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         /* flush used->idx update before we read avail->flags. */
 383         rte_mb();
 384
 385         /* Kick the guest if necessary. */
 386         if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
 387                         && (vq->callfd >= 0))
 388                 eventfd_write(vq->callfd, (eventfd_t)1);
 389 out:
 390         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 391                 vhost_user_iotlb_rd_unlock(vq);
 392
 393         return count;
 394 }
 395
 396 static __rte_always_inline int
 397 fill_vec_buf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 398                          uint32_t avail_idx, uint32_t *vec_idx,
 399                          struct buf_vector *buf_vec, uint16_t *desc_chain_head,
 400                          uint16_t *desc_chain_len)
 401 {
 402         uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
 403         uint32_t vec_id = *vec_idx;
 404         uint32_t len    = 0;
 405         struct vring_desc *descs = vq->desc;
 406
 407         *desc_chain_head = idx;
 408
 409         if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
 410                 descs = (struct vring_desc *)(uintptr_t)
 411                         vhost_iova_to_vva(dev, vq, vq->desc[idx].addr,
 412                                                 vq->desc[idx].len,
 413                                                 VHOST_ACCESS_RO);
 414                 if (unlikely(!descs))
 415                         return -1;
 416
 417                 idx = 0;
 418         }
 419
 420         while (1) {
 421                 if (unlikely(vec_id >= BUF_VECTOR_MAX || idx >= vq->size))
 422                         return -1;
 423
 424                 len += descs[idx].len;
 425                 buf_vec[vec_id].buf_addr = descs[idx].addr;
 426                 buf_vec[vec_id].buf_len  = descs[idx].len;
 427                 buf_vec[vec_id].desc_idx = idx;
 428                 vec_id++;
 429
 430                 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
 431                         break;
 432
 433                 idx = descs[idx].next;
 434         }
 435
 436         *desc_chain_len = len;
 437         *vec_idx = vec_id;
 438
 439         return 0;
 440 }
 441
 442 /*
 443  * Returns -1 on fail, 0 on success
 444  */
 445 static inline int
 446 reserve_avail_buf_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
 447                                 uint32_t size, struct buf_vector *buf_vec,
 448                                 uint16_t *num_buffers, uint16_t avail_head)
 449 {
 450         uint16_t cur_idx;
 451         uint32_t vec_idx = 0;
 452         uint16_t tries = 0;
 453
 454         uint16_t head_idx = 0;
 455         uint16_t len = 0;
 456
 457         *num_buffers = 0;
 458         cur_idx  = vq->last_avail_idx;
 459
 460         while (size > 0) {
 461                 if (unlikely(cur_idx == avail_head))
 462                         return -1;
 463
 464                 if (unlikely(fill_vec_buf(dev, vq, cur_idx, &vec_idx, buf_vec,
 465                                                 &head_idx, &len) < 0))
 466                         return -1;
 467                 len = RTE_MIN(len, size);
 468                 update_shadow_used_ring(vq, head_idx, len);
 469                 size -= len;
 470
 471                 cur_idx++;
 472                 tries++;
 473                 *num_buffers += 1;
 474
 475                 /*
 476                  * if we tried all available ring items, and still
 477                  * can't get enough buf, it means something abnormal
 478                  * happened.
