lib/librte_vhost/virtio_net.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2016 Intel Corporation
   3  */
   4
   5 #include <stdint.h>
   6 #include <stdbool.h>
   7 #include <linux/virtio_net.h>
   8
   9 #include <rte_mbuf.h>
  10 #include <rte_memcpy.h>
  11 #include <rte_ether.h>
  12 #include <rte_ip.h>
  13 #include <rte_vhost.h>
  14 #include <rte_tcp.h>
  15 #include <rte_udp.h>
  16 #include <rte_sctp.h>
  17 #include <rte_arp.h>
  18
  19 #include "iotlb.h"
  20 #include "vhost.h"
  21
  22 #define MAX_PKT_BURST 32
  23
  24 #define MAX_BATCH_LEN 256
  25
  26 static bool
  27 is_valid_virt_queue_idx(uint32_t idx, int is_tx, uint32_t nr_vring)
  28 {
  29         return (is_tx ^ (idx & 1)) == 0 && idx < nr_vring;
  30 }
  31
  32 static __rte_always_inline void
  33 do_flush_shadow_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq,
  34                           uint16_t to, uint16_t from, uint16_t size)
  35 {
  36         rte_memcpy(&vq->used->ring[to],
  37                         &vq->shadow_used_ring[from],
  38                         size * sizeof(struct vring_used_elem));
  39         vhost_log_used_vring(dev, vq,
  40                         offsetof(struct vring_used, ring[to]),
  41                         size * sizeof(struct vring_used_elem));
  42 }
  43
  44 static __rte_always_inline void
  45 flush_shadow_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq)
  46 {
  47         uint16_t used_idx = vq->last_used_idx & (vq->size - 1);
  48
  49         if (used_idx + vq->shadow_used_idx <= vq->size) {
  50                 do_flush_shadow_used_ring(dev, vq, used_idx, 0,
  51                                           vq->shadow_used_idx);
  52         } else {
  53                 uint16_t size;
  54
  55                 /* update used ring interval [used_idx, vq->size] */
  56                 size = vq->size - used_idx;
  57                 do_flush_shadow_used_ring(dev, vq, used_idx, 0, size);
  58
  59                 /* update the left half used ring interval [0, left_size] */
  60                 do_flush_shadow_used_ring(dev, vq, 0, size,
  61                                           vq->shadow_used_idx - size);
  62         }
  63         vq->last_used_idx += vq->shadow_used_idx;
  64
  65         rte_smp_wmb();
  66
  67         *(volatile uint16_t *)&vq->used->idx += vq->shadow_used_idx;
  68         vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
  69                 sizeof(vq->used->idx));
  70 }
  71
  72 static __rte_always_inline void
  73 update_shadow_used_ring(struct vhost_virtqueue *vq,
  74                          uint16_t desc_idx, uint16_t len)
  75 {
  76         uint16_t i = vq->shadow_used_idx++;
  77
  78         vq->shadow_used_ring[i].id  = desc_idx;
  79         vq->shadow_used_ring[i].len = len;
  80 }
  81
  82 static inline void
  83 do_data_copy_enqueue(struct virtio_net *dev, struct vhost_virtqueue *vq)
  84 {
  85         struct batch_copy_elem *elem = vq->batch_copy_elems;
  86         uint16_t count = vq->batch_copy_nb_elems;
  87         int i;
  88
  89         for (i = 0; i < count; i++) {
  90                 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
  91                 vhost_log_write(dev, elem[i].log_addr, elem[i].len);
  92                 PRINT_PACKET(dev, (uintptr_t)elem[i].dst, elem[i].len, 0);
  93         }
  94 }
  95
  96 static inline void
  97 do_data_copy_dequeue(struct vhost_virtqueue *vq)
  98 {
  99         struct batch_copy_elem *elem = vq->batch_copy_elems;
 100         uint16_t count = vq->batch_copy_nb_elems;
 101         int i;
 102
 103         for (i = 0; i < count; i++)
 104                 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
 105 }
 106
 107 /* avoid write operation when necessary, to lessen cache issues */
 108 #define ASSIGN_UNLESS_EQUAL(var, val) do {      \
 109         if ((var) != (val))                     \
 110                 (var) = (val);                  \
 111 } while (0)
 112
 113 static void
 114 virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
 115 {
 116         uint64_t csum_l4 = m_buf->ol_flags & PKT_TX_L4_MASK;
 117
 118         if (m_buf->ol_flags & PKT_TX_TCP_SEG)
 119                 csum_l4 |= PKT_TX_TCP_CKSUM;
 120
 121         if (csum_l4) {
 122                 net_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 123                 net_hdr->csum_start = m_buf->l2_len + m_buf->l3_len;
 124
 125                 switch (csum_l4) {
 126                 case PKT_TX_TCP_CKSUM:
 127                         net_hdr->csum_offset = (offsetof(struct tcp_hdr,
 128                                                 cksum));
 129                         break;
 130                 case PKT_TX_UDP_CKSUM:
 131                         net_hdr->csum_offset = (offsetof(struct udp_hdr,
 132                                                 dgram_cksum));
 133                         break;
 134                 case PKT_TX_SCTP_CKSUM:
 135                         net_hdr->csum_offset = (offsetof(struct sctp_hdr,
 136                                                 cksum));
 137                         break;
 138                 }
 139         } else {
 140                 ASSIGN_UNLESS_EQUAL(net_hdr->csum_start, 0);
 141                 ASSIGN_UNLESS_EQUAL(net_hdr->csum_offset, 0);
 142                 ASSIGN_UNLESS_EQUAL(net_hdr->flags, 0);
 143         }
 144
 145         /* IP cksum verification cannot be bypassed, then calculate here */
 146         if (m_buf->ol_flags & PKT_TX_IP_CKSUM) {
 147                 struct ipv4_hdr *ipv4_hdr;
 148
 149                 ipv4_hdr = rte_pktmbuf_mtod_offset(m_buf, struct ipv4_hdr *,
 150                                                    m_buf->l2_len);
 151                 ipv4_hdr->hdr_checksum = rte_ipv4_cksum(ipv4_hdr);
 152         }
 153
 154         if (m_buf->ol_flags & PKT_TX_TCP_SEG) {
 155                 if (m_buf->ol_flags & PKT_TX_IPV4)
 156                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 157                 else
 158                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 159                 net_hdr->gso_size = m_buf->tso_segsz;
 160                 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len
 161                                         + m_buf->l4_len;
 162         } else {
 163                 ASSIGN_UNLESS_EQUAL(net_hdr->gso_type, 0);
 164                 ASSIGN_UNLESS_EQUAL(net_hdr->gso_size, 0);
 165                 ASSIGN_UNLESS_EQUAL(net_hdr->hdr_len, 0);
 166         }
 167 }
 168
 169 static __rte_always_inline int
 170 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 171                   struct vring_desc *descs, struct rte_mbuf *m,
 172                   uint16_t desc_idx, uint32_t size)
 173 {
 174         uint32_t desc_avail, desc_offset;
 175         uint32_t mbuf_avail, mbuf_offset;
 176         uint32_t cpy_len;
 177         struct vring_desc *desc;
 178         uint64_t desc_addr;
 179         /* A counter to avoid desc dead loop chain */
 180         uint16_t nr_desc = 1;
 181         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 182         uint16_t copy_nb = vq->batch_copy_nb_elems;
 183         int error = 0;
 184
 185         desc = &descs[desc_idx];
 186         desc_addr = vhost_iova_to_vva(dev, vq, desc->addr,
 187                                         desc->len, VHOST_ACCESS_RW);
 188         /*
 189          * Checking of 'desc_addr' placed outside of 'unlikely' macro to avoid
 190          * performance issue with some versions of gcc (4.8.4 and 5.3.0) which
 191          * otherwise stores offset on the stack instead of in a register.
