lib/librte_vhost/virtio_net.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2016 Intel Corporation. All rights reserved.
   5  *   All rights reserved.
   6  *
   7  *   Redistribution and use in source and binary forms, with or without
   8  *   modification, are permitted provided that the following conditions
   9  *   are met:
  10  *
  11  *     * Redistributions of source code must retain the above copyright
  12  *       notice, this list of conditions and the following disclaimer.
  13  *     * Redistributions in binary form must reproduce the above copyright
  14  *       notice, this list of conditions and the following disclaimer in
  15  *       the documentation and/or other materials provided with the
  16  *       distribution.
  17  *     * Neither the name of Intel Corporation nor the names of its
  18  *       contributors may be used to endorse or promote products derived
  19  *       from this software without specific prior written permission.
  20  *
  21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  32  */
  33
  34 #include <stdint.h>
  35 #include <stdbool.h>
  36 #include <linux/virtio_net.h>
  37
  38 #include <rte_mbuf.h>
  39 #include <rte_memcpy.h>
  40 #include <rte_ether.h>
  41 #include <rte_ip.h>
  42 #include <rte_vhost.h>
  43 #include <rte_tcp.h>
  44 #include <rte_udp.h>
  45 #include <rte_sctp.h>
  46 #include <rte_arp.h>
  47
  48 #include "iotlb.h"
  49 #include "vhost.h"
  50
  51 #define MAX_PKT_BURST 32
  52
  53 #define MAX_BATCH_LEN 256
  54
  55 static bool
  56 is_valid_virt_queue_idx(uint32_t idx, int is_tx, uint32_t nr_vring)
  57 {
  58         return (is_tx ^ (idx & 1)) == 0 && idx < nr_vring;
  59 }
  60
  61 static __rte_always_inline void
  62 do_flush_shadow_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq,
  63                           uint16_t to, uint16_t from, uint16_t size)
  64 {
  65         rte_memcpy(&vq->used->ring[to],
  66                         &vq->shadow_used_ring[from],
  67                         size * sizeof(struct vring_used_elem));
  68         vhost_log_used_vring(dev, vq,
  69                         offsetof(struct vring_used, ring[to]),
  70                         size * sizeof(struct vring_used_elem));
  71 }
  72
  73 static __rte_always_inline void
  74 flush_shadow_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq)
  75 {
  76         uint16_t used_idx = vq->last_used_idx & (vq->size - 1);
  77
  78         if (used_idx + vq->shadow_used_idx <= vq->size) {
  79                 do_flush_shadow_used_ring(dev, vq, used_idx, 0,
  80                                           vq->shadow_used_idx);
  81         } else {
  82                 uint16_t size;
  83
  84                 /* update used ring interval [used_idx, vq->size] */
  85                 size = vq->size - used_idx;
  86                 do_flush_shadow_used_ring(dev, vq, used_idx, 0, size);
  87
  88                 /* update the left half used ring interval [0, left_size] */
  89                 do_flush_shadow_used_ring(dev, vq, 0, size,
  90                                           vq->shadow_used_idx - size);
  91         }
  92         vq->last_used_idx += vq->shadow_used_idx;
  93
  94         rte_smp_wmb();
  95
  96         *(volatile uint16_t *)&vq->used->idx += vq->shadow_used_idx;
  97         vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
  98                 sizeof(vq->used->idx));
  99 }
 100
 101 static __rte_always_inline void
 102 update_shadow_used_ring(struct vhost_virtqueue *vq,
 103                          uint16_t desc_idx, uint16_t len)
 104 {
 105         uint16_t i = vq->shadow_used_idx++;
 106
 107         vq->shadow_used_ring[i].id  = desc_idx;
 108         vq->shadow_used_ring[i].len = len;
 109 }
 110
 111 static inline void
 112 do_data_copy_enqueue(struct virtio_net *dev, struct vhost_virtqueue *vq)
 113 {
 114         struct batch_copy_elem *elem = vq->batch_copy_elems;
 115         uint16_t count = vq->batch_copy_nb_elems;
 116         int i;
 117
 118         for (i = 0; i < count; i++) {
 119                 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
 120                 vhost_log_write(dev, elem[i].log_addr, elem[i].len);
 121                 PRINT_PACKET(dev, (uintptr_t)elem[i].dst, elem[i].len, 0);
 122         }
 123 }
 124
 125 static inline void
 126 do_data_copy_dequeue(struct vhost_virtqueue *vq)
 127 {
 128         struct batch_copy_elem *elem = vq->batch_copy_elems;
 129         uint16_t count = vq->batch_copy_nb_elems;
 130         int i;
 131
 132         for (i = 0; i < count; i++)
 133                 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
 134 }
 135
 136 /* avoid write operation when necessary, to lessen cache issues */
 137 #define ASSIGN_UNLESS_EQUAL(var, val) do {      \
 138         if ((var) != (val))                     \
 139                 (var) = (val);                  \
 140 } while (0)
 141
 142 static void
 143 virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
 144 {
 145         uint64_t csum_l4 = m_buf->ol_flags & PKT_TX_L4_MASK;
 146
 147         if (m_buf->ol_flags & PKT_TX_TCP_SEG)
 148                 csum_l4 |= PKT_TX_TCP_CKSUM;
 149
 150         if (csum_l4) {
 151                 net_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 152                 net_hdr->csum_start = m_buf->l2_len + m_buf->l3_len;
 153
 154                 switch (csum_l4) {
 155                 case PKT_TX_TCP_CKSUM:
 156                         net_hdr->csum_offset = (offsetof(struct tcp_hdr,
 157                                                 cksum));
 158                         break;
 159                 case PKT_TX_UDP_CKSUM:
 160                         net_hdr->csum_offset = (offsetof(struct udp_hdr,
 161                                                 dgram_cksum));
 162                         break;
 163                 case PKT_TX_SCTP_CKSUM:
 164                         net_hdr->csum_offset = (offsetof(struct sctp_hdr,
 165                                                 cksum));
 166                         break;
 167                 }
 168         } else {
 169                 ASSIGN_UNLESS_EQUAL(net_hdr->csum_start, 0);
 170                 ASSIGN_UNLESS_EQUAL(net_hdr->csum_offset, 0);
 171                 ASSIGN_UNLESS_EQUAL(net_hdr->flags, 0);
 172         }