 479                  */
 480                 if (unlikely(tries >= vq->size))
 481                         return -1;
 482         }
 483
 484         return 0;
 485 }
 486
 487 static __rte_always_inline int
 488 copy_mbuf_to_desc_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
 489                             struct rte_mbuf *m, struct buf_vector *buf_vec,
 490                             uint16_t num_buffers)
 491 {
 492         uint32_t vec_idx = 0;
 493         uint64_t desc_addr;
 494         uint32_t mbuf_offset, mbuf_avail;
 495         uint32_t desc_offset, desc_avail;
 496         uint32_t cpy_len;
 497         uint64_t hdr_addr, hdr_phys_addr;
 498         struct rte_mbuf *hdr_mbuf;
 499         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 500         uint16_t copy_nb = vq->batch_copy_nb_elems;
 501         int error = 0;
 502
 503         if (unlikely(m == NULL)) {
 504                 error = -1;
 505                 goto out;
 506         }
 507
 508         desc_addr = vhost_iova_to_vva(dev, vq, buf_vec[vec_idx].buf_addr,
 509                                                 buf_vec[vec_idx].buf_len,
 510                                                 VHOST_ACCESS_RW);
 511         if (buf_vec[vec_idx].buf_len < dev->vhost_hlen || !desc_addr) {
 512                 error = -1;
 513                 goto out;
 514         }
 515
 516         hdr_mbuf = m;
 517         hdr_addr = desc_addr;
 518         hdr_phys_addr = buf_vec[vec_idx].buf_addr;
 519         rte_prefetch0((void *)(uintptr_t)hdr_addr);
 520
 521         LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
 522                 dev->vid, num_buffers);
 523
 524         desc_avail  = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
 525         desc_offset = dev->vhost_hlen;
 526
 527         mbuf_avail  = rte_pktmbuf_data_len(m);
 528         mbuf_offset = 0;
 529         while (mbuf_avail != 0 || m->next != NULL) {
 530                 /* done with current desc buf, get the next one */
 531                 if (desc_avail == 0) {
 532                         vec_idx++;
 533                         desc_addr =
 534                                 vhost_iova_to_vva(dev, vq,
 535                                         buf_vec[vec_idx].buf_addr,
 536                                         buf_vec[vec_idx].buf_len,
 537                                         VHOST_ACCESS_RW);
 538                         if (unlikely(!desc_addr)) {
 539                                 error = -1;
 540                                 goto out;
 541                         }
 542
 543                         /* Prefetch buffer address. */
 544                         rte_prefetch0((void *)(uintptr_t)desc_addr);
 545                         desc_offset = 0;
 546                         desc_avail  = buf_vec[vec_idx].buf_len;
 547                 }
 548
 549                 /* done with current mbuf, get the next one */
 550                 if (mbuf_avail == 0) {
 551                         m = m->next;
 552
 553                         mbuf_offset = 0;
 554                         mbuf_avail  = rte_pktmbuf_data_len(m);
 555                 }
 556
 557                 if (hdr_addr) {
 558                         struct virtio_net_hdr_mrg_rxbuf *hdr;
 559
 560                         hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)
 561                                 hdr_addr;
 562                         virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
 563                         ASSIGN_UNLESS_EQUAL(hdr->num_buffers, num_buffers);
 564
 565                         vhost_log_write(dev, hdr_phys_addr, dev->vhost_hlen);
 566                         PRINT_PACKET(dev, (uintptr_t)hdr_addr,
 567                                      dev->vhost_hlen, 0);
 568
 569                         hdr_addr = 0;
 570                 }
 571
 572                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 573
 574                 if (likely(cpy_len > MAX_BATCH_LEN || copy_nb >= vq->size)) {
 575                         rte_memcpy((void *)((uintptr_t)(desc_addr +
 576                                                         desc_offset)),
 577                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 578                                 cpy_len);
 579                         vhost_log_write(dev,
 580                                 buf_vec[vec_idx].buf_addr + desc_offset,
 581                                 cpy_len);
 582                         PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
 583                                 cpy_len, 0);
 584                 } else {
 585                         batch_copy[copy_nb].dst =
 586                                 (void *)((uintptr_t)(desc_addr + desc_offset));
 587                         batch_copy[copy_nb].src =
 588                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