 192          */
 193         if (unlikely(desc->len < dev->vhost_hlen) || !desc_addr) {
 194                 error = -1;
 195                 goto out;
 196         }
 197
 198         rte_prefetch0((void *)(uintptr_t)desc_addr);
 199
 200         virtio_enqueue_offload(m, (struct virtio_net_hdr *)(uintptr_t)desc_addr);
 201         vhost_log_write(dev, desc->addr, dev->vhost_hlen);
 202         PRINT_PACKET(dev, (uintptr_t)desc_addr, dev->vhost_hlen, 0);
 203
 204         desc_offset = dev->vhost_hlen;
 205         desc_avail  = desc->len - dev->vhost_hlen;
 206
 207         mbuf_avail  = rte_pktmbuf_data_len(m);
 208         mbuf_offset = 0;
 209         while (mbuf_avail != 0 || m->next != NULL) {
 210                 /* done with current mbuf, fetch next */
 211                 if (mbuf_avail == 0) {
 212                         m = m->next;
 213
 214                         mbuf_offset = 0;
 215                         mbuf_avail  = rte_pktmbuf_data_len(m);
 216                 }
 217
 218                 /* done with current desc buf, fetch next */
 219                 if (desc_avail == 0) {
 220                         if ((desc->flags & VRING_DESC_F_NEXT) == 0) {
 221                                 /* Room in vring buffer is not enough */
 222                                 error = -1;
 223                                 goto out;
 224                         }
 225                         if (unlikely(desc->next >= size || ++nr_desc > size)) {
 226                                 error = -1;
 227                                 goto out;
 228                         }
 229
 230                         desc = &descs[desc->next];
 231                         desc_addr = vhost_iova_to_vva(dev, vq, desc->addr,
 232                                                         desc->len,
 233                                                         VHOST_ACCESS_RW);
 234                         if (unlikely(!desc_addr)) {
 235                                 error = -1;
 236                                 goto out;
 237                         }
 238
 239                         desc_offset = 0;
 240                         desc_avail  = desc->len;
 241                 }
 242
 243                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 244                 if (likely(cpy_len > MAX_BATCH_LEN || copy_nb >= vq->size)) {
 245                         rte_memcpy((void *)((uintptr_t)(desc_addr +
 246                                                         desc_offset)),
 247                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 248                                 cpy_len);
 249                         vhost_log_write(dev, desc->addr + desc_offset, cpy_len);
 250                         PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
 251                                      cpy_len, 0);
 252                 } else {
 253                         batch_copy[copy_nb].dst =
 254                                 (void *)((uintptr_t)(desc_addr + desc_offset));
 255                         batch_copy[copy_nb].src =
 256                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
 257                         batch_copy[copy_nb].log_addr = desc->addr + desc_offset;
 258                         batch_copy[copy_nb].len = cpy_len;
 259                         copy_nb++;
 260                 }
 261
 262                 mbuf_avail  -= cpy_len;
 263                 mbuf_offset += cpy_len;
 264                 desc_avail  -= cpy_len;
 265                 desc_offset += cpy_len;
 266         }
 267
 268 out:
 269         vq->batch_copy_nb_elems = copy_nb;
 270
 271         return error;
 272 }
 273
 274 /**
 275  * This function adds buffers to the virtio devices RX virtqueue. Buffers can
 276  * be received from the physical port or from another virtio device. A packet
 277  * count is returned to indicate the number of packets that are successfully
 278  * added to the RX queue. This function works when the mbuf is scattered, but
 279  * it doesn't support the mergeable feature.
