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         uint16_t head_idx = vq->last_used_idx;
 101         uint16_t head_flags = 0;
 102
 103         /* Split loop in two to save memory barriers */
 104         for (i = 0; i < vq->shadow_used_idx; i++) {
 105                 vq->desc_packed[used_idx].id = vq->shadow_used_packed[i].id;
 106                 vq->desc_packed[used_idx].len = vq->shadow_used_packed[i].len;
 107
 108                 used_idx += vq->shadow_used_packed[i].count;
 109                 if (used_idx >= vq->size)
 110                         used_idx -= vq->size;
 111         }
 112
 113         rte_smp_wmb();
 114
 115         for (i = 0; i < vq->shadow_used_idx; i++) {
 116                 uint16_t flags;
 117
 118                 if (vq->shadow_used_packed[i].len)
 119                         flags = VRING_DESC_F_WRITE;
 120                 else
 121                         flags = 0;
 122
 123                 if (vq->used_wrap_counter) {
 124                         flags |= VRING_DESC_F_USED;
 125                         flags |= VRING_DESC_F_AVAIL;
 126                 } else {
 127                         flags &= ~VRING_DESC_F_USED;
 128                         flags &= ~VRING_DESC_F_AVAIL;
 129                 }
 130
 131                 if (i > 0) {
 132                         vq->desc_packed[vq->last_used_idx].flags = flags;
 133
 134                         vhost_log_cache_used_vring(dev, vq,
 135                                         vq->last_used_idx *
 136                                         sizeof(struct vring_packed_desc),
 137                                         sizeof(struct vring_packed_desc));
 138                 } else {
 139                         head_idx = vq->last_used_idx;
 140                         head_flags = flags;
 141                 }
 142
 143                 vq->last_used_idx += vq->shadow_used_packed[i].count;
 144                 if (vq->last_used_idx >= vq->size) {
 145                         vq->used_wrap_counter ^= 1;
 146                         vq->last_used_idx -= vq->size;
 147                 }
 148         }
 149
 150         vq->desc_packed[head_idx].flags = head_flags;
 151
 152         vhost_log_cache_used_vring(dev, vq,
 153                                 head_idx *
 154                                 sizeof(struct vring_packed_desc),
 155                                 sizeof(struct vring_packed_desc));
 156
 157         vq->shadow_used_idx = 0;
 158         vhost_log_cache_sync(dev, vq);
 159 }
 160
 161 static __rte_always_inline void
 162 update_shadow_used_ring_packed(struct vhost_virtqueue *vq,
 163                          uint16_t desc_idx, uint32_t len, uint16_t count)
 164 {
 165         uint16_t i = vq->shadow_used_idx++;
 166
 167         vq->shadow_used_packed[i].id  = desc_idx;
 168         vq->shadow_used_packed[i].len = len;
 169         vq->shadow_used_packed[i].count = count;
 170 }
 171
 172 static inline void
 173 do_data_copy_enqueue(struct virtio_net *dev, struct vhost_virtqueue *vq)
 174 {
 175         struct batch_copy_elem *elem = vq->batch_copy_elems;
 176         uint16_t count = vq->batch_copy_nb_elems;
 177         int i;
 178
 179         for (i = 0; i < count; i++) {
 180                 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
 181                 vhost_log_cache_write(dev, vq, elem[i].log_addr, elem[i].len);
 182                 PRINT_PACKET(dev, (uintptr_t)elem[i].dst, elem[i].len, 0);
 183         }
 184
 185         vq->batch_copy_nb_elems = 0;
 186 }
 187
 188 static inline void
 189 do_data_copy_dequeue(struct vhost_virtqueue *vq)
 190 {
 191         struct batch_copy_elem *elem = vq->batch_copy_elems;
 192         uint16_t count = vq->batch_copy_nb_elems;
 193         int i;
 194
 195         for (i = 0; i < count; i++)
 196                 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
 197
 198         vq->batch_copy_nb_elems = 0;
 199 }
 200
 201 /* avoid write operation when necessary, to lessen cache issues */
 202 #define ASSIGN_UNLESS_EQUAL(var, val) do {      \
 203         if ((var) != (val))                     \
 204                 (var) = (val);                  \
 205 } while (0)
 206
 207 static __rte_always_inline void
 208 virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
 209 {
 210         uint64_t csum_l4 = m_buf->ol_flags & PKT_TX_L4_MASK;
 211
 212         if (m_buf->ol_flags & PKT_TX_TCP_SEG)
 213                 csum_l4 |= PKT_TX_TCP_CKSUM;
 214
 215         if (csum_l4) {
 216                 net_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 217                 net_hdr->csum_start = m_buf->l2_len + m_buf->l3_len;
 218
 219                 switch (csum_l4) {
 220                 case PKT_TX_TCP_CKSUM:
 221                         net_hdr->csum_offset = (offsetof(struct rte_tcp_hdr,
 222                                                 cksum));
 223                         break;
 224                 case PKT_TX_UDP_CKSUM:
 225                         net_hdr->csum_offset = (offsetof(struct rte_udp_hdr,
 226                                                 dgram_cksum));
 227                         break;
 228                 case PKT_TX_SCTP_CKSUM:
 229                         net_hdr->csum_offset = (offsetof(struct rte_sctp_hdr,
 230                                                 cksum));
 231                         break;
 232                 }
 233         } else {
 234                 ASSIGN_UNLESS_EQUAL(net_hdr->csum_start, 0);
 235                 ASSIGN_UNLESS_EQUAL(net_hdr->csum_offset, 0);
 236                 ASSIGN_UNLESS_EQUAL(net_hdr->flags, 0);
 237         }
 238
 239         /* IP cksum verification cannot be bypassed, then calculate here */
 240         if (m_buf->ol_flags & PKT_TX_IP_CKSUM) {
 241                 struct rte_ipv4_hdr *ipv4_hdr;
 242
 243                 ipv4_hdr = rte_pktmbuf_mtod_offset(m_buf, struct rte_ipv4_hdr *,
 244                                                    m_buf->l2_len);
 245                 ipv4_hdr->hdr_checksum = rte_ipv4_cksum(ipv4_hdr);
 246         }
 247
 248         if (m_buf->ol_flags & PKT_TX_TCP_SEG) {
 249                 if (m_buf->ol_flags & PKT_TX_IPV4)
 250                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 251                 else
 252                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 253                 net_hdr->gso_size = m_buf->tso_segsz;
 254                 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len
 255                                         + m_buf->l4_len;
 256         } else if (m_buf->ol_flags & PKT_TX_UDP_SEG) {
 257                 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
 258                 net_hdr->gso_size = m_buf->tso_segsz;
 259                 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len +
 260                         m_buf->l4_len;
 261         } else {
 262                 ASSIGN_UNLESS_EQUAL(net_hdr->gso_type, 0);
 263                 ASSIGN_UNLESS_EQUAL(net_hdr->gso_size, 0);
 264                 ASSIGN_UNLESS_EQUAL(net_hdr->hdr_len, 0);
 265         }
 266 }
 267
 268 static __rte_always_inline int