 589                         batch_copy[copy_nb].log_addr =
 590                                 buf_vec[vec_idx].buf_addr + desc_offset;
 591                         batch_copy[copy_nb].len = cpy_len;
 592                         copy_nb++;
 593                 }
 594
 595                 mbuf_avail  -= cpy_len;
 596                 mbuf_offset += cpy_len;
 597                 desc_avail  -= cpy_len;
 598                 desc_offset += cpy_len;
 599         }
 600
 601 out:
 602         vq->batch_copy_nb_elems = copy_nb;
 603
 604         return error;
 605 }
 606
 607 static __rte_always_inline uint32_t
 608 virtio_dev_merge_rx(struct virtio_net *dev, uint16_t queue_id,
 609         struct rte_mbuf **pkts, uint32_t count)
 610 {
 611         struct vhost_virtqueue *vq;
 612         uint32_t pkt_idx = 0;
 613         uint16_t num_buffers;
 614         struct buf_vector buf_vec[BUF_VECTOR_MAX];
 615         uint16_t avail_head;
 616
 617         LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
 618         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
 619                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
 620                         dev->vid, __func__, queue_id);
 621                 return 0;
 622         }
 623
 624         vq = dev->virtqueue[queue_id];
 625         if (unlikely(vq->enabled == 0))
 626                 return 0;
 627
 628         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 629                 vhost_user_iotlb_rd_lock(vq);
 630
 631         if (unlikely(vq->access_ok == 0))
 632                 if (unlikely(vring_translate(dev, vq) < 0))
 633                         goto out;
 634
 635         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
 636         if (count == 0)
 637                 goto out;
 638
 639         vq->batch_copy_nb_elems = 0;
 640
 641         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
 642
 643         vq->shadow_used_idx = 0;
 644         avail_head = *((volatile uint16_t *)&vq->avail->idx);
 645         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
 646                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
 647
 648                 if (unlikely(reserve_avail_buf_mergeable(dev, vq,
 649                                                 pkt_len, buf_vec, &num_buffers,
 650                                                 avail_head) < 0)) {
 651                         LOG_DEBUG(VHOST_DATA,
 652                                 "(%d) failed to get enough desc from vring\n",
 653                                 dev->vid);
 654                         vq->shadow_used_idx -= num_buffers;
 655                         break;
 656                 }
 657
 658                 LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
 659                         dev->vid, vq->last_avail_idx,
 660                         vq->last_avail_idx + num_buffers);
 661
 662                 if (copy_mbuf_to_desc_mergeable(dev, vq, pkts[pkt_idx],
 663                                                 buf_vec, num_buffers) < 0) {
 664                         vq->shadow_used_idx -= num_buffers;
 665                         break;
 666                 }
 667
 668                 vq->last_avail_idx += num_buffers;
 669         }
 670
 671         do_data_copy_enqueue(dev, vq);
 672
 673         if (likely(vq->shadow_used_idx)) {
 674                 flush_shadow_used_ring(dev, vq);
 675
 676                 /* flush used->idx update before we read avail->flags. */
 677                 rte_mb();
 678
 679                 /* Kick the guest if necessary. */
 680                 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
 681                                 && (vq->callfd >= 0))
 682                         eventfd_write(vq->callfd, (eventfd_t)1);
 683         }
 684
 685 out:
 686         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 687                 vhost_user_iotlb_rd_unlock(vq);
 688
 689         return pkt_idx;
 690 }
 691
 692 uint16_t
 693 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
 694         struct rte_mbuf **pkts, uint16_t count)
 695 {
 696         struct virtio_net *dev = get_device(vid);
 697
 698         if (!dev)
 699                 return 0;
 700
 701         if (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF))
 702                 return virtio_dev_merge_rx(dev, queue_id, pkts, count);
 703         else
 704                 return virtio_dev_rx(dev, queue_id, pkts, count);
 705 }
 706
 707 static inline bool
 708 virtio_net_with_host_offload(struct virtio_net *dev)
 709 {
 710         if (dev->features &
 711                         ((1ULL << VIRTIO_NET_F_CSUM) |
 712                          (1ULL << VIRTIO_NET_F_HOST_ECN) |
 713                          (1ULL << VIRTIO_NET_F_HOST_TSO4) |
 714                          (1ULL << VIRTIO_NET_F_HOST_TSO6) |
 715                          (1ULL << VIRTIO_NET_F_HOST_UFO)))
 716                 return true;