 280  */
 281 static __rte_always_inline uint32_t
 282 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
 283               struct rte_mbuf **pkts, uint32_t count)
 284 {
 285         struct vhost_virtqueue *vq;
 286         uint16_t avail_idx, free_entries, start_idx;
 287         uint16_t desc_indexes[MAX_PKT_BURST];
 288         struct vring_desc *descs;
 289         uint16_t used_idx;
 290         uint32_t i, sz;
 291
 292         LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
 293         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
 294                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
 295                         dev->vid, __func__, queue_id);
 296                 return 0;
 297         }
 298
 299         vq = dev->virtqueue[queue_id];
 300         if (unlikely(vq->enabled == 0))
 301                 return 0;
 302
 303         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 304                 vhost_user_iotlb_rd_lock(vq);
 305
 306         if (unlikely(vq->access_ok == 0)) {
 307                 if (unlikely(vring_translate(dev, vq) < 0)) {
 308                         count = 0;
 309                         goto out;
 310                 }
 311         }
 312
 313         avail_idx = *((volatile uint16_t *)&vq->avail->idx);
 314         start_idx = vq->last_used_idx;
 315         free_entries = avail_idx - start_idx;
 316         count = RTE_MIN(count, free_entries);
 317         count = RTE_MIN(count, (uint32_t)MAX_PKT_BURST);
 318         if (count == 0)
 319                 goto out;
 320
 321         LOG_DEBUG(VHOST_DATA, "(%d) start_idx %d | end_idx %d\n",
 322                 dev->vid, start_idx, start_idx + count);
 323
 324         vq->batch_copy_nb_elems = 0;
 325
 326         /* Retrieve all of the desc indexes first to avoid caching issues. */
 327         rte_prefetch0(&vq->avail->ring[start_idx & (vq->size - 1)]);
 328         for (i = 0; i < count; i++) {
 329                 used_idx = (start_idx + i) & (vq->size - 1);
 330                 desc_indexes[i] = vq->avail->ring[used_idx];
 331                 vq->used->ring[used_idx].id = desc_indexes[i];
 332                 vq->used->ring[used_idx].len = pkts[i]->pkt_len +
 333                                                dev->vhost_hlen;
 334                 vhost_log_used_vring(dev, vq,
 335                         offsetof(struct vring_used, ring[used_idx]),
 336                         sizeof(vq->used->ring[used_idx]));
 337         }
 338
 339         rte_prefetch0(&vq->desc[desc_indexes[0]]);
 340         for (i = 0; i < count; i++) {
 341                 uint16_t desc_idx = desc_indexes[i];
 342                 int err;
 343
 344                 if (vq->desc[desc_idx].flags & VRING_DESC_F_INDIRECT) {
 345                         descs = (struct vring_desc *)(uintptr_t)
 346                                 vhost_iova_to_vva(dev,
 347                                                 vq, vq->desc[desc_idx].addr,
 348                                                 vq->desc[desc_idx].len,
 349                                                 VHOST_ACCESS_RO);
 350                         if (unlikely(!descs)) {
 351                                 count = i;
 352                                 break;
 353                         }
 354
 355                         desc_idx = 0;
 356                         sz = vq->desc[desc_idx].len / sizeof(*descs);
 357                 } else {
 358                         descs = vq->desc;
 359                         sz = vq->size;
 360                 }
 361
 362                 err = copy_mbuf_to_desc(dev, vq, descs, pkts[i], desc_idx, sz);
 363                 if (unlikely(err)) {
 364                         count = i;
 365                         break;
 366                 }
 367
 368                 if (i + 1 < count)
 369                         rte_prefetch0(&vq->desc[desc_indexes[i+1]]);
 370         }
 371
 372         do_data_copy_enqueue(dev, vq);
 373
 374         rte_smp_wmb();
 375
 376         *(volatile uint16_t *)&vq->used->idx += count;
 377         vq->last_used_idx += count;
 378         vhost_log_used_vring(dev, vq,
 379                 offsetof(struct vring_used, idx),
 380                 sizeof(vq->used->idx));
 381
 382         vhost_vring_call(vq);
 383 out:
 384         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 385                 vhost_user_iotlb_rd_unlock(vq);
 386
 387         return count;
 388 }
 389
 390 static __rte_always_inline int
 391 fill_vec_buf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 392                          uint32_t avail_idx, uint32_t *vec_idx,
 393                          struct buf_vector *buf_vec, uint16_t *desc_chain_head,
 394                          uint16_t *desc_chain_len)
 395 {
 396         uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
 397         uint32_t vec_id = *vec_idx;
 398         uint32_t len    = 0;
 399         struct vring_desc *descs = vq->desc;
 400
 401         *desc_chain_head = idx;
 402
 403         if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
 404                 descs = (struct vring_desc *)(uintptr_t)
 405                         vhost_iova_to_vva(dev, vq, vq->desc[idx].addr,
 406                                                 vq->desc[idx].len,
 407                                                 VHOST_ACCESS_RO);
 408                 if (unlikely(!descs))
 409                         return -1;
 410
 411                 idx = 0;
 412         }
 413
 414         while (1) {
 415                 if (unlikely(vec_id >= BUF_VECTOR_MAX || idx >= vq->size))
 416                         return -1;
 417
 418                 len += descs[idx].len;
 419                 buf_vec[vec_id].buf_addr = descs[idx].addr;
 420                 buf_vec[vec_id].buf_len  = descs[idx].len;
 421                 buf_vec[vec_id].desc_idx = idx;
 422                 vec_id++;
 423
 424                 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
 425                         break;
 426
 427                 idx = descs[idx].next;
 428         }
 429
 430         *desc_chain_len = len;
 431         *vec_idx = vec_id;
 432
 433         return 0;
 434 }
 435
 436 /*
 437  * Returns -1 on fail, 0 on success
 438  */
 439 static inline int
 440 reserve_avail_buf_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
 441                                 uint32_t size, struct buf_vector *buf_vec,
 442                                 uint16_t *num_buffers, uint16_t avail_head)
 443 {
 444         uint16_t cur_idx;
 445         uint32_t vec_idx = 0;
 446         uint16_t tries = 0;
 447
 448         uint16_t head_idx = 0;
 449         uint16_t len = 0;
 450
 451         *num_buffers = 0;
 452         cur_idx  = vq->last_avail_idx;
 453
 454         while (size > 0) {
 455                 if (unlikely(cur_idx == avail_head))
 456                         return -1;
 457
 458                 if (unlikely(fill_vec_buf(dev, vq, cur_idx, &vec_idx, buf_vec,
 459                                                 &head_idx, &len) < 0))
 460                         return -1;
 461                 len = RTE_MIN(len, size);
 462                 update_shadow_used_ring(vq, head_idx, len);
 463                 size -= len;
 464
 465                 cur_idx++;
 466                 tries++;
 467                 *num_buffers += 1;
 468
 469                 /*
 470                  * if we tried all available ring items, and still
 471                  * can't get enough buf, it means something abnormal
 472                  * happened.
