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         if (unlikely(!desc_is_avail(&descs[avail_idx], wrap_counter)))
 507                 return -1;
 508
 509         /*
 510          * The ordering between desc flags and desc
 511          * content reads need to be enforced.
 512          */
 513         rte_smp_rmb();
 514
 515         *desc_count = 0;
 516         *len = 0;
 517
 518         while (1) {
 519                 if (unlikely(vec_id >= BUF_VECTOR_MAX))
 520                         return -1;
 521
 522                 if (unlikely(*desc_count >= vq->size))
 523                         return -1;
 524
 525                 *desc_count += 1;
 526                 *buf_id = descs[avail_idx].id;
 527
 528                 if (descs[avail_idx].flags & VRING_DESC_F_INDIRECT) {
 529                         if (unlikely(fill_vec_buf_packed_indirect(dev, vq,
 530                                                         &descs[avail_idx],
 531                                                         &vec_id, buf_vec,
 532                                                         len, perm) < 0))
 533                                 return -1;
 534                 } else {
 535                         *len += descs[avail_idx].len;
 536
 537                         if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 538                                                         descs[avail_idx].addr,
 539                                                         descs[avail_idx].len,
 540                                                         perm)))
 541                                 return -1;
 542                 }
 543
 544                 if ((descs[avail_idx].flags & VRING_DESC_F_NEXT) == 0)
 545                         break;
 546
 547                 if (++avail_idx >= vq->size) {
 548                         avail_idx -= vq->size;
 549                         wrap_counter ^= 1;
 550                 }
 551         }
 552
 553         *vec_idx = vec_id;
 554
 555         return 0;
 556 }
 557
 558 /*
 559  * Returns -1 on fail, 0 on success
 560  */
 561 static inline int
 562 reserve_avail_buf_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
 563                                 uint32_t size, struct buf_vector *buf_vec,
 564                                 uint16_t *nr_vec, uint16_t *num_buffers,
 565                                 uint16_t *nr_descs)
 566 {
 567         uint16_t avail_idx;
 568         uint16_t vec_idx = 0;
 569         uint16_t max_tries, tries = 0;
 570
 571         uint16_t buf_id = 0;
 572         uint32_t len = 0;
 573         uint16_t desc_count;
 574
 575         *num_buffers = 0;
 576         avail_idx = vq->last_avail_idx;
 577
 578         if (rxvq_is_mergeable(dev))
 579                 max_tries = vq->size - 1;
 580         else
 581                 max_tries = 1;
 582
 583         while (size > 0) {
 584                 /*
 585                  * if we tried all available ring items, and still
 586                  * can't get enough buf, it means something abnormal
 587                  * happened.
 588                  */
 589                 if (unlikely(++tries > max_tries))
 590                         return -1;
 591
 592                 if (unlikely(fill_vec_buf_packed(dev, vq,
 593                                                 avail_idx, &desc_count,
 594                                                 buf_vec, &vec_idx,
 595                                                 &buf_id, &len,
 596                                                 VHOST_ACCESS_RW) < 0))
 597                         return -1;
 598
 599                 len = RTE_MIN(len, size);
 600                 update_shadow_used_ring_packed(vq, buf_id, len, desc_count);
 601                 size -= len;
 602
 603                 avail_idx += desc_count;
 604                 if (avail_idx >= vq->size)
 605                         avail_idx -= vq->size;
 606
 607                 *nr_descs += desc_count;
 608                 *num_buffers += 1;
 609         }
 610
 611         *nr_vec = vec_idx;
 612
 613         return 0;
 614 }
 615
 616 static __rte_noinline void
 617 copy_vnet_hdr_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 618                 struct buf_vector *buf_vec,
 619                 struct virtio_net_hdr_mrg_rxbuf *hdr)
 620 {
 621         uint64_t len;
 622         uint64_t remain = dev->vhost_hlen;
 623         uint64_t src = (uint64_t)(uintptr_t)hdr, dst;
 624         uint64_t iova = buf_vec->buf_iova;
 625
 626         while (remain) {
 627                 len = RTE_MIN(remain,
 628                                 buf_vec->buf_len);
 629                 dst = buf_vec->buf_addr;
 630                 rte_memcpy((void *)(uintptr_t)dst,
 631                                 (void *)(uintptr_t)src,
 632                                 len);
 633
 634                 PRINT_PACKET(dev, (uintptr_t)dst,
 635                                 (uint32_t)len, 0);
 636                 vhost_log_cache_write(dev, vq,
 637                                 iova, len);
 638
 639                 remain -= len;
 640                 iova += len;
 641                 src += len;
 642                 buf_vec++;
 643         }
 644 }
 645
 646 static __rte_always_inline int
 647 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 648                             struct rte_mbuf *m, struct buf_vector *buf_vec,
 649                             uint16_t nr_vec, uint16_t num_buffers)
 650 {
 651         uint32_t vec_idx = 0;
 652         uint32_t mbuf_offset, mbuf_avail;
 653         uint32_t buf_offset, buf_avail;
 654         uint64_t buf_addr, buf_iova, buf_len;
 655         uint32_t cpy_len;
 656         uint64_t hdr_addr;