 717
 718         return false;
 719 }
 720
 721 static void
 722 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
 723 {
 724         struct ipv4_hdr *ipv4_hdr;
 725         struct ipv6_hdr *ipv6_hdr;
 726         void *l3_hdr = NULL;
 727         struct ether_hdr *eth_hdr;
 728         uint16_t ethertype;
 729
 730         eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
 731
 732         m->l2_len = sizeof(struct ether_hdr);
 733         ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
 734
 735         if (ethertype == ETHER_TYPE_VLAN) {
 736                 struct vlan_hdr *vlan_hdr = (struct vlan_hdr *)(eth_hdr + 1);
 737
 738                 m->l2_len += sizeof(struct vlan_hdr);
 739                 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
 740         }
 741
 742         l3_hdr = (char *)eth_hdr + m->l2_len;
 743
 744         switch (ethertype) {
 745         case ETHER_TYPE_IPv4:
 746                 ipv4_hdr = l3_hdr;
 747                 *l4_proto = ipv4_hdr->next_proto_id;
 748                 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
 749                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 750                 m->ol_flags |= PKT_TX_IPV4;
 751                 break;
 752         case ETHER_TYPE_IPv6:
 753                 ipv6_hdr = l3_hdr;
 754                 *l4_proto = ipv6_hdr->proto;
 755                 m->l3_len = sizeof(struct ipv6_hdr);
 756                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 757                 m->ol_flags |= PKT_TX_IPV6;
 758                 break;
 759         default:
 760                 m->l3_len = 0;
 761                 *l4_proto = 0;
 762                 *l4_hdr = NULL;
 763                 break;
 764         }
 765 }
 766
 767 static __rte_always_inline void
 768 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
 769 {
 770         uint16_t l4_proto = 0;
 771         void *l4_hdr = NULL;
 772         struct tcp_hdr *tcp_hdr = NULL;
 773
 774         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
 775                 return;
 776
 777         parse_ethernet(m, &l4_proto, &l4_hdr);
 778         if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 779                 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
 780                         switch (hdr->csum_offset) {
 781                         case (offsetof(struct tcp_hdr, cksum)):
 782                                 if (l4_proto == IPPROTO_TCP)
 783                                         m->ol_flags |= PKT_TX_TCP_CKSUM;
 784                                 break;
 785                         case (offsetof(struct udp_hdr, dgram_cksum)):
 786                                 if (l4_proto == IPPROTO_UDP)
 787                                         m->ol_flags |= PKT_TX_UDP_CKSUM;
 788                                 break;
 789                         case (offsetof(struct sctp_hdr, cksum)):
 790                                 if (l4_proto == IPPROTO_SCTP)
 791                                         m->ol_flags |= PKT_TX_SCTP_CKSUM;
 792                                 break;
 793                         default:
 794                                 break;
 795                         }
 796                 }
 797         }
 798
 799         if (l4_hdr && hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 800                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 801                 case VIRTIO_NET_HDR_GSO_TCPV4:
 802                 case VIRTIO_NET_HDR_GSO_TCPV6:
 803                         tcp_hdr = l4_hdr;
 804                         m->ol_flags |= PKT_TX_TCP_SEG;
 805                         m->tso_segsz = hdr->gso_size;
 806                         m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
 807                         break;
 808                 default:
 809                         RTE_LOG(WARNING, VHOST_DATA,
 810                                 "unsupported gso type %u.\n", hdr->gso_type);
 811                         break;
 812                 }
 813         }
 814 }
 815
 816 #define RARP_PKT_SIZE   64
 817
 818 static int
 819 make_rarp_packet(struct rte_mbuf *rarp_mbuf, const struct ether_addr *mac)
 820 {
 821         struct ether_hdr *eth_hdr;
 822         struct arp_hdr  *rarp;
 823
 824         if (rarp_mbuf->buf_len < 64) {
 825                 RTE_LOG(WARNING, VHOST_DATA,
 826                         "failed to make RARP; mbuf size too small %u (< %d)\n",
 827                         rarp_mbuf->buf_len, RARP_PKT_SIZE);
 828                 return -1;
 829         }
 830
 831         /* Ethernet header. */
 832         eth_hdr = rte_pktmbuf_mtod_offset(rarp_mbuf, struct ether_hdr *, 0);
 833         memset(eth_hdr->d_addr.addr_bytes, 0xff, ETHER_ADDR_LEN);
 834         ether_addr_copy(mac, &eth_hdr->s_addr);
 835         eth_hdr->ether_type = htons(ETHER_TYPE_RARP);
 836
 837         /* RARP header. */
