuf_avail == 0) {
 226                         m = m->next;
 227
 228                         mbuf_offset = 0;
 229                         mbuf_avail  = rte_pktmbuf_data_len(m);
 230                 }
 231
 232                 /* done with current desc buf, fetch next */
 233                 if (desc_avail == 0) {
 234                         if ((desc->flags & VRING_DESC_F_NEXT) == 0) {
 235                                 /* Room in vring buffer is not enough */
 236                                 return -1;
 237                         }
 238                         if (unlikely(desc->next >= size || ++nr_desc > size))
 239                                 return -1;
 240
 241                         desc = &descs[desc->next];
 242                         desc_addr = rte_vhost_gpa_to_vva(dev->mem, desc->addr);
 243                         if (unlikely(!desc_addr))
 244                                 return -1;
 245
 246                         desc_offset = 0;
 247                         desc_avail  = desc->len;
 248                 }
 249
 250                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 251                 rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)),
 252                         rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 253                         cpy_len);
 254                 vhost_log_write(dev, desc->addr + desc_offset, cpy_len);
 255                 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
 256                              cpy_len, 0);
 257
 258                 mbuf_avail  -= cpy_len;
 259                 mbuf_offset += cpy_len;
 260                 desc_avail  -= cpy_len;
 261                 desc_offset += cpy_len;
 262         }
 263
 264         return 0;
 265 }
 266
 267 /**
 268  * This function adds buffers to the virtio devices RX virtqueue. Buffers can
 269  * be received from the physical port or from another virtio device. A packet
 270  * count is returned to indicate the number of packets that are succesfully
 271  * added to the RX queue. This function works when the mbuf is scattered, but
 272  * it doesn't support the mergeable feature.
 273  */
 274 static inline uint32_t __attribute__((always_inline))
 275 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
 276               struct rte_mbuf **pkts, uint32_t count)
 277 {
 278         struct vhost_virtqueue *vq;
 279         uint16_t avail_idx, free_entries, start_idx;
 280         uint16_t desc_indexes[MAX_PKT_BURST];
 281         struct vring_desc *descs;
 282         uint16_t used_idx;
 283         uint32_t i, sz;
 284
 285         LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
 286         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
 287                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
 288                         dev->vid, __func__, queue_id);
 289                 return 0;
 290         }
 291
 292         vq = dev->virtqueue[queue_id];
 293         if (unlikely(vq->enabled == 0))
 294                 return 0;
 295
 296         avail_idx = *((volatile uint16_t *)&vq->avail->idx);
 297         start_idx = vq->last_used_idx;
 298         free_entries = avail_idx - start_idx;
 299         count = RTE_MIN(count, free_entries);
 300         count = RTE_MIN(count, (uint32_t)MAX_PKT_BURST);
 301         if (count == 0)
 302                 return 0;
 303
 304         LOG_DEBUG(VHOST_DATA, "(%d) start_idx %d | end_idx %d\n",
 305                 dev->vid, start_idx, start_idx + count);
 306
 307         /* Retrieve all of the desc indexes first to avoid caching issues. */
 308         rte_prefetch0(&vq->avail->ring[start_idx & (vq->size - 1)]);
 309         for (i = 0; i < count; i++) {
 310                 used_idx = (start_idx + i) & (vq->size - 1);
 311                 desc_indexes[i] = vq->avail->ring[used_idx];
 312                 vq->used->ring[used_idx].id = desc_indexes[i];
 313                 vq->used->ring[used_idx].len = pkts[i]->pkt_len +
 314                                                dev->vhost_hlen;
 315                 vhost_log_used_vring(dev, vq,
 316                         offsetof(struct vring_used, ring[used_idx]),
 317                         sizeof(vq->used->ring[used_idx]));
 318         }
 319
 320         rte_prefetch0(&vq->desc[desc_indexes[0]]);
 321         for (i = 0; i < count; i++) {
 322                 uint16_t desc_idx = desc_indexes[i];
 323                 int err;
 324
 325                 if (vq->desc[desc_idx].flags & VRING_DESC_F_INDIRECT) {
 326                         descs = (struct vring_desc *)(uintptr_t)
 327                                 rte_vhost_gpa_to_vva(dev->mem,
 328                                         vq->desc[desc_idx].addr);