 173
 174         /* IP cksum verification cannot be bypassed, then calculate here */
 175         if (m_buf->ol_flags & PKT_TX_IP_CKSUM) {
 176                 struct ipv4_hdr *ipv4_hdr;
 177
 178                 ipv4_hdr = rte_pktmbuf_mtod_offset(m_buf, struct ipv4_hdr *,
 179                                                    m_buf->l2_len);
 180                 ipv4_hdr->hdr_checksum = rte_ipv4_cksum(ipv4_hdr);
 181         }
 182
 183         if (m_buf->ol_flags & PKT_TX_TCP_SEG) {
 184                 if (m_buf->ol_flags & PKT_TX_IPV4)
 185                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 186                 else
 187                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 188                 net_hdr->gso_size = m_buf->tso_segsz;
 189                 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len
 190                                         + m_buf->l4_len;
 191         } else {
 192                 ASSIGN_UNLESS_EQUAL(net_hdr->gso_type, 0);
 193                 ASSIGN_UNLESS_EQUAL(net_hdr->gso_size, 0);
 194                 ASSIGN_UNLESS_EQUAL(net_hdr->hdr_len, 0);
 195         }
 196 }
 197
 198 static __rte_always_inline int
 199 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 200                   struct vring_desc *descs, struct rte_mbuf *m,
 201                   uint16_t desc_idx, uint32_t size)
 202 {
 203         uint32_t desc_avail, desc_offset;
 204         uint32_t mbuf_avail, mbuf_offset;
 205         uint32_t cpy_len;
 206         struct vring_desc *desc;
 207         uint64_t desc_addr;
 208         /* A counter to avoid desc dead loop chain */
 209         uint16_t nr_desc = 1;
 210         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 211         uint16_t copy_nb = vq->batch_copy_nb_elems;
 212         int error = 0;
 213
 214         desc = &descs[desc_idx];
 215         desc_addr = vhost_iova_to_vva(dev, vq, desc->addr,
 216                                         desc->len, VHOST_ACCESS_RW);
 217         /*
 218          * Checking of 'desc_addr' placed outside of 'unlikely' macro to avoid
 219          * performance issue with some versions of gcc (4.8.4 and 5.3.0) which
 220          * otherwise stores offset on the stack instead of in a register.
 221          */
 222         if (unlikely(desc->len < dev->vhost_hlen) || !desc_addr) {
 223                 error = -1;
 224                 goto out;
 225         }
 226
 227         rte_prefetch0((void *)(uintptr_t)desc_addr);
 228
 229         virtio_enqueue_offload(m, (struct virtio_net_hdr *)(uintptr_t)desc_addr);
 230         vhost_log_write(dev, desc->addr, dev->vhost_hlen);
 231         PRINT_PACKET(dev, (uintptr_t)desc_addr, dev->vhost_hlen, 0);
 232
 233         desc_offset = dev->vhost_hlen;
 234         desc_avail  = desc->len - dev->vhost_hlen;
 235
 236         mbuf_avail  = rte_pktmbuf_data_len(m);
 237         mbuf_offset = 0;
 238         while (mbuf_avail != 0 || m->next != NULL) {
 239                 /* done with current mbuf, fetch next */
 240                 if (mbuf_avail == 0) {
 241                         m = m->next;
 242
 243                         mbuf_offset = 0;
 244                         mbuf_avail  = rte_pktmbuf_data_len(m);
 245                 }
 246
 247                 /* done with current desc buf, fetch next */
 248                 if (desc_avail == 0) {
 249                         if ((desc->flags & VRING_DESC_F_NEXT) == 0) {
 250                                 /* Room in vring buffer is not enough */
 251                                 error = -1;
 252                                 goto out;
 253                         }
 254                         if (unlikely(desc->next >= size || ++nr_desc > size)) {
 255                                 error = -1;
 256                                 goto out;
 257                         }
 258
 259                         desc = &descs[desc->next];
 260                         desc_addr = vhost_iova_to_vva(dev, vq, desc->addr,
 261                                                         desc->len,
 262                                                         VHOST_ACCESS_RW);
 263                         if (unlikely(!desc_addr)) {
 264                                 error = -1;
 265                                 goto out;
 266                         }
 267
 268                         desc_offset = 0;
 269                         desc_avail  = desc->len;
 270                 }
 271
 272                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 273                 if (likely(cpy_len > MAX_BATCH_LEN || copy_nb >= vq->size)) {
 274                         rte_memcpy((void *)((uintptr_t)(desc_addr +
 275                                                         desc_offset)),
 276                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 277                                 cpy_len);
 278                         vhost_log_write(dev, desc->addr + desc_offset, cpy_len);
 279                         PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
 280                                      cpy_len, 0);
 281                 } else {
 282                         batch_copy[copy_nb].dst =
 283                                 (void *)((uintptr_t)(desc_addr + desc_offset));
 284                         batch_copy[copy_nb].src =
 285                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
 286                         batch_copy[copy_nb].log_addr = desc->addr + desc_offset;
 287                         batch_copy[copy_nb].len = cpy_len;
 288                         copy_nb++;
 289                 }
 290
 291                 mbuf_avail  -= cpy_len;
 292                 mbuf_offset += cpy_len;
 293                 desc_avail  -= cpy_len;
 294                 desc_offset += cpy_len;
 295         }
 296
 297 out:
 298         vq->batch_copy_nb_elems = copy_nb;
 299
 300         return error;
 301 }
 302
 303 /**
 304  * This function adds buffers to the virtio devices RX virtqueue. Buffers can
 305  * be received from the physical port or from another virtio device. A packet
 306  * count is returned to indicate the number of packets that are successfully
 307  * added to the RX queue. This function works when the mbuf is scattered, but
 308  * it doesn't support the mergeable feature.