 838         rarp = (struct arp_hdr *)(eth_hdr + 1);
 839         rarp->arp_hrd = htons(ARP_HRD_ETHER);
 840         rarp->arp_pro = htons(ETHER_TYPE_IPv4);
 841         rarp->arp_hln = ETHER_ADDR_LEN;
 842         rarp->arp_pln = 4;
 843         rarp->arp_op  = htons(ARP_OP_REVREQUEST);
 844
 845         ether_addr_copy(mac, &rarp->arp_data.arp_sha);
 846         ether_addr_copy(mac, &rarp->arp_data.arp_tha);
 847         memset(&rarp->arp_data.arp_sip, 0x00, 4);
 848         memset(&rarp->arp_data.arp_tip, 0x00, 4);
 849
 850         rarp_mbuf->pkt_len  = rarp_mbuf->data_len = RARP_PKT_SIZE;
 851
 852         return 0;
 853 }
 854
 855 static __rte_always_inline void
 856 put_zmbuf(struct zcopy_mbuf *zmbuf)
 857 {
 858         zmbuf->in_use = 0;
 859 }
 860
 861 static __rte_always_inline int
 862 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 863                   struct vring_desc *descs, uint16_t max_desc,
 864                   struct rte_mbuf *m, uint16_t desc_idx,
 865                   struct rte_mempool *mbuf_pool)
 866 {
 867         struct vring_desc *desc;
 868         uint64_t desc_addr;
 869         uint32_t desc_avail, desc_offset;
 870         uint32_t mbuf_avail, mbuf_offset;
 871         uint32_t cpy_len;
 872         struct rte_mbuf *cur = m, *prev = m;
 873         struct virtio_net_hdr *hdr = NULL;
 874         /* A counter to avoid desc dead loop chain */
 875         uint32_t nr_desc = 1;
 876         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 877         uint16_t copy_nb = vq->batch_copy_nb_elems;
 878         int error = 0;
 879
 880         desc = &descs[desc_idx];
 881         if (unlikely((desc->len < dev->vhost_hlen)) ||
 882                         (desc->flags & VRING_DESC_F_INDIRECT)) {
 883                 error = -1;
 884                 goto out;
 885         }
 886
 887         desc_addr = vhost_iova_to_vva(dev,
 888                                         vq, desc->addr,
 889                                         desc->len,
 890                                         VHOST_ACCESS_RO);
 891         if (unlikely(!desc_addr)) {
 892                 error = -1;
 893                 goto out;
 894         }
 895
 896         if (virtio_net_with_host_offload(dev)) {
 897                 hdr = (struct virtio_net_hdr *)((uintptr_t)desc_addr);
 898                 rte_prefetch0(hdr);
 899         }
 900
 901         /*
 902          * A virtio driver normally uses at least 2 desc buffers
 903          * for Tx: the first for storing the header, and others
 904          * for storing the data.
 905          */
 906         if (likely((desc->len == dev->vhost_hlen) &&
 907                    (desc->flags & VRING_DESC_F_NEXT) != 0)) {
 908                 desc = &descs[desc->next];
 909                 if (unlikely(desc->flags & VRING_DESC_F_INDIRECT)) {
 910                         error = -1;
 911                         goto out;
 912                 }
 913
 914                 desc_addr = vhost_iova_to_vva(dev,
 915                                                         vq, desc->addr,
 916                                                         desc->len,
 917                                                         VHOST_ACCESS_RO);
 918                 if (unlikely(!desc_addr)) {
 919                         error = -1;
 920                         goto out;
 921                 }
 922
 923                 desc_offset = 0;
 924                 desc_avail  = desc->len;
 925                 nr_desc    += 1;
 926         } else {
 927                 desc_avail  = desc->len - dev->vhost_hlen;
 928                 desc_offset = dev->vhost_hlen;
 929         }
 930
 931         rte_prefetch0((void *)(uintptr_t)(desc_addr + desc_offset));
 932
 933         PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset), desc_avail, 0);
 934
 935         mbuf_offset = 0;
 936         mbuf_avail  = m->buf_len - RTE_PKTMBUF_HEADROOM;
 937         while (1) {
 938                 uint64_t hpa;
 939
 940                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 941
 942                 /*
 943                  * A desc buf might across two host physical pages that are
 944                  * not continuous. In such case (gpa_to_hpa returns 0), data
 945                  * will be copied even though zero copy is enabled.
 946                  */
 947                 if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
 948                                         desc->addr + desc_offset, cpy_len)))) {
 949                         cur->data_len = cpy_len;
 950                         cur->data_off = 0;
 951                         cur->buf_addr = (void *)(uintptr_t)desc_addr;
 952                         cur->buf_iova = hpa;
 953
 954                         /*
 955                          * In zero copy mode, one mbuf can only reference data
 956                          * for one or partial of one desc buff.