 329                         if (unlikely(!descs)) {
 330                                 count = i;
 331                                 break;
 332                         }
 333
 334                         desc_idx = 0;
 335                         sz = vq->desc[desc_idx].len / sizeof(*descs);
 336                 } else {
 337                         descs = vq->desc;
 338                         sz = vq->size;
 339                 }
 340
 341                 err = copy_mbuf_to_desc(dev, descs, pkts[i], desc_idx, sz);
 342                 if (unlikely(err)) {
 343                         used_idx = (start_idx + i) & (vq->size - 1);
 344                         vq->used->ring[used_idx].len = dev->vhost_hlen;
 345                         vhost_log_used_vring(dev, vq,
 346                                 offsetof(struct vring_used, ring[used_idx]),
 347                                 sizeof(vq->used->ring[used_idx]));
 348                 }
 349
 350                 if (i + 1 < count)
 351                         rte_prefetch0(&vq->desc[desc_indexes[i+1]]);
 352         }
 353
 354         rte_smp_wmb();
 355
 356         *(volatile uint16_t *)&vq->used->idx += count;
 357         vq->last_used_idx += count;
 358         vhost_log_used_vring(dev, vq,
 359                 offsetof(struct vring_used, idx),
 360                 sizeof(vq->used->idx));
 361
 362         /* flush used->idx update before we read avail->flags. */
 363         rte_mb();
 364
 365         /* Kick the guest if necessary. */
 366         if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
 367                         && (vq->callfd >= 0))
 368                 eventfd_write(vq->callfd, (eventfd_t)1);
 369         return count;
 370 }
 371
 372 static inline int __attribute__((always_inline))
 373 fill_vec_buf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 374                          uint32_t avail_idx, uint32_t *vec_idx,
 375                          struct buf_vector *buf_vec, uint16_t *desc_chain_head,
 376                          uint16_t *desc_chain_len)
 377 {
 378         uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
 379         uint32_t vec_id = *vec_idx;
 380         uint32_t len    = 0;
 381         struct vring_desc *descs = vq->desc;
 382
 383         *desc_chain_head = idx;
 384
 385         if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
 386                 descs = (struct vring_desc *)(uintptr_t)
 387                         rte_vhost_gpa_to_vva(dev->mem, vq->desc[idx].addr);
 388                 if (unlikely(!descs))
 389                         return -1;
 390
 391                 idx = 0;
 392         }
 393
 394         while (1) {
 395                 if (unlikely(vec_id >= BUF_VECTOR_MAX || idx >= vq->size))
 396                         return -1;
 397
 398                 len += descs[idx].len;
 399                 buf_vec[vec_id].buf_addr = descs[idx].addr;
 400                 buf_vec[vec_id].buf_len  = descs[idx].len;
 401                 buf_vec[vec_id].desc_idx = idx;
 402                 vec_id++;
 403
 404                 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
 405                         break;
 406
 407                 idx = descs[idx].next;
 408         }
 409
 410         *desc_chain_len = len;
 411         *vec_idx = vec_id;
 412
 413         return 0;
 414 }
 415
 416 /*
 417  * Returns -1 on fail, 0 on success
 418  */
 419 static inline int
 420 reserve_avail_buf_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
 421                                 uint32_t size, struct buf_vector *buf_vec,
 422                                 uint16_t *num_buffers, uint16_t avail_head)
 423 {
 424         uint16_t cur_idx;
 425         uint32_t vec_idx = 0;
 426         uint16_t tries = 0;
 427
 428         uint16_t head_idx = 0;
 429         uint16_t len = 0;
 430
 431         *num_buffers = 0;
 432         cur_idx  = vq->last_avail_idx;
 433
 434         while (size > 0) {
 435                 if (unlikely(cur_idx == avail_head))
 436                         return -1;
 437
 438                 if (unlikely(fill_vec_buf(dev, vq, cur_idx, &vec_idx, buf_vec,
 439                                                 &head_idx, &len) < 0))
 440                         return -1;
 441                 len = RTE_MIN(len, size);
 442                 update_shadow_used_ring(vq, head_idx, len);
 443                 size -= len;
 444
 445                 cur_idx++;
 446                 tries++;
 447                 *num_buffers += 1;
 448
 449                 /*
 450                  * if we tried all available ring items, and still
 451                  * can't get enough buf, it means something abnormal
 452                  * happened.