 269 map_one_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 270                 struct buf_vector *buf_vec, uint16_t *vec_idx,
 271                 uint64_t desc_iova, uint64_t desc_len, uint8_t perm)
 272 {
 273         uint16_t vec_id = *vec_idx;
 274
 275         while (desc_len) {
 276                 uint64_t desc_addr;
 277                 uint64_t desc_chunck_len = desc_len;
 278
 279                 if (unlikely(vec_id >= BUF_VECTOR_MAX))
 280                         return -1;
 281
 282                 desc_addr = vhost_iova_to_vva(dev, vq,
 283                                 desc_iova,
 284                                 &desc_chunck_len,
 285                                 perm);
 286                 if (unlikely(!desc_addr))
 287                         return -1;
 288
 289                 rte_prefetch0((void *)(uintptr_t)desc_addr);
 290
 291                 buf_vec[vec_id].buf_iova = desc_iova;
 292                 buf_vec[vec_id].buf_addr = desc_addr;
 293                 buf_vec[vec_id].buf_len  = desc_chunck_len;
 294
 295                 desc_len -= desc_chunck_len;
 296                 desc_iova += desc_chunck_len;
 297                 vec_id++;
 298         }
 299         *vec_idx = vec_id;
 300
 301         return 0;
 302 }
 303
 304 static __rte_always_inline int
 305 fill_vec_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 306                          uint32_t avail_idx, uint16_t *vec_idx,
 307                          struct buf_vector *buf_vec, uint16_t *desc_chain_head,
 308                          uint32_t *desc_chain_len, uint8_t perm)
 309 {
 310         uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
 311         uint16_t vec_id = *vec_idx;
 312         uint32_t len    = 0;
 313         uint64_t dlen;
 314         uint32_t nr_descs = vq->size;
 315         uint32_t cnt    = 0;
 316         struct vring_desc *descs = vq->desc;
 317         struct vring_desc *idesc = NULL;
 318
 319         if (unlikely(idx >= vq->size))
 320                 return -1;
 321
 322         *desc_chain_head = idx;
 323
 324         if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
 325                 dlen = vq->desc[idx].len;
 326                 nr_descs = dlen / sizeof(struct vring_desc);
 327                 if (unlikely(nr_descs > vq->size))
 328                         return -1;
 329
 330                 descs = (struct vring_desc *)(uintptr_t)
 331                         vhost_iova_to_vva(dev, vq, vq->desc[idx].addr,
 332                                                 &dlen,
 333                                                 VHOST_ACCESS_RO);
 334                 if (unlikely(!descs))
 335                         return -1;
 336
 337                 if (unlikely(dlen < vq->desc[idx].len)) {
 338                         /*
 339                          * The indirect desc table is not contiguous
 340                          * in process VA space, we have to copy it.
 341                          */
 342                         idesc = vhost_alloc_copy_ind_table(dev, vq,
 343                                         vq->desc[idx].addr, vq->desc[idx].len);
 344                         if (unlikely(!idesc))
 345                                 return -1;
 346
 347                         descs = idesc;
 348                 }
 349
 350                 idx = 0;
 351         }
 352
 353         while (1) {
 354                 if (unlikely(idx >= nr_descs || cnt++ >= nr_descs)) {
 355                         free_ind_table(idesc);
 356                         return -1;
 357                 }
 358
 359                 len += descs[idx].len;
 360
 361                 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 362                                                 descs[idx].addr, descs[idx].len,
 363                                                 perm))) {
 364                         free_ind_table(idesc);
 365                         return -1;
 366                 }
 367
 368                 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
 369                         break;
 370
 371                 idx = descs[idx].next;
 372         }
 373
 374         *desc_chain_len = len;
 375         *vec_idx = vec_id;
 376
 377         if (unlikely(!!idesc))
 378                 free_ind_table(idesc);
 379
 380         return 0;
 381 }
 382
 383 /*
 384  * Returns -1 on fail, 0 on success
 385  */
 386 static inline int
 387 reserve_avail_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 388                                 uint32_t size, struct buf_vector *buf_vec,
 389                                 uint16_t *num_buffers, uint16_t avail_head,
 390                                 uint16_t *nr_vec)
 391 {
 392         uint16_t cur_idx;
 393         uint16_t vec_idx = 0;
 394         uint16_t max_tries, tries = 0;
 395
 396         uint16_t head_idx = 0;
 397         uint32_t len = 0;
 398
 399         *num_buffers = 0;
 400         cur_idx  = vq->last_avail_idx;
 401
 402         if (rxvq_is_mergeable(dev))
 403                 max_tries = vq->size - 1;
 404         else
 405                 max_tries = 1;
 406
 407         while (size > 0) {
 408                 if (unlikely(cur_idx == avail_head))
 409                         return -1;
 410                 /*
 411                  * if we tried all available ring items, and still
 412                  * can't get enough buf, it means something abnormal
 413                  * happened.
 414                  */
 415                 if (unlikely(++tries > max_tries))
 416                         return -1;
 417
 418                 if (unlikely(fill_vec_buf_split(dev, vq, cur_idx,
 419                                                 &vec_idx, buf_vec,
 420                                                 &head_idx, &len,
 421                                                 VHOST_ACCESS_RW) < 0))
 422                         return -1;
 423                 len = RTE_MIN(len, size);
 424                 update_shadow_used_ring_split(vq, head_idx, len);
 425                 size -= len;
 426
 427                 cur_idx++;
 428                 *num_buffers += 1;
 429         }
 430
 431         *nr_vec = vec_idx;
 432
 433         return 0;
 434 }
 435
 436 static __rte_always_inline int
 437 fill_vec_buf_packed_indirect(struct virtio_net *dev,
 438                         struct vhost_virtqueue *vq,
 439                         struct vring_packed_desc *desc, uint16_t *vec_idx,
 440                         struct buf_vector *buf_vec, uint32_t *len, uint8_t perm)
 441 {
 442         uint16_t i;
 443         uint32_t nr_descs;
 444         uint16_t vec_id = *vec_idx;
 445         uint64_t dlen;
 446         struct vring_packed_desc *descs, *idescs = NULL;
 447
 448         dlen = desc->len;
 449         descs = (struct vring_packed_desc *)(uintptr_t)
 450                 vhost_iova_to_vva(dev, vq, desc->addr, &dlen, VHOST_ACCESS_RO);
 451         if (unlikely(!descs))
 452                 return -1;
 453
 454         if (unlikely(dlen < desc->len)) {
 455                 /*
 456                  * The indirect desc table is not contiguous
 457                  * in process VA space, we have to copy it.
