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 #include <rte_spinlock.h>
  19 #include <rte_malloc.h>
  20
  21 #include "iotlb.h"
  22 #include "vhost.h"
  23
  24 #define MAX_PKT_BURST 32
  25
  26 #define MAX_BATCH_LEN 256
  27
  28 static  __rte_always_inline bool
  29 rxvq_is_mergeable(struct virtio_net *dev)
  30 {
  31         return dev->features & (1ULL << VIRTIO_NET_F_MRG_RXBUF);
  32 }
  33
  34 static bool
  35 is_valid_virt_queue_idx(uint32_t idx, int is_tx, uint32_t nr_vring)
  36 {
  37         return (is_tx ^ (idx & 1)) == 0 && idx < nr_vring;
  38 }
  39
  40 static __rte_always_inline void *
  41 alloc_copy_ind_table(struct virtio_net *dev, struct vhost_virtqueue *vq,
  42                 uint64_t desc_addr, uint64_t desc_len)
  43 {
  44         void *idesc;
  45         uint64_t src, dst;
  46         uint64_t len, remain = desc_len;
  47
  48         idesc = rte_malloc(__func__, desc_len, 0);
  49         if (unlikely(!idesc))
  50                 return 0;
  51
  52         dst = (uint64_t)(uintptr_t)idesc;
  53
  54         while (remain) {
  55                 len = remain;
  56                 src = vhost_iova_to_vva(dev, vq, desc_addr, &len,
  57                                 VHOST_ACCESS_RO);
  58                 if (unlikely(!src || !len)) {
  59                         rte_free(idesc);
  60                         return 0;
  61                 }
  62
  63                 rte_memcpy((void *)(uintptr_t)dst, (void *)(uintptr_t)src, len);
  64
  65                 remain -= len;
  66                 dst += len;
  67                 desc_addr += len;
  68         }
  69
  70         return idesc;
  71 }
  72
  73 static __rte_always_inline void
  74 free_ind_table(void *idesc)
  75 {
  76         rte_free(idesc);
  77 }
  78
  79 static __rte_always_inline void
  80 do_flush_shadow_used_ring_split(struct virtio_net *dev,
  81                         struct vhost_virtqueue *vq,
  82                         uint16_t to, uint16_t from, uint16_t size)
  83 {
  84         rte_memcpy(&vq->used->ring[to],
  85                         &vq->shadow_used_split[from],
  86                         size * sizeof(struct vring_used_elem));
  87         vhost_log_cache_used_vring(dev, vq,
  88                         offsetof(struct vring_used, ring[to]),
  89                         size * sizeof(struct vring_used_elem));
  90 }
  91
  92 static __rte_always_inline void
  93 flush_shadow_used_ring_split(struct virtio_net *dev, struct vhost_virtqueue *vq)
  94 {
  95         uint16_t used_idx = vq->last_used_idx & (vq->size - 1);
  96
  97         if (used_idx + vq->shadow_used_idx <= vq->size) {
  98                 do_flush_shadow_used_ring_split(dev, vq, used_idx, 0,
  99                                           vq->shadow_used_idx);
 100         } else {
 101                 uint16_t size;
 102
 103                 /* update used ring interval [used_idx, vq->size] */
 104                 size = vq->size - used_idx;
 105                 do_flush_shadow_used_ring_split(dev, vq, used_idx, 0, size);
 106
 107                 /* update the left half used ring interval [0, left_size] */
 108                 do_flush_shadow_used_ring_split(dev, vq, 0, size,
 109                                           vq->shadow_used_idx - size);
 110         }
 111         vq->last_used_idx += vq->shadow_used_idx;
 112
 113         rte_smp_wmb();
 114
 115         vhost_log_cache_sync(dev, vq);
 116
 117         *(volatile uint16_t *)&vq->used->idx += vq->shadow_used_idx;
 118         vq->shadow_used_idx = 0;
 119         vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
 120                 sizeof(vq->used->idx));
 121 }
 122
 123 static __rte_always_inline void
 124 update_shadow_used_ring_split(struct vhost_virtqueue *vq,
 125                          uint16_t desc_idx, uint16_t len)
 126 {
 127         uint16_t i = vq->shadow_used_idx++;
 128
 129         vq->shadow_used_split[i].id  = desc_idx;
 130         vq->shadow_used_split[i].len = len;
 131 }
 132
 133 static __rte_unused __rte_always_inline void
 134 flush_shadow_used_ring_packed(struct virtio_net *dev,
 135                         struct vhost_virtqueue *vq)
 136 {
 137         int i;
 138         uint16_t used_idx = vq->last_used_idx;
 139
 140         /* Split loop in two to save memory barriers */
 141         for (i = 0; i < vq->shadow_used_idx; i++) {
 142                 vq->desc_packed[used_idx].id = vq->shadow_used_packed[i].id;
 143                 vq->desc_packed[used_idx].len = vq->shadow_used_packed[i].len;
 144
 145                 used_idx += vq->shadow_used_packed[i].count;
 146                 if (used_idx >= vq->size)
 147                         used_idx -= vq->size;
 148         }
 149
 150         rte_smp_wmb();
 151
 152         for (i = 0; i < vq->shadow_used_idx; i++) {
 153                 uint16_t flags;
 154
 155                 if (vq->shadow_used_packed[i].len)
 156                         flags = VRING_DESC_F_WRITE;
 157                 else
 158                         flags = 0;
 159
 160                 if (vq->used_wrap_counter) {
 161                         flags |= VRING_DESC_F_USED;
 162                         flags |= VRING_DESC_F_AVAIL;
 163                 } else {
 164                         flags &= ~VRING_DESC_F_USED;
 165                         flags &= ~VRING_DESC_F_AVAIL;
 166                 }
 167
 168                 vq->desc_packed[vq->last_used_idx].flags = flags;
 169
 170                 vhost_log_cache_used_vring(dev, vq,
 171                                         vq->last_used_idx *
 172                                         sizeof(struct vring_packed_desc),
 173                                         sizeof(struct vring_packed_desc));
 174
 175                 vq->last_used_idx += vq->shadow_used_packed[i].count;
 176                 if (vq->last_used_idx >= vq->size) {
 177                         vq->used_wrap_counter ^= 1;
 178                         vq->last_used_idx -= vq->size;
 179                 }
 180         }
 181
 182         rte_smp_wmb();
 183         vq->shadow_used_idx = 0;
 184         vhost_log_cache_sync(dev, vq);
 185 }
 186
 187 static __rte_unused __rte_always_inline void
 188 update_shadow_used_ring_packed(struct vhost_virtqueue *vq,
 189                          uint16_t desc_idx, uint16_t len, uint16_t count)
 190 {
 191         uint16_t i = vq->shadow_used_idx++;
 192
 193         vq->shadow_used_packed[i].id  = desc_idx;
 194         vq->shadow_used_packed[i].len = len;
 195         vq->shadow_used_packed[i].count = count;
 196 }
 197
 198 static inline void
 199 do_data_copy_enqueue(struct virtio_net *dev, struct vhost_virtqueue *vq)