 309  */
 310 static __rte_always_inline uint32_t
 311 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
 312               struct rte_mbuf **pkts, uint32_t count)
 313 {
 314         struct vhost_virtqueue *vq;
 315         uint16_t avail_idx, free_entries, start_idx;
 316         uint16_t desc_indexes[MAX_PKT_BURST];
 317         struct vring_desc *descs;
 318         uint16_t used_idx;
 319         uint32_t i, sz;
 320
 321         LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
 322         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
 323                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
 324                         dev->vid, __func__, queue_id);
 325                 return 0;
 326         }
 327
 328         vq = dev->virtqueue[queue_id];
 329         if (unlikely(vq->enabled == 0))
 330                 return 0;
 331
 332         avail_idx = *((volatile uint16_t *)&vq->avail->idx);
 333         start_idx = vq->last_used_idx;
 334         free_entries = avail_idx - start_idx;
 335         count = RTE_MIN(count, free_entries);
 336         count = RTE_MIN(count, (uint32_t)MAX_PKT_BURST);
 337         if (count == 0)
 338                 return 0;
 339
 340         LOG_DEBUG(VHOST_DATA, "(%d) start_idx %d | end_idx %d\n",
 341                 dev->vid, start_idx, start_idx + count);
 342
 343         vq->batch_copy_nb_elems = 0;
 344
 345         /* Retrieve all of the desc indexes first to avoid caching issues. */
 346         rte_prefetch0(&vq->avail->ring[start_idx & (vq->size - 1)]);
 347         for (i = 0; i < count; i++) {
 348                 used_idx = (start_idx + i) & (vq->size - 1);
 349                 desc_indexes[i] = vq->avail->ring[used_idx];
 350                 vq->used->ring[used_idx].id = desc_indexes[i];
 351                 vq->used->ring[used_idx].len = pkts[i]->pkt_len +
 352                                                dev->vhost_hlen;
 353                 vhost_log_used_vring(dev, vq,
 354                         offsetof(struct vring_used, ring[used_idx]),
 355                         sizeof(vq->used->ring[used_idx]));
 356         }
 357
 358         rte_prefetch0(&vq->desc[desc_indexes[0]]);
 359
 360         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 361                 vhost_user_iotlb_rd_lock(vq);
 362
 363         for (i = 0; i < count; i++) {
 364                 uint16_t desc_idx = desc_indexes[i];
 365                 int err;
 366
 367                 if (vq->desc[desc_idx].flags & VRING_DESC_F_INDIRECT) {
 368                         descs = (struct vring_desc *)(uintptr_t)
 369                                 vhost_iova_to_vva(dev,
 370                                                 vq, vq->desc[desc_idx].addr,
 371                                                 vq->desc[desc_idx].len,
 372                                                 VHOST_ACCESS_RO);
 373                         if (unlikely(!descs)) {
 374                                 count = i;
 375                                 break;
 376                         }
 377
 378                         desc_idx = 0;
 379                         sz = vq->desc[desc_idx].len / sizeof(*descs);
 380                 } else {
 381                         descs = vq->desc;
 382                         sz = vq->size;
 383                 }
 384
 385                 err = copy_mbuf_to_desc(dev, vq, descs, pkts[i], desc_idx, sz);
 386                 if (unlikely(err)) {
 387                         count = i;
 388                         break;
 389                 }
 390
 391                 if (i + 1 < count)
 392                         rte_prefetch0(&vq->desc[desc_indexes[i+1]]);
 393         }
 394
 395         do_data_copy_enqueue(dev, vq);
 396
 397         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 398                 vhost_user_iotlb_rd_unlock(vq);
 399
 400         rte_smp_wmb();
 401
 402         *(volatile uint16_t *)&vq->used->idx += count;
 403         vq->last_used_idx += count;
 404         vhost_log_used_vring(dev, vq,
 405                 offsetof(struct vring_used, idx),
 406                 sizeof(vq->used->idx));
 407
 408         /* flush used->idx update before we read avail->flags. */
 409         rte_mb();
 410
 411         /* Kick the guest if necessary. */
 412         if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
 413                         && (vq->callfd >= 0))
 414                 eventfd_write(vq->callfd, (eventfd_t)1);
 415         return count;
 416 }
 417
 418 static __rte_always_inline int
 419 fill_vec_buf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 420                          uint32_t avail_idx, uint32_t *vec_idx,
 421                          struct buf_vector *buf_vec, uint16_t *desc_chain_head,
 422                          uint16_t *desc_chain_len)
 423 {
 424         uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
 425         uint32_t vec_id = *vec_idx;
 426         uint32_t len    = 0;
 427         struct vring_desc *descs = vq->desc;
 428
 429         *desc_chain_head = idx;
 430
 431         if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
 432                 descs = (struct vring_desc *)(uintptr_t)
 433                         vhost_iova_to_vva(dev, vq, vq->desc[idx].addr,
 434                                                 vq->desc[idx].len,
 435                                                 VHOST_ACCESS_RO);
 436                 if (unlikely(!descs))
 437                         return -1;
 438
 439                 idx = 0;
 440         }
 441
 442         while (1) {
 443                 if (unlikely(vec_id >= BUF_VECTOR_MAX || idx >= vq->size))
 444                         return -1;
 445
 446                 len += descs[idx].len;
 447                 buf_vec[vec_id].buf_addr = descs[idx].addr;
 448                 buf_vec[vec_id].buf_len  = descs[idx].len;
 449                 buf_vec[vec_id].desc_idx = idx;
 450                 vec_id++;
 451
 452                 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
 453                         break;
 454
 455                 idx = descs[idx].next;
 456         }
 457
 458         *desc_chain_len = len;
 459         *vec_idx = vec_id;
 460
 461         return 0;
 462 }
 463
 464 /*
 465  * Returns -1 on fail, 0 on success
 466  */
 467 static inline int
 468 reserve_avail_buf_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
 469                                 uint32_t size, struct buf_vector *buf_vec,
 470                                 uint16_t *num_buffers, uint16_t avail_head)
 471 {
 472         uint16_t cur_idx;
 473         uint32_t vec_idx = 0;
 474         uint16_t tries = 0;
 475
 476         uint16_t head_idx = 0;
 477         uint16_t len = 0;
 478
 479         *num_buffers = 0;
 480         cur_idx  = vq->last_avail_idx;
 481
 482         while (size > 0) {
 483                 if (unlikely(cur_idx == avail_head))
 484                         return -1;
 485
 486                 if (unlikely(fill_vec_buf(dev, vq, cur_idx, &vec_idx, buf_vec,
 487                                                 &head_idx, &len) < 0))
 488                         return -1;
 489                 len = RTE_MIN(len, size);
 490                 update_shadow_used_ring(vq, head_idx, len);
 491                 size -= len;
 492
 493                 cur_idx++;
 494                 tries++;
 495                 *num_buffers += 1;
 496
 497                 /*
 498                  * if we tried all available ring items, and still
 499                  * can't get enough buf, it means something abnormal
 500                  * happened.
