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 do_flush_shadow_used_ring_split(struct virtio_net *dev,
  42                         struct vhost_virtqueue *vq,
  43                         uint16_t to, uint16_t from, uint16_t size)
  44 {
  45         rte_memcpy(&vq->used->ring[to],
  46                         &vq->shadow_used_split[from],
  47                         size * sizeof(struct vring_used_elem));
  48         vhost_log_cache_used_vring(dev, vq,
  49                         offsetof(struct vring_used, ring[to]),
  50                         size * sizeof(struct vring_used_elem));
  51 }
  52
  53 static __rte_always_inline void
  54 flush_shadow_used_ring_split(struct virtio_net *dev, struct vhost_virtqueue *vq)
  55 {
  56         uint16_t used_idx = vq->last_used_idx & (vq->size - 1);
  57
  58         if (used_idx + vq->shadow_used_idx <= vq->size) {
  59                 do_flush_shadow_used_ring_split(dev, vq, used_idx, 0,
  60                                           vq->shadow_used_idx);
  61         } else {
  62                 uint16_t size;
  63
  64                 /* update used ring interval [used_idx, vq->size] */
  65                 size = vq->size - used_idx;
  66                 do_flush_shadow_used_ring_split(dev, vq, used_idx, 0, size);
  67
  68                 /* update the left half used ring interval [0, left_size] */
  69                 do_flush_shadow_used_ring_split(dev, vq, 0, size,
  70                                           vq->shadow_used_idx - size);
  71         }
  72         vq->last_used_idx += vq->shadow_used_idx;
  73
  74         rte_smp_wmb();
  75
  76         vhost_log_cache_sync(dev, vq);
  77
  78         *(volatile uint16_t *)&vq->used->idx += vq->shadow_used_idx;
  79         vq->shadow_used_idx = 0;
  80         vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
  81                 sizeof(vq->used->idx));
  82 }
  83
  84 static __rte_always_inline void
  85 update_shadow_used_ring_split(struct vhost_virtqueue *vq,
  86                          uint16_t desc_idx, uint32_t len)
  87 {
  88         uint16_t i = vq->shadow_used_idx++;
  89
  90         vq->shadow_used_split[i].id  = desc_idx;
  91         vq->shadow_used_split[i].len = len;
  92 }
  93
  94 static __rte_always_inline void
  95 flush_shadow_used_ring_packed(struct virtio_net *dev,
  96                         struct vhost_virtqueue *vq)
  97 {
  98         int i;
  99         uint16_t used_idx = vq->last_used_idx;
 100
 101         /* Split loop in two to save memory barriers */
 102         for (i = 0; i < vq->shadow_used_idx; i++) {
 103                 vq->desc_packed[used_idx].id = vq->shadow_used_packed[i].id;
 104                 vq->desc_packed[used_idx].len = vq->shadow_used_packed[i].len;
 105
 106                 used_idx += vq->shadow_used_packed[i].count;
 107                 if (used_idx >= vq->size)
 108                         used_idx -= vq->size;
 109         }
 110
 111         rte_smp_wmb();
 112
 113         for (i = 0; i < vq->shadow_used_idx; i++) {
 114                 uint16_t flags;
 115
 116                 if (vq->shadow_used_packed[i].len)
 117                         flags = VRING_DESC_F_WRITE;
 118                 else
 119                         flags = 0;
 120
 121                 if (vq->used_wrap_counter) {
 122                         flags |= VRING_DESC_F_USED;
 123                         flags |= VRING_DESC_F_AVAIL;
 124                 } else {
 125                         flags &= ~VRING_DESC_F_USED;
 126                         flags &= ~VRING_DESC_F_AVAIL;
 127                 }
 128
 129                 vq->desc_packed[vq->last_used_idx].flags = flags;
 130
 131                 vhost_log_cache_used_vring(dev, vq,
 132                                         vq->last_used_idx *
 133                                         sizeof(struct vring_packed_desc),
 134                                         sizeof(struct vring_packed_desc));
 135
 136                 vq->last_used_idx += vq->shadow_used_packed[i].count;
 137                 if (vq->last_used_idx >= vq->size) {
 138                         vq->used_wrap_counter ^= 1;
 139                         vq->last_used_idx -= vq->size;
 140                 }
 141         }
 142
 143         rte_smp_wmb();
 144         vq->shadow_used_idx = 0;
 145         vhost_log_cache_sync(dev, vq);
 146 }
 147
 148 static __rte_always_inline void
 149 update_shadow_used_ring_packed(struct vhost_virtqueue *vq,
 150                          uint16_t desc_idx, uint32_t len, uint16_t count)
 151 {
 152         uint16_t i = vq->shadow_used_idx++;
 153
 154         vq->shadow_used_packed[i].id  = desc_idx;
 155         vq->shadow_used_packed[i].len = len;
 156         vq->shadow_used_packed[i].count = count;
 157 }
 158
 159 static inline void
 160 do_data_copy_enqueue(struct virtio_net *dev, struct vhost_virtqueue *vq)
 161 {
 162         struct batch_copy_elem *elem = vq->batch_copy_elems;
 163         uint16_t count = vq->batch_copy_nb_elems;
 164         int i;
 165
 166         for (i = 0; i < count; i++) {
 167                 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
 168                 vhost_log_cache_write(dev, vq, elem[i].log_addr, elem[i].len);
 169                 PRINT_PACKET(dev, (uintptr_t)elem[i].dst, elem[i].len, 0);
 170         }
 171
 172         vq->batch_copy_nb_elems = 0;
 173 }
 174
 175 static inline void
 176 do_data_copy_dequeue(struct vhost_virtqueue *vq)
 177 {
 178         struct batch_copy_elem *elem = vq->batch_copy_elems;
 179         uint16_t count = vq->batch_copy_nb_elems;
 180         int i;
 181
 182         for (i = 0; i < count; i++)
 183                 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
 184
 185         vq->batch_copy_nb_elems = 0;
 186 }
 187
 188 /* avoid write operation when necessary, to lessen cache issues */
 189 #define ASSIGN_UNLESS_EQUAL(var, val) do {      \
 190         if ((var) != (val))                     \
 191                 (var) = (val);                  \
 192 } while (0)
 193
 194 static __rte_always_inline void
 195 virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
 196 {
 197         uint64_t csum_l4 = m_buf->ol_flags & PKT_TX_L4_MASK;
 198
 199         if (m_buf->ol_flags & PKT_TX_TCP_SEG)
 200                 csum_l4 |= PKT_TX_TCP_CKSUM;
 201
 202         if (csum_l4) {
 203                 net_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 204                 net_hdr->csum_start = m_buf->l2_len + m_buf->l3_len;
 205
 206                 switch (csum_l4) {
 207                 case PKT_TX_TCP_CKSUM:
 208                         net_hdr->csum_offset = (offsetof(struct tcp_hdr,
 209                                                 cksum));
 210                         break;
 211                 case PKT_TX_UDP_CKSUM:
 212                         net_hdr->csum_offset = (offsetof(struct udp_hdr,
 213                                                 dgram_cksum));
 214                         break;
 215                 case PKT_TX_SCTP_CKSUM:
 216                         net_hdr->csum_offset = (offsetof(struct sctp_hdr,
 217                                                 cksum));
 218                         break;
 219                 }
 220         } else {
 221                 ASSIGN_UNLESS_EQUAL(net_hdr->csum_start, 0);
 222                 ASSIGN_UNLESS_EQUAL(net_hdr->csum_offset, 0);
 223                 ASSIGN_UNLESS_EQUAL(net_hdr->flags, 0);
 224         }
 225
 226         /* IP cksum verification cannot be bypassed, then calculate here */
 227         if (m_buf->ol_flags & PKT_TX_IP_CKSUM) {
 228                 struct ipv4_hdr *ipv4_hdr;
 229
 230                 ipv4_hdr = rte_pktmbuf_mtod_offset(m_buf, struct ipv4_hdr *,
 231                                                    m_buf->l2_len);
 232                 ipv4_hdr->hdr_checksum = rte_ipv4_cksum(ipv4_hdr);