 453                  */
 454                 if (unlikely(tries >= vq->size))
 455                         return -1;
 456         }
 457
 458         return 0;
 459 }
 460
 461 static inline int __attribute__((always_inline))
 462 copy_mbuf_to_desc_mergeable(struct virtio_net *dev, struct rte_mbuf *m,
 463                             struct buf_vector *buf_vec, uint16_t num_buffers)
 464 {
 465         struct virtio_net_hdr_mrg_rxbuf virtio_hdr = {{0, 0, 0, 0, 0, 0}, 0};
 466         uint32_t vec_idx = 0;
 467         uint64_t desc_addr;
 468         uint32_t mbuf_offset, mbuf_avail;
 469         uint32_t desc_offset, desc_avail;
 470         uint32_t cpy_len;
 471         uint64_t hdr_addr, hdr_phys_addr;
 472         struct rte_mbuf *hdr_mbuf;
 473
 474         if (unlikely(m == NULL))
 475                 return -1;
 476
 477         desc_addr = rte_vhost_gpa_to_vva(dev->mem, buf_vec[vec_idx].buf_addr);
 478         if (buf_vec[vec_idx].buf_len < dev->vhost_hlen || !desc_addr)
 479                 return -1;
 480
 481         hdr_mbuf = m;
 482         hdr_addr = desc_addr;
 483         hdr_phys_addr = buf_vec[vec_idx].buf_addr;
 484         rte_prefetch0((void *)(uintptr_t)hdr_addr);
 485
 486         virtio_hdr.num_buffers = num_buffers;
 487         LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
 488                 dev->vid, num_buffers);
 489
 490         desc_avail  = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
 491         desc_offset = dev->vhost_hlen;
 492
 493         mbuf_avail  = rte_pktmbuf_data_len(m);
 494         mbuf_offset = 0;
 495         while (mbuf_avail != 0 || m->next != NULL) {
 496                 /* done with current desc buf, get the next one */
 497                 if (desc_avail == 0) {
 498                         vec_idx++;
 499                         desc_addr = rte_vhost_gpa_to_vva(dev->mem,
 500                                         buf_vec[vec_idx].buf_addr);
 501                         if (unlikely(!desc_addr))
 502                                 return -1;
 503
 504                         /* Prefetch buffer address. */
 505                         rte_prefetch0((void *)(uintptr_t)desc_addr);
 506                         desc_offset = 0;
 507                         desc_avail  = buf_vec[vec_idx].buf_len;
 508                 }
 509
 510                 /* done with current mbuf, get the next one */
 511                 if (mbuf_avail == 0) {
 512                         m = m->next;
 513
 514                         mbuf_offset = 0;
 515                         mbuf_avail  = rte_pktmbuf_data_len(m);
 516                 }
 517
 518                 if (hdr_addr) {
 519                         virtio_enqueue_offload(hdr_mbuf, &virtio_hdr.hdr);
 520                         copy_virtio_net_hdr(dev, hdr_addr, virtio_hdr);
 521                         vhost_log_write(dev, hdr_phys_addr, dev->vhost_hlen);
 522                         PRINT_PACKET(dev, (uintptr_t)hdr_addr,
 523                                      dev->vhost_hlen, 0);
 524
 525                         hdr_addr = 0;
 526                 }
 527
 528                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 529                 rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)),
 530                         rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 531                         cpy_len);
 532                 vhost_log_write(dev, buf_vec[vec_idx].buf_addr + desc_offset,
 533                         cpy_len);
 534                 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
 535                         cpy_len, 0);
 536
 537                 mbuf_avail  -= cpy_len;
 538                 mbuf_offset += cpy_len;
 539                 desc_avail  -= cpy_len;
 540                 desc_offset += cpy_len;
 541         }
 542
 543         return 0;
 544 }
 545
 546 static inline uint32_t __attribute__((always_inline))
 547 virtio_dev_merge_rx(struct virtio_net *dev, uint16_t queue_id,
 548         struct rte_mbuf **pkts, uint32_t count)
 549 {
 550         struct vhost_virtqueue *vq;
 551         uint32_t pkt_idx = 0;
 552         uint16_t num_buffers;
 553         struct buf_vector buf_vec[BUF_VECTOR_MAX];
 554         uint16_t avail_head;
 555
 556         LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
 557         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
 558                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
 559                         dev->vid, __func__, queue_id);
 560                 return 0;
 561         }
 562
 563         vq = dev->virtqueue[queue_id];
 564         if (unlikely(vq->enabled == 0))
 565                 return 0;
 566
 567         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
 568         if (count == 0)
 569                 return 0;
 570
 571         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