 501                  */
 502                 if (unlikely(tries >= vq->size))
 503                         return -1;
 504         }
 505
 506         return 0;
 507 }
 508
 509 static __rte_always_inline int
 510 copy_mbuf_to_desc_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
 511                             struct rte_mbuf *m, struct buf_vector *buf_vec,
 512                             uint16_t num_buffers)
 513 {
 514         uint32_t vec_idx = 0;
 515         uint64_t desc_addr;
 516         uint32_t mbuf_offset, mbuf_avail;
 517         uint32_t desc_offset, desc_avail;
 518         uint32_t cpy_len;
 519         uint64_t hdr_addr, hdr_phys_addr;
 520         struct rte_mbuf *hdr_mbuf;
 521         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 522         uint16_t copy_nb = vq->batch_copy_nb_elems;
 523         int error = 0;
 524
 525         if (unlikely(m == NULL)) {
 526                 error = -1;
 527                 goto out;
 528         }
 529
 530         desc_addr = vhost_iova_to_vva(dev, vq, buf_vec[vec_idx].buf_addr,
 531                                                 buf_vec[vec_idx].buf_len,
 532                                                 VHOST_ACCESS_RW);
 533         if (buf_vec[vec_idx].buf_len < dev->vhost_hlen || !desc_addr) {
 534                 error = -1;
 535                 goto out;
 536         }
 537
 538         hdr_mbuf = m;
 539         hdr_addr = desc_addr;
 540         hdr_phys_addr = buf_vec[vec_idx].buf_addr;
 541         rte_prefetch0((void *)(uintptr_t)hdr_addr);
 542
 543         LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
 544                 dev->vid, num_buffers);
 545
 546         desc_avail  = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
 547         desc_offset = dev->vhost_hlen;
 548
 549         mbuf_avail  = rte_pktmbuf_data_len(m);
 550         mbuf_offset = 0;
 551         while (mbuf_avail != 0 || m->next != NULL) {
 552                 /* done with current desc buf, get the next one */
 553                 if (desc_avail == 0) {
 554                         vec_idx++;
 555                         desc_addr =
 556                                 vhost_iova_to_vva(dev, vq,
 557                                         buf_vec[vec_idx].buf_addr,
 558                                         buf_vec[vec_idx].buf_len,
 559                                         VHOST_ACCESS_RW);
 560                         if (unlikely(!desc_addr)) {
 561                                 error = -1;
 562                                 goto out;
 563                         }
 564
 565                         /* Prefetch buffer address. */
 566                         rte_prefetch0((void *)(uintptr_t)desc_addr);
 567                         desc_offset = 0;
 568                         desc_avail  = buf_vec[vec_idx].buf_len;
 569                 }
 570
 571                 /* done with current mbuf, get the next one */
 572                 if (mbuf_avail == 0) {
 573                         m = m->next;
 574
 575                         mbuf_offset = 0;
 576                         mbuf_avail  = rte_pktmbuf_data_len(m);
 577                 }
 578
 579                 if (hdr_addr) {
 580                         struct virtio_net_hdr_mrg_rxbuf *hdr;
 581
 582                         hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)
 583                                 hdr_addr;
 584                         virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
 585                         ASSIGN_UNLESS_EQUAL(hdr->num_buffers, num_buffers);
 586
 587                         vhost_log_write(dev, hdr_phys_addr, dev->vhost_hlen);
 588                         PRINT_PACKET(dev, (uintptr_t)hdr_addr,
 589                                      dev->vhost_hlen, 0);
 590
 591                         hdr_addr = 0;
 592                 }
 593
 594                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 595
 596                 if (likely(cpy_len > MAX_BATCH_LEN || copy_nb >= vq->size)) {
 597                         rte_memcpy((void *)((uintptr_t)(desc_addr +
 598                                                         desc_offset)),
 599                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 600                                 cpy_len);
 601                         vhost_log_write(dev,
 602                                 buf_vec[vec_idx].buf_addr + desc_offset,
 603                                 cpy_len);
 604                         PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
 605                                 cpy_len, 0);
 606                 } else {
 607                         batch_copy[copy_nb].dst =
 608                                 (void *)((uintptr_t)(desc_addr + desc_offset));
 609                         batch_copy[copy_nb].src =
 610                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
 611                         batch_copy[copy_nb].log_addr =
 612                                 buf_vec[vec_idx].buf_addr + desc_offset;
 613                         batch_copy[copy_nb].len = cpy_len;
 614                         copy_nb++;
 615                 }
 616
 617                 mbuf_avail  -= cpy_len;
 618                 mbuf_offset += cpy_len;
 619                 desc_avail  -= cpy_len;
 620                 desc_offset += cpy_len;
 621         }
 622
 623 out:
 624         vq->batch_copy_nb_elems = copy_nb;
 625
 626         return error;
 627 }
 628
 629 static __rte_always_inline uint32_t
 630 virtio_dev_merge_rx(struct virtio_net *dev, uint16_t queue_id,
 631         struct rte_mbuf **pkts, uint32_t count)
 632 {
 633         struct vhost_virtqueue *vq;
 634         uint32_t pkt_idx = 0;
 635         uint16_t num_buffers;
 636         struct buf_vector buf_vec[BUF_VECTOR_MAX];
 637         uint16_t avail_head;
 638
 639         LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
 640         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
 641                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
 642                         dev->vid, __func__, queue_id);
 643                 return 0;
 644         }
 645
 646         vq = dev->virtqueue[queue_id];
 647         if (unlikely(vq->enabled == 0))
 648                 return 0;
 649
 650         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
 651         if (count == 0)
 652                 return 0;
 653
 654         vq->batch_copy_nb_elems = 0;
 655
 656         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
 657
 658         vq->shadow_used_idx = 0;
 659         avail_head = *((volatile uint16_t *)&vq->avail->idx);
 660
 661         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 662                 vhost_user_iotlb_rd_lock(vq);
 663
 664         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
 665                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
 666
 667                 if (unlikely(reserve_avail_buf_mergeable(dev, vq,
 668                                                 pkt_len, buf_vec, &num_buffers,
 669                                                 avail_head) < 0)) {
 670                         LOG_DEBUG(VHOST_DATA,
 671                                 "(%d) failed to get enough desc from vring\n",
 672                                 dev->vid);
 673                         vq->shadow_used_idx -= num_buffers;
 674                         break;
 675                 }
 676
 677                 LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
 678                         dev->vid, vq->last_avail_idx,