 200 {
 201         struct batch_copy_elem *elem = vq->batch_copy_elems;
 202         uint16_t count = vq->batch_copy_nb_elems;
 203         int i;
 204
 205         for (i = 0; i < count; i++) {
 206                 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
 207                 vhost_log_cache_write(dev, vq, elem[i].log_addr, elem[i].len);
 208                 PRINT_PACKET(dev, (uintptr_t)elem[i].dst, elem[i].len, 0);
 209         }
 210
 211         vq->batch_copy_nb_elems = 0;
 212 }
 213
 214 static inline void
 215 do_data_copy_dequeue(struct vhost_virtqueue *vq)
 216 {
 217         struct batch_copy_elem *elem = vq->batch_copy_elems;
 218         uint16_t count = vq->batch_copy_nb_elems;
 219         int i;
 220
 221         for (i = 0; i < count; i++)
 222                 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
 223
 224         vq->batch_copy_nb_elems = 0;
 225 }
 226
 227 /* avoid write operation when necessary, to lessen cache issues */
 228 #define ASSIGN_UNLESS_EQUAL(var, val) do {      \
 229         if ((var) != (val))                     \
 230                 (var) = (val);                  \
 231 } while (0)
 232
 233 static __rte_always_inline void
 234 virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
 235 {
 236         uint64_t csum_l4 = m_buf->ol_flags & PKT_TX_L4_MASK;
 237
 238         if (m_buf->ol_flags & PKT_TX_TCP_SEG)
 239                 csum_l4 |= PKT_TX_TCP_CKSUM;
 240
 241         if (csum_l4) {
 242                 net_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 243                 net_hdr->csum_start = m_buf->l2_len + m_buf->l3_len;
 244
 245                 switch (csum_l4) {
 246                 case PKT_TX_TCP_CKSUM:
 247                         net_hdr->csum_offset = (offsetof(struct tcp_hdr,
 248                                                 cksum));
 249                         break;
 250                 case PKT_TX_UDP_CKSUM:
 251                         net_hdr->csum_offset = (offsetof(struct udp_hdr,
 252                                                 dgram_cksum));
 253                         break;
 254                 case PKT_TX_SCTP_CKSUM:
 255                         net_hdr->csum_offset = (offsetof(struct sctp_hdr,
 256                                                 cksum));
 257                         break;
 258                 }
 259         } else {
 260                 ASSIGN_UNLESS_EQUAL(net_hdr->csum_start, 0);
 261                 ASSIGN_UNLESS_EQUAL(net_hdr->csum_offset, 0);
 262                 ASSIGN_UNLESS_EQUAL(net_hdr->flags, 0);
 263         }
 264
 265         /* IP cksum verification cannot be bypassed, then calculate here */
 266         if (m_buf->ol_flags & PKT_TX_IP_CKSUM) {
 267                 struct ipv4_hdr *ipv4_hdr;
 268
 269                 ipv4_hdr = rte_pktmbuf_mtod_offset(m_buf, struct ipv4_hdr *,
 270                                                    m_buf->l2_len);
 271                 ipv4_hdr->hdr_checksum = rte_ipv4_cksum(ipv4_hdr);
 272         }
 273
 274         if (m_buf->ol_flags & PKT_TX_TCP_SEG) {
 275                 if (m_buf->ol_flags & PKT_TX_IPV4)
 276                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 277                 else
 278                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 279                 net_hdr->gso_size = m_buf->tso_segsz;
 280                 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len
 281                                         + m_buf->l4_len;
 282         } else if (m_buf->ol_flags & PKT_TX_UDP_SEG) {
 283                 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
 284                 net_hdr->gso_size = m_buf->tso_segsz;
 285                 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len +
 286                         m_buf->l4_len;
 287         } else {
 288                 ASSIGN_UNLESS_EQUAL(net_hdr->gso_type, 0);
 289                 ASSIGN_UNLESS_EQUAL(net_hdr->gso_size, 0);
 290                 ASSIGN_UNLESS_EQUAL(net_hdr->hdr_len, 0);
 291         }
 292 }
 293
 294 static __rte_always_inline int
 295 map_one_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 296                 struct buf_vector *buf_vec, uint16_t *vec_idx,
 297                 uint64_t desc_iova, uint64_t desc_len, uint8_t perm)
 298 {
 299         uint16_t vec_id = *vec_idx;
 300
 301         while (desc_len) {
 302                 uint64_t desc_addr;
 303                 uint64_t desc_chunck_len = desc_len;
 304
 305                 if (unlikely(vec_id >= BUF_VECTOR_MAX))
 306                         return -1;
 307
 308                 desc_addr = vhost_iova_to_vva(dev, vq,
 309                                 desc_iova,
 310                                 &desc_chunck_len,
 311                                 perm);
 312                 if (unlikely(!desc_addr))
 313                         return -1;
 314
 315                 buf_vec[vec_id].buf_iova = desc_iova;
 316                 buf_vec[vec_id].buf_addr = desc_addr;
 317                 buf_vec[vec_id].buf_len  = desc_chunck_len;
 318
 319                 desc_len -= desc_chunck_len;
 320                 desc_iova += desc_chunck_len;
 321                 vec_id++;
 322         }
 323         *vec_idx = vec_id;
 324
 325         return 0;
 326 }
 327
 328 static __rte_always_inline int
 329 fill_vec_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 330                          uint32_t avail_idx, uint16_t *vec_idx,
 331                          struct buf_vector *buf_vec, uint16_t *desc_chain_head,
 332                          uint16_t *desc_chain_len, uint8_t perm)
 333 {
 334         uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
 335         uint16_t vec_id = *vec_idx;
 336         uint32_t len    = 0;
 337         uint64_t dlen;
 338         struct vring_desc *descs = vq->desc;
 339         struct vring_desc *idesc = NULL;
 340
 341         *desc_chain_head = idx;
 342
 343         if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
 344                 dlen = vq->desc[idx].len;
 345                 descs = (struct vring_desc *)(uintptr_t)
 346                         vhost_iova_to_vva(dev, vq, vq->desc[idx].addr,
 347                                                 &dlen,
 348                                                 VHOST_ACCESS_RO);
 349                 if (unlikely(!descs))
 350                         return -1;
 351
 352                 if (unlikely(dlen < vq->desc[idx].len)) {
 353                         /*
 354                          * The indirect desc table is not contiguous
 355                          * in process VA space, we have to copy it.