 657         struct rte_mbuf *hdr_mbuf;
 658         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 659         struct virtio_net_hdr_mrg_rxbuf tmp_hdr, *hdr = NULL;
 660         int error = 0;
 661
 662         if (unlikely(m == NULL)) {
 663                 error = -1;
 664                 goto out;
 665         }
 666
 667         buf_addr = buf_vec[vec_idx].buf_addr;
 668         buf_iova = buf_vec[vec_idx].buf_iova;
 669         buf_len = buf_vec[vec_idx].buf_len;
 670
 671         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
 672                 error = -1;
 673                 goto out;
 674         }
 675
 676         hdr_mbuf = m;
 677         hdr_addr = buf_addr;
 678         if (unlikely(buf_len < dev->vhost_hlen))
 679                 hdr = &tmp_hdr;
 680         else
 681                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)hdr_addr;
 682
 683         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
 684                 dev->vid, num_buffers);
 685
 686         if (unlikely(buf_len < dev->vhost_hlen)) {
 687                 buf_offset = dev->vhost_hlen - buf_len;
 688                 vec_idx++;
 689                 buf_addr = buf_vec[vec_idx].buf_addr;
 690                 buf_iova = buf_vec[vec_idx].buf_iova;
 691                 buf_len = buf_vec[vec_idx].buf_len;
 692                 buf_avail = buf_len - buf_offset;
 693         } else {
 694                 buf_offset = dev->vhost_hlen;
 695                 buf_avail = buf_len - dev->vhost_hlen;
 696         }
 697
 698         mbuf_avail  = rte_pktmbuf_data_len(m);
 699         mbuf_offset = 0;
 700         while (mbuf_avail != 0 || m->next != NULL) {
 701                 /* done with current buf, get the next one */
 702                 if (buf_avail == 0) {
 703                         vec_idx++;
 704                         if (unlikely(vec_idx >= nr_vec)) {
 705                                 error = -1;
 706                                 goto out;
 707                         }
 708
 709                         buf_addr = buf_vec[vec_idx].buf_addr;
 710                         buf_iova = buf_vec[vec_idx].buf_iova;
 711                         buf_len = buf_vec[vec_idx].buf_len;
 712
 713                         buf_offset = 0;
 714                         buf_avail  = buf_len;
 715                 }
 716
 717                 /* done with current mbuf, get the next one */
 718                 if (mbuf_avail == 0) {
 719                         m = m->next;
 720
 721                         mbuf_offset = 0;
 722                         mbuf_avail  = rte_pktmbuf_data_len(m);
 723                 }
 724
 725                 if (hdr_addr) {
 726                         virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
 727                         if (rxvq_is_mergeable(dev))
 728                                 ASSIGN_UNLESS_EQUAL(hdr->num_buffers,
 729                                                 num_buffers);
 730
 731                         if (unlikely(hdr == &tmp_hdr)) {
 732                                 copy_vnet_hdr_to_desc(dev, vq, buf_vec, hdr);
 733                         } else {
 734                                 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
 735                                                 dev->vhost_hlen, 0);
 736                                 vhost_log_cache_write(dev, vq,
 737                                                 buf_vec[0].buf_iova,
 738                                                 dev->vhost_hlen);
 739                         }
 740
 741                         hdr_addr = 0;
 742                 }
 743
 744                 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
 745
 746                 if (likely(cpy_len > MAX_BATCH_LEN ||
 747                                         vq->batch_copy_nb_elems >= vq->size)) {
 748                         rte_memcpy((void *)((uintptr_t)(buf_addr + buf_offset)),
 749                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 750                                 cpy_len);
 751                         vhost_log_cache_write(dev, vq, buf_iova + buf_offset,
 752                                         cpy_len);
 753                         PRINT_PACKET(dev, (uintptr_t)(buf_addr + buf_offset),
 754                                 cpy_len, 0);
 755                 } else {
 756                         batch_copy[vq->batch_copy_nb_elems].dst =
 757                                 (void *)((uintptr_t)(buf_addr + buf_offset));
 758                         batch_copy[vq->batch_copy_nb_elems].src =
 759                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
 760                         batch_copy[vq->batch_copy_nb_elems].log_addr =
 761                                 buf_iova + buf_offset;
 762                         batch_copy[vq->batch_copy_nb_elems].len = cpy_len;
 763                         vq->batch_copy_nb_elems++;
 764                 }
 765
 766                 mbuf_avail  -= cpy_len;
 767                 mbuf_offset += cpy_len;
 768                 buf_avail  -= cpy_len;
 769                 buf_offset += cpy_len;
 770         }
 771
 772 out:
 773
 774         return error;
 775 }
 776
 777 static __rte_noinline uint32_t
 778 virtio_dev_rx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 779         struct rte_mbuf **pkts, uint32_t count)
 780 {
 781         uint32_t pkt_idx = 0;
 782         uint16_t num_buffers;
 783         struct buf_vector buf_vec[BUF_VECTOR_MAX];
 784         uint16_t avail_head;
 785
 786         avail_head = *((volatile uint16_t *)&vq->avail->idx);
 787
 788         /*
 789          * The ordering between avail index and
 790          * desc reads needs to be enforced.