 233         }
 234
 235         if (m_buf->ol_flags & PKT_TX_TCP_SEG) {
 236                 if (m_buf->ol_flags & PKT_TX_IPV4)
 237                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 238                 else
 239                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 240                 net_hdr->gso_size = m_buf->tso_segsz;
 241                 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len
 242                                         + m_buf->l4_len;
 243         } else if (m_buf->ol_flags & PKT_TX_UDP_SEG) {
 244                 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
 245                 net_hdr->gso_size = m_buf->tso_segsz;
 246                 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len +
 247                         m_buf->l4_len;
 248         } else {
 249                 ASSIGN_UNLESS_EQUAL(net_hdr->gso_type, 0);
 250                 ASSIGN_UNLESS_EQUAL(net_hdr->gso_size, 0);
 251                 ASSIGN_UNLESS_EQUAL(net_hdr->hdr_len, 0);
 252         }
 253 }
 254
 255 static __rte_always_inline int
 256 map_one_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 257                 struct buf_vector *buf_vec, uint16_t *vec_idx,
 258                 uint64_t desc_iova, uint64_t desc_len, uint8_t perm)
 259 {
 260         uint16_t vec_id = *vec_idx;
 261
 262         while (desc_len) {
 263                 uint64_t desc_addr;
 264                 uint64_t desc_chunck_len = desc_len;
 265
 266                 if (unlikely(vec_id >= BUF_VECTOR_MAX))
 267                         return -1;
 268
 269                 desc_addr = vhost_iova_to_vva(dev, vq,
 270                                 desc_iova,
 271                                 &desc_chunck_len,
 272                                 perm);
 273                 if (unlikely(!desc_addr))
 274                         return -1;
 275
 276                 buf_vec[vec_id].buf_iova = desc_iova;
 277                 buf_vec[vec_id].buf_addr = desc_addr;
 278                 buf_vec[vec_id].buf_len  = desc_chunck_len;
 279
 280                 desc_len -= desc_chunck_len;
 281                 desc_iova += desc_chunck_len;
 282                 vec_id++;
 283         }
 284         *vec_idx = vec_id;
 285
 286         return 0;
 287 }
 288
 289 static __rte_always_inline int
 290 fill_vec_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 291                          uint32_t avail_idx, uint16_t *vec_idx,
 292                          struct buf_vector *buf_vec, uint16_t *desc_chain_head,
 293                          uint32_t *desc_chain_len, uint8_t perm)
 294 {
 295         uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
 296         uint16_t vec_id = *vec_idx;
 297         uint32_t len    = 0;
 298         uint64_t dlen;
 299         struct vring_desc *descs = vq->desc;
 300         struct vring_desc *idesc = NULL;
 301
 302         *desc_chain_head = idx;
 303
 304         if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
 305                 dlen = vq->desc[idx].len;
 306                 descs = (struct vring_desc *)(uintptr_t)
 307                         vhost_iova_to_vva(dev, vq, vq->desc[idx].addr,
 308                                                 &dlen,
 309                                                 VHOST_ACCESS_RO);
 310                 if (unlikely(!descs))
 311                         return -1;
 312
 313                 if (unlikely(dlen < vq->desc[idx].len)) {
 314                         /*
 315                          * The indirect desc table is not contiguous
 316                          * in process VA space, we have to copy it.
 317                          */
 318                         idesc = alloc_copy_ind_table(dev, vq,
 319                                         vq->desc[idx].addr, vq->desc[idx].len);
 320                         if (unlikely(!idesc))
 321                                 return -1;
 322
 323                         descs = idesc;
 324                 }
 325
 326                 idx = 0;
 327         }
 328
 329         while (1) {
 330                 if (unlikely(idx >= vq->size)) {
 331                         free_ind_table(idesc);
 332                         return -1;
 333                 }
 334
 335                 len += descs[idx].len;
 336
 337                 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 338                                                 descs[idx].addr, descs[idx].len,
 339                                                 perm))) {
 340                         free_ind_table(idesc);
 341                         return -1;
 342                 }
 343
 344                 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
 345                         break;
 346
 347                 idx = descs[idx].next;
 348         }
 349
 350         *desc_chain_len = len;
 351         *vec_idx = vec_id;
 352
 353         if (unlikely(!!idesc))
 354                 free_ind_table(idesc);
 355
 356         return 0;
 357 }
 358
 359 /*
 360  * Returns -1 on fail, 0 on success
 361  */
 362 static inline int
 363 reserve_avail_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 364                                 uint32_t size, struct buf_vector *buf_vec,
 365                                 uint16_t *num_buffers, uint16_t avail_head,
 366                                 uint16_t *nr_vec)
 367 {
 368         uint16_t cur_idx;
 369         uint16_t vec_idx = 0;
 370         uint16_t max_tries, tries = 0;
 371
 372         uint16_t head_idx = 0;
 373         uint32_t len = 0;
 374
 375         *num_buffers = 0;
 376         cur_idx  = vq->last_avail_idx;
 377
 378         if (rxvq_is_mergeable(dev))
 379                 max_tries = vq->size - 1;
 380         else
 381                 max_tries = 1;
 382
 383         while (size > 0) {
 384                 if (unlikely(cur_idx == avail_head))
 385                         return -1;
 386                 /*
 387                  * if we tried all available ring items, and still
 388                  * can't get enough buf, it means something abnormal
 389                  * happened.
 390                  */
 391                 if (unlikely(++tries > max_tries))
 392                         return -1;
 393
 394                 if (unlikely(fill_vec_buf_split(dev, vq, cur_idx,
 395                                                 &vec_idx, buf_vec,
 396                                                 &head_idx, &len,
 397                                                 VHOST_ACCESS_RW) < 0))
 398                         return -1;
 399                 len = RTE_MIN(len, size);
 400                 update_shadow_used_ring_split(vq, head_idx, len);
 401                 size -= len;
 402
 403                 cur_idx++;
 404                 *num_buffers += 1;
 405         }
 406
 407         *nr_vec = vec_idx;
 408
 409         return 0;
 410 }
 411
 412 static __rte_always_inline int
 413 fill_vec_buf_packed_indirect(struct virtio_net *dev,
 414                         struct vhost_virtqueue *vq,
 415                         struct vring_packed_desc *desc, uint16_t *vec_idx,
 416                         struct buf_vector *buf_vec, uint32_t *len, uint8_t perm)
 417 {
 418         uint16_t i;
 419         uint32_t nr_descs;
 420         uint16_t vec_id = *vec_idx;
 421         uint64_t dlen;
 422         struct vring_packed_desc *descs, *idescs = NULL;
 423
 424         dlen = desc->len;
 425         descs = (struct vring_packed_desc *)(uintptr_t)
 426                 vhost_iova_to_vva(dev, vq, desc->addr, &dlen, VHOST_ACCESS_RO);
 427         if (unlikely(!descs))
 428                 return -1;
 429
 430         if (unlikely(dlen < desc->len)) {
 431                 /*
 432                  * The indirect desc table is not contiguous
 433                  * in process VA space, we have to copy it.