 473                  */
 474                 if (unlikely(tries >= vq->size))
 475                         return -1;
 476         }
 477
 478         return 0;
 479 }
 480
 481 static __rte_always_inline int
 482 copy_mbuf_to_desc_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
 483                             struct rte_mbuf *m, struct buf_vector *buf_vec,
 484                             uint16_t num_buffers)
 485 {
 486         uint32_t vec_idx = 0;
 487         uint64_t desc_addr;
 488         uint32_t mbuf_offset, mbuf_avail;
 489         uint32_t desc_offset, desc_avail;
 490         uint32_t cpy_len;
 491         uint64_t hdr_addr, hdr_phys_addr;
 492         struct rte_mbuf *hdr_mbuf;
 493         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 494         uint16_t copy_nb = vq->batch_copy_nb_elems;
 495         int error = 0;
 496
 497         if (unlikely(m == NULL)) {
 498                 error = -1;
 499                 goto out;
 500         }
 501
 502         desc_addr = vhost_iova_to_vva(dev, vq, buf_vec[vec_idx].buf_addr,
 503                                                 buf_vec[vec_idx].buf_len,
 504                                                 VHOST_ACCESS_RW);
 505         if (buf_vec[vec_idx].buf_len < dev->vhost_hlen || !desc_addr) {
 506                 error = -1;
 507                 goto out;
 508         }
 509
 510         hdr_mbuf = m;
 511         hdr_addr = desc_addr;
 512         hdr_phys_addr = buf_vec[vec_idx].buf_addr;
 513         rte_prefetch0((void *)(uintptr_t)hdr_addr);
 514
 515         LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
 516                 dev->vid, num_buffers);
 517
 518         desc_avail  = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
 519         desc_offset = dev->vhost_hlen;
 520
 521         mbuf_avail  = rte_pktmbuf_data_len(m);
 522         mbuf_offset = 0;
 523         while (mbuf_avail != 0 || m->next != NULL) {
 524                 /* done with current desc buf, get the next one */
 525                 if (desc_avail == 0) {
 526                         vec_idx++;
 527                         desc_addr =
 528                                 vhost_iova_to_vva(dev, vq,
 529                                         buf_vec[vec_idx].buf_addr,
 530                                         buf_vec[vec_idx].buf_len,
 531                                         VHOST_ACCESS_RW);
 532                         if (unlikely(!desc_addr)) {
 533                                 error = -1;
 534                                 goto out;
 535                         }
 536
 537                         /* Prefetch buffer address. */
 538                         rte_prefetch0((void *)(uintptr_t)desc_addr);
 539                         desc_offset = 0;
 540                         desc_avail  = buf_vec[vec_idx].buf_len;
 541                 }
 542
 543                 /* done with current mbuf, get the next one */
 544                 if (mbuf_avail == 0) {
 545                         m = m->next;
 546
 547                         mbuf_offset = 0;
 548                         mbuf_avail  = rte_pktmbuf_data_len(m);
 549                 }
 550
 551                 if (hdr_addr) {
 552                         struct virtio_net_hdr_mrg_rxbuf *hdr;
 553
 554                         hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)
 555                                 hdr_addr;
 556                         virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
 557                         ASSIGN_UNLESS_EQUAL(hdr->num_buffers, num_buffers);
 558
 559                         vhost_log_write(dev, hdr_phys_addr, dev->vhost_hlen);
 560                         PRINT_PACKET(dev, (uintptr_t)hdr_addr,
 561                                      dev->vhost_hlen, 0);
 562
 563                         hdr_addr = 0;
 564                 }
 565
 566                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 567
 568                 if (likely(cpy_len > MAX_BATCH_LEN || copy_nb >= vq->size)) {
 569                         rte_memcpy((void *)((uintptr_t)(desc_addr +
 570                                                         desc_offset)),
 571                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 572                                 cpy_len);
 573                         vhost_log_write(dev,
 574                                 buf_vec[vec_idx].buf_addr + desc_offset,
 575                                 cpy_len);
 576                         PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
 577                                 cpy_len, 0);
 578                 } else {
 579                         batch_copy[copy_nb].dst =
 580                                 (void *)((uintptr_t)(desc_addr + desc_offset));
 581                         batch_copy[copy_nb].src =
 582                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
 583                         batch_copy[copy_nb].log_addr =
 584                                 buf_vec[vec_idx].buf_addr + desc_offset;
 585                         batch_copy[copy_nb].len = cpy_len;
 586                         copy_nb++;
 587                 }
 588
 589                 mbuf_avail  -= cpy_len;
 590                 mbuf_offset += cpy_len;
 591                 desc_avail  -= cpy_len;
 592                 desc_offset += cpy_len;
 593         }
 594
 595 out:
 596         vq->batch_copy_nb_elems = copy_nb;
 597
 598         return error;
 599 }
 600
 601 static __rte_always_inline uint32_t
 602 virtio_dev_merge_rx(struct virtio_net *dev, uint16_t queue_id,
 603         struct rte_mbuf **pkts, uint32_t count)
 604 {
 605         struct vhost_virtqueue *vq;
 606         uint32_t pkt_idx = 0;
 607         uint16_t num_buffers;
 608         struct buf_vector buf_vec[BUF_VECTOR_MAX];
 609         uint16_t avail_head;
 610
 611         LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
 612         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
 613                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
 614                         dev->vid, __func__, queue_id);
 615                 return 0;
 616         }
 617
 618         vq = dev->virtqueue[queue_id];
 619         if (unlikely(vq->enabled == 0))
 620                 return 0;
 621
 622         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 623                 vhost_user_iotlb_rd_lock(vq);
 624
 625         if (unlikely(vq->access_ok == 0))
 626                 if (unlikely(vring_translate(dev, vq) < 0))
 627                         goto out;
 628
 629         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
 630         if (count == 0)
 631                 goto out;
 632
 633         vq->batch_copy_nb_elems = 0;
 634
 635         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
 636
 637         vq->shadow_used_idx = 0;
 638         avail_head = *((volatile uint16_t *)&vq->avail->idx);
 639         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
 640                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
 641
 642                 if (unlikely(reserve_avail_buf_mergeable(dev, vq,
 643                                                 pkt_len, buf_vec, &num_buffers,
 644                                                 avail_head) < 0)) {
 645                         LOG_DEBUG(VHOST_DATA,
 646                                 "(%d) failed to get enough desc from vring\n",
 647                                 dev->vid);
 648                         vq->shadow_used_idx -= num_buffers;