 458                  */
 459                 idescs = vhost_alloc_copy_ind_table(dev,
 460                                 vq, desc->addr, desc->len);
 461                 if (unlikely(!idescs))
 462                         return -1;
 463
 464                 descs = idescs;
 465         }
 466
 467         nr_descs =  desc->len / sizeof(struct vring_packed_desc);
 468         if (unlikely(nr_descs >= vq->size)) {
 469                 free_ind_table(idescs);
 470                 return -1;
 471         }
 472
 473         for (i = 0; i < nr_descs; i++) {
 474                 if (unlikely(vec_id >= BUF_VECTOR_MAX)) {
 475                         free_ind_table(idescs);
 476                         return -1;
 477                 }
 478
 479                 *len += descs[i].len;
 480                 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 481                                                 descs[i].addr, descs[i].len,
 482                                                 perm)))
 483                         return -1;
 484         }
 485         *vec_idx = vec_id;
 486
 487         if (unlikely(!!idescs))
 488                 free_ind_table(idescs);
 489
 490         return 0;
 491 }
 492
 493 static __rte_always_inline int
 494 fill_vec_buf_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
 495                                 uint16_t avail_idx, uint16_t *desc_count,
 496                                 struct buf_vector *buf_vec, uint16_t *vec_idx,
 497                                 uint16_t *buf_id, uint32_t *len, uint8_t perm)
 498 {
 499         bool wrap_counter = vq->avail_wrap_counter;
 500         struct vring_packed_desc *descs = vq->desc_packed;
 501         uint16_t vec_id = *vec_idx;
 502
 503         if (avail_idx < vq->last_avail_idx)
 504                 wrap_counter ^= 1;
 505
 506         /*
 507          * Perform a load-acquire barrier in desc_is_avail to
 508          * enforce the ordering between desc flags and desc
 509          * content.
 510          */
 511         if (unlikely(!desc_is_avail(&descs[avail_idx], wrap_counter)))
 512                 return -1;
 513
 514         *desc_count = 0;
 515         *len = 0;
 516
 517         while (1) {
 518                 if (unlikely(vec_id >= BUF_VECTOR_MAX))
 519                         return -1;
 520
 521                 if (unlikely(*desc_count >= vq->size))
 522                         return -1;
 523
 524                 *desc_count += 1;
 525                 *buf_id = descs[avail_idx].id;
 526
 527                 if (descs[avail_idx].flags & VRING_DESC_F_INDIRECT) {
 528                         if (unlikely(fill_vec_buf_packed_indirect(dev, vq,
 529                                                         &descs[avail_idx],
 530                                                         &vec_id, buf_vec,
 531                                                         len, perm) < 0))
 532                                 return -1;
 533                 } else {
 534                         *len += descs[avail_idx].len;
 535
 536                         if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 537                                                         descs[avail_idx].addr,
 538                                                         descs[avail_idx].len,
 539                                                         perm)))
 540                                 return -1;
 541                 }
 542
 543                 if ((descs[avail_idx].flags & VRING_DESC_F_NEXT) == 0)
 544                         break;
 545
 546                 if (++avail_idx >= vq->size) {
 547                         avail_idx -= vq->size;
 548                         wrap_counter ^= 1;
 549                 }
 550         }
 551
 552         *vec_idx = vec_id;
 553
 554         return 0;
 555 }
 556
 557 /*
 558  * Returns -1 on fail, 0 on success
 559  */
 560 static inline int
 561 reserve_avail_buf_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
 562                                 uint32_t size, struct buf_vector *buf_vec,
 563                                 uint16_t *nr_vec, uint16_t *num_buffers,
 564                                 uint16_t *nr_descs)
 565 {
 566         uint16_t avail_idx;
 567         uint16_t vec_idx = 0;
 568         uint16_t max_tries, tries = 0;
 569
 570         uint16_t buf_id = 0;
 571         uint32_t len = 0;
 572         uint16_t desc_count;
 573
 574         *num_buffers = 0;
 575         avail_idx = vq->last_avail_idx;
 576
 577         if (rxvq_is_mergeable(dev))
 578                 max_tries = vq->size - 1;
 579         else
 580                 max_tries = 1;
 581
 582         while (size > 0) {
 583                 /*
 584                  * if we tried all available ring items, and still
 585                  * can't get enough buf, it means something abnormal
 586                  * happened.
 587                  */
 588                 if (unlikely(++tries > max_tries))
 589                         return -1;
 590
 591                 if (unlikely(fill_vec_buf_packed(dev, vq,
 592                                                 avail_idx, &desc_count,
 593                                                 buf_vec, &vec_idx,
 594                                                 &buf_id, &len,
 595                                                 VHOST_ACCESS_RW) < 0))
 596                         return -1;
 597
 598                 len = RTE_MIN(len, size);
 599                 update_shadow_used_ring_packed(vq, buf_id, len, desc_count);
 600                 size -= len;
 601
 602                 avail_idx += desc_count;
 603                 if (avail_idx >= vq->size)
 604                         avail_idx -= vq->size;
 605
 606                 *nr_descs += desc_count;
 607                 *num_buffers += 1;
 608         }
 609
 610         *nr_vec = vec_idx;
 611
 612         return 0;
 613 }
 614
 615 static __rte_noinline void
 616 copy_vnet_hdr_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 617                 struct buf_vector *buf_vec,
 618                 struct virtio_net_hdr_mrg_rxbuf *hdr)
 619 {
 620         uint64_t len;
 621         uint64_t remain = dev->vhost_hlen;
 622         uint64_t src = (uint64_t)(uintptr_t)hdr, dst;
 623         uint64_t iova = buf_vec->buf_iova;
 624
 625         while (remain) {
 626                 len = RTE_MIN(remain,
 627                                 buf_vec->buf_len);
 628                 dst = buf_vec->buf_addr;
 629                 rte_memcpy((void *)(uintptr_t)dst,
 630                                 (void *)(uintptr_t)src,
 631                                 len);
 632
 633                 PRINT_PACKET(dev, (uintptr_t)dst,
 634                                 (uint32_t)len, 0);
 635                 vhost_log_cache_write(dev, vq,
 636                                 iova, len);
 637
 638                 remain -= len;
 639                 iova += len;
 640                 src += len;
 641                 buf_vec++;
 642         }
 643 }
 644
 645 static __rte_always_inline int
 646 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 647                             struct rte_mbuf *m, struct buf_vector *buf_vec,
 648                             uint16_t nr_vec, uint16_t num_buffers)
 649 {
 650         uint32_t vec_idx = 0;
 651         uint32_t mbuf_offset, mbuf_avail;
 652         uint32_t buf_offset, buf_avail;
 653         uint64_t buf_addr, buf_iova, buf_len;
 654         uint32_t cpy_len;
 655         uint64_t hdr_addr;
 656         struct rte_mbuf *hdr_mbuf;
 657         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 658         struct virtio_net_hdr_mrg_rxbuf tmp_hdr, *hdr = NULL;
 659         int error = 0;
 660
 661         if (unlikely(m == NULL)) {
 662                 error = -1;
 663                 goto out;
 664         }
 665
 666         buf_addr = buf_vec[vec_idx].buf_addr;