 434                  */
 435                 idescs = alloc_copy_ind_table(dev, vq, desc->addr, desc->len);
 436                 if (unlikely(!idescs))
 437                         return -1;
 438
 439                 descs = idescs;
 440         }
 441
 442         nr_descs =  desc->len / sizeof(struct vring_packed_desc);
 443         if (unlikely(nr_descs >= vq->size)) {
 444                 free_ind_table(idescs);
 445                 return -1;
 446         }
 447
 448         for (i = 0; i < nr_descs; i++) {
 449                 if (unlikely(vec_id >= BUF_VECTOR_MAX)) {
 450                         free_ind_table(idescs);
 451                         return -1;
 452                 }
 453
 454                 *len += descs[i].len;
 455                 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 456                                                 descs[i].addr, descs[i].len,
 457                                                 perm)))
 458                         return -1;
 459         }
 460         *vec_idx = vec_id;
 461
 462         if (unlikely(!!idescs))
 463                 free_ind_table(idescs);
 464
 465         return 0;
 466 }
 467
 468 static __rte_always_inline int
 469 fill_vec_buf_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
 470                                 uint16_t avail_idx, uint16_t *desc_count,
 471                                 struct buf_vector *buf_vec, uint16_t *vec_idx,
 472                                 uint16_t *buf_id, uint32_t *len, uint8_t perm)
 473 {
 474         bool wrap_counter = vq->avail_wrap_counter;
 475         struct vring_packed_desc *descs = vq->desc_packed;
 476         uint16_t vec_id = *vec_idx;
 477
 478         if (avail_idx < vq->last_avail_idx)
 479                 wrap_counter ^= 1;
 480
 481         if (unlikely(!desc_is_avail(&descs[avail_idx], wrap_counter)))
 482                 return -1;
 483
 484         /*
 485          * The ordering between desc flags and desc
 486          * content reads need to be enforced.
 487          */
 488         rte_smp_rmb();
 489
 490         *desc_count = 0;
 491         *len = 0;
 492
 493         while (1) {
 494                 if (unlikely(vec_id >= BUF_VECTOR_MAX))
 495                         return -1;
 496
 497                 *desc_count += 1;
 498                 *buf_id = descs[avail_idx].id;
 499
 500                 if (descs[avail_idx].flags & VRING_DESC_F_INDIRECT) {
 501                         if (unlikely(fill_vec_buf_packed_indirect(dev, vq,
 502                                                         &descs[avail_idx],
 503                                                         &vec_id, buf_vec,
 504                                                         len, perm) < 0))
 505                                 return -1;
 506                 } else {
 507                         *len += descs[avail_idx].len;
 508
 509                         if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 510                                                         descs[avail_idx].addr,
 511                                                         descs[avail_idx].len,
 512                                                         perm)))
 513                                 return -1;
 514                 }
 515
 516                 if ((descs[avail_idx].flags & VRING_DESC_F_NEXT) == 0)
 517                         break;
 518
 519                 if (++avail_idx >= vq->size) {
 520                         avail_idx -= vq->size;
 521                         wrap_counter ^= 1;
 522                 }
 523         }
 524
 525         *vec_idx = vec_id;
 526
 527         return 0;
 528 }
 529
 530 /*
 531  * Returns -1 on fail, 0 on success
 532  */
 533 static inline int
 534 reserve_avail_buf_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
 535                                 uint32_t size, struct buf_vector *buf_vec,
 536                                 uint16_t *nr_vec, uint16_t *num_buffers,
 537                                 uint16_t *nr_descs)
 538 {
 539         uint16_t avail_idx;
 540         uint16_t vec_idx = 0;
 541         uint16_t max_tries, tries = 0;
 542
 543         uint16_t buf_id = 0;
 544         uint32_t len = 0;
 545         uint16_t desc_count;
 546
 547         *num_buffers = 0;
 548         avail_idx = vq->last_avail_idx;
 549
 550         if (rxvq_is_mergeable(dev))
 551                 max_tries = vq->size - 1;
 552         else
 553                 max_tries = 1;
 554
 555         while (size > 0) {
 556                 /*
 557                  * if we tried all available ring items, and still
 558                  * can't get enough buf, it means something abnormal
 559                  * happened.
 560                  */
 561                 if (unlikely(++tries > max_tries))
 562                         return -1;
 563
 564                 if (unlikely(fill_vec_buf_packed(dev, vq,
 565                                                 avail_idx, &desc_count,
 566                                                 buf_vec, &vec_idx,
 567                                                 &buf_id, &len,
 568                                                 VHOST_ACCESS_RW) < 0))
 569                         return -1;
 570
 571                 len = RTE_MIN(len, size);
 572                 update_shadow_used_ring_packed(vq, buf_id, len, desc_count);
 573                 size -= len;
 574
 575                 avail_idx += desc_count;
 576                 if (avail_idx >= vq->size)
 577                         avail_idx -= vq->size;
 578
 579                 *nr_descs += desc_count;
 580                 *num_buffers += 1;
 581         }
 582
 583         *nr_vec = vec_idx;
 584
 585         return 0;
 586 }
 587
 588 static __rte_always_inline int
 589 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 590                             struct rte_mbuf *m, struct buf_vector *buf_vec,
 591                             uint16_t nr_vec, uint16_t num_buffers)
 592 {
 593         uint32_t vec_idx = 0;
 594         uint32_t mbuf_offset, mbuf_avail;
 595         uint32_t buf_offset, buf_avail;
 596         uint64_t buf_addr, buf_iova, buf_len;
 597         uint32_t cpy_len;
 598         uint64_t hdr_addr;
 599         struct rte_mbuf *hdr_mbuf;
 600         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 601         struct virtio_net_hdr_mrg_rxbuf tmp_hdr, *hdr = NULL;
 602         int error = 0;
 603
 604         if (unlikely(m == NULL)) {
 605                 error = -1;
 606                 goto out;
 607         }
 608
 609         buf_addr = buf_vec[vec_idx].buf_addr;
 610         buf_iova = buf_vec[vec_idx].buf_iova;
 611         buf_len = buf_vec[vec_idx].buf_len;
 612
 613         if (nr_vec > 1)
 614                 rte_prefetch0((void *)(uintptr_t)buf_vec[1].buf_addr);
 615
 616         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
 617                 error = -1;
 618                 goto out;
 619         }
 620
 621         hdr_mbuf = m;
 622         hdr_addr = buf_addr;
 623         if (unlikely(buf_len < dev->vhost_hlen))
 624                 hdr = &tmp_hdr;
 625         else
 626                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)hdr_addr;
 627
 628         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
 629                 dev->vid, num_buffers);
 630
 631         if (unlikely(buf_len < dev->vhost_hlen)) {
 632                 buf_offset = dev->vhost_hlen - buf_len;
 633                 vec_idx++;
 634                 buf_addr = buf_vec[vec_idx].buf_addr;
 635                 buf_iova = buf_vec[vec_idx].buf_iova;
 636                 buf_len = buf_vec[vec_idx].buf_len;
 637                 buf_avail = buf_len - buf_offset;
 638         } else {
 639                 buf_offset = dev->vhost_hlen;
 640                 buf_avail = buf_len - dev->vhost_hlen;
 641         }
 642
 643         mbuf_avail  = rte_pktmbuf_data_len(m);
 644         mbuf_offset = 0;
 645         while (mbuf_avail != 0 || m->next != NULL) {
 646                 /* done with current buf, get the next one */
 647                 if (buf_avail == 0) {
 648                         vec_idx++;
 649                         if (unlikely(vec_idx >= nr_vec)) {
 650                                 error = -1;
 651                                 goto out;
 652                         }
 653
 654                         buf_addr = buf_vec[vec_idx].buf_addr;
 655                         buf_iova = buf_vec[vec_idx].buf_iova;
 656                         buf_len = buf_vec[vec_idx].buf_len;
 657
 658                         /* Prefetch next buffer address. */
 659                         if (vec_idx + 1 < nr_vec)
 660                                 rte_prefetch0((void *)(uintptr_t)
 661                                                 buf_vec[vec_idx + 1].buf_addr);
 662                         buf_offset = 0;
 663                         buf_avail  = buf_len;