 356                          */
 357                         idesc = alloc_copy_ind_table(dev, vq,
 358                                         vq->desc[idx].addr, vq->desc[idx].len);
 359                         if (unlikely(!idesc))
 360                                 return -1;
 361
 362                         descs = idesc;
 363                 }
 364
 365                 idx = 0;
 366         }
 367
 368         while (1) {
 369                 if (unlikely(idx >= vq->size)) {
 370                         free_ind_table(idesc);
 371                         return -1;
 372                 }
 373
 374                 len += descs[idx].len;
 375
 376                 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 377                                                 descs[idx].addr, descs[idx].len,
 378                                                 perm))) {
 379                         free_ind_table(idesc);
 380                         return -1;
 381                 }
 382
 383                 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
 384                         break;
 385
 386                 idx = descs[idx].next;
 387         }
 388
 389         *desc_chain_len = len;
 390         *vec_idx = vec_id;
 391
 392         if (unlikely(!!idesc))
 393                 free_ind_table(idesc);
 394
 395         return 0;
 396 }
 397
 398 /*
 399  * Returns -1 on fail, 0 on success
 400  */
 401 static inline int
 402 reserve_avail_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 403                                 uint32_t size, struct buf_vector *buf_vec,
 404                                 uint16_t *num_buffers, uint16_t avail_head,
 405                                 uint16_t *nr_vec)
 406 {
 407         uint16_t cur_idx;
 408         uint16_t vec_idx = 0;
 409         uint16_t max_tries, tries = 0;
 410
 411         uint16_t head_idx = 0;
 412         uint16_t len = 0;
 413
 414         *num_buffers = 0;
 415         cur_idx  = vq->last_avail_idx;
 416
 417         if (rxvq_is_mergeable(dev))
 418                 max_tries = vq->size;
 419         else
 420                 max_tries = 1;
 421
 422         while (size > 0) {
 423                 if (unlikely(cur_idx == avail_head))
 424                         return -1;
 425
 426                 if (unlikely(fill_vec_buf_split(dev, vq, cur_idx,
 427                                                 &vec_idx, buf_vec,
 428                                                 &head_idx, &len,
 429                                                 VHOST_ACCESS_RW) < 0))
 430                         return -1;
 431                 len = RTE_MIN(len, size);
 432                 update_shadow_used_ring_split(vq, head_idx, len);
 433                 size -= len;
 434
 435                 cur_idx++;
 436                 tries++;
 437                 *num_buffers += 1;
 438
 439                 /*
 440                  * if we tried all available ring items, and still
 441                  * can't get enough buf, it means something abnormal
 442                  * happened.
 443                  */
 444                 if (unlikely(tries > max_tries))
 445                         return -1;
 446         }
 447
 448         *nr_vec = vec_idx;
 449
 450         return 0;
 451 }
 452
 453 static __rte_always_inline int
 454 fill_vec_buf_packed_indirect(struct virtio_net *dev,
 455                         struct vhost_virtqueue *vq,
 456                         struct vring_packed_desc *desc, uint16_t *vec_idx,
 457                         struct buf_vector *buf_vec, uint16_t *len, uint8_t perm)
 458 {
 459         uint16_t i;
 460         uint32_t nr_descs;
 461         uint16_t vec_id = *vec_idx;
 462         uint64_t dlen;
 463         struct vring_packed_desc *descs, *idescs = NULL;
 464
 465         dlen = desc->len;
 466         descs = (struct vring_packed_desc *)(uintptr_t)
 467                 vhost_iova_to_vva(dev, vq, desc->addr, &dlen, VHOST_ACCESS_RO);
 468         if (unlikely(!descs))
 469                 return -1;
 470
 471         if (unlikely(dlen < desc->len)) {
 472                 /*
 473                  * The indirect desc table is not contiguous
 474                  * in process VA space, we have to copy it.
 475                  */
 476                 idescs = alloc_copy_ind_table(dev, vq, desc->addr, desc->len);
 477                 if (unlikely(!idescs))
 478                         return -1;
 479
 480                 descs = idescs;
 481         }
 482
 483         nr_descs =  desc->len / sizeof(struct vring_packed_desc);
 484         if (unlikely(nr_descs >= vq->size)) {
 485                 free_ind_table(idescs);
 486                 return -1;
 487         }
 488
 489         for (i = 0; i < nr_descs; i++) {
 490                 if (unlikely(vec_id >= BUF_VECTOR_MAX)) {
 491                         free_ind_table(idescs);
 492                         return -1;
 493                 }
 494
 495                 *len += descs[i].len;
 496                 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 497                                                 descs[i].addr, descs[i].len,
 498                                                 perm)))
 499                         return -1;
 500         }
 501         *vec_idx = vec_id;
 502
 503         if (unlikely(!!idescs))
 504                 free_ind_table(idescs);
 505
 506         return 0;
 507 }
 508
 509 static __rte_unused __rte_always_inline int
 510 fill_vec_buf_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
 511                                 uint16_t avail_idx, uint16_t *desc_count,
 512                                 struct buf_vector *buf_vec, uint16_t *vec_idx,
 513                                 uint16_t *buf_id, uint16_t *len, uint8_t perm)
 514 {
 515         bool wrap_counter = vq->avail_wrap_counter;
 516         struct vring_packed_desc *descs = vq->desc_packed;
 517         uint16_t vec_id = *vec_idx;
 518
 519         if (avail_idx < vq->last_avail_idx)
 520                 wrap_counter ^= 1;
 521
 522         if (unlikely(!desc_is_avail(&descs[avail_idx], wrap_counter)))
 523                 return -1;
 524
 525         *desc_count = 0;
 526
 527         while (1) {
 528                 if (unlikely(vec_id >= BUF_VECTOR_MAX))
 529                         return -1;
 530
 531                 *desc_count += 1;
 532                 *buf_id = descs[avail_idx].id;
 533
 534                 if (descs[avail_idx].flags & VRING_DESC_F_INDIRECT) {
 535                         if (unlikely(fill_vec_buf_packed_indirect(dev, vq,
 536                                                         &descs[avail_idx],
 537                                                         &vec_id, buf_vec,
 538                                                         len, perm) < 0))
 539                                 return -1;
 540                 } else {
 541                         *len += descs[avail_idx].len;
 542
 543                         if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 544                                                         descs[avail_idx].addr,
 545                                                         descs[avail_idx].len,