 667         buf_iova = buf_vec[vec_idx].buf_iova;
 668         buf_len = buf_vec[vec_idx].buf_len;
 669
 670         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
 671                 error = -1;
 672                 goto out;
 673         }
 674
 675         hdr_mbuf = m;
 676         hdr_addr = buf_addr;
 677         if (unlikely(buf_len < dev->vhost_hlen))
 678                 hdr = &tmp_hdr;
 679         else
 680                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)hdr_addr;
 681
 682         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
 683                 dev->vid, num_buffers);
 684
 685         if (unlikely(buf_len < dev->vhost_hlen)) {
 686                 buf_offset = dev->vhost_hlen - buf_len;
 687                 vec_idx++;
 688                 buf_addr = buf_vec[vec_idx].buf_addr;
 689                 buf_iova = buf_vec[vec_idx].buf_iova;
 690                 buf_len = buf_vec[vec_idx].buf_len;
 691                 buf_avail = buf_len - buf_offset;
 692         } else {
 693                 buf_offset = dev->vhost_hlen;
 694                 buf_avail = buf_len - dev->vhost_hlen;
 695         }
 696
 697         mbuf_avail  = rte_pktmbuf_data_len(m);
 698         mbuf_offset = 0;
 699         while (mbuf_avail != 0 || m->next != NULL) {
 700                 /* done with current buf, get the next one */
 701                 if (buf_avail == 0) {
 702                         vec_idx++;
 703                         if (unlikely(vec_idx >= nr_vec)) {
 704                                 error = -1;
 705                                 goto out;
 706                         }
 707
 708                         buf_addr = buf_vec[vec_idx].buf_addr;
 709                         buf_iova = buf_vec[vec_idx].buf_iova;
 710                         buf_len = buf_vec[vec_idx].buf_len;
 711
 712                         buf_offset = 0;
 713                         buf_avail  = buf_len;
 714                 }
 715
 716                 /* done with current mbuf, get the next one */
 717                 if (mbuf_avail == 0) {
 718                         m = m->next;
 719
 720                         mbuf_offset = 0;
 721                         mbuf_avail  = rte_pktmbuf_data_len(m);
 722                 }
 723
 724                 if (hdr_addr) {
 725                         virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
 726                         if (rxvq_is_mergeable(dev))
 727                                 ASSIGN_UNLESS_EQUAL(hdr->num_buffers,
 728                                                 num_buffers);
 729
 730                         if (unlikely(hdr == &tmp_hdr)) {
 731                                 copy_vnet_hdr_to_desc(dev, vq, buf_vec, hdr);
 732                         } else {
 733                                 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
 734                                                 dev->vhost_hlen, 0);
 735                                 vhost_log_cache_write(dev, vq,
 736                                                 buf_vec[0].buf_iova,
 737                                                 dev->vhost_hlen);
 738                         }
 739
 740                         hdr_addr = 0;
 741                 }
 742
 743                 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
 744
 745                 if (likely(cpy_len > MAX_BATCH_LEN ||
 746                                         vq->batch_copy_nb_elems >= vq->size)) {
 747                         rte_memcpy((void *)((uintptr_t)(buf_addr + buf_offset)),
 748                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 749                                 cpy_len);
 750                         vhost_log_cache_write(dev, vq, buf_iova + buf_offset,
 751                                         cpy_len);
 752                         PRINT_PACKET(dev, (uintptr_t)(buf_addr + buf_offset),
 753                                 cpy_len, 0);
 754                 } else {
 755                         batch_copy[vq->batch_copy_nb_elems].dst =
 756                                 (void *)((uintptr_t)(buf_addr + buf_offset));
 757                         batch_copy[vq->batch_copy_nb_elems].src =
 758                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
 759                         batch_copy[vq->batch_copy_nb_elems].log_addr =
 760                                 buf_iova + buf_offset;
 761                         batch_copy[vq->batch_copy_nb_elems].len = cpy_len;
 762                         vq->batch_copy_nb_elems++;
 763                 }
 764
 765                 mbuf_avail  -= cpy_len;
 766                 mbuf_offset += cpy_len;
 767                 buf_avail  -= cpy_len;
 768                 buf_offset += cpy_len;
 769         }
 770
 771 out:
 772
 773         return error;
 774 }
 775
 776 static __rte_noinline uint32_t
 777 virtio_dev_rx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 778         struct rte_mbuf **pkts, uint32_t count)
 779 {
 780         uint32_t pkt_idx = 0;
 781         uint16_t num_buffers;
 782         struct buf_vector buf_vec[BUF_VECTOR_MAX];
 783         uint16_t avail_head;
 784
 785         avail_head = *((volatile uint16_t *)&vq->avail->idx);
 786
 787         /*
 788          * The ordering between avail index and
 789          * desc reads needs to be enforced.
 790          */
 791         rte_smp_rmb();
 792
 793         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
 794
 795         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
 796                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
 797                 uint16_t nr_vec = 0;
 798
 799                 if (unlikely(reserve_avail_buf_split(dev, vq,
 800                                                 pkt_len, buf_vec, &num_buffers,
 801                                                 avail_head, &nr_vec) < 0)) {
 802                         VHOST_LOG_DEBUG(VHOST_DATA,
 803                                 "(%d) failed to get enough desc from vring\n",
 804                                 dev->vid);
 805                         vq->shadow_used_idx -= num_buffers;
 806                         break;
 807                 }
 808
 809                 VHOST_LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
 810                         dev->vid, vq->last_avail_idx,
 811                         vq->last_avail_idx + num_buffers);
 812
 813                 if (copy_mbuf_to_desc(dev, vq, pkts[pkt_idx],
 814                                                 buf_vec, nr_vec,
 815                                                 num_buffers) < 0) {
 816                         vq->shadow_used_idx -= num_buffers;
 817                         break;
 818                 }
 819
 820                 vq->last_avail_idx += num_buffers;
 821         }
 822
 823         do_data_copy_enqueue(dev, vq);
 824
 825         if (likely(vq->shadow_used_idx)) {
 826                 flush_shadow_used_ring_split(dev, vq);
 827                 vhost_vring_call_split(dev, vq);
 828         }
 829
 830         return pkt_idx;
 831 }
 832
 833 static __rte_noinline uint32_t
 834 virtio_dev_rx_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
 835         struct rte_mbuf **pkts, uint32_t count)
 836 {
 837         uint32_t pkt_idx = 0;
 838         uint16_t num_buffers;
 839         struct buf_vector buf_vec[BUF_VECTOR_MAX];
 840
 841         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
 842                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
 843                 uint16_t nr_vec = 0;
 844                 uint16_t nr_descs = 0;
 845
 846                 if (unlikely(reserve_avail_buf_packed(dev, vq,
 847                                                 pkt_len, buf_vec, &nr_vec,
 848                                                 &num_buffers, &nr_descs) < 0)) {
 849                         VHOST_LOG_DEBUG(VHOST_DATA,