 664                 }
 665
 666                 /* done with current mbuf, get the next one */
 667                 if (mbuf_avail == 0) {
 668                         m = m->next;
 669
 670                         mbuf_offset = 0;
 671                         mbuf_avail  = rte_pktmbuf_data_len(m);
 672                 }
 673
 674                 if (hdr_addr) {
 675                         virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
 676                         if (rxvq_is_mergeable(dev))
 677                                 ASSIGN_UNLESS_EQUAL(hdr->num_buffers,
 678                                                 num_buffers);
 679
 680                         if (unlikely(hdr == &tmp_hdr)) {
 681                                 uint64_t len;
 682                                 uint64_t remain = dev->vhost_hlen;
 683                                 uint64_t src = (uint64_t)(uintptr_t)hdr, dst;
 684                                 uint64_t iova = buf_vec[0].buf_iova;
 685                                 uint16_t hdr_vec_idx = 0;
 686
 687                                 while (remain) {
 688                                         len = RTE_MIN(remain,
 689                                                 buf_vec[hdr_vec_idx].buf_len);
 690                                         dst = buf_vec[hdr_vec_idx].buf_addr;
 691                                         rte_memcpy((void *)(uintptr_t)dst,
 692                                                         (void *)(uintptr_t)src,
 693                                                         len);
 694
 695                                         PRINT_PACKET(dev, (uintptr_t)dst,
 696                                                         (uint32_t)len, 0);
 697                                         vhost_log_cache_write(dev, vq,
 698                                                         iova, len);
 699
 700                                         remain -= len;
 701                                         iova += len;
 702                                         src += len;
 703                                         hdr_vec_idx++;
 704                                 }
 705                         } else {
 706                                 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
 707                                                 dev->vhost_hlen, 0);
 708                                 vhost_log_cache_write(dev, vq,
 709                                                 buf_vec[0].buf_iova,
 710                                                 dev->vhost_hlen);
 711                         }
 712
 713                         hdr_addr = 0;
 714                 }
 715
 716                 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
 717
 718                 if (likely(cpy_len > MAX_BATCH_LEN ||
 719                                         vq->batch_copy_nb_elems >= vq->size)) {
 720                         rte_memcpy((void *)((uintptr_t)(buf_addr + buf_offset)),
 721                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 722                                 cpy_len);
 723                         vhost_log_cache_write(dev, vq, buf_iova + buf_offset,
 724                                         cpy_len);
 725                         PRINT_PACKET(dev, (uintptr_t)(buf_addr + buf_offset),
 726                                 cpy_len, 0);
 727                 } else {
 728                         batch_copy[vq->batch_copy_nb_elems].dst =
 729                                 (void *)((uintptr_t)(buf_addr + buf_offset));
 730                         batch_copy[vq->batch_copy_nb_elems].src =
 731                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
 732                         batch_copy[vq->batch_copy_nb_elems].log_addr =
 733                                 buf_iova + buf_offset;
 734                         batch_copy[vq->batch_copy_nb_elems].len = cpy_len;
 735                         vq->batch_copy_nb_elems++;
 736                 }
 737
 738                 mbuf_avail  -= cpy_len;
 739                 mbuf_offset += cpy_len;
 740                 buf_avail  -= cpy_len;
 741                 buf_offset += cpy_len;
 742         }
 743
 744 out:
 745
 746         return error;
 747 }
 748
 749 static __rte_always_inline uint32_t
 750 virtio_dev_rx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 751         struct rte_mbuf **pkts, uint32_t count)
 752 {
 753         uint32_t pkt_idx = 0;
 754         uint16_t num_buffers;
 755         struct buf_vector buf_vec[BUF_VECTOR_MAX];
 756         uint16_t avail_head;
 757
 758         avail_head = *((volatile uint16_t *)&vq->avail->idx);
 759
 760         /*
 761          * The ordering between avail index and
 762          * desc reads needs to be enforced.
 763          */
 764         rte_smp_rmb();
 765
 766         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
 767
 768         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
 769                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
 770                 uint16_t nr_vec = 0;
 771
 772                 if (unlikely(reserve_avail_buf_split(dev, vq,
 773                                                 pkt_len, buf_vec, &num_buffers,
 774                                                 avail_head, &nr_vec) < 0)) {
 775                         VHOST_LOG_DEBUG(VHOST_DATA,
 776                                 "(%d) failed to get enough desc from vring\n",
 777                                 dev->vid);
 778                         vq->shadow_used_idx -= num_buffers;
 779                         break;
 780                 }
 781
 782                 rte_prefetch0((void *)(uintptr_t)buf_vec[0].buf_addr);
 783
 784                 VHOST_LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
 785                         dev->vid, vq->last_avail_idx,
 786                         vq->last_avail_idx + num_buffers);
 787
 788                 if (copy_mbuf_to_desc(dev, vq, pkts[pkt_idx],
 789                                                 buf_vec, nr_vec,
 790                                                 num_buffers) < 0) {
 791                         vq->shadow_used_idx -= num_buffers;
 792                         break;
 793                 }
 794
 795                 vq->last_avail_idx += num_buffers;
 796         }
 797
 798         do_data_copy_enqueue(dev, vq);
 799
 800         if (likely(vq->shadow_used_idx)) {
 801                 flush_shadow_used_ring_split(dev, vq);
 802                 vhost_vring_call_split(dev, vq);
 803         }
 804
 805         return pkt_idx;
 806 }
 807
 808 static __rte_always_inline uint32_t
 809 virtio_dev_rx_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
 810         struct rte_mbuf **pkts, uint32_t count)
 811 {
 812         uint32_t pkt_idx = 0;
 813         uint16_t num_buffers;
 814         struct buf_vector buf_vec[BUF_VECTOR_MAX];
 815
 816         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
 817                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
 818                 uint16_t nr_vec = 0;
 819                 uint16_t nr_descs = 0;
 820
 821                 if (unlikely(reserve_avail_buf_packed(dev, vq,
 822                                                 pkt_len, buf_vec, &nr_vec,
 823                                                 &num_buffers, &nr_descs) < 0)) {
 824                         VHOST_LOG_DEBUG(VHOST_DATA,
 825                                 "(%d) failed to get enough desc from vring\n",
 826                                 dev->vid);
 827                         vq->shadow_used_idx -= num_buffers;
 828                         break;
 829                 }
 830
 831                 rte_prefetch0((void *)(uintptr_t)buf_vec[0].buf_addr);
 832
 833                 VHOST_LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
 834                         dev->vid, vq->last_avail_idx,
 835                         vq->last_avail_idx + num_buffers);
 836
 837                 if (copy_mbuf_to_desc(dev, vq, pkts[pkt_idx],
 838                                                 buf_vec, nr_vec,
 839                                                 num_buffers) < 0) {
 840                         vq->shadow_used_idx -= num_buffers;
 841                         break;
 842                 }
 843
 844                 vq->last_avail_idx += nr_descs;
 845                 if (vq->last_avail_idx >= vq->size) {
 846                         vq->last_avail_idx -= vq->size;
 847                         vq->avail_wrap_counter ^= 1;
 848                 }
 849         }
 850
 851         do_data_copy_enqueue(dev, vq);
 852
 853         if (likely(vq->shadow_used_idx)) {
 854                 flush_shadow_used_ring_packed(dev, vq);
 855                 vhost_vring_call_packed(dev, vq);
 856         }
 857
 858         return pkt_idx;
 859 }
 860
 861 static __rte_always_inline uint32_t