 791          */
 792         rte_smp_rmb();
 793
 794         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
 795
 796         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
 797                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
 798                 uint16_t nr_vec = 0;
 799
 800                 if (unlikely(reserve_avail_buf_split(dev, vq,
 801                                                 pkt_len, buf_vec, &num_buffers,
 802                                                 avail_head, &nr_vec) < 0)) {
 803                         VHOST_LOG_DEBUG(VHOST_DATA,
 804                                 "(%d) failed to get enough desc from vring\n",
 805                                 dev->vid);
 806                         vq->shadow_used_idx -= num_buffers;
 807                         break;
 808                 }
 809
 810                 VHOST_LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
 811                         dev->vid, vq->last_avail_idx,
 812                         vq->last_avail_idx + num_buffers);
 813
 814                 if (copy_mbuf_to_desc(dev, vq, pkts[pkt_idx],
 815                                                 buf_vec, nr_vec,
 816                                                 num_buffers) < 0) {
 817                         vq->shadow_used_idx -= num_buffers;
 818                         break;
 819                 }
 820
 821                 vq->last_avail_idx += num_buffers;
 822         }
 823
 824         do_data_copy_enqueue(dev, vq);
 825
 826         if (likely(vq->shadow_used_idx)) {
 827                 flush_shadow_used_ring_split(dev, vq);
 828                 vhost_vring_call_split(dev, vq);
 829         }
 830
 831         return pkt_idx;
 832 }
 833
 834 static __rte_noinline uint32_t
 835 virtio_dev_rx_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
 836         struct rte_mbuf **pkts, uint32_t count)
 837 {
 838         uint32_t pkt_idx = 0;
 839         uint16_t num_buffers;
 840         struct buf_vector buf_vec[BUF_VECTOR_MAX];
 841
 842         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
 843                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
 844                 uint16_t nr_vec = 0;
 845                 uint16_t nr_descs = 0;
 846
 847                 if (unlikely(reserve_avail_buf_packed(dev, vq,
 848                                                 pkt_len, buf_vec, &nr_vec,
 849                                                 &num_buffers, &nr_descs) < 0)) {
 850                         VHOST_LOG_DEBUG(VHOST_DATA,
 851                                 "(%d) failed to get enough desc from vring\n",
 852                                 dev->vid);
 853                         vq->shadow_used_idx -= num_buffers;
 854                         break;
 855                 }
 856
 857                 VHOST_LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
 858                         dev->vid, vq->last_avail_idx,
 859                         vq->last_avail_idx + num_buffers);
 860
 861                 if (copy_mbuf_to_desc(dev, vq, pkts[pkt_idx],
 862                                                 buf_vec, nr_vec,
 863                                                 num_buffers) < 0) {
 864                         vq->shadow_used_idx -= num_buffers;
 865                         break;
 866                 }
 867
 868                 vq->last_avail_idx += nr_descs;
 869                 if (vq->last_avail_idx >= vq->size) {
 870                         vq->last_avail_idx -= vq->size;
 871                         vq->avail_wrap_counter ^= 1;
 872                 }
 873         }
 874
 875         do_data_copy_enqueue(dev, vq);
 876
 877         if (likely(vq->shadow_used_idx)) {
 878                 flush_shadow_used_ring_packed(dev, vq);
 879                 vhost_vring_call_packed(dev, vq);
 880         }
 881
 882         return pkt_idx;
 883 }
 884
 885 static __rte_always_inline uint32_t
 886 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
 887         struct rte_mbuf **pkts, uint32_t count)
 888 {
 889         struct vhost_virtqueue *vq;
 890         uint32_t nb_tx = 0;
 891
 892         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
 893         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
 894                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
 895                         dev->vid, __func__, queue_id);
 896                 return 0;
 897         }
 898
 899         vq = dev->virtqueue[queue_id];
 900