 850                                 "(%d) failed to get enough desc from vring\n",
 851                                 dev->vid);
 852                         vq->shadow_used_idx -= num_buffers;
 853                         break;
 854                 }
 855
 856                 VHOST_LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
 857                         dev->vid, vq->last_avail_idx,
 858                         vq->last_avail_idx + num_buffers);
 859
 860                 if (copy_mbuf_to_desc(dev, vq, pkts[pkt_idx],
 861                                                 buf_vec, nr_vec,
 862                                                 num_buffers) < 0) {
 863                         vq->shadow_used_idx -= num_buffers;
 864                         break;
 865                 }
 866
 867                 vq->last_avail_idx += nr_descs;
 868                 if (vq->last_avail_idx >= vq->size) {
 869                         vq->last_avail_idx -= vq->size;
 870                         vq->avail_wrap_counter ^= 1;
 871                 }
 872         }
 873
 874         do_data_copy_enqueue(dev, vq);
 875
 876         if (likely(vq->shadow_used_idx)) {
 877                 flush_shadow_used_ring_packed(dev, vq);
 878                 vhost_vring_call_packed(dev, vq);
 879         }
 880
 881         return pkt_idx;
 882 }
 883
 884 static __rte_always_inline uint32_t
 885 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
 886         struct rte_mbuf **pkts, uint32_t count)
 887 {
 888         struct vhost_virtqueue *vq;
 889         uint32_t nb_tx = 0;
 890
 891         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
 892         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
 893                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
 894                         dev->vid, __func__, queue_id);
 895                 return 0;
 896         }
 897
 898         vq = dev->virtqueue[queue_id];
 899
 900         rte_spinlock_lock(&vq->access_lock);
 901
 902         if (unlikely(vq->enabled == 0))
 903                 goto out_access_unlock;
 904
 905         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 906                 vhost_user_iotlb_rd_lock(vq);
 907
 908         if (unlikely(vq->access_ok == 0))
 909                 if (unlikely(vring_translate(dev, vq) < 0))
 910                         goto out;
 911
 912         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
 913         if (count == 0)
 914                 goto out;
 915
 916         if (vq_is_packed(dev))
 917                 nb_tx = virtio_dev_rx_packed(dev, vq, pkts, count);
 918         else
 919                 nb_tx = virtio_dev_rx_split(dev, vq, pkts, count);
 920
 921 out:
 922         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 923                 vhost_user_iotlb_rd_unlock(vq);
 924
 925 out_access_unlock:
 926         rte_spinlock_unlock(&vq->access_lock);
 927
 928         return nb_tx;
 929 }
 930
 931 uint16_t
 932 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
 933         struct rte_mbuf **pkts, uint16_t count)
 934 {
 935         struct virtio_net *dev = get_device(vid);
 936
 937         if (!dev)
 938                 return 0;
 939
 940         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
 941                 RTE_LOG(ERR, VHOST_DATA,
 942                         "(%d) %s: built-in vhost net backend is disabled.\n",
 943                         dev->vid, __func__);
 944                 return 0;
 945         }
 946
 947         return virtio_dev_rx(dev, queue_id, pkts, count);
 948 }
 949
 950 static inline bool
 951 virtio_net_with_host_offload(struct virtio_net *dev)
 952 {
 953         if (dev->features &
 954                         ((1ULL << VIRTIO_NET_F_CSUM) |
 955                          (1ULL << VIRTIO_NET_F_HOST_ECN) |
 956                          (1ULL << VIRTIO_NET_F_HOST_TSO4) |
 957                          (1ULL << VIRTIO_NET_F_HOST_TSO6) |
 958                          (1ULL << VIRTIO_NET_F_HOST_UFO)))
 959                 return true;
 960
 961         return false;
 962 }
 963
 964 static void
 965 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
 966 {
 967         struct rte_ipv4_hdr *ipv4_hdr;
 968         struct rte_ipv6_hdr *ipv6_hdr;
 969         void *l3_hdr = NULL;
 970         struct rte_ether_hdr *eth_hdr;
 971         uint16_t ethertype;
 972
 973         eth_hdr = rte_pktmbuf_mtod(m, struct rte_ether_hdr *);
 974
 975         m->l2_len = sizeof(struct rte_ether_hdr);
 976         ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
 977
 978         if (ethertype == RTE_ETHER_TYPE_VLAN) {
 979                 struct rte_vlan_hdr *vlan_hdr =
 980                         (struct rte_vlan_hdr *)(eth_hdr + 1);
 981
 982                 m->l2_len += sizeof(struct rte_vlan_hdr);
 983                 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
 984         }
 985
 986         l3_hdr = (char *)eth_hdr + m->l2_len;
 987
 988         switch (ethertype) {
 989         case RTE_ETHER_TYPE_IPV4:
 990                 ipv4_hdr = l3_hdr;
 991                 *l4_proto = ipv4_hdr->next_proto_id;
 992                 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
 993                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 994                 m->ol_flags |= PKT_TX_IPV4;
 995                 break;
 996         case RTE_ETHER_TYPE_IPV6:
 997                 ipv6_hdr = l3_hdr;
 998                 *l4_proto = ipv6_hdr->proto;
 999                 m->l3_len = sizeof(struct rte_ipv6_hdr);
1000                 *l4_hdr = (char *)l3_hdr + m->l3_len;
1001                 m->ol_flags |= PKT_TX_IPV6;
1002                 break;
1003         default:
1004                 m->l3_len = 0;
1005                 *l4_proto = 0;
1006                 *l4_hdr = NULL;
1007                 break;
1008         }
1009 }
1010
1011 static __rte_always_inline void
1012 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
1013 {
1014         uint16_t l4_proto = 0;
1015         void *l4_hdr = NULL;
1016         struct rte_tcp_hdr *tcp_hdr = NULL;
1017
1018         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
1019                 return;
1020
1021         parse_ethernet(m, &l4_proto, &l4_hdr);
1022         if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
1023                 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
1024                         switch (hdr->csum_offset) {
1025                         case (offsetof(struct rte_tcp_hdr, cksum)):
1026                                 if (l4_proto == IPPROTO_TCP)
1027                                         m->ol_flags |= PKT_TX_TCP_CKSUM;
1028                                 break;
1029                         case (offsetof(struct rte_udp_hdr, dgram_cksum)):
1030                                 if (l4_proto == IPPROTO_UDP)
1031                                         m->ol_flags |= PKT_TX_UDP_CKSUM;
1032                                 break;
1033                         case (offsetof(struct rte_sctp_hdr, cksum)):
1034                                 if (l4_proto == IPPROTO_SCTP)
1035                                         m->ol_flags |= PKT_TX_SCTP_CKSUM;
1036                                 break;
1037                         default:
1038                                 break;
1039                         }
1040                 }
1041         }
1042
1043         if (l4_hdr && hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
1044                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
1045                 case VIRTIO_NET_HDR_GSO_TCPV4:
1046                 case VIRTIO_NET_HDR_GSO_TCPV6:
1047                         tcp_hdr = l4_hdr;
1048                         m->ol_flags |= PKT_TX_TCP_SEG;
1049                         m->tso_segsz = hdr->gso_size;
1050                         m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