 862 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
 863         struct rte_mbuf **pkts, uint32_t count)
 864 {
 865         struct vhost_virtqueue *vq;
 866         uint32_t nb_tx = 0;
 867
 868         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
 869         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
 870                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
 871                         dev->vid, __func__, queue_id);
 872                 return 0;
 873         }
 874
 875         vq = dev->virtqueue[queue_id];
 876
 877         rte_spinlock_lock(&vq->access_lock);
 878
 879         if (unlikely(vq->enabled == 0))
 880                 goto out_access_unlock;
 881
 882         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 883                 vhost_user_iotlb_rd_lock(vq);
 884
 885         if (unlikely(vq->access_ok == 0))
 886                 if (unlikely(vring_translate(dev, vq) < 0))
 887                         goto out;
 888
 889         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
 890         if (count == 0)
 891                 goto out;
 892
 893         if (vq_is_packed(dev))
 894                 nb_tx = virtio_dev_rx_packed(dev, vq, pkts, count);
 895         else
 896                 nb_tx = virtio_dev_rx_split(dev, vq, pkts, count);
 897
 898 out:
 899         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 900                 vhost_user_iotlb_rd_unlock(vq);
 901
 902 out_access_unlock:
 903         rte_spinlock_unlock(&vq->access_lock);
 904
 905         return nb_tx;
 906 }
 907
 908 uint16_t
 909 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
 910         struct rte_mbuf **pkts, uint16_t count)
 911 {
 912         struct virtio_net *dev = get_device(vid);
 913
 914         if (!dev)
 915                 return 0;
 916
 917         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
 918                 RTE_LOG(ERR, VHOST_DATA,
 919                         "(%d) %s: built-in vhost net backend is disabled.\n",
 920                         dev->vid, __func__);
 921                 return 0;
 922         }
 923
 924         return virtio_dev_rx(dev, queue_id, pkts, count);
 925 }
 926
 927 static inline bool
 928 virtio_net_with_host_offload(struct virtio_net *dev)
 929 {
 930         if (dev->features &
 931                         ((1ULL << VIRTIO_NET_F_CSUM) |
 932                          (1ULL << VIRTIO_NET_F_HOST_ECN) |
 933                          (1ULL << VIRTIO_NET_F_HOST_TSO4) |
 934                          (1ULL << VIRTIO_NET_F_HOST_TSO6) |
 935                          (1ULL << VIRTIO_NET_F_HOST_UFO)))
 936                 return true;
 937
 938         return false;
 939 }
 940
 941 static void
 942 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
 943 {
 944         struct ipv4_hdr *ipv4_hdr;
 945         struct ipv6_hdr *ipv6_hdr;
 946         void *l3_hdr = NULL;
 947         struct ether_hdr *eth_hdr;
 948         uint16_t ethertype;
 949
 950         eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
 951
 952         m->l2_len = sizeof(struct ether_hdr);
 953         ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
 954
 955         if (ethertype == ETHER_TYPE_VLAN) {
 956                 struct vlan_hdr *vlan_hdr = (struct vlan_hdr *)(eth_hdr + 1);
 957
 958                 m->l2_len += sizeof(struct vlan_hdr);
 959                 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
 960         }
 961
 962         l3_hdr = (char *)eth_hdr + m->l2_len;
 963
 964         switch (ethertype) {
 965         case ETHER_TYPE_IPv4:
 966                 ipv4_hdr = l3_hdr;
 967                 *l4_proto = ipv4_hdr->next_proto_id;
 968                 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
 969                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 970                 m->ol_flags |= PKT_TX_IPV4;
 971                 break;
 972         case ETHER_TYPE_IPv6:
 973                 ipv6_hdr = l3_hdr;
 974                 *l4_proto = ipv6_hdr->proto;
 975                 m->l3_len = sizeof(struct ipv6_hdr);
 976                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 977                 m->ol_flags |= PKT_TX_IPV6;
 978                 break;
 979         default:
 980                 m->l3_len = 0;
 981                 *l4_proto = 0;
 982                 *l4_hdr = NULL;
 983                 break;
 984         }
 985 }
 986
 987 static __rte_always_inline void
 988 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
 989 {
 990         uint16_t l4_proto = 0;
 991         void *l4_hdr = NULL;
 992         struct tcp_hdr *tcp_hdr = NULL;
 993
 994         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
 995                 return;
 996
 997         parse_ethernet(m, &l4_proto, &l4_hdr);
 998         if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 999                 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
1000                         switch (hdr->csum_offset) {
1001                         case (offsetof(struct tcp_hdr, cksum)):
1002                                 if (l4_proto == IPPROTO_TCP)
1003                                         m->ol_flags |= PKT_TX_TCP_CKSUM;
1004                                 break;
1005                         case (offsetof(struct udp_hdr, dgram_cksum)):
1006                                 if (l4_proto == IPPROTO_UDP)
1007                                         m->ol_flags |= PKT_TX_UDP_CKSUM;
1008                                 break;
1009                         case (offsetof(struct sctp_hdr, cksum)):
1010                                 if (l4_proto == IPPROTO_SCTP)
1011                                         m->ol_flags |= PKT_TX_SCTP_CKSUM;
1012                                 break;
1013                         default:
1014                                 break;
1015                         }
1016                 }
1017         }
1018
1019         if (l4_hdr && hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
1020                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
1021                 case VIRTIO_NET_HDR_GSO_TCPV4:
1022                 case VIRTIO_NET_HDR_GSO_TCPV6:
1023                         tcp_hdr = l4_hdr;
1024                         m->ol_flags |= PKT_TX_TCP_SEG;
1025                         m->tso_segsz = hdr->gso_size;
1026                         m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
1027                         break;
1028                 case VIRTIO_NET_HDR_GSO_UDP:
1029                         m->ol_flags |= PKT_TX_UDP_SEG;
1030                         m->tso_segsz = hdr->gso_size;
1031                         m->l4_len = sizeof(struct udp_hdr);
1032                         break;
1033                 default:
1034                         RTE_LOG(WARNING, VHOST_DATA,
1035                                 "unsupported gso type %u.\n", hdr->gso_type);
1036                         break;
1037                 }
1038         }
1039 }
1040
1041 static __rte_always_inline void
1042 put_zmbuf(struct zcopy_mbuf *zmbuf)
1043 {
1044         zmbuf->in_use = 0;
1045 }
1046
1047 static __rte_always_inline int
1048 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
1049                   struct buf_vector *buf_vec, uint16_t nr_vec,
1050                   struct rte_mbuf *m, struct rte_mempool *mbuf_pool)
1051 {
1052         uint32_t buf_avail, buf_offset;
1053         uint64_t buf_addr, buf_iova, buf_len;
1054         uint32_t mbuf_avail, mbuf_offset;
1055         uint32_t cpy_len;
1056         struct rte_mbuf *cur = m, *prev = m;
1057         struct virtio_net_hdr tmp_hdr;
1058         struct virtio_net_hdr *hdr = NULL;
1059         /* A counter to avoid desc dead loop chain */
1060         uint16_t vec_idx = 0;
1061         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
1062         int error = 0;
1063
1064         buf_addr = buf_vec[vec_idx].buf_addr;
1065         buf_iova = buf_vec[vec_idx].buf_iova;
1066         buf_len = buf_vec[vec_idx].buf_len;
1067
1068         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
1069                 error = -1;
1070                 goto out;
1071         }
1072
1073         if (likely(nr_vec > 1))
1074                 rte_prefetch0((void *)(uintptr_t)buf_vec[1].buf_addr);
1075
1076         if (virtio_net_with_host_offload(dev)) {
1077                 if (unlikely(buf_len < sizeof(struct virtio_net_hdr))) {
1078                         uint64_t len;
1079                         uint64_t remain = sizeof(struct virtio_net_hdr);
1080                         uint64_t src;
1081                         uint64_t dst = (uint64_t)(uintptr_t)&tmp_hdr;
1082                         uint16_t hdr_vec_idx = 0;
1083
1084                         /*
1085                          * No luck, the virtio-net header doesn't fit
1086                          * in a contiguous virtual area.