 957                          */
 958                         mbuf_avail = cpy_len;
 959                 } else {
 960                         if (likely(cpy_len > MAX_BATCH_LEN ||
 961                                    copy_nb >= vq->size ||
 962                                    (hdr && cur == m))) {
 963                                 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
 964                                                                    mbuf_offset),
 965                                            (void *)((uintptr_t)(desc_addr +
 966                                                                 desc_offset)),
 967                                            cpy_len);
 968                         } else {
 969                                 batch_copy[copy_nb].dst =
 970                                         rte_pktmbuf_mtod_offset(cur, void *,
 971                                                                 mbuf_offset);
 972                                 batch_copy[copy_nb].src =
 973                                         (void *)((uintptr_t)(desc_addr +
 974                                                              desc_offset));
 975                                 batch_copy[copy_nb].len = cpy_len;
 976                                 copy_nb++;
 977                         }
 978                 }
 979
 980                 mbuf_avail  -= cpy_len;
 981                 mbuf_offset += cpy_len;
 982                 desc_avail  -= cpy_len;
 983                 desc_offset += cpy_len;
 984
 985                 /* This desc reaches to its end, get the next one */
 986                 if (desc_avail == 0) {
 987                         if ((desc->flags & VRING_DESC_F_NEXT) == 0)
 988                                 break;
 989
 990                         if (unlikely(desc->next >= max_desc ||
 991                                      ++nr_desc > max_desc)) {
 992                                 error = -1;
 993                                 goto out;
 994                         }
 995                         desc = &descs[desc->next];
 996                         if (unlikely(desc->flags & VRING_DESC_F_INDIRECT)) {
 997                                 error = -1;
 998                                 goto out;
 999                         }
1000
1001                         desc_addr = vhost_iova_to_vva(dev,
1002                                                         vq, desc->addr,
1003                                                         desc->len,
1004                                                         VHOST_ACCESS_RO);
1005                         if (unlikely(!desc_addr)) {
1006                                 error = -1;
1007                                 goto out;
1008                         }
1009
1010                         rte_prefetch0((void *)(uintptr_t)desc_addr);
1011
1012                         desc_offset = 0;
1013                         desc_avail  = desc->len;
1014
1015                         PRINT_PACKET(dev, (uintptr_t)desc_addr, desc->len, 0);
1016                 }
1017
1018                 /*
1019                  * This mbuf reaches to its end, get a new one
1020                  * to hold more data.
1021                  */
1022                 if (mbuf_avail == 0) {
1023                         cur = rte_pktmbuf_alloc(mbuf_pool);
1024                         if (unlikely(cur == NULL)) {
1025                                 RTE_LOG(ERR, VHOST_DATA, "Failed to "
1026                                         "allocate memory for mbuf.\n");
1027                                 error = -1;
1028                                 goto out;
1029                         }
1030                         if (unlikely(dev->dequeue_zero_copy))
1031                                 rte_mbuf_refcnt_update(cur, 1);
1032
1033                         prev->next = cur;
1034                         prev->data_len = mbuf_offset;
1035                         m->nb_segs += 1;
1036                         m->pkt_len += mbuf_offset;
1037                         prev = cur;
1038
1039                         mbuf_offset = 0;
1040                         mbuf_avail  = cur->buf_len - RTE_PKTMBUF_HEADROOM;
1041                 }
1042         }
1043
1044         prev->data_len = mbuf_offset;
1045         m->pkt_len    += mbuf_offset;
1046
1047         if (hdr)
1048                 vhost_dequeue_offload(hdr, m);
1049
1050 out:
1051         vq->batch_copy_nb_elems = copy_nb;
1052
1053         return error;
1054 }
1055
1056 static __rte_always_inline void
1057 update_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq,
1058                  uint32_t used_idx, uint32_t desc_idx)
1059 {
1060         vq->used->ring[used_idx].id  = desc_idx;
1061         vq->used->ring[used_idx].len = 0;
1062         vhost_log_used_vring(dev, vq,
1063                         offsetof(struct vring_used, ring[used_idx]),
1064                         sizeof(vq->used->ring[used_idx]));
1065 }
1066