 679                         vq->last_avail_idx + num_buffers);
 680
 681                 if (copy_mbuf_to_desc_mergeable(dev, vq, pkts[pkt_idx],
 682                                                 buf_vec, num_buffers) < 0) {
 683                         vq->shadow_used_idx -= num_buffers;
 684                         break;
 685                 }
 686
 687                 vq->last_avail_idx += num_buffers;
 688         }
 689
 690         do_data_copy_enqueue(dev, vq);
 691
 692         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 693                 vhost_user_iotlb_rd_unlock(vq);
 694
 695         if (likely(vq->shadow_used_idx)) {
 696                 flush_shadow_used_ring(dev, vq);
 697
 698                 /* flush used->idx update before we read avail->flags. */
 699                 rte_mb();
 700
 701                 /* Kick the guest if necessary. */
 702                 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
 703                                 && (vq->callfd >= 0))
 704                         eventfd_write(vq->callfd, (eventfd_t)1);
 705         }
 706
 707         return pkt_idx;
 708 }
 709
 710 uint16_t
 711 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
 712         struct rte_mbuf **pkts, uint16_t count)
 713 {
 714         struct virtio_net *dev = get_device(vid);
 715
 716         if (!dev)
 717                 return 0;
 718
 719         if (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF))
 720                 return virtio_dev_merge_rx(dev, queue_id, pkts, count);
 721         else
 722                 return virtio_dev_rx(dev, queue_id, pkts, count);
 723 }
 724
 725 static inline bool
 726 virtio_net_with_host_offload(struct virtio_net *dev)
 727 {
 728         if (dev->features &
 729                         ((1ULL << VIRTIO_NET_F_CSUM) |
 730                          (1ULL << VIRTIO_NET_F_HOST_ECN) |
 731                          (1ULL << VIRTIO_NET_F_HOST_TSO4) |
 732                          (1ULL << VIRTIO_NET_F_HOST_TSO6) |
 733                          (1ULL << VIRTIO_NET_F_HOST_UFO)))
 734                 return true;
 735
 736         return false;
 737 }
 738
 739 static void
 740 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
 741 {
 742         struct ipv4_hdr *ipv4_hdr;
 743         struct ipv6_hdr *ipv6_hdr;
 744         void *l3_hdr = NULL;
 745         struct ether_hdr *eth_hdr;
 746         uint16_t ethertype;
 747
 748         eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
 749
 750         m->l2_len = sizeof(struct ether_hdr);
 751         ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
 752
 753         if (ethertype == ETHER_TYPE_VLAN) {
 754                 struct vlan_hdr *vlan_hdr = (struct vlan_hdr *)(eth_hdr + 1);
 755
 756                 m->l2_len += sizeof(struct vlan_hdr);
 757                 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
 758         }
 759
 760         l3_hdr = (char *)eth_hdr + m->l2_len;
 761
 762         switch (ethertype) {
 763         case ETHER_TYPE_IPv4:
 764                 ipv4_hdr = l3_hdr;
 765                 *l4_proto = ipv4_hdr->next_proto_id;
 766                 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
 767                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 768                 m->ol_flags |= PKT_TX_IPV4;
 769                 break;
 770         case ETHER_TYPE_IPv6:
 771                 ipv6_hdr = l3_hdr;
 772                 *l4_proto = ipv6_hdr->proto;
 773                 m->l3_len = sizeof(struct ipv6_hdr);
 774                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 775                 m->ol_flags |= PKT_TX_IPV6;
 776                 break;
 777         default:
 778                 m->l3_len = 0;
 779                 *l4_proto = 0;
 780                 *l4_hdr = NULL;
 781                 break;
 782         }
 783 }
 784
 785 static __rte_always_inline void
 786 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
 787 {
 788         uint16_t l4_proto = 0;
 789         void *l4_hdr = NULL;
 790         struct tcp_hdr *tcp_hdr = NULL;
 791
 792         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
 793                 return;
 794
 795         parse_ethernet(m, &l4_proto, &l4_hdr);
 796         if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 797                 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
 798                         switch (hdr->csum_offset) {
 799                         case (offsetof(struct tcp_hdr, cksum)):
 800                                 if (l4_proto == IPPROTO_TCP)
 801                                         m->ol_flags |= PKT_TX_TCP_CKSUM;
 802                                 break;
 803                         case (offsetof(struct udp_hdr, dgram_cksum)):
 804                                 if (l4_proto == IPPROTO_UDP)
 805                                         m->ol_flags |= PKT_TX_UDP_CKSUM;
 806                                 break;
 807                         case (offsetof(struct sctp_hdr, cksum)):
 808                                 if (l4_proto == IPPROTO_SCTP)
 809                                         m->ol_flags |= PKT_TX_SCTP_CKSUM;
 810                                 break;
 811                         default:
 812                                 break;
 813                         }
 814                 }
 815         }
 816
 817         if (l4_hdr && hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 818                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 819                 case VIRTIO_NET_HDR_GSO_TCPV4:
 820                 case VIRTIO_NET_HDR_GSO_TCPV6:
 821                         tcp_hdr = l4_hdr;
 822                         m->ol_flags |= PKT_TX_TCP_SEG;
 823                         m->tso_segsz = hdr->gso_size;
 824                         m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
 825                         break;
 826                 default:
 827                         RTE_LOG(WARNING, VHOST_DATA,
 828                                 "unsupported gso type %u.\n", hdr->gso_type);
 829                         break;
 830                 }
 831         }
 832 }
 833
 834 #define RARP_PKT_SIZE   64
 835
 836 static int
 837 make_rarp_packet(struct rte_mbuf *rarp_mbuf, const struct ether_addr *mac)
 838 {
 839         struct ether_hdr *eth_hdr;
 840         struct arp_hdr  *rarp;
 841
 842         if (rarp_mbuf->buf_len < 64) {
 843                 RTE_LOG(WARNING, VHOST_DATA,
 844                         "failed to make RARP; mbuf size too small %u (< %d)\n",
 845                         rarp_mbuf->buf_len, RARP_PKT_SIZE);
 846                 return -1;
 847         }
 848
 849         /* Ethernet header. */
 850         eth_hdr = rte_pktmbuf_mtod_offset(rarp_mbuf, struct ether_hdr *, 0);
 851         memset(eth_hdr->d_addr.addr_bytes, 0xff, ETHER_ADDR_LEN);
 852         ether_addr_copy(mac, &eth_hdr->s_addr);
 853         eth_hdr->ether_type = htons(ETHER_TYPE_RARP);
 854
 855         /* RARP header. */
 856         rarp = (struct arp_hdr *)(eth_hdr + 1);
 857         rarp->arp_hrd = htons(ARP_HRD_ETHER);
 858         rarp->arp_pro = htons(ETHER_TYPE_IPv4);
 859         rarp->arp_hln = ETHER_ADDR_LEN;
 860         rarp->arp_pln = 4;
 861         rarp->arp_op  = htons(ARP_OP_REVREQUEST);
 862
 863         ether_addr_copy(mac, &rarp->arp_data.arp_sha);
 864         ether_addr_copy(mac, &rarp->arp_data.arp_tha);
 865         memset(&rarp->arp_data.arp_sip, 0x00, 4);
 866         memset(&rarp->arp_data.arp_tip, 0x00, 4);
 867