 901         rte_spinlock_lock(&vq->access_lock);
 902
 903         if (unlikely(vq->enabled == 0))
 904                 goto out_access_unlock;
 905
 906         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 907                 vhost_user_iotlb_rd_lock(vq);
 908
 909         if (unlikely(vq->access_ok == 0))
 910                 if (unlikely(vring_translate(dev, vq) < 0))
 911                         goto out;
 912
 913         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
 914         if (count == 0)
 915                 goto out;
 916
 917         if (vq_is_packed(dev))
 918                 nb_tx = virtio_dev_rx_packed(dev, vq, pkts, count);
 919         else
 920                 nb_tx = virtio_dev_rx_split(dev, vq, pkts, count);
 921
 922 out:
 923         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 924                 vhost_user_iotlb_rd_unlock(vq);
 925
 926 out_access_unlock:
 927         rte_spinlock_unlock(&vq->access_lock);
 928
 929         return nb_tx;
 930 }
 931
 932 uint16_t
 933 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
 934         struct rte_mbuf **pkts, uint16_t count)
 935 {
 936         struct virtio_net *dev = get_device(vid);
 937
 938         if (!dev)
 939                 return 0;
 940
 941         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
 942                 RTE_LOG(ERR, VHOST_DATA,
 943                         "(%d) %s: built-in vhost net backend is disabled.\n",
 944                         dev->vid, __func__);
 945                 return 0;
 946         }
 947
 948         return virtio_dev_rx(dev, queue_id, pkts, count);
 949 }
 950
 951 static inline bool
 952 virtio_net_with_host_offload(struct virtio_net *dev)
 953 {
 954         if (dev->features &
 955                         ((1ULL << VIRTIO_NET_F_CSUM) |
 956                          (1ULL << VIRTIO_NET_F_HOST_ECN) |
 957                          (1ULL << VIRTIO_NET_F_HOST_TSO4) |
 958                          (1ULL << VIRTIO_NET_F_HOST_TSO6) |
 959                          (1ULL << VIRTIO_NET_F_HOST_UFO)))
 960                 return true;
 961
 962         return false;
 963 }
 964
 965 static void
 966 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
 967 {
 968         struct rte_ipv4_hdr *ipv4_hdr;
 969         struct rte_ipv6_hdr *ipv6_hdr;
 970         void *l3_hdr = NULL;
 971         struct rte_ether_hdr *eth_hdr;
 972         uint16_t ethertype;
 973
 974         eth_hdr = rte_pktmbuf_mtod(m, struct rte_ether_hdr *);
 975
 976         m->l2_len = sizeof(struct rte_ether_hdr);
 977         ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
 978
 979         if (ethertype == RTE_ETHER_TYPE_VLAN) {
 980                 struct rte_vlan_hdr *vlan_hdr =
 981                         (struct rte_vlan_hdr *)(eth_hdr + 1);
 982
 983                 m->l2_len += sizeof(struct rte_vlan_hdr);
 984                 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
 985         }
 986
 987         l3_hdr = (char *)eth_hdr + m->l2_len;
 988
 989         switch (ethertype) {
 990         case RTE_ETHER_TYPE_IPV4:
 991                 ipv4_hdr = l3_hdr;
 992                 *l4_proto = ipv4_hdr->next_proto_id;
 993                 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
 994                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 995                 m->ol_flags |= PKT_TX_IPV4;
 996                 break;
 997         case RTE_ETHER_TYPE_IPV6:
 998                 ipv6_hdr = l3_hdr;
 999                 *l4_proto = ipv6_hdr->proto;
1000                 m->l3_len = sizeof(struct rte_ipv6_hdr);
1001                 *l4_hdr = (char *)l3_hdr + m->l3_len;
1002                 m->ol_flags |= PKT_TX_IPV6;
1003                 break;
1004         default:
1005                 m->l3_len = 0;
1006                 *l4_proto = 0;
1007                 *l4_hdr = NULL;
1008                 break;
1009         }
1010 }
1011
1012 static __rte_always_inline void
1013 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
1014 {
1015         uint16_t l4_proto = 0;
1016         void *l4_hdr = NULL;
1017         struct rte_tcp_hdr *tcp_hdr = NULL;
1018
1019         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
1020                 return;
1021
1022         parse_ethernet(m, &l4_proto, &l4_hdr);
1023         if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
1024                 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
1025                         switch (hdr->csum_offset) {
1026                         case (offsetof(struct rte_tcp_hdr, cksum)):