1087                          */
1088                         while (remain) {
1089                                 len = RTE_MIN(remain,
1090                                         buf_vec[hdr_vec_idx].buf_len);
1091                                 src = buf_vec[hdr_vec_idx].buf_addr;
1092                                 rte_memcpy((void *)(uintptr_t)dst,
1093                                                    (void *)(uintptr_t)src, len);
1094
1095                                 remain -= len;
1096                                 dst += len;
1097                                 hdr_vec_idx++;
1098                         }
1099
1100                         hdr = &tmp_hdr;
1101                 } else {
1102                         hdr = (struct virtio_net_hdr *)((uintptr_t)buf_addr);
1103                         rte_prefetch0(hdr);
1104                 }
1105         }
1106
1107         /*
1108          * A virtio driver normally uses at least 2 desc buffers
1109          * for Tx: the first for storing the header, and others
1110          * for storing the data.
1111          */
1112         if (unlikely(buf_len < dev->vhost_hlen)) {
1113                 buf_offset = dev->vhost_hlen - buf_len;
1114                 vec_idx++;
1115                 buf_addr = buf_vec[vec_idx].buf_addr;
1116                 buf_iova = buf_vec[vec_idx].buf_iova;
1117                 buf_len = buf_vec[vec_idx].buf_len;
1118                 buf_avail  = buf_len - buf_offset;
1119         } else if (buf_len == dev->vhost_hlen) {
1120                 if (unlikely(++vec_idx >= nr_vec))
1121                         goto out;
1122                 buf_addr = buf_vec[vec_idx].buf_addr;
1123                 buf_iova = buf_vec[vec_idx].buf_iova;
1124                 buf_len = buf_vec[vec_idx].buf_len;
1125
1126                 buf_offset = 0;
1127                 buf_avail = buf_len;
1128         } else {
1129                 buf_offset = dev->vhost_hlen;
1130                 buf_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
1131         }
1132
1133         rte_prefetch0((void *)(uintptr_t)
1134                         (buf_addr + buf_offset));
1135
1136         PRINT_PACKET(dev,
1137                         (uintptr_t)(buf_addr + buf_offset),
1138                         (uint32_t)buf_avail, 0);
1139
1140         mbuf_offset = 0;
1141         mbuf_avail  = m->buf_len - RTE_PKTMBUF_HEADROOM;
1142         while (1) {
1143                 uint64_t hpa;
1144
1145                 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
1146
1147                 /*
1148                  * A desc buf might across two host physical pages that are
1149                  * not continuous. In such case (gpa_to_hpa returns 0), data
1150                  * will be copied even though zero copy is enabled.
1151                  */
1152                 if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
1153                                         buf_iova + buf_offset, cpy_len)))) {
1154                         cur->data_len = cpy_len;
1155                         cur->data_off = 0;
1156                         cur->buf_addr =
1157                                 (void *)(uintptr_t)(buf_addr + buf_offset);
1158                         cur->buf_iova = hpa;
1159
1160                         /*
1161                          * In zero copy mode, one mbuf can only reference data
1162                          * for one or partial of one desc buff.
1163                          */
1164                         mbuf_avail = cpy_len;
1165                 } else {
1166                         if (likely(cpy_len > MAX_BATCH_LEN ||
1167                                    vq->batch_copy_nb_elems >= vq->size ||
1168                                    (hdr && cur == m))) {
1169                                 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
1170                                                                    mbuf_offset),
1171                                            (void *)((uintptr_t)(buf_addr +
1172                                                            buf_offset)),
1173                                            cpy_len);
1174                         } else {
1175                                 batch_copy[vq->batch_copy_nb_elems].dst =
1176                                         rte_pktmbuf_mtod_offset(cur, void *,
1177                                                                 mbuf_offset);
1178                                 batch_copy[vq->batch_copy_nb_elems].src =
1179                                         (void *)((uintptr_t)(buf_addr +
1180                                                                 buf_offset));
1181                                 batch_copy[vq->batch_copy_nb_elems].len =
1182                                         cpy_len;
1183                                 vq->batch_copy_nb_elems++;
1184                         }
1185                 }
1186
1187                 mbuf_avail  -= cpy_len;
1188                 mbuf_offset += cpy_len;
1189                 buf_avail -= cpy_len;
1190                 buf_offset += cpy_len;
1191
1192                 /* This buf reaches to its end, get the next one */
1193                 if (buf_avail == 0) {
1194                         if (++vec_idx >= nr_vec)
1195                                 break;
1196
1197                         buf_addr = buf_vec[vec_idx].buf_addr;
1198                         buf_iova = buf_vec[vec_idx].buf_iova;
1199                         buf_len = buf_vec[vec_idx].buf_len;
1200
1201                         /*
1202                          * Prefecth desc n + 1 buffer while
1203                          * desc n buffer is processed.
1204                          */
1205                         if (vec_idx + 1 < nr_vec)
1206                                 rte_prefetch0((void *)(uintptr_t)
1207                                                 buf_vec[vec_idx + 1].buf_addr);
1208
1209                         buf_offset = 0;
1210                         buf_avail  = buf_len;
1211
1212                         PRINT_PACKET(dev, (uintptr_t)buf_addr,
1213                                         (uint32_t)buf_avail, 0);
1214                 }
1215
1216                 /*
1217                  * This mbuf reaches to its end, get a new one
1218                  * to hold more data.