 868         rarp_mbuf->pkt_len  = rarp_mbuf->data_len = RARP_PKT_SIZE;
 869
 870         return 0;
 871 }
 872
 873 static __rte_always_inline void
 874 put_zmbuf(struct zcopy_mbuf *zmbuf)
 875 {
 876         zmbuf->in_use = 0;
 877 }
 878
 879 static __rte_always_inline int
 880 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 881                   struct vring_desc *descs, uint16_t max_desc,
 882                   struct rte_mbuf *m, uint16_t desc_idx,
 883                   struct rte_mempool *mbuf_pool)
 884 {
 885         struct vring_desc *desc;
 886         uint64_t desc_addr;
 887         uint32_t desc_avail, desc_offset;
 888         uint32_t mbuf_avail, mbuf_offset;
 889         uint32_t cpy_len;
 890         struct rte_mbuf *cur = m, *prev = m;
 891         struct virtio_net_hdr *hdr = NULL;
 892         /* A counter to avoid desc dead loop chain */
 893         uint32_t nr_desc = 1;
 894         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 895         uint16_t copy_nb = vq->batch_copy_nb_elems;
 896         int error = 0;
 897
 898         desc = &descs[desc_idx];
 899         if (unlikely((desc->len < dev->vhost_hlen)) ||
 900                         (desc->flags & VRING_DESC_F_INDIRECT)) {
 901                 error = -1;
 902                 goto out;
 903         }
 904
 905         desc_addr = vhost_iova_to_vva(dev,
 906                                         vq, desc->addr,
 907                                         desc->len,
 908                                         VHOST_ACCESS_RO);
 909         if (unlikely(!desc_addr)) {
 910                 error = -1;
 911                 goto out;
 912         }
 913
 914         if (virtio_net_with_host_offload(dev)) {
 915                 hdr = (struct virtio_net_hdr *)((uintptr_t)desc_addr);
 916                 rte_prefetch0(hdr);
 917         }
 918
 919         /*
 920          * A virtio driver normally uses at least 2 desc buffers
 921          * for Tx: the first for storing the header, and others
 922          * for storing the data.
 923          */
 924         if (likely((desc->len == dev->vhost_hlen) &&
 925                    (desc->flags & VRING_DESC_F_NEXT) != 0)) {
 926                 desc = &descs[desc->next];
 927                 if (unlikely(desc->flags & VRING_DESC_F_INDIRECT)) {
 928                         error = -1;
 929                         goto out;
 930                 }
 931
 932                 desc_addr = vhost_iova_to_vva(dev,
 933                                                         vq, desc->addr,
 934                                                         desc->len,
 935                                                         VHOST_ACCESS_RO);
 936                 if (unlikely(!desc_addr)) {
 937                         error = -1;
 938                         goto out;
 939                 }
 940
 941                 desc_offset = 0;
 942                 desc_avail  = desc->len;
 943                 nr_desc    += 1;
 944         } else {
 945                 desc_avail  = desc->len - dev->vhost_hlen;
 946                 desc_offset = dev->vhost_hlen;
 947         }
 948
 949         rte_prefetch0((void *)(uintptr_t)(desc_addr + desc_offset));
 950
 951         PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset), desc_avail, 0);
 952
 953         mbuf_offset = 0;
 954         mbuf_avail  = m->buf_len - RTE_PKTMBUF_HEADROOM;
 955         while (1) {
 956                 uint64_t hpa;
 957
 958                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 959
 960                 /*
 961                  * A desc buf might across two host physical pages that are
 962                  * not continuous. In such case (gpa_to_hpa returns 0), data
 963                  * will be copied even though zero copy is enabled.
 964                  */
 965                 if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
 966                                         desc->addr + desc_offset, cpy_len)))) {
 967                         cur->data_len = cpy_len;
 968                         cur->data_off = 0;
 969                         cur->buf_addr = (void *)(uintptr_t)desc_addr;
 970                         cur->buf_physaddr = hpa;
 971
 972                         /*
 973                          * In zero copy mode, one mbuf can only reference data
 974                          * for one or partial of one desc buff.
 975                          */
 976                         mbuf_avail = cpy_len;
 977                 } else {
 978                         if (likely(cpy_len > MAX_BATCH_LEN ||
 979                                    copy_nb >= vq->size)) {
 980                                 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
 981                                                                    mbuf_offset),
 982                                            (void *)((uintptr_t)(desc_addr +
 983                                                                 desc_offset)),
 984                                            cpy_len);
 985                         } else {
 986                                 batch_copy[copy_nb].dst =
 987                                         rte_pktmbuf_mtod_offset(cur, void *,
 988                                                                 mbuf_offset);
 989                                 batch_copy[copy_nb].src =
 990                                         (void *)((uintptr_t)(desc_addr +
 991                                                              desc_offset));
 992                                 batch_copy[copy_nb].len = cpy_len;
 993                                 copy_nb++;
 994                         }
 995                 }
 996
 997                 mbuf_avail  -= cpy_len;
 998                 mbuf_offset += cpy_len;
 999                 desc_avail  -= cpy_len;
1000                 desc_offset += cpy_len;
1001
1002                 /* This desc reaches to its end, get the next one */
1003                 if (desc_avail == 0) {
1004                         if ((desc->flags & VRING_DESC_F_NEXT) == 0)
1005                                 break;
1006
1007                         if (unlikely(desc->next >= max_desc ||
1008                                      ++nr_desc > max_desc)) {
1009                                 error = -1;
1010                                 goto out;
1011                         }
1012                         desc = &descs[desc->next];
1013                         if (unlikely(desc->flags & VRING_DESC_F_INDIRECT)) {
1014                                 error = -1;
1015                                 goto out;
1016                         }
1017
1018                         desc_addr = vhost_iova_to_vva(dev,
1019                                                         vq, desc->addr,
1020                                                         desc->len,
1021                                                         VHOST_ACCESS_RO);
1022                         if (unlikely(!desc_addr)) {
1023                                 error = -1;
1024                                 goto out;
1025                         }
1026
1027                         rte_prefetch0((void *)(uintptr_t)desc_addr);
1028
1029                         desc_offset = 0;
1030                         desc_avail  = desc->len;
1031
1032                         PRINT_PACKET(dev, (uintptr_t)desc_addr, desc->len, 0);
1033                 }
1034
1035                 /*
1036                  * This mbuf reaches to its end, get a new one
1037                  * to hold more data.