1027                                 if (l4_proto == IPPROTO_TCP)
1028                                         m->ol_flags |= PKT_TX_TCP_CKSUM;
1029                                 break;
1030                         case (offsetof(struct rte_udp_hdr, dgram_cksum)):
1031                                 if (l4_proto == IPPROTO_UDP)
1032                                         m->ol_flags |= PKT_TX_UDP_CKSUM;
1033                                 break;
1034                         case (offsetof(struct rte_sctp_hdr, cksum)):
1035                                 if (l4_proto == IPPROTO_SCTP)
1036                                         m->ol_flags |= PKT_TX_SCTP_CKSUM;
1037                                 break;
1038                         default:
1039                                 break;
1040                         }
1041                 }
1042         }
1043
1044         if (l4_hdr && hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
1045                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
1046                 case VIRTIO_NET_HDR_GSO_TCPV4:
1047                 case VIRTIO_NET_HDR_GSO_TCPV6:
1048                         tcp_hdr = l4_hdr;
1049                         m->ol_flags |= PKT_TX_TCP_SEG;
1050                         m->tso_segsz = hdr->gso_size;
1051                         m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
1052                         break;
1053                 case VIRTIO_NET_HDR_GSO_UDP:
1054                         m->ol_flags |= PKT_TX_UDP_SEG;
1055                         m->tso_segsz = hdr->gso_size;
1056                         m->l4_len = sizeof(struct rte_udp_hdr);
1057                         break;
1058                 default:
1059                         RTE_LOG(WARNING, VHOST_DATA,
1060                                 "unsupported gso type %u.\n", hdr->gso_type);
1061                         break;
1062                 }
1063         }
1064 }
1065
1066 static __rte_noinline void
1067 copy_vnet_hdr_from_desc(struct virtio_net_hdr *hdr,
1068                 struct buf_vector *buf_vec)
1069 {
1070         uint64_t len;
1071         uint64_t remain = sizeof(struct virtio_net_hdr);
1072         uint64_t src;
1073         uint64_t dst = (uint64_t)(uintptr_t)hdr;
1074
1075         while (remain) {
1076                 len = RTE_MIN(remain, buf_vec->buf_len);
1077                 src = buf_vec->buf_addr;
1078                 rte_memcpy((void *)(uintptr_t)dst,
1079                                 (void *)(uintptr_t)src, len);
1080
1081                 remain -= len;
1082                 dst += len;
1083                 buf_vec++;
1084         }
1085 }
1086
1087 static __rte_always_inline int
1088 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
1089                   struct buf_vector *buf_vec, uint16_t nr_vec,
1090                   struct rte_mbuf *m, struct rte_mempool *mbuf_pool)
1091 {
1092         uint32_t buf_avail, buf_offset;
1093         uint64_t buf_addr, buf_iova, buf_len;
1094         uint32_t mbuf_avail, mbuf_offset;
1095         uint32_t cpy_len;
1096         struct rte_mbuf *cur = m, *prev = m;
1097         struct virtio_net_hdr tmp_hdr;
1098         struct virtio_net_hdr *hdr = NULL;
1099         /* A counter to avoid desc dead loop chain */
1100         uint16_t vec_idx = 0;
1101         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
1102         int error = 0;
1103
1104         buf_addr = buf_vec[vec_idx].buf_addr;
1105         buf_iova = buf_vec[vec_idx].buf_iova;
1106         buf_len = buf_vec[vec_idx].buf_len;
1107
1108         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
1109                 error = -1;
1110                 goto out;
1111         }
1112
1113         if (virtio_net_with_host_offload(dev)) {
1114                 if (unlikely(buf_len < sizeof(struct virtio_net_hdr))) {
1115                         /*
1116                          * No luck, the virtio-net header doesn't fit
1117                          * in a contiguous virtual area.
1118                          */
1119                         copy_vnet_hdr_from_desc(&tmp_hdr, buf_vec);
1120                         hdr = &tmp_hdr;
1121                 } else {
1122                         hdr = (struct virtio_net_hdr *)((uintptr_t)buf_addr);
1123                 }
1124         }
1125
1126         /*
1127          * A virtio driver normally uses at least 2 desc buffers
1128          * for Tx: the first for storing the header, and others
1129          * for storing the data.