 546                                                         perm)))
 547                                 return -1;
 548                 }
 549
 550                 if ((descs[avail_idx].flags & VRING_DESC_F_NEXT) == 0)
 551                         break;
 552
 553                 if (++avail_idx >= vq->size) {
 554                         avail_idx -= vq->size;
 555                         wrap_counter ^= 1;
 556                 }
 557         }
 558
 559         *vec_idx = vec_id;
 560
 561         return 0;
 562 }
 563
 564 static __rte_always_inline int
 565 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 566                             struct rte_mbuf *m, struct buf_vector *buf_vec,
 567                             uint16_t nr_vec, uint16_t num_buffers)
 568 {
 569         uint32_t vec_idx = 0;
 570         uint32_t mbuf_offset, mbuf_avail;
 571         uint32_t buf_offset, buf_avail;
 572         uint64_t buf_addr, buf_iova, buf_len;
 573         uint32_t cpy_len;
 574         uint64_t hdr_addr;
 575         struct rte_mbuf *hdr_mbuf;
 576         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 577         struct virtio_net_hdr_mrg_rxbuf tmp_hdr, *hdr = NULL;
 578         int error = 0;
 579
 580         if (unlikely(m == NULL)) {
 581                 error = -1;
 582                 goto out;
 583         }
 584
 585         buf_addr = buf_vec[vec_idx].buf_addr;
 586         buf_iova = buf_vec[vec_idx].buf_iova;
 587         buf_len = buf_vec[vec_idx].buf_len;
 588
 589         if (nr_vec > 1)
 590                 rte_prefetch0((void *)(uintptr_t)buf_vec[1].buf_addr);
 591
 592         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
 593                 error = -1;
 594                 goto out;
 595         }
 596
 597         hdr_mbuf = m;
 598         hdr_addr = buf_addr;
 599         if (unlikely(buf_len < dev->vhost_hlen))
 600                 hdr = &tmp_hdr;
 601         else
 602                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)hdr_addr;
 603
 604         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
 605                 dev->vid, num_buffers);
 606
 607         if (unlikely(buf_len < dev->vhost_hlen)) {
 608                 buf_offset = dev->vhost_hlen - buf_len;
 609                 vec_idx++;
 610                 buf_addr = buf_vec[vec_idx].buf_addr;
 611                 buf_iova = buf_vec[vec_idx].buf_iova;
 612                 buf_len = buf_vec[vec_idx].buf_len;
 613                 buf_avail = buf_len - buf_offset;
 614         } else {
 615                 buf_offset = dev->vhost_hlen;
 616                 buf_avail = buf_len - dev->vhost_hlen;
 617         }
 618
 619         mbuf_avail  = rte_pktmbuf_data_len(m);
 620         mbuf_offset = 0;
 621         while (mbuf_avail != 0 || m->next != NULL) {
 622                 /* done with current buf, get the next one */
 623                 if (buf_avail == 0) {
 624                         vec_idx++;
 625                         if (unlikely(vec_idx >= nr_vec)) {
 626                                 error = -1;
 627                                 goto out;
 628                         }
 629
 630                         buf_addr = buf_vec[vec_idx].buf_addr;
 631                         buf_iova = buf_vec[vec_idx].buf_iova;
 632                         buf_len = buf_vec[vec_idx].buf_len;
 633
 634                         /* Prefetch next buffer address. */
 635                         if (vec_idx + 1 < nr_vec)
 636                                 rte_prefetch0((void *)(uintptr_t)
 637                                                 buf_vec[vec_idx + 1].buf_addr);
 638                         buf_offset = 0;
 639                         buf_avail  = buf_len;
 640                 }
 641
 642                 /* done with current mbuf, get the next one */
 643                 if (mbuf_avail == 0) {
 644                         m = m->next;
 645
 646                         mbuf_offset = 0;
 647                         mbuf_avail  = rte_pktmbuf_data_len(m);
 648                 }
 649
 650                 if (hdr_addr) {
 651                         virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
 652                         if (rxvq_is_mergeable(dev))
 653                                 ASSIGN_UNLESS_EQUAL(hdr->num_buffers,
 654                                                 num_buffers);
 655
 656                         if (unlikely(hdr == &tmp_hdr)) {
 657                                 uint64_t len;
 658                                 uint64_t remain = dev->vhost_hlen;
 659                                 uint64_t src = (uint64_t)(uintptr_t)hdr, dst;
 660                                 uint64_t iova = buf_vec[0].buf_iova;
 661                                 uint16_t hdr_vec_idx = 0;
 662
 663                                 while (remain) {
 664                                         len = remain;
 665                                         dst = buf_vec[hdr_vec_idx].buf_addr;
 666                                         rte_memcpy((void *)(uintptr_t)dst,
 667                                                         (void *)(uintptr_t)src,
 668                                                         len);
 669
 670                                         PRINT_PACKET(dev, (uintptr_t)dst,
 671                                                         (uint32_t)len, 0);
 672                                         vhost_log_cache_write(dev, vq,
 673                                                         iova, len);
 674
 675                                         remain -= len;
 676                                         iova += len;
 677                                         src += len;
 678                                         hdr_vec_idx++;
 679                                 }
 680                         } else {
 681                                 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
 682                                                 dev->vhost_hlen, 0);
 683                                 vhost_log_cache_write(dev, vq,
 684                                                 buf_vec[0].buf_iova,
 685                                                 dev->vhost_hlen);
 686                         }
 687
 688                         hdr_addr = 0;
 689                 }
 690
 691                 cpy_len = RTE_MIN(buf_len, mbuf_avail);
 692
 693                 if (likely(cpy_len > MAX_BATCH_LEN ||
 694                                         vq->batch_copy_nb_elems >= vq->size)) {
 695                         rte_memcpy((void *)((uintptr_t)(buf_addr + buf_offset)),
 696                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 697                                 cpy_len);
 698                         vhost_log_cache_write(dev, vq, buf_iova + buf_offset,
 699                                         cpy_len);
 700                         PRINT_PACKET(dev, (uintptr_t)(buf_addr + buf_offset),
 701                                 cpy_len, 0);
 702                 } else {
 703                         batch_copy[vq->batch_copy_nb_elems].dst =
 704                                 (void *)((uintptr_t)(buf_addr + buf_offset));