1038                  */
1039                 if (mbuf_avail == 0) {
1040                         cur = rte_pktmbuf_alloc(mbuf_pool);
1041                         if (unlikely(cur == NULL)) {
1042                                 RTE_LOG(ERR, VHOST_DATA, "Failed to "
1043                                         "allocate memory for mbuf.\n");
1044                                 error = -1;
1045                                 goto out;
1046                         }
1047                         if (unlikely(dev->dequeue_zero_copy))
1048                                 rte_mbuf_refcnt_update(cur, 1);
1049
1050                         prev->next = cur;
1051                         prev->data_len = mbuf_offset;
1052                         m->nb_segs += 1;
1053                         m->pkt_len += mbuf_offset;
1054                         prev = cur;
1055
1056                         mbuf_offset = 0;
1057                         mbuf_avail  = cur->buf_len - RTE_PKTMBUF_HEADROOM;
1058                 }
1059         }
1060
1061         prev->data_len = mbuf_offset;
1062         m->pkt_len    += mbuf_offset;
1063
1064         if (hdr)
1065                 vhost_dequeue_offload(hdr, m);
1066
1067 out:
1068         vq->batch_copy_nb_elems = copy_nb;
1069
1070         return error;
1071 }
1072
1073 static __rte_always_inline void
1074 update_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq,
1075                  uint32_t used_idx, uint32_t desc_idx)
1076 {
1077         vq->used->ring[used_idx].id  = desc_idx;
1078         vq->used->ring[used_idx].len = 0;
1079         vhost_log_used_vring(dev, vq,
1080                         offsetof(struct vring_used, ring[used_idx]),
1081                         sizeof(vq->used->ring[used_idx]));
1082 }
1083
1084 static __rte_always_inline void
1085 update_used_idx(struct virtio_net *dev, struct vhost_virtqueue *vq,
1086                 uint32_t count)
1087 {
1088         if (unlikely(count == 0))
1089                 return;
1090
1091         rte_smp_wmb();
1092         rte_smp_rmb();
1093
1094         vq->used->idx += count;
1095         vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
1096                         sizeof(vq->used->idx));
1097
1098         /* Kick guest if required. */
1099         if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
1100                         && (vq->callfd >= 0))
1101                 eventfd_write(vq->callfd, (eventfd_t)1);
1102 }
1103
1104 static __rte_always_inline struct zcopy_mbuf *
1105 get_zmbuf(struct vhost_virtqueue *vq)
1106 {
1107         uint16_t i;
1108         uint16_t last;
1109         int tries = 0;
1110
1111         /* search [last_zmbuf_idx, zmbuf_size) */
1112         i = vq->last_zmbuf_idx;
1113         last = vq->zmbuf_size;
1114
1115 again:
1116         for (; i < last; i++) {
1117                 if (vq->zmbufs[i].in_use == 0) {
1118                         vq->last_zmbuf_idx = i + 1;
1119                         vq->zmbufs[i].in_use = 1;
1120                         return &vq->zmbufs[i];
1121                 }
1122         }
1123
1124         tries++;
1125         if (tries == 1) {
1126                 /* search [0, last_zmbuf_idx) */
1127                 i = 0;
1128                 last = vq->last_zmbuf_idx;
1129                 goto again;
1130         }
1131
1132         return NULL;
1133 }
1134
1135 static __rte_always_inline bool
1136 mbuf_is_consumed(struct rte_mbuf *m)
1137 {
1138         while (m) {
1139                 if (rte_mbuf_refcnt_read(m) > 1)
1140                         return false;
1141                 m = m->next;
1142         }
1143
1144         return true;
1145 }
1146
1147 uint16_t
1148 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
1149         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1150 {
1151         struct virtio_net *dev;
1152         struct rte_mbuf *rarp_mbuf = NULL;
1153         struct vhost_virtqueue *vq;
1154         uint32_t desc_indexes[MAX_PKT_BURST];
1155         uint32_t used_idx;
1156         uint32_t i = 0;
1157         uint16_t free_entries;
1158         uint16_t avail_idx;
1159
1160         dev = get_device(vid);
1161         if (!dev)
1162                 return 0;
1163
1164         if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
1165                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
1166                         dev->vid, __func__, queue_id);
1167                 return 0;
1168         }
1169
1170         vq = dev->virtqueue[queue_id];
1171         if (unlikely(vq->enabled == 0))
1172                 return 0;
1173
1174         vq->batch_copy_nb_elems = 0;
1175
1176         if (unlikely(dev->dequeue_zero_copy)) {
1177                 struct zcopy_mbuf *zmbuf, *next;
1178                 int nr_updated = 0;
1179
1180                 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1181                      zmbuf != NULL; zmbuf = next) {
1182                         next = TAILQ_NEXT(zmbuf, next);
1183
1184                         if (mbuf_is_consumed(zmbuf->mbuf)) {
1185                                 used_idx = vq->last_used_idx++ & (vq->size - 1);
1186                                 update_used_ring(dev, vq, used_idx,
1187                                                  zmbuf->desc_idx);
1188                                 nr_updated += 1;
1189
1190                                 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1191                                 rte_pktmbuf_free(zmbuf->mbuf);
1192                                 put_zmbuf(zmbuf);
1193                                 vq->nr_zmbuf -= 1;
1194                         }
1195                 }
1196
1197                 update_used_idx(dev, vq, nr_updated);
1198         }
1199
1200         /*
1201          * Construct a RARP broadcast packet, and inject it to the "pkts"
1202          * array, to looks like that guest actually send such packet.