1219                  */
1220                 if (mbuf_avail == 0) {
1221                         cur = rte_pktmbuf_alloc(mbuf_pool);
1222                         if (unlikely(cur == NULL)) {
1223                                 RTE_LOG(ERR, VHOST_DATA, "Failed to "
1224                                         "allocate memory for mbuf.\n");
1225                                 error = -1;
1226                                 goto out;
1227                         }
1228                         if (unlikely(dev->dequeue_zero_copy))
1229                                 rte_mbuf_refcnt_update(cur, 1);
1230
1231                         prev->next = cur;
1232                         prev->data_len = mbuf_offset;
1233                         m->nb_segs += 1;
1234                         m->pkt_len += mbuf_offset;
1235                         prev = cur;
1236
1237                         mbuf_offset = 0;
1238                         mbuf_avail  = cur->buf_len - RTE_PKTMBUF_HEADROOM;
1239                 }
1240         }
1241
1242         prev->data_len = mbuf_offset;
1243         m->pkt_len    += mbuf_offset;
1244
1245         if (hdr)
1246                 vhost_dequeue_offload(hdr, m);
1247
1248 out:
1249
1250         return error;
1251 }
1252
1253 static __rte_always_inline struct zcopy_mbuf *
1254 get_zmbuf(struct vhost_virtqueue *vq)
1255 {
1256         uint16_t i;
1257         uint16_t last;
1258         int tries = 0;
1259
1260         /* search [last_zmbuf_idx, zmbuf_size) */
1261         i = vq->last_zmbuf_idx;
1262         last = vq->zmbuf_size;
1263
1264 again:
1265         for (; i < last; i++) {
1266                 if (vq->zmbufs[i].in_use == 0) {
1267                         vq->last_zmbuf_idx = i + 1;
1268                         vq->zmbufs[i].in_use = 1;
1269                         return &vq->zmbufs[i];
1270                 }
1271         }
1272
1273         tries++;
1274         if (tries == 1) {
1275                 /* search [0, last_zmbuf_idx) */
1276                 i = 0;
1277                 last = vq->last_zmbuf_idx;
1278                 goto again;
1279         }
1280
1281         return NULL;
1282 }
1283
1284 static __rte_always_inline bool
1285 mbuf_is_consumed(struct rte_mbuf *m)
1286 {
1287         while (m) {
1288                 if (rte_mbuf_refcnt_read(m) > 1)
1289                         return false;
1290                 m = m->next;
1291         }
1292
1293         return true;
1294 }
1295
1296 static __rte_always_inline void
1297 restore_mbuf(struct rte_mbuf *m)
1298 {
1299         uint32_t mbuf_size, priv_size;
1300
1301         while (m) {
1302                 priv_size = rte_pktmbuf_priv_size(m->pool);
1303                 mbuf_size = sizeof(struct rte_mbuf) + priv_size;
1304                 /* start of buffer is after mbuf structure and priv data */
1305
1306                 m->buf_addr = (char *)m + mbuf_size;
1307                 m->buf_iova = rte_mempool_virt2iova(m) + mbuf_size;
1308                 m = m->next;
1309         }
1310 }
1311
1312 static __rte_always_inline uint16_t
1313 virtio_dev_tx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
1314         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1315 {
1316         uint16_t i;
1317         uint16_t free_entries;
1318
1319         if (unlikely(dev->dequeue_zero_copy)) {
1320                 struct zcopy_mbuf *zmbuf, *next;
1321
1322                 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1323                      zmbuf != NULL; zmbuf = next) {
1324                         next = TAILQ_NEXT(zmbuf, next);
1325
1326                         if (mbuf_is_consumed(zmbuf->mbuf)) {
1327                                 update_shadow_used_ring_split(vq,
1328                                                 zmbuf->desc_idx, 0);
1329                                 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1330                                 restore_mbuf(zmbuf->mbuf);
1331                                 rte_pktmbuf_free(zmbuf->mbuf);
1332                                 put_zmbuf(zmbuf);
1333                                 vq->nr_zmbuf -= 1;
1334                         }
1335                 }
1336
1337                 if (likely(vq->shadow_used_idx)) {
1338                         flush_shadow_used_ring_split(dev, vq);
1339                         vhost_vring_call_split(dev, vq);
1340                 }
1341         }
1342
1343         free_entries = *((volatile uint16_t *)&vq->avail->idx) -
1344                         vq->last_avail_idx;
1345         if (free_entries == 0)
1346                 return 0;
1347
1348         /*
1349          * The ordering between avail index and
1350          * desc reads needs to be enforced.
1351          */
1352         rte_smp_rmb();
1353
1354         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
1355
1356         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1357
1358         count = RTE_MIN(count, MAX_PKT_BURST);
1359         count = RTE_MIN(count, free_entries);
1360         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1361                         dev->vid, count);
1362
1363         for (i = 0; i < count; i++) {
1364                 struct buf_vector buf_vec[BUF_VECTOR_MAX];
1365                 uint16_t head_idx;
1366                 uint32_t dummy_len;
1367                 uint16_t nr_vec = 0;
1368                 int err;
1369
1370                 if (unlikely(fill_vec_buf_split(dev, vq,
1371                                                 vq->last_avail_idx + i,
1372                                                 &nr_vec, buf_vec,
1373                                                 &head_idx, &dummy_len,
1374                                                 VHOST_ACCESS_RO) < 0))
1375                         break;
1376
1377                 if (likely(dev->dequeue_zero_copy == 0))
1378                         update_shadow_used_ring_split(vq, head_idx, 0);
1379
1380                 rte_prefetch0((void *)(uintptr_t)buf_vec[0].buf_addr);
1381
1382                 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
1383                 if (unlikely(pkts[i] == NULL)) {
1384                         RTE_LOG(ERR, VHOST_DATA,
1385                                 "Failed to allocate memory for mbuf.\n");
1386                         break;
1387                 }
1388
1389                 err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts[i],
1390                                 mbuf_pool);
1391                 if (unlikely(err)) {
1392                         rte_pktmbuf_free(pkts[i]);
1393                         break;
1394                 }
1395
1396                 if (unlikely(dev->dequeue_zero_copy)) {
1397                         struct zcopy_mbuf *zmbuf;
1398
1399                         zmbuf = get_zmbuf(vq);
1400                         if (!zmbuf) {
1401                                 rte_pktmbuf_free(pkts[i]);
1402                                 break;
1403                         }
1404                         zmbuf->mbuf = pkts[i];
1405                         zmbuf->desc_idx = head_idx;
1406
1407                         /*
1408                          * Pin lock the mbuf; we will check later to see
1409                          * whether the mbuf is freed (when we are the last
1410                          * user) or not. If that's the case, we then could
1411                          * update the used ring safely.