 705                         batch_copy[vq->batch_copy_nb_elems].src =
 706                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
 707                         batch_copy[vq->batch_copy_nb_elems].log_addr =
 708                                 buf_iova + buf_offset;
 709                         batch_copy[vq->batch_copy_nb_elems].len = cpy_len;
 710                         vq->batch_copy_nb_elems++;
 711                 }
 712
 713                 mbuf_avail  -= cpy_len;
 714                 mbuf_offset += cpy_len;
 715                 buf_avail  -= cpy_len;
 716                 buf_offset += cpy_len;
 717         }
 718
 719 out:
 720
 721         return error;
 722 }
 723
 724 static __rte_always_inline uint32_t
 725 virtio_dev_rx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 726         struct rte_mbuf **pkts, uint32_t count)
 727 {
 728         uint32_t pkt_idx = 0;
 729         uint16_t num_buffers;
 730         struct buf_vector buf_vec[BUF_VECTOR_MAX];
 731         uint16_t avail_head;
 732
 733         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
 734         avail_head = *((volatile uint16_t *)&vq->avail->idx);
 735
 736         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
 737                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
 738                 uint16_t nr_vec = 0;
 739
 740                 if (unlikely(reserve_avail_buf_split(dev, vq,
 741                                                 pkt_len, buf_vec, &num_buffers,
 742                                                 avail_head, &nr_vec) < 0)) {
 743                         VHOST_LOG_DEBUG(VHOST_DATA,
 744                                 "(%d) failed to get enough desc from vring\n",
 745                                 dev->vid);
 746                         vq->shadow_used_idx -= num_buffers;
 747                         break;
 748                 }
 749
 750                 rte_prefetch0((void *)(uintptr_t)buf_vec[0].buf_addr);
 751
 752                 VHOST_LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
 753                         dev->vid, vq->last_avail_idx,
 754                         vq->last_avail_idx + num_buffers);
 755
 756                 if (copy_mbuf_to_desc(dev, vq, pkts[pkt_idx],
 757                                                 buf_vec, nr_vec,
 758                                                 num_buffers) < 0) {
 759                         vq->shadow_used_idx -= num_buffers;
 760                         break;
 761                 }
 762
 763                 vq->last_avail_idx += num_buffers;
 764         }
 765
 766         do_data_copy_enqueue(dev, vq);
 767
 768         if (likely(vq->shadow_used_idx)) {
 769                 flush_shadow_used_ring_split(dev, vq);
 770                 vhost_vring_call(dev, vq);
 771         }
 772
 773         return pkt_idx;
 774 }
 775
 776 static __rte_always_inline uint32_t
 777 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
 778         struct rte_mbuf **pkts, uint32_t count)
 779 {
 780         struct vhost_virtqueue *vq;
 781
 782         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
 783         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
 784                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
 785                         dev->vid, __func__, queue_id);
 786                 return 0;
 787         }
 788
 789         vq = dev->virtqueue[queue_id];
 790
 791         rte_spinlock_lock(&vq->access_lock);
 792
 793         if (unlikely(vq->enabled == 0))
 794                 goto out_access_unlock;
 795
 796         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 797                 vhost_user_iotlb_rd_lock(vq);
 798
 799         if (unlikely(vq->access_ok == 0))
 800                 if (unlikely(vring_translate(dev, vq) < 0))
 801                         goto out;
 802
 803         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
 804         if (count == 0)
 805                 goto out;
 806
 807         count = virtio_dev_rx_split(dev, vq, pkts, count);
 808
 809 out:
 810         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 811                 vhost_user_iotlb_rd_unlock(vq);
 812
 813 out_access_unlock:
 814         rte_spinlock_unlock(&vq->access_lock);
 815
 816         return count;
 817 }
 818
 819 uint16_t
 820 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
 821         struct rte_mbuf **pkts, uint16_t count)
 822 {
 823         struct virtio_net *dev = get_device(vid);
 824
 825         if (!dev)
 826                 return 0;
 827
 828         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
 829                 RTE_LOG(ERR, VHOST_DATA,
 830                         "(%d) %s: built-in vhost net backend is disabled.\n",
 831                         dev->vid, __func__);
 832                 return 0;
 833         }
 834
 835         return virtio_dev_rx(dev, queue_id, pkts, count);
 836 }
 837
 838 static inline bool
 839 virtio_net_with_host_offload(struct virtio_net *dev)
 840 {
 841         if (dev->features &
 842                         ((1ULL << VIRTIO_NET_F_CSUM) |
 843                          (1ULL << VIRTIO_NET_F_HOST_ECN) |
 844                          (1ULL << VIRTIO_NET_F_HOST_TSO4) |
 845                          (1ULL << VIRTIO_NET_F_HOST_TSO6) |
 846                          (1ULL << VIRTIO_NET_F_HOST_UFO)))
 847                 return true;
 848
 849         return false;
 850 }
 851
 852 static void
 853 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
 854 {
 855         struct ipv4_hdr *ipv4_hdr;
 856         struct ipv6_hdr *ipv6_hdr;
 857         void *l3_hdr = NULL;
 858         struct ether_hdr *eth_hdr;
 859         uint16_t ethertype;
 860
 861         eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
 862
 863         m->l2_len = sizeof(struct ether_hdr);
 864         ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
 865
 866         if (ethertype == ETHER_TYPE_VLAN) {
 867                 struct vlan_hdr *vlan_hdr = (struct vlan_hdr *)(eth_hdr + 1);
 868
 869                 m->l2_len += sizeof(struct vlan_hdr);
 870                 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
 871         }
 872
 873         l3_hdr = (char *)eth_hdr + m->l2_len;
 874
 875         switch (ethertype) {
 876         case ETHER_TYPE_IPv4:
 877                 ipv4_hdr = l3_hdr;
 878                 *l4_proto = ipv4_hdr->next_proto_id;
 879                 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
 880                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 881                 m->ol_flags |= PKT_TX_IPV4;
 882                 break;
 883         case ETHER_TYPE_IPv6:
 884                 ipv6_hdr = l3_hdr;
 885                 *l4_proto = ipv6_hdr->proto;
 886                 m->l3_len = sizeof(struct ipv6_hdr);
 887                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 888                 m->ol_flags |= PKT_TX_IPV6;
 889                 break;
 890         default:
 891                 m->l3_len = 0;
 892                 *l4_proto = 0;
 893                 *l4_hdr = NULL;
 894                 break;
 895         }
 896 }
 897