1203          *
1204          * Check user_send_rarp() for more information.
1205          *
1206          * broadcast_rarp shares a cacheline in the virtio_net structure
1207          * with some fields that are accessed during enqueue and
1208          * rte_atomic16_cmpset() causes a write if using cmpxchg. This could
1209          * result in false sharing between enqueue and dequeue.
1210          *
1211          * Prevent unnecessary false sharing by reading broadcast_rarp first
1212          * and only performing cmpset if the read indicates it is likely to
1213          * be set.
1214          */
1215
1216         if (unlikely(rte_atomic16_read(&dev->broadcast_rarp) &&
1217                         rte_atomic16_cmpset((volatile uint16_t *)
1218                                 &dev->broadcast_rarp.cnt, 1, 0))) {
1219
1220                 rarp_mbuf = rte_pktmbuf_alloc(mbuf_pool);
1221                 if (rarp_mbuf == NULL) {
1222                         RTE_LOG(ERR, VHOST_DATA,
1223                                 "Failed to allocate memory for mbuf.\n");
1224                         return 0;
1225                 }
1226
1227                 if (make_rarp_packet(rarp_mbuf, &dev->mac)) {
1228                         rte_pktmbuf_free(rarp_mbuf);
1229                         rarp_mbuf = NULL;
1230                 } else {
1231                         count -= 1;
1232                 }
1233         }
1234
1235         free_entries = *((volatile uint16_t *)&vq->avail->idx) -
1236                         vq->last_avail_idx;
1237         if (free_entries == 0)
1238                 goto out;
1239
1240         LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1241
1242         /* Prefetch available and used ring */
1243         avail_idx = vq->last_avail_idx & (vq->size - 1);
1244         used_idx  = vq->last_used_idx  & (vq->size - 1);
1245         rte_prefetch0(&vq->avail->ring[avail_idx]);
1246         rte_prefetch0(&vq->used->ring[used_idx]);
1247
1248         count = RTE_MIN(count, MAX_PKT_BURST);
1249         count = RTE_MIN(count, free_entries);
1250         LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1251                         dev->vid, count);
1252
1253         /* Retrieve all of the head indexes first to avoid caching issues. */
1254         for (i = 0; i < count; i++) {
1255                 avail_idx = (vq->last_avail_idx + i) & (vq->size - 1);
1256                 used_idx  = (vq->last_used_idx  + i) & (vq->size - 1);
1257                 desc_indexes[i] = vq->avail->ring[avail_idx];
1258
1259                 if (likely(dev->dequeue_zero_copy == 0))
1260                         update_used_ring(dev, vq, used_idx, desc_indexes[i]);
1261         }
1262
1263         /* Prefetch descriptor index. */
1264         rte_prefetch0(&vq->desc[desc_indexes[0]]);
1265
1266         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1267                 vhost_user_iotlb_rd_lock(vq);
1268
1269         for (i = 0; i < count; i++) {
1270                 struct vring_desc *desc;
1271                 uint16_t sz, idx;
1272                 int err;
1273
1274                 if (likely(i + 1 < count))
1275                         rte_prefetch0(&vq->desc[desc_indexes[i + 1]]);
1276
1277                 if (vq->desc[desc_indexes[i]].flags & VRING_DESC_F_INDIRECT) {
1278                         desc = (struct vring_desc *)(uintptr_t)
1279                                 vhost_iova_to_vva(dev, vq,
1280                                                 vq->desc[desc_indexes[i]].addr,
1281                                                 sizeof(*desc),
1282                                                 VHOST_ACCESS_RO);
1283                         if (unlikely(!desc))
1284                                 break;
1285
1286                         rte_prefetch0(desc);
1287                         sz = vq->desc[desc_indexes[i]].len / sizeof(*desc);
1288                         idx = 0;
1289                 } else {
1290                         desc = vq->desc;
1291                         sz = vq->size;
1292                         idx = desc_indexes[i];
1293                 }
1294
1295                 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
1296                 if (unlikely(pkts[i] == NULL)) {
1297                         RTE_LOG(ERR, VHOST_DATA,
1298                                 "Failed to allocate memory for mbuf.\n");
1299                         break;
1300                 }
1301
1302                 err = copy_desc_to_mbuf(dev, vq, desc, sz, pkts[i], idx,
1303                                         mbuf_pool);
1304                 if (unlikely(err)) {
1305                         rte_pktmbuf_free(pkts[i]);
1306                         break;
1307                 }
1308
1309                 if (unlikely(dev->dequeue_zero_copy)) {
1310                         struct zcopy_mbuf *zmbuf;
1311
1312                         zmbuf = get_zmbuf(vq);
1313                         if (!zmbuf) {
1314                                 rte_pktmbuf_free(pkts[i]);
1315                                 break;
1316                         }
1317                         zmbuf->mbuf = pkts[i];
1318                         zmbuf->desc_idx = desc_indexes[i];
1319
1320                         /*
1321                          * Pin lock the mbuf; we will check later to see
1322                          * whether the mbuf is freed (when we are the last
1323                          * user) or not. If that's the case, we then could
1324                          * update the used ring safely.
1325                          */
1326                         rte_mbuf_refcnt_update(pkts[i], 1);
1327
1328                         vq->nr_zmbuf += 1;
1329                         TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1330                 }
1331         }
1332         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1333                 vhost_user_iotlb_rd_unlock(vq);
1334
1335         vq->last_avail_idx += i;
1336
1337         if (likely(dev->dequeue_zero_copy == 0)) {
1338                 do_data_copy_dequeue(vq);
1339                 vq->last_used_idx += i;
1340                 update_used_idx(dev, vq, i);
1341         }
1342
1343 out:
1344         if (unlikely(rarp_mbuf != NULL)) {
1345                 /*
1346                  * Inject it to the head of "pkts" array, so that switch's mac
1347                  * learning table will get updated first.
1348                  */
1349                 memmove(&pkts[1], pkts, i * sizeof(struct rte_mbuf *));
1350                 pkts[0] = rarp_mbuf;
1351                 i += 1;
1352         }
1353
1354         return i;
1355 }