 649                         break;
 650                 }
 651
 652                 LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
 653                         dev->vid, vq->last_avail_idx,
 654                         vq->last_avail_idx + num_buffers);
 655
 656                 if (copy_mbuf_to_desc_mergeable(dev, vq, pkts[pkt_idx],
 657                                                 buf_vec, num_buffers) < 0) {
 658                         vq->shadow_used_idx -= num_buffers;
 659                         break;
 660                 }
 661
 662                 vq->last_avail_idx += num_buffers;
 663         }
 664
 665         do_data_copy_enqueue(dev, vq);
 666
 667         if (likely(vq->shadow_used_idx)) {
 668                 flush_shadow_used_ring(dev, vq);
 669                 vhost_vring_call(vq);
 670         }
 671
 672 out:
 673         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 674                 vhost_user_iotlb_rd_unlock(vq);
 675
 676         return pkt_idx;
 677 }
 678
 679 uint16_t
 680 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
 681         struct rte_mbuf **pkts, uint16_t count)
 682 {
 683         struct virtio_net *dev = get_device(vid);
 684
 685         if (!dev)
 686                 return 0;
 687
 688         if (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF))
 689                 return virtio_dev_merge_rx(dev, queue_id, pkts, count);
 690         else
 691                 return virtio_dev_rx(dev, queue_id, pkts, count);
 692 }
 693
 694 static inline bool
 695 virtio_net_with_host_offload(struct virtio_net *dev)
 696 {
 697         if (dev->features &
 698                         ((1ULL << VIRTIO_NET_F_CSUM) |
 699                          (1ULL << VIRTIO_NET_F_HOST_ECN) |
 700                          (1ULL << VIRTIO_NET_F_HOST_TSO4) |
 701                          (1ULL << VIRTIO_NET_F_HOST_TSO6) |
 702                          (1ULL << VIRTIO_NET_F_HOST_UFO)))
 703                 return true;
 704
 705         return false;
 706 }
 707
 708 static void
 709 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
 710 {
 711         struct ipv4_hdr *ipv4_hdr;
 712         struct ipv6_hdr *ipv6_hdr;
 713         void *l3_hdr = NULL;
 714         struct ether_hdr *eth_hdr;
 715         uint16_t ethertype;
 716
 717         eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
 718
 719         m->l2_len = sizeof(struct ether_hdr);
 720         ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
 721
 722         if (ethertype == ETHER_TYPE_VLAN) {
 723                 struct vlan_hdr *vlan_hdr = (struct vlan_hdr *)(eth_hdr + 1);
 724
 725                 m->l2_len += sizeof(struct vlan_hdr);
 726                 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
 727         }
 728
 729         l3_hdr = (char *)eth_hdr + m->l2_len;
 730
 731         switch (ethertype) {
 732         case ETHER_TYPE_IPv4:
 733                 ipv4_hdr = l3_hdr;
 734                 *l4_proto = ipv4_hdr->next_proto_id;
 735                 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
 736                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 737                 m->ol_flags |= PKT_TX_IPV4;
 738                 break;
 739         case ETHER_TYPE_IPv6:
 740                 ipv6_hdr = l3_hdr;
 741                 *l4_proto = ipv6_hdr->proto;
 742                 m->l3_len = sizeof(struct ipv6_hdr);
 743                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 744                 m->ol_flags |= PKT_TX_IPV6;
 745                 break;
 746         default:
 747                 m->l3_len = 0;
 748                 *l4_proto = 0;
 749                 *l4_hdr = NULL;
 750                 break;
 751         }
 752 }
 753
 754 static __rte_always_inline void
 755 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
 756 {
 757         uint16_t l4_proto = 0;
 758         void *l4_hdr = NULL;
 759         struct tcp_hdr *tcp_hdr = NULL;
 760
 761         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
 762                 return;
 763
 764         parse_ethernet(m, &l4_proto, &l4_hdr);
 765         if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 766                 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
 767                         switch (hdr->csum_offset) {
 768                         case (offsetof(struct tcp_hdr, cksum)):
 769                                 if (l4_proto == IPPROTO_TCP)
 770                                         m->ol_flags |= PKT_TX_TCP_CKSUM;
 771                                 break;
 772                         case (offsetof(struct udp_hdr, dgram_cksum)):
 773                                 if (l4_proto == IPPROTO_UDP)
 774                                         m->ol_flags |= PKT_TX_UDP_CKSUM;
 775                                 break;
 776                         case (offsetof(struct sctp_hdr, cksum)):
 777                                 if (l4_proto == IPPROTO_SCTP)
 778                                         m->ol_flags |= PKT_TX_SCTP_CKSUM;
 779                                 break;
 780                         default:
 781                                 break;
 782                         }
 783                 }
 784         }
 785
 786         if (l4_hdr && hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 787                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 788                 case VIRTIO_NET_HDR_GSO_TCPV4:
 789                 case VIRTIO_NET_HDR_GSO_TCPV6:
 790                         tcp_hdr = l4_hdr;
 791                         m->ol_flags |= PKT_TX_TCP_SEG;
 792                         m->tso_segsz = hdr->gso_size;
 793                         m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
 794                         break;
 795                 default:
 796                         RTE_LOG(WARNING, VHOST_DATA,
 797                                 "unsupported gso type %u.\n", hdr->gso_type);
 798                         break;
 799                 }
 800         }
 801 }
 802
 803 #define RARP_PKT_SIZE   64
 804
 805 static int
 806 make_rarp_packet(struct rte_mbuf *rarp_mbuf, const struct ether_addr *mac)
 807 {
 808         struct ether_hdr *eth_hdr;
 809         struct arp_hdr  *rarp;
 810
 811         if (rarp_mbuf->buf_len < 64) {
 812                 RTE_LOG(WARNING, VHOST_DATA,
 813                         "failed to make RARP; mbuf size too small %u (< %d)\n",
 814                         rarp_mbuf->buf_len, RARP_PKT_SIZE);
 815                 return -1;
 816         }
 817
 818         /* Ethernet header. */
 819         eth_hdr = rte_pktmbuf_mtod_offset(rarp_mbuf, struct ether_hdr *, 0);
 820         memset(eth_hdr->d_addr.addr_bytes, 0xff, ETHER_ADDR_LEN);
 821         ether_addr_copy(mac, &eth_hdr->s_addr);
 822         eth_hdr->ether_type = htons(ETHER_TYPE_RARP);
 823
 824         /* RARP header. */
 825         rarp = (struct arp_hdr *)(eth_hdr + 1);
 826         rarp->arp_hrd = htons(ARP_HRD_ETHER);
 827         rarp->arp_pro = htons(ETHER_TYPE_IPv4);
 828         rarp->arp_hln = ETHER_ADDR_LEN;
 829         rarp->arp_pln = 4;
 830         rarp->arp_op  = htons(ARP_OP_REVREQUEST);
 831
 832         ether_addr_copy(mac, &rarp->arp_data.arp_sha);
 833         ether_addr_copy(mac, &rarp->arp_data.arp_tha);
 834         memset(&rarp->arp_data.arp_sip, 0x00, 4);
 835         memset(&rarp->arp_data.arp_tip, 0x00, 4);
 836
 837         rarp_mbuf->pkt_len  = rarp_mbuf->data_len = RARP_PKT_SIZE;
 838
 839         return 0;
 840 }
 841
 842 static __rte_always_inline void