 572
 573         vq->shadow_used_idx = 0;
 574         avail_head = *((volatile uint16_t *)&vq->avail->idx);
 575         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
 576                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
 577
 578                 if (unlikely(reserve_avail_buf_mergeable(dev, vq,
 579                                                 pkt_len, buf_vec, &num_buffers,
 580                                                 avail_head) < 0)) {
 581                         LOG_DEBUG(VHOST_DATA,
 582                                 "(%d) failed to get enough desc from vring\n",
 583                                 dev->vid);
 584                         vq->shadow_used_idx -= num_buffers;
 585                         break;
 586                 }
 587
 588                 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_mergeable(dev, pkts[pkt_idx],
 593                                                 buf_vec, num_buffers) < 0) {
 594                         vq->shadow_used_idx -= num_buffers;
 595                         break;
 596                 }
 597
 598                 vq->last_avail_idx += num_buffers;
 599         }
 600
 601         if (likely(vq->shadow_used_idx)) {
 602                 flush_shadow_used_ring(dev, vq);
 603
 604                 /* flush used->idx update before we read avail->flags. */
 605                 rte_mb();
 606
 607                 /* Kick the guest if necessary. */
 608                 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
 609                                 && (vq->callfd >= 0))
 610                         eventfd_write(vq->callfd, (eventfd_t)1);
 611         }
 612
 613         return pkt_idx;
 614 }
 615
 616 uint16_t
 617 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
 618         struct rte_mbuf **pkts, uint16_t count)
 619 {
 620         struct virtio_net *dev = get_device(vid);
 621
 622         if (!dev)
 623                 return 0;
 624
 625         if (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF))
 626                 return virtio_dev_merge_rx(dev, queue_id, pkts, count);
 627         else
 628                 return virtio_dev_rx(dev, queue_id, pkts, count);
 629 }
 630
 631 static inline bool
 632 virtio_net_with_host_offload(struct virtio_net *dev)
 633 {
 634         if (dev->features &
 635                         (VIRTIO_NET_F_CSUM | VIRTIO_NET_F_HOST_ECN |
 636                          VIRTIO_NET_F_HOST_TSO4 | VIRTIO_NET_F_HOST_TSO6 |
 637                          VIRTIO_NET_F_HOST_UFO))
 638                 return true;
 639
 640         return false;
 641 }
 642
 643 static void
 644 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
 645 {
 646         struct ipv4_hdr *ipv4_hdr;
 647         struct ipv6_hdr *ipv6_hdr;
 648         void *l3_hdr = NULL;
 649         struct ether_hdr *eth_hdr;
 650         uint16_t ethertype;
 651
 652         eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
 653
 654         m->l2_len = sizeof(struct ether_hdr);
 655         ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
 656
 657         if (ethertype == ETHER_TYPE_VLAN) {
 658                 struct vlan_hdr *vlan_hdr = (struct vlan_hdr *)(eth_hdr + 1);
 659
 660                 m->l2_len += sizeof(struct vlan_hdr);
 661                 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
 662         }
 663
 664         l3_hdr = (char *)eth_hdr + m->l2_len;
 665
 666         switch (ethertype) {
 667         case ETHER_TYPE_IPv4:
 668                 ipv4_hdr = (struct ipv4_hdr *)l3_hdr;
 669                 *l4_proto = ipv4_hdr->next_proto_id;
 670                 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
 671                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 672                 m->ol_flags |= PKT_TX_IPV4;
 673                 break;
 674         case ETHER_TYPE_IPv6:
 675                 ipv6_hdr = (struct ipv6_hdr *)l3_hdr;
 676                 *l4_proto = ipv6_hdr->proto;
 677                 m->l3_len = sizeof(struct ipv6_hdr);
 678                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 679                 m->ol_flags |= PKT_TX_IPV6;
 680                 break;
 681         default:
 682                 m->l3_len = 0;
 683                 *l4_proto = 0;
 684                 *l4_hdr = NULL;
 685                 break;
 686         }
 687 }
 688
 689 static inline void __attribute__((always_inline))
 690 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
 691 {
 692         uint16_t l4_proto = 0;
 693         void *l4_hdr = NULL;
 694         struct tcp_hdr *tcp_hdr = NULL;
 695
 696         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
 697                 return;
 698
 699         parse_ethernet(m, &l4_proto, &l4_hdr);
 700         if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 701                 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