1051                         break;
1052                 case VIRTIO_NET_HDR_GSO_UDP:
1053                         m->ol_flags |= PKT_TX_UDP_SEG;
1054                         m->tso_segsz = hdr->gso_size;
1055                         m->l4_len = sizeof(struct rte_udp_hdr);
1056                         break;
1057                 default:
1058                         RTE_LOG(WARNING, VHOST_DATA,
1059                                 "unsupported gso type %u.\n", hdr->gso_type);
1060                         break;
1061                 }
1062         }
1063 }
1064
1065 static __rte_noinline void
1066 copy_vnet_hdr_from_desc(struct virtio_net_hdr *hdr,
1067                 struct buf_vector *buf_vec)
1068 {
1069         uint64_t len;
1070         uint64_t remain = sizeof(struct virtio_net_hdr);
1071         uint64_t src;
1072         uint64_t dst = (uint64_t)(uintptr_t)hdr;
1073
1074         while (remain) {
1075                 len = RTE_MIN(remain, buf_vec->buf_len);
1076                 src = buf_vec->buf_addr;
1077                 rte_memcpy((void *)(uintptr_t)dst,
1078                                 (void *)(uintptr_t)src, len);
1079
1080                 remain -= len;
1081                 dst += len;
1082                 buf_vec++;
1083         }
1084 }
1085
1086 static __rte_always_inline int
1087 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
1088                   struct buf_vector *buf_vec, uint16_t nr_vec,
1089                   struct rte_mbuf *m, struct rte_mempool *mbuf_pool)
1090 {
1091         uint32_t buf_avail, buf_offset;
1092         uint64_t buf_addr, buf_iova, buf_len;
1093         uint32_t mbuf_avail, mbuf_offset;
1094         uint32_t cpy_len;
1095         struct rte_mbuf *cur = m, *prev = m;
1096         struct virtio_net_hdr tmp_hdr;
1097         struct virtio_net_hdr *hdr = NULL;
1098         /* A counter to avoid desc dead loop chain */
1099         uint16_t vec_idx = 0;
1100         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
1101         int error = 0;
1102
1103         buf_addr = buf_vec[vec_idx].buf_addr;
1104         buf_iova = buf_vec[vec_idx].buf_iova;
1105         buf_len = buf_vec[vec_idx].buf_len;
1106
1107         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
1108                 error = -1;
1109                 goto out;
1110         }
1111
1112         if (virtio_net_with_host_offload(dev)) {
1113                 if (unlikely(buf_len < sizeof(struct virtio_net_hdr))) {
1114                         /*
1115                          * No luck, the virtio-net header doesn't fit
1116                          * in a contiguous virtual area.
1117                          */
1118                         copy_vnet_hdr_from_desc(&tmp_hdr, buf_vec);
1119                         hdr = &tmp_hdr;
1120                 } else {
1121                         hdr = (struct virtio_net_hdr *)((uintptr_t)buf_addr);
1122                 }
1123         }
1124
1125         /*
1126          * A virtio driver normally uses at least 2 desc buffers
1127          * for Tx: the first for storing the header, and others
1128          * for storing the data.
1129          */
1130         if (unlikely(buf_len < dev->vhost_hlen)) {
1131                 buf_offset = dev->vhost_hlen - buf_len;
1132                 vec_idx++;
1133                 buf_addr = buf_vec[vec_idx].buf_addr;
1134                 buf_iova = buf_vec[vec_idx].buf_iova;
1135                 buf_len = buf_vec[vec_idx].buf_len;
1136                 buf_avail  = buf_len - buf_offset;
1137         } else if (buf_len == dev->vhost_hlen) {
1138                 if (unlikely(++vec_idx >= nr_vec))
1139                         goto out;
1140                 buf_addr = buf_vec[vec_idx].buf_addr;
1141                 buf_iova = buf_vec[vec_idx].buf_iova;
1142                 buf_len = buf_vec[vec_idx].buf_len;
1143
1144                 buf_offset = 0;
1145                 buf_avail = buf_len;
1146         } else {
1147                 buf_offset = dev->vhost_hlen;
1148                 buf_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
1149         }
1150
1151         PRINT_PACKET(dev,
1152                         (uintptr_t)(buf_addr + buf_offset),
1153                         (uint32_t)buf_avail, 0);
1154
1155         mbuf_offset = 0;
1156         mbuf_avail  = m->buf_len - RTE_PKTMBUF_HEADROOM;
1157         while (1) {
1158                 uint64_t hpa;
1159
1160                 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
1161
1162                 /*
1163                  * A desc buf might across two host physical pages that are
1164                  * not continuous. In such case (gpa_to_hpa returns 0), data
1165                  * will be copied even though zero copy is enabled.
1166                  */
1167                 if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
1168                                         buf_iova + buf_offset, cpy_len)))) {
1169                         cur->data_len = cpy_len;
1170                         cur->data_off = 0;
1171                         cur->buf_addr =
1172                                 (void *)(uintptr_t)(buf_addr + buf_offset);
1173                         cur->buf_iova = hpa;
1174
1175                         /*
1176                          * In zero copy mode, one mbuf can only reference data
1177                          * for one or partial of one desc buff.
1178                          */
1179                         mbuf_avail = cpy_len;
1180                 } else {
1181                         if (likely(cpy_len > MAX_BATCH_LEN ||
1182                                    vq->batch_copy_nb_elems >= vq->size ||
1183                                    (hdr && cur == m))) {
1184                                 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
1185                                                                    mbuf_offset),
1186                                            (void *)((uintptr_t)(buf_addr +
1187                                                            buf_offset)),
1188                                            cpy_len);
1189                         } else {
1190                                 batch_copy[vq->batch_copy_nb_elems].dst =
1191                                         rte_pktmbuf_mtod_offset(cur, void *,
1192                                                                 mbuf_offset);
1193                                 batch_copy[vq->batch_copy_nb_elems].src =
1194                                         (void *)((uintptr_t)(buf_addr +
1195                                                                 buf_offset));
1196                                 batch_copy[vq->batch_copy_nb_elems].len =
1197                                         cpy_len;
1198                                 vq->batch_copy_nb_elems++;
1199                         }
1200                 }
1201
1202                 mbuf_avail  -= cpy_len;
1203                 mbuf_offset += cpy_len;
1204                 buf_avail -= cpy_len;
1205                 buf_offset += cpy_len;
1206
1207                 /* This buf reaches to its end, get the next one */
1208                 if (buf_avail == 0) {
1209                         if (++vec_idx >= nr_vec)
1210                                 break;
1211
1212                         buf_addr = buf_vec[vec_idx].buf_addr;
1213                         buf_iova = buf_vec[vec_idx].buf_iova;
1214                         buf_len = buf_vec[vec_idx].buf_len;
1215
1216                         buf_offset = 0;
1217                         buf_avail  = buf_len;
1218
1219                         PRINT_PACKET(dev, (uintptr_t)buf_addr,
1220                                         (uint32_t)buf_avail, 0);
1221                 }
1222
1223                 /*
1224                  * This mbuf reaches to its end, get a new one
1225                  * to hold more data.