 843 put_zmbuf(struct zcopy_mbuf *zmbuf)
 844 {
 845         zmbuf->in_use = 0;
 846 }
 847
 848 static __rte_always_inline int
 849 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 850                   struct vring_desc *descs, uint16_t max_desc,
 851                   struct rte_mbuf *m, uint16_t desc_idx,
 852                   struct rte_mempool *mbuf_pool)
 853 {
 854         struct vring_desc *desc;
 855         uint64_t desc_addr;
 856         uint32_t desc_avail, desc_offset;
 857         uint32_t mbuf_avail, mbuf_offset;
 858         uint32_t cpy_len;
 859         struct rte_mbuf *cur = m, *prev = m;
 860         struct virtio_net_hdr *hdr = NULL;
 861         /* A counter to avoid desc dead loop chain */
 862         uint32_t nr_desc = 1;
 863         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 864         uint16_t copy_nb = vq->batch_copy_nb_elems;
 865         int error = 0;
 866
 867         desc = &descs[desc_idx];
 868         if (unlikely((desc->len < dev->vhost_hlen)) ||
 869                         (desc->flags & VRING_DESC_F_INDIRECT)) {
 870                 error = -1;
 871                 goto out;
 872         }
 873
 874         desc_addr = vhost_iova_to_vva(dev,
 875                                         vq, desc->addr,
 876                                         desc->len,
 877                                         VHOST_ACCESS_RO);
 878         if (unlikely(!desc_addr)) {
 879                 error = -1;
 880                 goto out;
 881         }
 882
 883         if (virtio_net_with_host_offload(dev)) {
 884                 hdr = (struct virtio_net_hdr *)((uintptr_t)desc_addr);
 885                 rte_prefetch0(hdr);
 886         }
 887
 888         /*
 889          * A virtio driver normally uses at least 2 desc buffers
 890          * for Tx: the first for storing the header, and others
 891          * for storing the data.
 892          */
 893         if (likely((desc->len == dev->vhost_hlen) &&
 894                    (desc->flags & VRING_DESC_F_NEXT) != 0)) {
 895                 desc = &descs[desc->next];
 896                 if (unlikely(desc->flags & VRING_DESC_F_INDIRECT)) {
 897                         error = -1;
 898                         goto out;
 899                 }
 900
 901                 desc_addr = vhost_iova_to_vva(dev,
 902                                                         vq, desc->addr,
 903                                                         desc->len,
 904                                                         VHOST_ACCESS_RO);
 905                 if (unlikely(!desc_addr)) {
 906                         error = -1;
 907                         goto out;
 908                 }
 909
 910                 desc_offset = 0;
 911                 desc_avail  = desc->len;
 912                 nr_desc    += 1;
 913         } else {
 914                 desc_avail  = desc->len - dev->vhost_hlen;
 915                 desc_offset = dev->vhost_hlen;
 916         }
 917
 918         rte_prefetch0((void *)(uintptr_t)(desc_addr + desc_offset));
 919
 920         PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset), desc_avail, 0);
 921
 922         mbuf_offset = 0;
 923         mbuf_avail  = m->buf_len - RTE_PKTMBUF_HEADROOM;
 924         while (1) {
 925                 uint64_t hpa;
 926
 927                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 928
 929                 /*
 930                  * A desc buf might across two host physical pages that are
 931                  * not continuous. In such case (gpa_to_hpa returns 0), data
 932                  * will be copied even though zero copy is enabled.
 933                  */
 934                 if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
 935                                         desc->addr + desc_offset, cpy_len)))) {
 936                         cur->data_len = cpy_len;
 937                         cur->data_off = 0;
 938                         cur->buf_addr = (void *)(uintptr_t)(desc_addr
 939                                 + desc_offset);
 940                         cur->buf_iova = hpa;
 941
 942                         /*
 943                          * In zero copy mode, one mbuf can only reference data
 944                          * for one or partial of one desc buff.
 945                          */
 946                         mbuf_avail = cpy_len;
 947                 } else {
 948                         if (likely(cpy_len > MAX_BATCH_LEN ||
 949                                    copy_nb >= vq->size ||
 950                                    (hdr && cur == m))) {
 951                                 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
 952                                                                    mbuf_offset),
 953                                            (void *)((uintptr_t)(desc_addr +
 954                                                                 desc_offset)),
 955                                            cpy_len);
 956                         } else {
 957                                 batch_copy[copy_nb].dst =
 958                                         rte_pktmbuf_mtod_offset(cur, void *,
 959                                                                 mbuf_offset);
 960                                 batch_copy[copy_nb].src =
 961                                         (void *)((uintptr_t)(desc_addr +
 962                                                              desc_offset));
 963                                 batch_copy[copy_nb].len = cpy_len;
 964                                 copy_nb++;
 965                         }
 966                 }
 967
 968                 mbuf_avail  -= cpy_len;
 969                 mbuf_offset += cpy_len;
 970                 desc_avail  -= cpy_len;
 971                 desc_offset += cpy_len;
 972
 973                 /* This desc reaches to its end, get the next one */
 974                 if (desc_avail == 0) {
 975                         if ((desc->flags & VRING_DESC_F_NEXT) == 0)
 976                                 break;
 977
 978                         if (unlikely(desc->next >= max_desc ||
 979                                      ++nr_desc > max_desc)) {
 980                                 error = -1;
 981                                 goto out;
 982                         }
 983                         desc = &descs[desc->next];
 984                         if (unlikely(desc->flags & VRING_DESC_F_INDIRECT)) {
 985                                 error = -1;
 986                                 goto out;
 987                         }
 988
 989                         desc_addr = vhost_iova_to_vva(dev,
 990                                                         vq, desc->addr,
 991                                                         desc->len,
 992                                                         VHOST_ACCESS_RO);
 993                         if (unlikely(!desc_addr)) {
 994                                 error = -1;
 995                                 goto out;
 996                         }
 997
 998                         rte_prefetch0((void *)(uintptr_t)desc_addr);
 999
1000                         desc_offset = 0;
1001                         desc_avail  = desc->len;
1002
1003                         PRINT_PACKET(dev, (uintptr_t)desc_addr, desc->len, 0);
1004                 }
1005
1006                 /*
1007                  * This mbuf reaches to its end, get a new one
1008                  * to hold more data.