 702                         switch (hdr->csum_offset) {
 703                         case (offsetof(struct tcp_hdr, cksum)):
 704                                 if (l4_proto == IPPROTO_TCP)
 705                                         m->ol_flags |= PKT_TX_TCP_CKSUM;
 706                                 break;
 707                         case (offsetof(struct udp_hdr, dgram_cksum)):
 708                                 if (l4_proto == IPPROTO_UDP)
 709                                         m->ol_flags |= PKT_TX_UDP_CKSUM;
 710                                 break;
 711                         case (offsetof(struct sctp_hdr, cksum)):
 712                                 if (l4_proto == IPPROTO_SCTP)
 713                                         m->ol_flags |= PKT_TX_SCTP_CKSUM;
 714                                 break;
 715                         default:
 716                                 break;
 717                         }
 718                 }
 719         }
 720
 721         if (l4_hdr && hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 722                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 723                 case VIRTIO_NET_HDR_GSO_TCPV4:
 724                 case VIRTIO_NET_HDR_GSO_TCPV6:
 725                         tcp_hdr = (struct tcp_hdr *)l4_hdr;
 726                         m->ol_flags |= PKT_TX_TCP_SEG;
 727                         m->tso_segsz = hdr->gso_size;
 728                         m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
 729                         break;
 730                 default:
 731                         RTE_LOG(WARNING, VHOST_DATA,
 732                                 "unsupported gso type %u.\n", hdr->gso_type);
 733                         break;
 734                 }
 735         }
 736 }
 737
 738 #define RARP_PKT_SIZE   64
 739
 740 static int
 741 make_rarp_packet(struct rte_mbuf *rarp_mbuf, const struct ether_addr *mac)
 742 {
 743         struct ether_hdr *eth_hdr;
 744         struct arp_hdr  *rarp;
 745
 746         if (rarp_mbuf->buf_len < 64) {
 747                 RTE_LOG(WARNING, VHOST_DATA,
 748                         "failed to make RARP; mbuf size too small %u (< %d)\n",
 749                         rarp_mbuf->buf_len, RARP_PKT_SIZE);
 750                 return -1;
 751         }
 752
 753         /* Ethernet header. */
 754         eth_hdr = rte_pktmbuf_mtod_offset(rarp_mbuf, struct ether_hdr *, 0);
 755         memset(eth_hdr->d_addr.addr_bytes, 0xff, ETHER_ADDR_LEN);
 756         ether_addr_copy(mac, &eth_hdr->s_addr);
 757         eth_hdr->ether_type = htons(ETHER_TYPE_RARP);
 758
 759         /* RARP header. */
 760         rarp = (struct arp_hdr *)(eth_hdr + 1);
 761         rarp->arp_hrd = htons(ARP_HRD_ETHER);
 762         rarp->arp_pro = htons(ETHER_TYPE_IPv4);
 763         rarp->arp_hln = ETHER_ADDR_LEN;
 764         rarp->arp_pln = 4;
 765         rarp->arp_op  = htons(ARP_OP_REVREQUEST);
 766
 767         ether_addr_copy(mac, &rarp->arp_data.arp_sha);
 768         ether_addr_copy(mac, &rarp->arp_data.arp_tha);
 769         memset(&rarp->arp_data.arp_sip, 0x00, 4);
 770         memset(&rarp->arp_data.arp_tip, 0x00, 4);
 771
 772         rarp_mbuf->pkt_len  = rarp_mbuf->data_len = RARP_PKT_SIZE;
 773
 774         return 0;
 775 }
 776
 777 static inline void __attribute__((always_inline))
 778 put_zmbuf(struct zcopy_mbuf *zmbuf)
 779 {
 780         zmbuf->in_use = 0;
 781 }
 782
 783 static inline int __attribute__((always_inline))
 784 copy_desc_to_mbuf(struct virtio_net *dev, struct vring_desc *descs,
 785                   uint16_t max_desc, struct rte_mbuf *m, uint16_t desc_idx,
 786                   struct rte_mempool *mbuf_pool)
 787 {
 788         struct vring_desc *desc;
 789         uint64_t desc_addr;
 790         uint32_t desc_avail, desc_offset;
 791         uint32_t mbuf_avail, mbuf_offset;
 792         uint32_t cpy_len;
 793         struct rte_mbuf *cur = m, *prev = m;
 794         struct virtio_net_hdr *hdr = NULL;
 795         /* A counter to avoid desc dead loop chain */
 796         uint32_t nr_desc = 1;
 797
 798         desc = &descs[desc_idx];
 799         if (unlikely((desc->len < dev->vhost_hlen)) ||
 800                         (desc->flags & VRING_DESC_F_INDIRECT))
 801                 return -1;
 802
 803         desc_addr = rte_vhost_gpa_to_vva(dev->mem, desc->addr);
 804         if (unlikely(!desc_addr))
 805                 return -1;
 806
 807         if (virtio_net_with_host_offload(dev)) {
 808                 hdr = (struct virtio_net_hdr *)((uintptr_t)desc_addr);
 809                 rte_prefetch0(hdr);
 810         }
 811
 812         /*
 813          * A virtio driver normally uses at least 2 desc buffers
 814          * for Tx: the first for storing the header, and others
 815          * for storing the data.