 898 static __rte_always_inline void
 899 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
 900 {
 901         uint16_t l4_proto = 0;
 902         void *l4_hdr = NULL;
 903         struct tcp_hdr *tcp_hdr = NULL;
 904
 905         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
 906                 return;
 907
 908         parse_ethernet(m, &l4_proto, &l4_hdr);
 909         if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 910                 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
 911                         switch (hdr->csum_offset) {
 912                         case (offsetof(struct tcp_hdr, cksum)):
 913                                 if (l4_proto == IPPROTO_TCP)
 914                                         m->ol_flags |= PKT_TX_TCP_CKSUM;
 915                                 break;
 916                         case (offsetof(struct udp_hdr, dgram_cksum)):
 917                                 if (l4_proto == IPPROTO_UDP)
 918                                         m->ol_flags |= PKT_TX_UDP_CKSUM;
 919                                 break;
 920                         case (offsetof(struct sctp_hdr, cksum)):
 921                                 if (l4_proto == IPPROTO_SCTP)
 922                                         m->ol_flags |= PKT_TX_SCTP_CKSUM;
 923                                 break;
 924                         default:
 925                                 break;
 926                         }
 927                 }
 928         }
 929
 930         if (l4_hdr && hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 931                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 932                 case VIRTIO_NET_HDR_GSO_TCPV4:
 933                 case VIRTIO_NET_HDR_GSO_TCPV6:
 934                         tcp_hdr = l4_hdr;
 935                         m->ol_flags |= PKT_TX_TCP_SEG;
 936                         m->tso_segsz = hdr->gso_size;
 937                         m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
 938                         break;
 939                 case VIRTIO_NET_HDR_GSO_UDP:
 940                         m->ol_flags |= PKT_TX_UDP_SEG;
 941                         m->tso_segsz = hdr->gso_size;
 942                         m->l4_len = sizeof(struct udp_hdr);
 943                         break;
 944                 default:
 945                         RTE_LOG(WARNING, VHOST_DATA,
 946                                 "unsupported gso type %u.\n", hdr->gso_type);
 947                         break;
 948                 }
 949         }
 950 }
 951
 952 static __rte_always_inline void
 953 put_zmbuf(struct zcopy_mbuf *zmbuf)
 954 {
 955         zmbuf->in_use = 0;
 956 }
 957
 958 static __rte_always_inline int
 959 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 960                   struct buf_vector *buf_vec, uint16_t nr_vec,
 961                   struct rte_mbuf *m, struct rte_mempool *mbuf_pool)
 962 {
 963         uint32_t buf_avail, buf_offset;
 964         uint64_t buf_addr, buf_iova, buf_len;
 965         uint32_t mbuf_avail, mbuf_offset;
 966         uint32_t cpy_len;
 967         struct rte_mbuf *cur = m, *prev = m;
 968         struct virtio_net_hdr tmp_hdr;
 969         struct virtio_net_hdr *hdr = NULL;
 970         /* A counter to avoid desc dead loop chain */
 971         uint16_t vec_idx = 0;
 972         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 973         int error = 0;
 974
 975         buf_addr = buf_vec[vec_idx].buf_addr;
 976         buf_iova = buf_vec[vec_idx].buf_iova;
 977         buf_len = buf_vec[vec_idx].buf_len;
 978
 979         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
 980                 error = -1;
 981                 goto out;
 982         }
 983
 984         if (likely(nr_vec > 1))
 985                 rte_prefetch0((void *)(uintptr_t)buf_vec[1].buf_addr);
 986
 987         if (virtio_net_with_host_offload(dev)) {
 988                 if (unlikely(buf_len < sizeof(struct virtio_net_hdr))) {
 989                         uint64_t len;
 990                         uint64_t remain = sizeof(struct virtio_net_hdr);
 991                         uint64_t src;
 992                         uint64_t dst = (uint64_t)(uintptr_t)&tmp_hdr;
 993                         uint16_t hdr_vec_idx = 0;
 994
 995                         /*
 996                          * No luck, the virtio-net header doesn't fit
 997                          * in a contiguous virtual area.
 998                          */
 999                         while (remain) {
1000                                 len = remain;
1001                                 src = buf_vec[hdr_vec_idx].buf_addr;
1002                                 rte_memcpy((void *)(uintptr_t)dst,
1003                                                    (void *)(uintptr_t)src, len);
1004
1005                                 remain -= len;
1006                                 dst += len;
1007                                 hdr_vec_idx++;
1008                         }
1009
1010                         hdr = &tmp_hdr;
1011                 } else {
1012                         hdr = (struct virtio_net_hdr *)((uintptr_t)buf_addr);
1013                         rte_prefetch0(hdr);
1014                 }
1015         }
1016
1017         /*
1018          * A virtio driver normally uses at least 2 desc buffers
1019          * for Tx: the first for storing the header, and others
1020          * for storing the data.
1021          */
1022         if (unlikely(buf_len < dev->vhost_hlen)) {
1023                 buf_offset = dev->vhost_hlen - buf_len;
1024                 vec_idx++;
1025                 buf_addr = buf_vec[vec_idx].buf_addr;
1026                 buf_iova = buf_vec[vec_idx].buf_iova;
1027                 buf_len = buf_vec[vec_idx].buf_len;
1028                 buf_avail  = buf_len - buf_offset;
1029         } else if (buf_len == dev->vhost_hlen) {
1030                 if (unlikely(++vec_idx >= nr_vec))
1031                         goto out;
1032                 buf_addr = buf_vec[vec_idx].buf_addr;
1033                 buf_iova = buf_vec[vec_idx].buf_iova;
1034                 buf_len = buf_vec[vec_idx].buf_len;
1035
1036                 buf_offset = 0;
1037                 buf_avail = buf_len;
1038         } else {
1039                 buf_offset = dev->vhost_hlen;
1040                 buf_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
1041         }
1042
1043         rte_prefetch0((void *)(uintptr_t)
1044                         (buf_addr + buf_offset));
1045
1046         PRINT_PACKET(dev,
1047                         (uintptr_t)(buf_addr + buf_offset),
1048                         (uint32_t)buf_avail, 0);
1049
1050         mbuf_offset = 0;
1051         mbuf_avail  = m->buf_len - RTE_PKTMBUF_HEADROOM;
1052         while (1) {
1053                 uint64_t hpa;
1054
1055                 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
1056
1057                 /*
1058                  * A desc buf might across two host physical pages that are
1059                  * not continuous. In such case (gpa_to_hpa returns 0), data
1060                  * will be copied even though zero copy is enabled.