1009                  */
1010                 if (mbuf_avail == 0) {
1011                         cur = rte_pktmbuf_alloc(mbuf_pool);
1012                         if (unlikely(cur == NULL)) {
1013                                 RTE_LOG(ERR, VHOST_DATA, "Failed to "
1014                                         "allocate memory for mbuf.\n");
1015                                 error = -1;
1016                                 goto out;
1017                         }
1018                         if (unlikely(dev->dequeue_zero_copy))
1019                                 rte_mbuf_refcnt_update(cur, 1);
1020
1021                         prev->next = cur;
1022                         prev->data_len = mbuf_offset;
1023                         m->nb_segs += 1;
1024                         m->pkt_len += mbuf_offset;
1025                         prev = cur;
1026
1027                         mbuf_offset = 0;
1028                         mbuf_avail  = cur->buf_len - RTE_PKTMBUF_HEADROOM;
1029                 }
1030         }
1031
1032         prev->data_len = mbuf_offset;
1033         m->pkt_len    += mbuf_offset;
1034
1035         if (hdr)
1036                 vhost_dequeue_offload(hdr, m);
1037
1038 out:
1039         vq->batch_copy_nb_elems = copy_nb;
1040
1041         return error;
1042 }
1043
1044 static __rte_always_inline void
1045 update_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq,
1046                  uint32_t used_idx, uint32_t desc_idx)
1047 {
1048         vq->used->ring[used_idx].id  = desc_idx;
1049         vq->used->ring[used_idx].len = 0;
1050         vhost_log_used_vring(dev, vq,
1051                         offsetof(struct vring_used, ring[used_idx]),
1052                         sizeof(vq->used->ring[used_idx]));
1053 }
1054
1055 static __rte_always_inline void
1056 update_used_idx(struct virtio_net *dev, struct vhost_virtqueue *vq,
1057                 uint32_t count)
1058 {
1059         if (unlikely(count == 0))
1060                 return;
1061
1062         rte_smp_wmb();
1063         rte_smp_rmb();
1064
1065         vq->used->idx += count;
1066         vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
1067                         sizeof(vq->used->idx));
1068         vhost_vring_call(vq);
1069 }
1070
1071 static __rte_always_inline struct zcopy_mbuf *
1072 get_zmbuf(struct vhost_virtqueue *vq)
1073 {
1074         uint16_t i;
1075         uint16_t last;
1076         int tries = 0;
1077
1078         /* search [last_zmbuf_idx, zmbuf_size) */
1079         i = vq->last_zmbuf_idx;
1080         last = vq->zmbuf_size;
1081
1082 again:
1083         for (; i < last; i++) {
1084                 if (vq->zmbufs[i].in_use == 0) {
1085                         vq->last_zmbuf_idx = i + 1;
1086                         vq->zmbufs[i].in_use = 1;
1087                         return &vq->zmbufs[i];
1088                 }
1089         }
1090
1091         tries++;
1092         if (tries == 1) {
1093                 /* search [0, last_zmbuf_idx) */
1094                 i = 0;
1095                 last = vq->last_zmbuf_idx;
1096                 goto again;
1097         }
1098
1099         return NULL;
1100 }
1101
1102 static __rte_always_inline bool
1103 mbuf_is_consumed(struct rte_mbuf *m)
1104 {
1105         while (m) {
1106                 if (rte_mbuf_refcnt_read(m) > 1)
1107                         return false;
1108                 m = m->next;
1109         }
1110
1111         return true;
1112 }
1113
1114 uint16_t
1115 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
1116         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1117 {
1118         struct virtio_net *dev;
1119         struct rte_mbuf *rarp_mbuf = NULL;
1120         struct vhost_virtqueue *vq;
1121         uint32_t desc_indexes[MAX_PKT_BURST];
1122         uint32_t used_idx;
1123         uint32_t i = 0;
1124         uint16_t free_entries;
1125         uint16_t avail_idx;
1126
1127         dev = get_device(vid);
1128         if (!dev)
1129                 return 0;
1130
1131         if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
1132                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
1133                         dev->vid, __func__, queue_id);
1134                 return 0;
1135         }
1136
1137         vq = dev->virtqueue[queue_id];
1138         if (unlikely(vq->enabled == 0))
1139                 return 0;
1140
1141         vq->batch_copy_nb_elems = 0;
1142
1143         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1144                 vhost_user_iotlb_rd_lock(vq);
1145
1146         if (unlikely(vq->access_ok == 0))
1147                 if (unlikely(vring_translate(dev, vq) < 0))
1148                         goto out;
1149
1150         if (unlikely(dev->dequeue_zero_copy)) {
1151                 struct zcopy_mbuf *zmbuf, *next;
1152                 int nr_updated = 0;
1153
1154                 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1155                      zmbuf != NULL; zmbuf = next) {
1156                         next = TAILQ_NEXT(zmbuf, next);
1157
1158                         if (mbuf_is_consumed(zmbuf->mbuf)) {
1159                                 used_idx = vq->last_used_idx++ & (vq->size - 1);
1160                                 update_used_ring(dev, vq, used_idx,
1161                                                  zmbuf->desc_idx);
1162                                 nr_updated += 1;
1163
1164                                 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1165                                 rte_pktmbuf_free(zmbuf->mbuf);
1166                                 put_zmbuf(zmbuf);
1167                                 vq->nr_zmbuf -= 1;
1168                         }
1169                 }
1170
1171                 update_used_idx(dev, vq, nr_updated);
1172         }
1173
1174         /*
1175          * Construct a RARP broadcast packet, and inject it to the "pkts"
1176          * array, to looks like that guest actually send such packet.