 816          */
 817         if (likely((desc->len == dev->vhost_hlen) &&
 818                    (desc->flags & VRING_DESC_F_NEXT) != 0)) {
 819                 desc = &descs[desc->next];
 820                 if (unlikely(desc->flags & VRING_DESC_F_INDIRECT))
 821                         return -1;
 822
 823                 desc_addr = rte_vhost_gpa_to_vva(dev->mem, desc->addr);
 824                 if (unlikely(!desc_addr))
 825                         return -1;
 826
 827                 desc_offset = 0;
 828                 desc_avail  = desc->len;
 829                 nr_desc    += 1;
 830         } else {
 831                 desc_avail  = desc->len - dev->vhost_hlen;
 832                 desc_offset = dev->vhost_hlen;
 833         }
 834
 835         rte_prefetch0((void *)(uintptr_t)(desc_addr + desc_offset));
 836
 837         PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset), desc_avail, 0);
 838
 839         mbuf_offset = 0;
 840         mbuf_avail  = m->buf_len - RTE_PKTMBUF_HEADROOM;
 841         while (1) {
 842                 uint64_t hpa;
 843
 844                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 845
 846                 /*
 847                  * A desc buf might across two host physical pages that are
 848                  * not continuous. In such case (gpa_to_hpa returns 0), data
 849                  * will be copied even though zero copy is enabled.
 850                  */
 851                 if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
 852                                         desc->addr + desc_offset, cpy_len)))) {
 853                         cur->data_len = cpy_len;
 854                         cur->data_off = 0;
 855                         cur->buf_addr = (void *)(uintptr_t)desc_addr;
 856                         cur->buf_physaddr = hpa;
 857
 858                         /*
 859                          * In zero copy mode, one mbuf can only reference data
 860                          * for one or partial of one desc buff.
 861                          */
 862                         mbuf_avail = cpy_len;
 863                 } else {
 864                         rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
 865                                                            mbuf_offset),
 866                                 (void *)((uintptr_t)(desc_addr + desc_offset)),
 867                                 cpy_len);
 868                 }
 869
 870                 mbuf_avail  -= cpy_len;
 871                 mbuf_offset += cpy_len;
 872                 desc_avail  -= cpy_len;
 873                 desc_offset += cpy_len;
 874
 875                 /* This desc reaches to its end, get the next one */
 876                 if (desc_avail == 0) {
 877                         if ((desc->flags & VRING_DESC_F_NEXT) == 0)
 878                                 break;
 879
 880                         if (unlikely(desc->next >= max_desc ||
 881                                      ++nr_desc > max_desc))
 882                                 return -1;
 883                         desc = &descs[desc->next];
 884                         if (unlikely(desc->flags & VRING_DESC_F_INDIRECT))
 885                                 return -1;
 886
 887                         desc_addr = rte_vhost_gpa_to_vva(dev->mem, desc->addr);
 888                         if (unlikely(!desc_addr))
 889                                 return -1;
 890
 891                         rte_prefetch0((void *)(uintptr_t)desc_addr);
 892
 893                         desc_offset = 0;
 894                         desc_avail  = desc->len;
 895
 896                         PRINT_PACKET(dev, (uintptr_t)desc_addr, desc->len, 0);
 897                 }
 898
 899                 /*
 900                  * This mbuf reaches to its end, get a new one
 901                  * to hold more data.
 902                  */
 903                 if (mbuf_avail == 0) {
 904                         cur = rte_pktmbuf_alloc(mbuf_pool);
 905                         if (unlikely(cur == NULL)) {
 906                                 RTE_LOG(ERR, VHOST_DATA, "Failed to "
 907                                         "allocate memory for mbuf.\n");
 908                                 return -1;
 909                         }
 910
 911                         prev->next = cur;
 912                         prev->data_len = mbuf_offset;
 913                         m->nb_segs += 1;
 914                         m->pkt_len += mbuf_offset;
 915                         prev = cur;
 916
 917                         mbuf_offset = 0;
 918                         mbuf_avail  = cur->buf_len - RTE_PKTMBUF_HEADROOM;
 919                 }
 920         }
 921
 922         prev->data_len = mbuf_offset;
 923         m->pkt_len    += mbuf_offset;
 924
 925         if (hdr)
 926                 vhost_dequeue_offload(hdr, m);
 927
 928         return 0;
 929 }
 930
 931 static inline void __attribute__((always_inline))
 932 update_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq,