1061                  */
1062                 if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
1063                                         buf_iova + buf_offset, cpy_len)))) {
1064                         cur->data_len = cpy_len;
1065                         cur->data_off = 0;
1066                         cur->buf_addr =
1067                                 (void *)(uintptr_t)(buf_addr + buf_offset);
1068                         cur->buf_iova = hpa;
1069
1070                         /*
1071                          * In zero copy mode, one mbuf can only reference data
1072                          * for one or partial of one desc buff.
1073                          */
1074                         mbuf_avail = cpy_len;
1075                 } else {
1076                         if (likely(cpy_len > MAX_BATCH_LEN ||
1077                                    vq->batch_copy_nb_elems >= vq->size ||
1078                                    (hdr && cur == m))) {
1079                                 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
1080                                                                    mbuf_offset),
1081                                            (void *)((uintptr_t)(buf_addr +
1082                                                            buf_offset)),
1083                                            cpy_len);
1084                         } else {
1085                                 batch_copy[vq->batch_copy_nb_elems].dst =
1086                                         rte_pktmbuf_mtod_offset(cur, void *,
1087                                                                 mbuf_offset);
1088                                 batch_copy[vq->batch_copy_nb_elems].src =
1089                                         (void *)((uintptr_t)(buf_addr +
1090                                                                 buf_offset));
1091                                 batch_copy[vq->batch_copy_nb_elems].len =
1092                                         cpy_len;
1093                                 vq->batch_copy_nb_elems++;
1094                         }
1095                 }
1096
1097                 mbuf_avail  -= cpy_len;
1098                 mbuf_offset += cpy_len;
1099                 buf_avail -= cpy_len;
1100                 buf_offset += cpy_len;
1101
1102                 /* This buf reaches to its end, get the next one */
1103                 if (buf_avail == 0) {
1104                         if (++vec_idx >= nr_vec)
1105                                 break;
1106
1107                         buf_addr = buf_vec[vec_idx].buf_addr;
1108                         buf_iova = buf_vec[vec_idx].buf_iova;
1109                         buf_len = buf_vec[vec_idx].buf_len;
1110
1111                         /*
1112                          * Prefecth desc n + 1 buffer while
1113                          * desc n buffer is processed.
1114                          */
1115                         if (vec_idx + 1 < nr_vec)
1116                                 rte_prefetch0((void *)(uintptr_t)
1117                                                 buf_vec[vec_idx + 1].buf_addr);
1118
1119                         buf_offset = 0;
1120                         buf_avail  = buf_len;
1121
1122                         PRINT_PACKET(dev, (uintptr_t)buf_addr,
1123                                         (uint32_t)buf_avail, 0);
1124                 }
1125
1126                 /*
1127                  * This mbuf reaches to its end, get a new one
1128                  * to hold more data.
1129                  */
1130                 if (mbuf_avail == 0) {
1131                         cur = rte_pktmbuf_alloc(mbuf_pool);
1132                         if (unlikely(cur == NULL)) {
1133                                 RTE_LOG(ERR, VHOST_DATA, "Failed to "
1134                                         "allocate memory for mbuf.\n");
1135                                 error = -1;
1136                                 goto out;
1137                         }
1138                         if (unlikely(dev->dequeue_zero_copy))
1139                                 rte_mbuf_refcnt_update(cur, 1);
1140
1141                         prev->next = cur;
1142                         prev->data_len = mbuf_offset;
1143                         m->nb_segs += 1;
1144                         m->pkt_len += mbuf_offset;
1145                         prev = cur;
1146
1147                         mbuf_offset = 0;
1148                         mbuf_avail  = cur->buf_len - RTE_PKTMBUF_HEADROOM;
1149                 }
1150         }
1151
1152         prev->data_len = mbuf_offset;
1153         m->pkt_len    += mbuf_offset;
1154
1155         if (hdr)
1156                 vhost_dequeue_offload(hdr, m);
1157
1158 out:
1159
1160         return error;
1161 }
1162
1163 static __rte_always_inline struct zcopy_mbuf *
1164 get_zmbuf(struct vhost_virtqueue *vq)
1165 {
1166         uint16_t i;
1167         uint16_t last;
1168         int tries = 0;
1169
1170         /* search [last_zmbuf_idx, zmbuf_size) */
1171         i = vq->last_zmbuf_idx;
1172         last = vq->zmbuf_size;
1173
1174 again:
1175         for (; i < last; i++) {
1176                 if (vq->zmbufs[i].in_use == 0) {
1177                         vq->last_zmbuf_idx = i + 1;
1178                         vq->zmbufs[i].in_use = 1;
1179                         return &vq->zmbufs[i];
1180                 }
1181         }
1182
1183         tries++;
1184         if (tries == 1) {
1185                 /* search [0, last_zmbuf_idx) */
1186                 i = 0;
1187                 last = vq->last_zmbuf_idx;
1188                 goto again;
1189         }
1190
1191         return NULL;
1192 }
1193
1194 static __rte_always_inline bool
1195 mbuf_is_consumed(struct rte_mbuf *m)
1196 {
1197         while (m) {
1198                 if (rte_mbuf_refcnt_read(m) > 1)
1199                         return false;
1200                 m = m->next;
1201         }
1202
1203         return true;
1204 }
1205
1206 static __rte_always_inline void
1207 restore_mbuf(struct rte_mbuf *m)
1208 {
1209         uint32_t mbuf_size, priv_size;
1210
1211         while (m) {
1212                 priv_size = rte_pktmbuf_priv_size(m->pool);
1213                 mbuf_size = sizeof(struct rte_mbuf) + priv_size;
1214                 /* start of buffer is after mbuf structure and priv data */
1215
1216                 m->buf_addr = (char *)m + mbuf_size;
1217                 m->buf_iova = rte_mempool_virt2iova(m) + mbuf_size;
1218                 m = m->next;
1219         }
1220 }
1221
1222 static __rte_always_inline uint16_t
1223 virtio_dev_tx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
1224         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1225 {
1226         uint16_t i;
1227         uint16_t free_entries;
1228
1229         if (unlikely(dev->dequeue_zero_copy)) {
1230                 struct zcopy_mbuf *zmbuf, *next;
1231                 int nr_updated = 0;
1232
1233                 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1234                      zmbuf != NULL; zmbuf = next) {
1235                         next = TAILQ_NEXT(zmbuf, next);
1236
1237                         if (mbuf_is_consumed(zmbuf->mbuf)) {
1238                                 update_shadow_used_ring_split(vq,
1239                                                 zmbuf->desc_idx, 0);
1240                                 nr_updated += 1;
1241