1130          */
1131         if (unlikely(buf_len < dev->vhost_hlen)) {
1132                 buf_offset = dev->vhost_hlen - buf_len;
1133                 vec_idx++;
1134                 buf_addr = buf_vec[vec_idx].buf_addr;
1135                 buf_iova = buf_vec[vec_idx].buf_iova;
1136                 buf_len = buf_vec[vec_idx].buf_len;
1137                 buf_avail  = buf_len - buf_offset;
1138         } else if (buf_len == dev->vhost_hlen) {
1139                 if (unlikely(++vec_idx >= nr_vec))
1140                         goto out;
1141                 buf_addr = buf_vec[vec_idx].buf_addr;
1142                 buf_iova = buf_vec[vec_idx].buf_iova;
1143                 buf_len = buf_vec[vec_idx].buf_len;
1144
1145                 buf_offset = 0;
1146                 buf_avail = buf_len;
1147         } else {
1148                 buf_offset = dev->vhost_hlen;
1149                 buf_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
1150         }
1151
1152         PRINT_PACKET(dev,
1153                         (uintptr_t)(buf_addr + buf_offset),
1154                         (uint32_t)buf_avail, 0);
1155
1156         mbuf_offset = 0;
1157         mbuf_avail  = m->buf_len - RTE_PKTMBUF_HEADROOM;
1158         while (1) {
1159                 uint64_t hpa;
1160
1161                 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
1162
1163                 /*
1164                  * A desc buf might across two host physical pages that are
1165                  * not continuous. In such case (gpa_to_hpa returns 0), data
1166                  * will be copied even though zero copy is enabled.
1167                  */
1168                 if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
1169                                         buf_iova + buf_offset, cpy_len)))) {
1170                         cur->data_len = cpy_len;
1171                         cur->data_off = 0;
1172                         cur->buf_addr =
1173                                 (void *)(uintptr_t)(buf_addr + buf_offset);
1174                         cur->buf_iova = hpa;
1175
1176                         /*
1177                          * In zero copy mode, one mbuf can only reference data
1178                          * for one or partial of one desc buff.
1179                          */
1180                         mbuf_avail = cpy_len;
1181                 } else {
1182                         if (likely(cpy_len > MAX_BATCH_LEN ||
1183                                    vq->batch_copy_nb_elems >= vq->size ||
1184                                    (hdr && cur == m))) {
1185                                 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
1186                                                                    mbuf_offset),
1187                                            (void *)((uintptr_t)(buf_addr +
1188                                                            buf_offset)),
1189                                            cpy_len);
1190                         } else {
1191                                 batch_copy[vq->batch_copy_nb_elems].dst =
1192                                         rte_pktmbuf_mtod_offset(cur, void *,
1193                                                                 mbuf_offset);
1194                                 batch_copy[vq->batch_copy_nb_elems].src =
1195                                         (void *)((uintptr_t)(buf_addr +
1196                                                                 buf_offset));
1197                                 batch_copy[vq->batch_copy_nb_elems].len =
1198                                         cpy_len;
1199                                 vq->batch_copy_nb_elems++;
1200                         }
1201                 }
1202
1203                 mbuf_avail  -= cpy_len;
1204                 mbuf_offset += cpy_len;
1205                 buf_avail -= cpy_len;
1206                 buf_offset += cpy_len;
1207
1208                 /* This buf reaches to its end, get the next one */
1209                 if (buf_avail == 0) {
1210                         if (++vec_idx >= nr_vec)
1211                                 break;
1212
1213                         buf_addr = buf_vec[vec_idx].buf_addr;
1214                         buf_iova = buf_vec[vec_idx].buf_iova;
1215                         buf_len = buf_vec[vec_idx].buf_len;
1216
1217                         buf_offset = 0;
1218                         buf_avail  = buf_len;
1219
1220                         PRINT_PACKET(dev, (uintptr_t)buf_addr,
1221                                         (uint32_t)buf_avail, 0);
1222                 }
1223
1224                 /*
1225                  * This mbuf reaches to its end, get a new one
1226                  * to hold more data.