 933                  uint32_t used_idx, uint32_t desc_idx)
 934 {
 935         vq->used->ring[used_idx].id  = desc_idx;
 936         vq->used->ring[used_idx].len = 0;
 937         vhost_log_used_vring(dev, vq,
 938                         offsetof(struct vring_used, ring[used_idx]),
 939                         sizeof(vq->used->ring[used_idx]));
 940 }
 941
 942 static inline void __attribute__((always_inline))
 943 update_used_idx(struct virtio_net *dev, struct vhost_virtqueue *vq,
 944                 uint32_t count)
 945 {
 946         if (unlikely(count == 0))
 947                 return;
 948
 949         rte_smp_wmb();
 950         rte_smp_rmb();
 951
 952         vq->used->idx += count;
 953         vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
 954                         sizeof(vq->used->idx));
 955
 956         /* Kick guest if required. */
 957         if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
 958                         && (vq->callfd >= 0))
 959                 eventfd_write(vq->callfd, (eventfd_t)1);
 960 }
 961
 962 static inline struct zcopy_mbuf *__attribute__((always_inline))
 963 get_zmbuf(struct vhost_virtqueue *vq)
 964 {
 965         uint16_t i;
 966         uint16_t last;
 967         int tries = 0;
 968
 969         /* search [last_zmbuf_idx, zmbuf_size) */
 970         i = vq->last_zmbuf_idx;
 971         last = vq->zmbuf_size;
 972
 973 again:
 974         for (; i < last; i++) {
 975                 if (vq->zmbufs[i].in_use == 0) {
 976                         vq->last_zmbuf_idx = i + 1;
 977                         vq->zmbufs[i].in_use = 1;
 978                         return &vq->zmbufs[i];
 979                 }
 980         }
 981
 982         tries++;
 983         if (tries == 1) {
 984                 /* search [0, last_zmbuf_idx) */
 985                 i = 0;
 986                 last = vq->last_zmbuf_idx;
 987                 goto again;
 988         }
 989
 990         return NULL;
 991 }
 992
 993 static inline bool __attribute__((always_inline))
 994 mbuf_is_consumed(struct rte_mbuf *m)
 995 {
 996         while (m) {
 997                 if (rte_mbuf_refcnt_read(m) > 1)
 998                         return false;
 999                 m = m->next;
1000         }
1001
1002         return true;
1003 }
1004
1005 uint16_t
1006 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
1007         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1008 {
1009         struct virtio_net *dev;
1010         struct rte_mbuf *rarp_mbuf = NULL;
1011         struct vhost_virtqueue *vq;
1012         uint32_t desc_indexes[MAX_PKT_BURST];
1013         uint32_t used_idx;
1014         uint32_t i = 0;
1015         uint16_t free_entries;
1016         uint16_t avail_idx;
1017
1018         dev = get_device(vid);
1019         if (!dev)
1020                 return 0;
1021
1022         if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
1023                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
1024                         dev->vid, __func__, queue_id);
1025                 return 0;
1026         }
1027
1028         vq = dev->virtqueue[queue_id];
1029         if (unlikely(vq->enabled == 0))
1030                 return 0;
1031
1032         if (unlikely(dev->dequeue_zero_copy)) {
1033                 struct zcopy_mbuf *zmbuf, *next;
1034                 int nr_updated = 0;
1035
1036                 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1037                      zmbuf != NULL; zmbuf = next) {
1038                         next = TAILQ_NEXT(zmbuf, next);
1039
1040                         if (mbuf_is_consumed(zmbuf->mbuf)) {
1041                                 used_idx = vq->last_used_idx++ & (vq->size - 1);
1042                                 update_used_ring(dev, vq, used_idx,
1043                                                  zmbuf->desc_idx);
1044                                 nr_updated += 1;
1045
1046                                 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1047                                 rte_pktmbuf_free(zmbuf->mbuf);
1048                                 put_zmbuf(zmbuf);
1049                                 vq->nr_zmbuf -= 1;
1050                         }
1051                 }
1052
1053                 update_used_idx(dev, vq, nr_updated);
1054         }
1055
1056         /*
1057          * Construct a RARP broadcast packet, and inject it to the "pkts"
1058          * array, to looks like that guest actually send such packet.
1059          *
1060          * Check user_send_rarp() for more information.
1061          *
1062          * broadcast_rarp shares a cacheline in the virtio_net structure
1063          * with some fields that are accessed during enqueue and
1064          * rte_atomic16_cmpset() causes a write if using cmpxchg. This could
1065          * result in false sharing between enqueue and dequeue.
1066          *
1067          * Prevent unnecessary false sharing by reading broadcast_rarp first
1068          * and only performing cmpset if the read indicates it is likely to
1069          * be set.