1412                          */
1413                         rte_mbuf_refcnt_update(pkts[i], 1);
1414
1415                         vq->nr_zmbuf += 1;
1416                         TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1417                 }
1418         }
1419         vq->last_avail_idx += i;
1420
1421         if (likely(dev->dequeue_zero_copy == 0)) {
1422                 do_data_copy_dequeue(vq);
1423                 if (unlikely(i < count))
1424                         vq->shadow_used_idx = i;
1425                 if (likely(vq->shadow_used_idx)) {
1426                         flush_shadow_used_ring_split(dev, vq);
1427                         vhost_vring_call_split(dev, vq);
1428                 }
1429         }
1430
1431         return i;
1432 }
1433
1434 static __rte_always_inline uint16_t
1435 virtio_dev_tx_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
1436         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1437 {
1438         uint16_t i;
1439
1440         if (unlikely(dev->dequeue_zero_copy)) {
1441                 struct zcopy_mbuf *zmbuf, *next;
1442
1443                 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1444                      zmbuf != NULL; zmbuf = next) {
1445                         next = TAILQ_NEXT(zmbuf, next);
1446
1447                         if (mbuf_is_consumed(zmbuf->mbuf)) {
1448                                 update_shadow_used_ring_packed(vq,
1449                                                 zmbuf->desc_idx,
1450                                                 0,
1451                                                 zmbuf->desc_count);
1452
1453                                 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1454                                 restore_mbuf(zmbuf->mbuf);
1455                                 rte_pktmbuf_free(zmbuf->mbuf);
1456                                 put_zmbuf(zmbuf);
1457                                 vq->nr_zmbuf -= 1;
1458                         }
1459                 }
1460
1461                 if (likely(vq->shadow_used_idx)) {
1462                         flush_shadow_used_ring_packed(dev, vq);
1463                         vhost_vring_call_packed(dev, vq);
1464                 }
1465         }
1466
1467         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1468
1469         count = RTE_MIN(count, MAX_PKT_BURST);
1470         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1471                         dev->vid, count);
1472
1473         for (i = 0; i < count; i++) {
1474                 struct buf_vector buf_vec[BUF_VECTOR_MAX];
1475                 uint16_t buf_id;
1476                 uint32_t dummy_len;
1477                 uint16_t desc_count, nr_vec = 0;
1478                 int err;
1479
1480                 if (unlikely(fill_vec_buf_packed(dev, vq,
1481                                                 vq->last_avail_idx, &desc_count,
1482                                                 buf_vec, &nr_vec,
1483                                                 &buf_id, &dummy_len,
1484                                                 VHOST_ACCESS_RO) < 0))
1485                         break;
1486
1487                 if (likely(dev->dequeue_zero_copy == 0))
1488                         update_shadow_used_ring_packed(vq, buf_id, 0,
1489                                         desc_count);
1490
1491                 rte_prefetch0((void *)(uintptr_t)buf_vec[0].buf_addr);
1492
1493                 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
1494                 if (unlikely(pkts[i] == NULL)) {
1495                         RTE_LOG(ERR, VHOST_DATA,
1496                                 "Failed to allocate memory for mbuf.\n");
1497                         break;
1498                 }
1499
1500                 err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts[i],
1501                                 mbuf_pool);
1502                 if (unlikely(err)) {
1503                         rte_pktmbuf_free(pkts[i]);
1504                         break;
1505                 }
1506
1507                 if (unlikely(dev->dequeue_zero_copy)) {
1508                         struct zcopy_mbuf *zmbuf;
1509
1510                         zmbuf = get_zmbuf(vq);
1511                         if (!zmbuf) {
1512                                 rte_pktmbuf_free(pkts[i]);
1513                                 break;
1514                         }
1515                         zmbuf->mbuf = pkts[i];
1516                         zmbuf->desc_idx = buf_id;
1517                         zmbuf->desc_count = desc_count;
1518
1519                         /*
1520                          * Pin lock the mbuf; we will check later to see
1521                          * whether the mbuf is freed (when we are the last
1522                          * user) or not. If that's the case, we then could
1523                          * update the used ring safely.
1524                          */
1525                         rte_mbuf_refcnt_update(pkts[i], 1);
1526
1527                         vq->nr_zmbuf += 1;
1528                         TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1529                 }
1530
1531                 vq->last_avail_idx += desc_count;
1532                 if (vq->last_avail_idx >= vq->size) {
1533                         vq->last_avail_idx -= vq->size;
1534                         vq->avail_wrap_counter ^= 1;
1535                 }
1536         }
1537
1538         if (likely(dev->dequeue_zero_copy == 0)) {
1539                 do_data_copy_dequeue(vq);
1540                 if (unlikely(i < count))
1541                         vq->shadow_used_idx = i;
1542                 if (likely(vq->shadow_used_idx)) {
1543                         flush_shadow_used_ring_packed(dev, vq);
1544                         vhost_vring_call_packed(dev, vq);
1545                 }
1546         }
1547
1548         return i;
1549 }
1550
1551 uint16_t
1552 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
1553         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1554 {
1555         struct virtio_net *dev;
1556         struct rte_mbuf *rarp_mbuf = NULL;
1557         struct vhost_virtqueue *vq;
1558
1559         dev = get_device(vid);
1560         if (!dev)
1561                 return 0;
1562
1563         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
1564                 RTE_LOG(ERR, VHOST_DATA,
1565                         "(%d) %s: built-in vhost net backend is disabled.\n",
1566                         dev->vid, __func__);
1567                 return 0;
1568         }
1569
1570         if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
1571                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
1572                         dev->vid, __func__, queue_id);
1573                 return 0;
1574         }
1575
1576         vq = dev->virtqueue[queue_id];
1577
1578         if (unlikely(rte_spinlock_trylock(&vq->access_lock) == 0))
1579                 return 0;
1580
1581         if (unlikely(vq->enabled == 0)) {
1582                 count = 0;
1583                 goto out_access_unlock;
1584         }
1585
1586         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1587                 vhost_user_iotlb_rd_lock(vq);
1588
1589         if (unlikely(vq->access_ok == 0))
1590                 if (unlikely(vring_translate(dev, vq) < 0)) {
1591                         count = 0;
1592                         goto out;
1593                 }
1594
1595         /*
1596          * Construct a RARP broadcast packet, and inject it to the "pkts"
1597          * array, to looks like that guest actually send such packet.
1598          *
1599          * Check user_send_rarp() for more information.
1600          *
1601          * broadcast_rarp shares a cacheline in the virtio_net structure
1602          * with some fields that are accessed during enqueue and
1603          * rte_atomic16_cmpset() causes a write if using cmpxchg. This could
1604          * result in false sharing between enqueue and dequeue.
1605          *
1606          * Prevent unnecessary false sharing by reading broadcast_rarp first
1607          * and only performing cmpset if the read indicates it is likely to
1608          * be set.
1609          */
1610         if (unlikely(rte_atomic16_read(&dev->broadcast_rarp) &&
1611                         rte_atomic16_cmpset((volatile uint16_t *)
1612                                 &dev->broadcast_rarp.cnt, 1, 0))) {
1613
1614                 rarp_mbuf = rte_net_make_rarp_packet(mbuf_pool, &dev->mac);
1615                 if (rarp_mbuf == NULL) {
1616                         RTE_LOG(ERR, VHOST_DATA,
1617                                 "Failed to make RARP packet.\n");
1618                         count = 0;
1619                         goto out;
1620                 }
1621                 count -= 1;
1622         }
1623
1624         if (vq_is_packed(dev))
1625                 count = virtio_dev_tx_packed(dev, vq, mbuf_pool, pkts, count);
1626         else
1627                 count = virtio_dev_tx_split(dev, vq, mbuf_pool, pkts, count);
1628
1629 out:
1630         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1631                 vhost_user_iotlb_rd_unlock(vq);
1632
1633 out_access_unlock:
1634         rte_spinlock_unlock(&vq->access_lock);
1635
1636         if (unlikely(rarp_mbuf != NULL)) {
1637                 /*
1638                  * Inject it to the head of "pkts" array, so that switch's mac
1639                  * learning table will get updated first.
1640                  */
1641                 memmove(&pkts[1], pkts, count * sizeof(struct rte_mbuf *));
1642                 pkts[0] = rarp_mbuf;
1643                 count += 1;
1644         }
1645
1646         return count;
1647 }