1177          *
1178          * Check user_send_rarp() for more information.
1179          *
1180          * broadcast_rarp shares a cacheline in the virtio_net structure
1181          * with some fields that are accessed during enqueue and
1182          * rte_atomic16_cmpset() causes a write if using cmpxchg. This could
1183          * result in false sharing between enqueue and dequeue.
1184          *
1185          * Prevent unnecessary false sharing by reading broadcast_rarp first
1186          * and only performing cmpset if the read indicates it is likely to
1187          * be set.
1188          */
1189
1190         if (unlikely(rte_atomic16_read(&dev->broadcast_rarp) &&
1191                         rte_atomic16_cmpset((volatile uint16_t *)
1192                                 &dev->broadcast_rarp.cnt, 1, 0))) {
1193
1194                 rarp_mbuf = rte_pktmbuf_alloc(mbuf_pool);
1195                 if (rarp_mbuf == NULL) {
1196                         RTE_LOG(ERR, VHOST_DATA,
1197                                 "Failed to allocate memory for mbuf.\n");
1198                         return 0;
1199                 }
1200
1201                 if (make_rarp_packet(rarp_mbuf, &dev->mac)) {
1202                         rte_pktmbuf_free(rarp_mbuf);
1203                         rarp_mbuf = NULL;
1204                 } else {
1205                         count -= 1;
1206                 }
1207         }
1208
1209         free_entries = *((volatile uint16_t *)&vq->avail->idx) -
1210                         vq->last_avail_idx;
1211         if (free_entries == 0)
1212                 goto out;
1213
1214         LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1215
1216         /* Prefetch available and used ring */
1217         avail_idx = vq->last_avail_idx & (vq->size - 1);
1218         used_idx  = vq->last_used_idx  & (vq->size - 1);
1219         rte_prefetch0(&vq->avail->ring[avail_idx]);
1220         rte_prefetch0(&vq->used->ring[used_idx]);
1221
1222         count = RTE_MIN(count, MAX_PKT_BURST);
1223         count = RTE_MIN(count, free_entries);
1224         LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1225                         dev->vid, count);
1226
1227         /* Retrieve all of the head indexes first to avoid caching issues. */
1228         for (i = 0; i < count; i++) {
1229                 avail_idx = (vq->last_avail_idx + i) & (vq->size - 1);
1230                 used_idx  = (vq->last_used_idx  + i) & (vq->size - 1);
1231                 desc_indexes[i] = vq->avail->ring[avail_idx];
1232
1233                 if (likely(dev->dequeue_zero_copy == 0))
1234                         update_used_ring(dev, vq, used_idx, desc_indexes[i]);
1235         }
1236
1237         /* Prefetch descriptor index. */
1238         rte_prefetch0(&vq->desc[desc_indexes[0]]);
1239         for (i = 0; i < count; i++) {
1240                 struct vring_desc *desc;
1241                 uint16_t sz, idx;
1242                 int err;
1243
1244                 if (likely(i + 1 < count))
1245                         rte_prefetch0(&vq->desc[desc_indexes[i + 1]]);
1246
1247                 if (vq->desc[desc_indexes[i]].flags & VRING_DESC_F_INDIRECT) {
1248                         desc = (struct vring_desc *)(uintptr_t)
1249                                 vhost_iova_to_vva(dev, vq,
1250                                                 vq->desc[desc_indexes[i]].addr,
1251                                                 sizeof(*desc),
1252                                                 VHOST_ACCESS_RO);
1253                         if (unlikely(!desc))
1254                                 break;
1255
1256                         rte_prefetch0(desc);
1257                         sz = vq->desc[desc_indexes[i]].len / sizeof(*desc);
1258                         idx = 0;
1259                 } else {
1260                         desc = vq->desc;
1261                         sz = vq->size;
1262                         idx = desc_indexes[i];
1263                 }
1264
1265                 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
1266                 if (unlikely(pkts[i] == NULL)) {
1267                         RTE_LOG(ERR, VHOST_DATA,
1268                                 "Failed to allocate memory for mbuf.\n");
1269                         break;
1270                 }
1271
1272                 err = copy_desc_to_mbuf(dev, vq, desc, sz, pkts[i], idx,
1273                                         mbuf_pool);
1274                 if (unlikely(err)) {
1275                         rte_pktmbuf_free(pkts[i]);
1276                         break;
1277                 }
1278
1279                 if (unlikely(dev->dequeue_zero_copy)) {
1280                         struct zcopy_mbuf *zmbuf;
1281
1282                         zmbuf = get_zmbuf(vq);
1283                         if (!zmbuf) {
1284                                 rte_pktmbuf_free(pkts[i]);
1285                                 break;
1286                         }
1287                         zmbuf->mbuf = pkts[i];
1288                         zmbuf->desc_idx = desc_indexes[i];
1289
1290                         /*
1291                          * Pin lock the mbuf; we will check later to see
1292                          * whether the mbuf is freed (when we are the last
1293                          * user) or not. If that's the case, we then could
1294                          * update the used ring safely.
1295                          */
1296                         rte_mbuf_refcnt_update(pkts[i], 1);
1297
1298                         vq->nr_zmbuf += 1;
1299                         TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1300                 }
1301         }
1302         vq->last_avail_idx += i;
1303
1304         if (likely(dev->dequeue_zero_copy == 0)) {
1305                 do_data_copy_dequeue(vq);
1306                 vq->last_used_idx += i;
1307                 update_used_idx(dev, vq, i);
1308         }
1309
1310 out:
1311         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1312                 vhost_user_iotlb_rd_unlock(vq);
1313
1314         if (unlikely(rarp_mbuf != NULL)) {
1315                 /*
1316                  * Inject it to the head of "pkts" array, so that switch's mac
1317                  * learning table will get updated first.
1318                  */
1319                 memmove(&pkts[1], pkts, i * sizeof(struct rte_mbuf *));
1320                 pkts[0] = rarp_mbuf;
1321                 i += 1;
1322         }
1323
1324         return i;
1325 }