1067 static __rte_always_inline void
1068 update_used_idx(struct virtio_net *dev, struct vhost_virtqueue *vq,
1069                 uint32_t count)
1070 {
1071         if (unlikely(count == 0))
1072                 return;
1073
1074         rte_smp_wmb();
1075         rte_smp_rmb();
1076
1077         vq->used->idx += count;
1078         vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
1079                         sizeof(vq->used->idx));
1080
1081         /* Kick guest if required. */
1082         if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
1083                         && (vq->callfd >= 0))
1084                 eventfd_write(vq->callfd, (eventfd_t)1);
1085 }
1086
1087 static __rte_always_inline struct zcopy_mbuf *
1088 get_zmbuf(struct vhost_virtqueue *vq)
1089 {
1090         uint16_t i;
1091         uint16_t last;
1092         int tries = 0;
1093
1094         /* search [last_zmbuf_idx, zmbuf_size) */
1095         i = vq->last_zmbuf_idx;
1096         last = vq->zmbuf_size;
1097
1098 again:
1099         for (; i < last; i++) {
1100                 if (vq->zmbufs[i].in_use == 0) {
1101                         vq->last_zmbuf_idx = i + 1;
1102                         vq->zmbufs[i].in_use = 1;
1103                         return &vq->zmbufs[i];
1104                 }
1105         }
1106
1107         tries++;
1108         if (tries == 1) {
1109                 /* search [0, last_zmbuf_idx) */
1110                 i = 0;
1111                 last = vq->last_zmbuf_idx;
1112                 goto again;
1113         }
1114
1115         return NULL;
1116 }
1117
1118 static __rte_always_inline bool
1119 mbuf_is_consumed(struct rte_mbuf *m)
1120 {
1121         while (m) {
1122                 if (rte_mbuf_refcnt_read(m) > 1)
1123                         return false;
1124                 m = m->next;
1125         }
1126
1127         return true;
1128 }
1129
1130 uint16_t
1131 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
1132         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1133 {
1134         struct virtio_net *dev;
1135         struct rte_mbuf *rarp_mbuf = NULL;
1136         struct vhost_virtqueue *vq;
1137         uint32_t desc_indexes[MAX_PKT_BURST];
1138         uint32_t used_idx;
1139         uint32_t i = 0;
1140         uint16_t free_entries;
1141         uint16_t avail_idx;
1142
1143         dev = get_device(vid);
1144         if (!dev)
1145                 return 0;
1146
1147         if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
1148                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
1149                         dev->vid, __func__, queue_id);
1150                 return 0;
1151         }
1152
1153         vq = dev->virtqueue[queue_id];
1154         if (unlikely(vq->enabled == 0))
1155                 return 0;
1156
1157         vq->batch_copy_nb_elems = 0;
1158
1159         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1160                 vhost_user_iotlb_rd_lock(vq);
1161
1162         if (unlikely(vq->access_ok == 0))
1163                 if (unlikely(vring_translate(dev, vq) < 0))
1164                         goto out;
1165
1166         if (unlikely(dev->dequeue_zero_copy)) {
1167                 struct zcopy_mbuf *zmbuf, *next;
1168                 int nr_updated = 0;
1169
1170                 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1171                      zmbuf != NULL; zmbuf = next) {
1172                         next = TAILQ_NEXT(zmbuf, next);
1173
1174                         if (mbuf_is_consumed(zmbuf->mbuf)) {
1175                                 used_idx = vq->last_used_idx++ & (vq->size - 1);
1176                                 update_used_ring(dev, vq, used_idx,
1177                                                  zmbuf->desc_idx);
1178                                 nr_updated += 1;
1179
1180                                 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1181                                 rte_pktmbuf_free(zmbuf->mbuf);
1182                                 put_zmbuf(zmbuf);
1183                                 vq->nr_zmbuf -= 1;
1184                         }
1185                 }
1186
1187                 update_used_idx(dev, vq, nr_updated);
1188         }
1189
1190         /*
1191          * Construct a RARP broadcast packet, and inject it to the "pkts"
1192          * array, to looks like that guest actually send such packet.
1193          *
1194          * Check user_send_rarp() for more information.
1195          *
1196          * broadcast_rarp shares a cacheline in the virtio_net structure
1197          * with some fields that are accessed during enqueue and
1198          * rte_atomic16_cmpset() causes a write if using cmpxchg. This could
1199          * result in false sharing between enqueue and dequeue.
1200          *
1201          * Prevent unnecessary false sharing by reading broadcast_rarp first
1202          * and only performing cmpset if the read indicates it is likely to
1203          * be set.