1226                  */
1227                 if (mbuf_avail == 0) {
1228                         cur = rte_pktmbuf_alloc(mbuf_pool);
1229                         if (unlikely(cur == NULL)) {
1230                                 RTE_LOG(ERR, VHOST_DATA, "Failed to "
1231                                         "allocate memory for mbuf.\n");
1232                                 error = -1;
1233                                 goto out;
1234                         }
1235                         if (unlikely(dev->dequeue_zero_copy))
1236                                 rte_mbuf_refcnt_update(cur, 1);
1237
1238                         prev->next = cur;
1239                         prev->data_len = mbuf_offset;
1240                         m->nb_segs += 1;
1241                         m->pkt_len += mbuf_offset;
1242                         prev = cur;
1243
1244                         mbuf_offset = 0;
1245                         mbuf_avail  = cur->buf_len - RTE_PKTMBUF_HEADROOM;
1246                 }
1247         }
1248
1249         prev->data_len = mbuf_offset;
1250         m->pkt_len    += mbuf_offset;
1251
1252         if (hdr)
1253                 vhost_dequeue_offload(hdr, m);
1254
1255 out:
1256
1257         return error;
1258 }
1259
1260 static __rte_always_inline struct zcopy_mbuf *
1261 get_zmbuf(struct vhost_virtqueue *vq)
1262 {
1263         uint16_t i;
1264         uint16_t last;
1265         int tries = 0;
1266
1267         /* search [last_zmbuf_idx, zmbuf_size) */
1268         i = vq->last_zmbuf_idx;
1269         last = vq->zmbuf_size;
1270
1271 again:
1272         for (; i < last; i++) {
1273                 if (vq->zmbufs[i].in_use == 0) {
1274                         vq->last_zmbuf_idx = i + 1;
1275                         vq->zmbufs[i].in_use = 1;
1276                         return &vq->zmbufs[i];
1277                 }
1278         }
1279
1280         tries++;
1281         if (tries == 1) {
1282                 /* search [0, last_zmbuf_idx) */
1283                 i = 0;
1284                 last = vq->last_zmbuf_idx;
1285                 goto again;
1286         }
1287
1288         return NULL;
1289 }
1290
1291 static __rte_noinline uint16_t
1292 virtio_dev_tx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
1293         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1294 {
1295         uint16_t i;
1296         uint16_t free_entries;
1297
1298         if (unlikely(dev->dequeue_zero_copy)) {
1299                 struct zcopy_mbuf *zmbuf, *next;
1300
1301                 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1302                      zmbuf != NULL; zmbuf = next) {
1303                         next = TAILQ_NEXT(zmbuf, next);
1304
1305                         if (mbuf_is_consumed(zmbuf->mbuf)) {
1306                                 update_shadow_used_ring_split(vq,
1307                                                 zmbuf->desc_idx, 0);
1308                                 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1309                                 restore_mbuf(zmbuf->mbuf);
1310                                 rte_pktmbuf_free(zmbuf->mbuf);
1311                                 put_zmbuf(zmbuf);
1312                                 vq->nr_zmbuf -= 1;
1313                         }
1314                 }
1315
1316                 if (likely(vq->shadow_used_idx)) {
1317                         flush_shadow_used_ring_split(dev, vq);
1318                         vhost_vring_call_split(dev, vq);
1319                 }
1320         }
1321
1322         free_entries = *((volatile uint16_t *)&vq->avail->idx) -
1323                         vq->last_avail_idx;
1324         if (free_entries == 0)
1325                 return 0;
1326
1327         /*
1328          * The ordering between avail index and
1329          * desc reads needs to be enforced.
1330          */
1331         rte_smp_rmb();
1332
1333         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
1334
1335         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1336
1337         count = RTE_MIN(count, MAX_PKT_BURST);
1338         count = RTE_MIN(count, free_entries);
1339         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1340                         dev->vid, count);
1341
1342         for (i = 0; i < count; i++) {
1343                 struct buf_vector buf_vec[BUF_VECTOR_MAX];
1344                 uint16_t head_idx;
1345                 uint32_t dummy_len;
1346                 uint16_t nr_vec = 0;
1347                 int err;
1348
1349                 if (unlikely(fill_vec_buf_split(dev, vq,
1350                                                 vq->last_avail_idx + i,
1351                                                 &nr_vec, buf_vec,
1352                                                 &head_idx, &dummy_len,
1353                                                 VHOST_ACCESS_RO) < 0))
1354                         break;
1355
1356                 if (likely(dev->dequeue_zero_copy == 0))
1357                         update_shadow_used_ring_split(vq, head_idx, 0);
1358
1359                 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
1360                 if (unlikely(pkts[i] == NULL)) {
1361                         RTE_LOG(ERR, VHOST_DATA,
1362                                 "Failed to allocate memory for mbuf.\n");
1363                         break;
1364                 }
1365
1366                 err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts[i],
1367                                 mbuf_pool);
1368                 if (unlikely(err)) {
1369                         rte_pktmbuf_free(pkts[i]);
1370                         break;
1371                 }
1372
1373                 if (unlikely(dev->dequeue_zero_copy)) {
1374                         struct zcopy_mbuf *zmbuf;
1375
1376                         zmbuf = get_zmbuf(vq);
1377                         if (!zmbuf) {
1378                                 rte_pktmbuf_free(pkts[i]);
1379                                 break;
1380                         }
1381                         zmbuf->mbuf = pkts[i];
1382                         zmbuf->desc_idx = head_idx;
1383
1384                         /*
1385                          * Pin lock the mbuf; we will check later to see
1386                          * whether the mbuf is freed (when we are the last
1387                          * user) or not. If that's the case, we then could
1388                          * update the used ring safely.