1227                  */
1228                 if (mbuf_avail == 0) {
1229                         cur = rte_pktmbuf_alloc(mbuf_pool);
1230                         if (unlikely(cur == NULL)) {
1231                                 RTE_LOG(ERR, VHOST_DATA, "Failed to "
1232                                         "allocate memory for mbuf.\n");
1233                                 error = -1;
1234                                 goto out;
1235                         }
1236                         if (unlikely(dev->dequeue_zero_copy))
1237                                 rte_mbuf_refcnt_update(cur, 1);
1238
1239                         prev->next = cur;
1240                         prev->data_len = mbuf_offset;
1241                         m->nb_segs += 1;
1242                         m->pkt_len += mbuf_offset;
1243                         prev = cur;
1244
1245                         mbuf_offset = 0;
1246                         mbuf_avail  = cur->buf_len - RTE_PKTMBUF_HEADROOM;
1247                 }
1248         }
1249
1250         prev->data_len = mbuf_offset;
1251         m->pkt_len    += mbuf_offset;
1252
1253         if (hdr)
1254                 vhost_dequeue_offload(hdr, m);
1255
1256 out:
1257
1258         return error;
1259 }
1260
1261 static __rte_always_inline struct zcopy_mbuf *
1262 get_zmbuf(struct vhost_virtqueue *vq)
1263 {
1264         uint16_t i;
1265         uint16_t last;
1266         int tries = 0;
1267
1268         /* search [last_zmbuf_idx, zmbuf_size) */
1269         i = vq->last_zmbuf_idx;
1270         last = vq->zmbuf_size;
1271
1272 again:
1273         for (; i < last; i++) {
1274                 if (vq->zmbufs[i].in_use == 0) {
1275                         vq->last_zmbuf_idx = i + 1;
1276                         vq->zmbufs[i].in_use = 1;
1277                         return &vq->zmbufs[i];
1278                 }
1279         }
1280
1281         tries++;
1282         if (tries == 1) {
1283                 /* search [0, last_zmbuf_idx) */
1284                 i = 0;
1285                 last = vq->last_zmbuf_idx;
1286                 goto again;
1287         }
1288
1289         return NULL;
1290 }
1291
1292 static __rte_noinline uint16_t
1293 virtio_dev_tx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
1294         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1295 {
1296         uint16_t i;
1297         uint16_t free_entries;
1298
1299         if (unlikely(dev->dequeue_zero_copy)) {
1300                 struct zcopy_mbuf *zmbuf, *next;
1301
1302                 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1303                      zmbuf != NULL; zmbuf = next) {
1304                         next = TAILQ_NEXT(zmbuf, next);
1305
1306                         if (mbuf_is_consumed(zmbuf->mbuf)) {
1307                                 update_shadow_used_ring_split(vq,
1308                                                 zmbuf->desc_idx, 0);
1309                                 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1310                                 restore_mbuf(zmbuf->mbuf);
1311                                 rte_pktmbuf_free(zmbuf->mbuf);
1312                                 put_zmbuf(zmbuf);
1313                                 vq->nr_zmbuf -= 1;
1314                         }
1315                 }
1316
1317                 if (likely(vq->shadow_used_idx)) {
1318                         flush_shadow_used_ring_split(dev, vq);
1319                         vhost_vring_call_split(dev, vq);
1320                 }
1321         }
1322
1323         free_entries = *((volatile uint16_t *)&vq->avail->idx) -
1324                         vq->last_avail_idx;
1325         if (free_entries == 0)
1326                 return 0;
1327
1328         /*
1329          * The ordering between avail index and
1330          * desc reads needs to be enforced.
1331          */
1332         rte_smp_rmb();
1333
1334         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
1335
1336         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1337
1338         count = RTE_MIN(count, MAX_PKT_BURST);
1339         count = RTE_MIN(count, free_entries);
1340         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1341                         dev->vid, count);
1342
1343         for (i = 0; i < count; i++) {
1344                 struct buf_vector buf_vec[BUF_VECTOR_MAX];
1345                 uint16_t head_idx;
1346                 uint32_t dummy_len;
1347                 uint16_t nr_vec = 0;
1348                 int err;
1349
1350                 if (unlikely(fill_vec_buf_split(dev, vq,
1351                                                 vq->last_avail_idx + i,
1352                                                 &nr_vec, buf_vec,
1353                                                 &head_idx, &dummy_len,
1354                                                 VHOST_ACCESS_RO) < 0))
1355                         break;
1356
1357                 if (likely(dev->dequeue_zero_copy == 0))
1358                         update_shadow_used_ring_split(vq, head_idx, 0);
1359
1360                 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
1361                 if (unlikely(pkts[i] == NULL)) {
1362                         RTE_LOG(ERR, VHOST_DATA,
1363                                 "Failed to allocate memory for mbuf.\n");
1364                         break;
1365                 }
1366
1367                 err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts[i],
1368                                 mbuf_pool);
1369                 if (unlikely(err)) {
1370                         rte_pktmbuf_free(pkts[i]);
1371                         break;
1372                 }
1373
1374                 if (unlikely(dev->dequeue_zero_copy)) {
1375                         struct zcopy_mbuf *zmbuf;
1376
1377                         zmbuf = get_zmbuf(vq);
1378                         if (!zmbuf) {
1379                                 rte_pktmbuf_free(pkts[i]);
1380                                 break;
1381                         }
1382                         zmbuf->mbuf = pkts[i];
1383                         zmbuf->desc_idx = head_idx;
1384
1385                         /*
1386                          * Pin lock the mbuf; we will check later to see
1387                          * whether the mbuf is freed (when we are the last
1388                          * user) or not. If that's the case, we then could
1389                          * update the used ring safely.