1204          */
1205
1206         if (unlikely(rte_atomic16_read(&dev->broadcast_rarp) &&
1207                         rte_atomic16_cmpset((volatile uint16_t *)
1208                                 &dev->broadcast_rarp.cnt, 1, 0))) {
1209
1210                 rarp_mbuf = rte_pktmbuf_alloc(mbuf_pool);
1211                 if (rarp_mbuf == NULL) {
1212                         RTE_LOG(ERR, VHOST_DATA,
1213                                 "Failed to allocate memory for mbuf.\n");
1214                         return 0;
1215                 }
1216
1217                 if (make_rarp_packet(rarp_mbuf, &dev->mac)) {
1218                         rte_pktmbuf_free(rarp_mbuf);
1219                         rarp_mbuf = NULL;
1220                 } else {
1221                         count -= 1;
1222                 }
1223         }
1224
1225         free_entries = *((volatile uint16_t *)&vq->avail->idx) -
1226                         vq->last_avail_idx;
1227         if (free_entries == 0)
1228                 goto out;
1229
1230         LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1231
1232         /* Prefetch available and used ring */
1233         avail_idx = vq->last_avail_idx & (vq->size - 1);
1234         used_idx  = vq->last_used_idx  & (vq->size - 1);
1235         rte_prefetch0(&vq->avail->ring[avail_idx]);
1236         rte_prefetch0(&vq->used->ring[used_idx]);
1237
1238         count = RTE_MIN(count, MAX_PKT_BURST);
1239         count = RTE_MIN(count, free_entries);
1240         LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1241                         dev->vid, count);
1242
1243         /* Retrieve all of the head indexes first to avoid caching issues. */
1244         for (i = 0; i < count; i++) {
1245                 avail_idx = (vq->last_avail_idx + i) & (vq->size - 1);
1246                 used_idx  = (vq->last_used_idx  + i) & (vq->size - 1);
1247                 desc_indexes[i] = vq->avail->ring[avail_idx];
1248
1249                 if (likely(dev->dequeue_zero_copy == 0))
1250                         update_used_ring(dev, vq, used_idx, desc_indexes[i]);
1251         }
1252
1253         /* Prefetch descriptor index. */
1254         rte_prefetch0(&vq->desc[desc_indexes[0]]);
1255         for (i = 0; i < count; i++) {
1256                 struct vring_desc *desc;
1257                 uint16_t sz, idx;
1258                 int err;
1259
1260                 if (likely(i + 1 < count))
1261                         rte_prefetch0(&vq->desc[desc_indexes[i + 1]]);
1262
1263                 if (vq->desc[desc_indexes[i]].flags & VRING_DESC_F_INDIRECT) {
1264                         desc = (struct vring_desc *)(uintptr_t)
1265                                 vhost_iova_to_vva(dev, vq,
1266                                                 vq->desc[desc_indexes[i]].addr,
1267                                                 sizeof(*desc),
1268                                                 VHOST_ACCESS_RO);
1269                         if (unlikely(!desc))
1270                                 break;
1271
1272                         rte_prefetch0(desc);
1273                         sz = vq->desc[desc_indexes[i]].len / sizeof(*desc);
1274                         idx = 0;
1275                 } else {
1276                         desc = vq->desc;
1277                         sz = vq->size;
1278                         idx = desc_indexes[i];
1279                 }
1280
1281                 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
1282                 if (unlikely(pkts[i] == NULL)) {
1283                         RTE_LOG(ERR, VHOST_DATA,
1284                                 "Failed to allocate memory for mbuf.\n");
1285                         break;
1286                 }
1287
1288                 err = copy_desc_to_mbuf(dev, vq, desc, sz, pkts[i], idx,
1289                                         mbuf_pool);
1290                 if (unlikely(err)) {
1291                         rte_pktmbuf_free(pkts[i]);
1292                         break;
1293                 }
1294
1295                 if (unlikely(dev->dequeue_zero_copy)) {
1296                         struct zcopy_mbuf *zmbuf;
1297
1298                         zmbuf = get_zmbuf(vq);
1299                         if (!zmbuf) {
1300                                 rte_pktmbuf_free(pkts[i]);
1301                                 break;
1302                         }
1303                         zmbuf->mbuf = pkts[i];
1304                         zmbuf->desc_idx = desc_indexes[i];
1305
1306                         /*
1307                          * Pin lock the mbuf; we will check later to see
1308                          * whether the mbuf is freed (when we are the last
1309                          * user) or not. If that's the case, we then could
1310                          * update the used ring safely.
1311                          */
1312                         rte_mbuf_refcnt_update(pkts[i], 1);
1313
1314                         vq->nr_zmbuf += 1;
1315                         TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1316                 }
1317         }
1318         vq->last_avail_idx += i;
1319
1320         if (likely(dev->dequeue_zero_copy == 0)) {
1321                 do_data_copy_dequeue(vq);
1322                 vq->last_used_idx += i;
1323                 update_used_idx(dev, vq, i);
1324         }
1325
1326 out:
1327         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1328                 vhost_user_iotlb_rd_unlock(vq);
1329
1330         if (unlikely(rarp_mbuf != NULL)) {
1331                 /*
1332                  * Inject it to the head of "pkts" array, so that switch's mac
1333                  * learning table will get updated first.
1334                  */
1335                 memmove(&pkts[1], pkts, i * sizeof(struct rte_mbuf *));
1336                 pkts[0] = rarp_mbuf;
1337                 i += 1;
1338         }
1339
1340         return i;
1341 }