1389                          */
1390                         rte_mbuf_refcnt_update(pkts[i], 1);
1391
1392                         vq->nr_zmbuf += 1;
1393                         TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1394                 }
1395         }
1396         vq->last_avail_idx += i;
1397
1398         if (likely(dev->dequeue_zero_copy == 0)) {
1399                 do_data_copy_dequeue(vq);
1400                 if (unlikely(i < count))
1401                         vq->shadow_used_idx = i;
1402                 if (likely(vq->shadow_used_idx)) {
1403                         flush_shadow_used_ring_split(dev, vq);
1404                         vhost_vring_call_split(dev, vq);
1405                 }
1406         }
1407
1408         return i;
1409 }
1410
1411 static __rte_noinline uint16_t
1412 virtio_dev_tx_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
1413         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1414 {
1415         uint16_t i;
1416
1417         if (unlikely(dev->dequeue_zero_copy)) {
1418                 struct zcopy_mbuf *zmbuf, *next;
1419
1420                 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1421                      zmbuf != NULL; zmbuf = next) {
1422                         next = TAILQ_NEXT(zmbuf, next);
1423
1424                         if (mbuf_is_consumed(zmbuf->mbuf)) {
1425                                 update_shadow_used_ring_packed(vq,
1426                                                 zmbuf->desc_idx,
1427                                                 0,
1428                                                 zmbuf->desc_count);
1429
1430                                 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1431                                 restore_mbuf(zmbuf->mbuf);
1432                                 rte_pktmbuf_free(zmbuf->mbuf);
1433                                 put_zmbuf(zmbuf);
1434                                 vq->nr_zmbuf -= 1;
1435                         }
1436                 }
1437
1438                 if (likely(vq->shadow_used_idx)) {
1439                         flush_shadow_used_ring_packed(dev, vq);
1440                         vhost_vring_call_packed(dev, vq);
1441                 }
1442         }
1443
1444         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1445
1446         count = RTE_MIN(count, MAX_PKT_BURST);
1447         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1448                         dev->vid, count);
1449
1450         for (i = 0; i < count; i++) {
1451                 struct buf_vector buf_vec[BUF_VECTOR_MAX];
1452                 uint16_t buf_id;
1453                 uint32_t dummy_len;
1454                 uint16_t desc_count, nr_vec = 0;
1455                 int err;
1456
1457                 if (unlikely(fill_vec_buf_packed(dev, vq,
1458                                                 vq->last_avail_idx, &desc_count,
1459                                                 buf_vec, &nr_vec,
1460                                                 &buf_id, &dummy_len,
1461                                                 VHOST_ACCESS_RO) < 0))
1462                         break;
1463
1464                 if (likely(dev->dequeue_zero_copy == 0))
1465                         update_shadow_used_ring_packed(vq, buf_id, 0,
1466                                         desc_count);
1467
1468                 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
1469                 if (unlikely(pkts[i] == NULL)) {
1470                         RTE_LOG(ERR, VHOST_DATA,
1471                                 "Failed to allocate memory for mbuf.\n");
1472                         break;
1473                 }
1474
1475                 err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts[i],
1476                                 mbuf_pool);
1477                 if (unlikely(err)) {
1478                         rte_pktmbuf_free(pkts[i]);
1479                         break;
1480                 }
1481
1482                 if (unlikely(dev->dequeue_zero_copy)) {
1483                         struct zcopy_mbuf *zmbuf;
1484
1485                         zmbuf = get_zmbuf(vq);
1486                         if (!zmbuf) {
1487                                 rte_pktmbuf_free(pkts[i]);
1488                                 break;
1489                         }
1490                         zmbuf->mbuf = pkts[i];
1491                         zmbuf->desc_idx = buf_id;
1492                         zmbuf->desc_count = desc_count;
1493
1494                         /*
1495                          * Pin lock the mbuf; we will check later to see
1496                          * whether the mbuf is freed (when we are the last
1497                          * user) or not. If that's the case, we then could
1498                          * update the used ring safely.
1499                          */
1500                         rte_mbuf_refcnt_update(pkts[i], 1);
1501
1502                         vq->nr_zmbuf += 1;
1503                         TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1504                 }
1505
1506                 vq->last_avail_idx += desc_count;
1507                 if (vq->last_avail_idx >= vq->size) {
1508                         vq->last_avail_idx -= vq->size;
1509                         vq->avail_wrap_counter ^= 1;
1510                 }
1511         }
1512
1513         if (likely(dev->dequeue_zero_copy == 0)) {
1514                 do_data_copy_dequeue(vq);
1515                 if (unlikely(i < count))
1516                         vq->shadow_used_idx = i;
1517                 if (likely(vq->shadow_used_idx)) {
1518                         flush_shadow_used_ring_packed(dev, vq);
1519                         vhost_vring_call_packed(dev, vq);
1520                 }
1521         }
1522
1523         return i;
1524 }
1525
1526 uint16_t
1527 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
1528         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1529 {
1530         struct virtio_net *dev;
1531         struct rte_mbuf *rarp_mbuf = NULL;
1532         struct vhost_virtqueue *vq;
1533
1534         dev = get_device(vid);
1535         if (!dev)
1536                 return 0;
1537
1538         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
1539                 RTE_LOG(ERR, VHOST_DATA,
1540                         "(%d) %s: built-in vhost net backend is disabled.\n",
1541                         dev->vid, __func__);
1542                 return 0;
1543         }
1544
1545         if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
1546                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
1547                         dev->vid, __func__, queue_id);
1548                 return 0;
1549         }
1550
1551         vq = dev->virtqueue[queue_id];
1552
1553         if (unlikely(rte_spinlock_trylock(&vq->access_lock) == 0))
1554                 return 0;
1555
1556         if (unlikely(vq->enabled == 0)) {
1557                 count = 0;
1558                 goto out_access_unlock;
1559         }
1560
1561         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1562                 vhost_user_iotlb_rd_lock(vq);
1563
1564         if (unlikely(vq->access_ok == 0))
1565                 if (unlikely(vring_translate(dev, vq) < 0)) {
1566                         count = 0;
1567                         goto out;
1568                 }
1569
1570         /*
1571          * Construct a RARP broadcast packet, and inject it to the "pkts"
1572          * array, to looks like that guest actually send such packet.
1573          *
1574          * Check user_send_rarp() for more information.
1575          *
1576          * broadcast_rarp shares a cacheline in the virtio_net structure
1577          * with some fields that are accessed during enqueue and
1578          * rte_atomic16_cmpset() causes a write if using cmpxchg. This could
1579          * result in false sharing between enqueue and dequeue.
1580          *
1581          * Prevent unnecessary false sharing by reading broadcast_rarp first
1582          * and only performing cmpset if the read indicates it is likely to
1583          * be set.
1584          */
1585         if (unlikely(rte_atomic16_read(&dev->broadcast_rarp) &&
1586                         rte_atomic16_cmpset((volatile uint16_t *)
1587                                 &dev->broadcast_rarp.cnt, 1, 0))) {
1588
1589                 rarp_mbuf = rte_net_make_rarp_packet(mbuf_pool, &dev->mac);
1590                 if (rarp_mbuf == NULL) {
1591                         RTE_LOG(ERR, VHOST_DATA,
1592                                 "Failed to make RARP packet.\n");
1593                         count = 0;
1594                         goto out;
1595                 }
1596                 count -= 1;
1597         }
1598
1599         if (vq_is_packed(dev))
1600                 count = virtio_dev_tx_packed(dev, vq, mbuf_pool, pkts, count);
1601         else
1602                 count = virtio_dev_tx_split(dev, vq, mbuf_pool, pkts, count);
1603
1604 out:
1605         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1606                 vhost_user_iotlb_rd_unlock(vq);
1607
1608 out_access_unlock:
1609         rte_spinlock_unlock(&vq->access_lock);
1610
1611         if (unlikely(rarp_mbuf != NULL)) {
1612                 /*
1613                  * Inject it to the head of "pkts" array, so that switch's mac
1614                  * learning table will get updated first.
1615                  */
1616                 memmove(&pkts[1], pkts, count * sizeof(struct rte_mbuf *));
1617                 pkts[0] = rarp_mbuf;
1618                 count += 1;
1619         }
1620
1621         return count;
1622 }