1070          */
1071
1072         if (unlikely(rte_atomic16_read(&dev->broadcast_rarp) &&
1073                         rte_atomic16_cmpset((volatile uint16_t *)
1074                                 &dev->broadcast_rarp.cnt, 1, 0))) {
1075
1076                 rarp_mbuf = rte_pktmbuf_alloc(mbuf_pool);
1077                 if (rarp_mbuf == NULL) {
1078                         RTE_LOG(ERR, VHOST_DATA,
1079                                 "Failed to allocate memory for mbuf.\n");
1080                         return 0;
1081                 }
1082
1083                 if (make_rarp_packet(rarp_mbuf, &dev->mac)) {
1084                         rte_pktmbuf_free(rarp_mbuf);
1085                         rarp_mbuf = NULL;
1086                 } else {
1087                         count -= 1;
1088                 }
1089         }
1090
1091         free_entries = *((volatile uint16_t *)&vq->avail->idx) -
1092                         vq->last_avail_idx;
1093         if (free_entries == 0)
1094                 goto out;
1095
1096         LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1097
1098         /* Prefetch available and used ring */
1099         avail_idx = vq->last_avail_idx & (vq->size - 1);
1100         used_idx  = vq->last_used_idx  & (vq->size - 1);
1101         rte_prefetch0(&vq->avail->ring[avail_idx]);
1102         rte_prefetch0(&vq->used->ring[used_idx]);
1103
1104         count = RTE_MIN(count, MAX_PKT_BURST);
1105         count = RTE_MIN(count, free_entries);
1106         LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1107                         dev->vid, count);
1108
1109         /* Retrieve all of the head indexes first to avoid caching issues. */
1110         for (i = 0; i < count; i++) {
1111                 avail_idx = (vq->last_avail_idx + i) & (vq->size - 1);
1112                 used_idx  = (vq->last_used_idx  + i) & (vq->size - 1);
1113                 desc_indexes[i] = vq->avail->ring[avail_idx];
1114
1115                 if (likely(dev->dequeue_zero_copy == 0))
1116                         update_used_ring(dev, vq, used_idx, desc_indexes[i]);
1117         }
1118
1119         /* Prefetch descriptor index. */
1120         rte_prefetch0(&vq->desc[desc_indexes[0]]);
1121         for (i = 0; i < count; i++) {
1122                 struct vring_desc *desc;
1123                 uint16_t sz, idx;
1124                 int err;
1125
1126                 if (likely(i + 1 < count))
1127                         rte_prefetch0(&vq->desc[desc_indexes[i + 1]]);
1128
1129                 if (vq->desc[desc_indexes[i]].flags & VRING_DESC_F_INDIRECT) {
1130                         desc = (struct vring_desc *)(uintptr_t)
1131                                 rte_vhost_gpa_to_vva(dev->mem,
1132                                         vq->desc[desc_indexes[i]].addr);
1133                         if (unlikely(!desc))
1134                                 break;
1135
1136                         rte_prefetch0(desc);
1137                         sz = vq->desc[desc_indexes[i]].len / sizeof(*desc);
1138                         idx = 0;
1139                 } else {
1140                         desc = vq->desc;
1141                         sz = vq->size;
1142                         idx = desc_indexes[i];
1143                 }
1144
1145                 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
1146                 if (unlikely(pkts[i] == NULL)) {
1147                         RTE_LOG(ERR, VHOST_DATA,
1148                                 "Failed to allocate memory for mbuf.\n");
1149                         break;
1150                 }
1151
1152                 err = copy_desc_to_mbuf(dev, desc, sz, pkts[i], idx, mbuf_pool);
1153                 if (unlikely(err)) {
1154                         rte_pktmbuf_free(pkts[i]);
1155                         break;
1156                 }
1157
1158                 if (unlikely(dev->dequeue_zero_copy)) {
1159                         struct zcopy_mbuf *zmbuf;
1160
1161                         zmbuf = get_zmbuf(vq);
1162                         if (!zmbuf) {
1163                                 rte_pktmbuf_free(pkts[i]);
1164                                 break;
1165                         }
1166                         zmbuf->mbuf = pkts[i];
1167                         zmbuf->desc_idx = desc_indexes[i];
1168
1169                         /*
1170                          * Pin lock the mbuf; we will check later to see
1171                          * whether the mbuf is freed (when we are the last
1172                          * user) or not. If that's the case, we then could
1173                          * update the used ring safely.
1174                          */
1175                         rte_mbuf_refcnt_update(pkts[i], 1);
1176
1177                         vq->nr_zmbuf += 1;
1178                         TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1179                 }
1180         }
1181         vq->last_avail_idx += i;
1182
1183         if (likely(dev->dequeue_zero_copy == 0)) {
1184                 vq->last_used_idx += i;
1185                 update_used_idx(dev, vq, i);
1186         }
1187
1188 out:
1189         if (unlikely(rarp_mbuf != NULL)) {
1190                 /*
1191                  * Inject it to the head of "pkts" array, so that switch's mac
1192                  * learning table will get updated first.
1193                  */
1194                 memmove(&pkts[1], pkts, i * sizeof(struct rte_mbuf *));
1195                 pkts[0] = rarp_mbuf;
1196                 i += 1;
1197         }
1198
1199         return i;
1200 }