1242                                 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1243                                 restore_mbuf(zmbuf->mbuf);
1244                                 rte_pktmbuf_free(zmbuf->mbuf);
1245                                 put_zmbuf(zmbuf);
1246                                 vq->nr_zmbuf -= 1;
1247                         }
1248                 }
1249
1250                 flush_shadow_used_ring_split(dev, vq);
1251                 vhost_vring_call(dev, vq);
1252         }
1253
1254         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
1255
1256         free_entries = *((volatile uint16_t *)&vq->avail->idx) -
1257                         vq->last_avail_idx;
1258         if (free_entries == 0)
1259                 return 0;
1260
1261         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1262
1263         count = RTE_MIN(count, MAX_PKT_BURST);
1264         count = RTE_MIN(count, free_entries);
1265         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1266                         dev->vid, count);
1267
1268         for (i = 0; i < count; i++) {
1269                 struct buf_vector buf_vec[BUF_VECTOR_MAX];
1270                 uint16_t head_idx, dummy_len;
1271                 uint16_t nr_vec = 0;
1272                 int err;
1273
1274                 if (unlikely(fill_vec_buf_split(dev, vq,
1275                                                 vq->last_avail_idx + i,
1276                                                 &nr_vec, buf_vec,
1277                                                 &head_idx, &dummy_len,
1278                                                 VHOST_ACCESS_RO) < 0))
1279                         break;
1280
1281                 if (likely(dev->dequeue_zero_copy == 0))
1282                         update_shadow_used_ring_split(vq, head_idx, 0);
1283
1284                 rte_prefetch0((void *)(uintptr_t)buf_vec[0].buf_addr);
1285
1286                 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
1287                 if (unlikely(pkts[i] == NULL)) {
1288                         RTE_LOG(ERR, VHOST_DATA,
1289                                 "Failed to allocate memory for mbuf.\n");
1290                         break;
1291                 }
1292
1293                 err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts[i],
1294                                 mbuf_pool);
1295                 if (unlikely(err)) {
1296                         rte_pktmbuf_free(pkts[i]);
1297                         break;
1298                 }
1299
1300                 if (unlikely(dev->dequeue_zero_copy)) {
1301                         struct zcopy_mbuf *zmbuf;
1302
1303                         zmbuf = get_zmbuf(vq);
1304                         if (!zmbuf) {
1305                                 rte_pktmbuf_free(pkts[i]);
1306                                 break;
1307                         }
1308                         zmbuf->mbuf = pkts[i];
1309                         zmbuf->desc_idx = head_idx;
1310
1311                         /*
1312                          * Pin lock the mbuf; we will check later to see
1313                          * whether the mbuf is freed (when we are the last
1314                          * user) or not. If that's the case, we then could
1315                          * update the used ring safely.
1316                          */
1317                         rte_mbuf_refcnt_update(pkts[i], 1);
1318
1319                         vq->nr_zmbuf += 1;
1320                         TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1321                 }
1322         }
1323         vq->last_avail_idx += i;
1324
1325         if (likely(dev->dequeue_zero_copy == 0)) {
1326                 do_data_copy_dequeue(vq);
1327                 if (unlikely(i < count))
1328                         vq->shadow_used_idx = i;
1329                 flush_shadow_used_ring_split(dev, vq);
1330                 vhost_vring_call(dev, vq);
1331         }
1332
1333         return i;
1334 }
1335
1336 uint16_t
1337 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
1338         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1339 {
1340         struct virtio_net *dev;
1341         struct rte_mbuf *rarp_mbuf = NULL;
1342         struct vhost_virtqueue *vq;
1343
1344         dev = get_device(vid);
1345         if (!dev)
1346                 return 0;
1347
1348         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
1349                 RTE_LOG(ERR, VHOST_DATA,
1350                         "(%d) %s: built-in vhost net backend is disabled.\n",
1351                         dev->vid, __func__);
1352                 return 0;
1353         }
1354
1355         if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
1356                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
1357                         dev->vid, __func__, queue_id);
1358                 return 0;
1359         }
1360
1361         vq = dev->virtqueue[queue_id];
1362
1363         if (unlikely(rte_spinlock_trylock(&vq->access_lock) == 0))
1364                 return 0;
1365
1366         if (unlikely(vq->enabled == 0))
1367                 goto out_access_unlock;
1368
1369         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1370                 vhost_user_iotlb_rd_lock(vq);
1371
1372         if (unlikely(vq->access_ok == 0))
1373                 if (unlikely(vring_translate(dev, vq) < 0))
1374                         goto out;
1375
1376         /*
1377          * Construct a RARP broadcast packet, and inject it to the "pkts"
1378          * array, to looks like that guest actually send such packet.
1379          *
1380          * Check user_send_rarp() for more information.
1381          *
1382          * broadcast_rarp shares a cacheline in the virtio_net structure
1383          * with some fields that are accessed during enqueue and
1384          * rte_atomic16_cmpset() causes a write if using cmpxchg. This could
1385          * result in false sharing between enqueue and dequeue.
1386          *
1387          * Prevent unnecessary false sharing by reading broadcast_rarp first
1388          * and only performing cmpset if the read indicates it is likely to
1389          * be set.
1390          */
1391         if (unlikely(rte_atomic16_read(&dev->broadcast_rarp) &&
1392                         rte_atomic16_cmpset((volatile uint16_t *)
1393                                 &dev->broadcast_rarp.cnt, 1, 0))) {
1394
1395                 rarp_mbuf = rte_net_make_rarp_packet(mbuf_pool, &dev->mac);
1396                 if (rarp_mbuf == NULL) {
1397                         RTE_LOG(ERR, VHOST_DATA,
1398                                 "Failed to make RARP packet.\n");
1399                         return 0;
1400                 }
1401                 count -= 1;
1402         }
1403
1404         count = virtio_dev_tx_split(dev, vq, mbuf_pool, pkts, count);
1405
1406 out:
1407         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1408                 vhost_user_iotlb_rd_unlock(vq);
1409
1410 out_access_unlock:
1411         rte_spinlock_unlock(&vq->access_lock);
1412
1413         if (unlikely(rarp_mbuf != NULL)) {
1414                 /*
1415                  * Inject it to the head of "pkts" array, so that switch's mac
1416                  * learning table will get updated first.
1417                  */
1418                 memmove(&pkts[1], pkts, count * sizeof(struct rte_mbuf *));
1419                 pkts[0] = rarp_mbuf;
1420                 count += 1;
1421         }
1422
1423         return count;
1424 }