1390                          */
1391                         rte_mbuf_refcnt_update(pkts[i], 1);
1392
1393                         vq->nr_zmbuf += 1;
1394                         TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1395                 }
1396         }
1397         vq->last_avail_idx += i;
1398
1399         if (likely(dev->dequeue_zero_copy == 0)) {
1400                 do_data_copy_dequeue(vq);
1401                 if (unlikely(i < count))
1402                         vq->shadow_used_idx = i;
1403                 if (likely(vq->shadow_used_idx)) {
1404                         flush_shadow_used_ring_split(dev, vq);
1405                         vhost_vring_call_split(dev, vq);
1406                 }
1407         }
1408
1409         return i;
1410 }
1411
1412 static __rte_noinline uint16_t
1413 virtio_dev_tx_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
1414         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1415 {
1416         uint16_t i;
1417
1418         if (unlikely(dev->dequeue_zero_copy)) {
1419                 struct zcopy_mbuf *zmbuf, *next;
1420
1421                 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1422                      zmbuf != NULL; zmbuf = next) {
1423                         next = TAILQ_NEXT(zmbuf, next);
1424
1425                         if (mbuf_is_consumed(zmbuf->mbuf)) {
1426                                 update_shadow_used_ring_packed(vq,
1427                                                 zmbuf->desc_idx,
1428                                                 0,
1429                                                 zmbuf->desc_count);
1430
1431                                 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1432                                 restore_mbuf(zmbuf->mbuf);
1433                                 rte_pktmbuf_free(zmbuf->mbuf);
1434                                 put_zmbuf(zmbuf);
1435                                 vq->nr_zmbuf -= 1;
1436                         }
1437                 }
1438
1439                 if (likely(vq->shadow_used_idx)) {
1440                         flush_shadow_used_ring_packed(dev, vq);
1441                         vhost_vring_call_packed(dev, vq);
1442                 }
1443         }
1444
1445         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1446
1447         count = RTE_MIN(count, MAX_PKT_BURST);
1448         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1449                         dev->vid, count);
1450
1451         for (i = 0; i < count; i++) {
1452                 struct buf_vector buf_vec[BUF_VECTOR_MAX];
1453                 uint16_t buf_id;
1454                 uint32_t dummy_len;
1455                 uint16_t desc_count, nr_vec = 0;
1456                 int err;
1457
1458                 if (unlikely(fill_vec_buf_packed(dev, vq,
1459                                                 vq->last_avail_idx, &desc_count,
1460                                                 buf_vec, &nr_vec,
1461                                                 &buf_id, &dummy_len,
1462                                                 VHOST_ACCESS_RO) < 0))
1463                         break;
1464
1465                 if (likely(dev->dequeue_zero_copy == 0))
1466                         update_shadow_used_ring_packed(vq, buf_id, 0,
1467                                         desc_count);
1468
1469                 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
1470                 if (unlikely(pkts[i] == NULL)) {
1471                         RTE_LOG(ERR, VHOST_DATA,
1472                                 "Failed to allocate memory for mbuf.\n");
1473                         break;
1474                 }
1475
1476                 err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts[i],
1477                                 mbuf_pool);
1478                 if (unlikely(err)) {
1479                         rte_pktmbuf_free(pkts[i]);
1480                         break;
1481                 }
1482
1483                 if (unlikely(dev->dequeue_zero_copy)) {
1484                         struct zcopy_mbuf *zmbuf;
1485
1486                         zmbuf = get_zmbuf(vq);
1487                         if (!zmbuf) {
1488                                 rte_pktmbuf_free(pkts[i]);
1489                                 break;
1490                         }
1491                         zmbuf->mbuf = pkts[i];
1492                         zmbuf->desc_idx = buf_id;
1493                         zmbuf->desc_count = desc_count;
1494
1495                         /*
1496                          * Pin lock the mbuf; we will check later to see
1497                          * whether the mbuf is freed (when we are the last
1498                          * user) or not. If that's the case, we then could
1499                          * update the used ring safely.
1500                          */
1501                         rte_mbuf_refcnt_update(pkts[i], 1);
1502
1503                         vq->nr_zmbuf += 1;
1504                         TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1505                 }
1506
1507                 vq->last_avail_idx += desc_count;
1508                 if (vq->last_avail_idx >= vq->size) {
1509                         vq->last_avail_idx -= vq->size;
1510                         vq->avail_wrap_counter ^= 1;
1511                 }
1512         }
1513
1514         if (likely(dev->dequeue_zero_copy == 0)) {
1515                 do_data_copy_dequeue(vq);
1516                 if (unlikely(i < count))
1517                         vq->shadow_used_idx = i;
1518                 if (likely(vq->shadow_used_idx)) {
1519                         flush_shadow_used_ring_packed(dev, vq);
1520                         vhost_vring_call_packed(dev, vq);
1521                 }
1522         }
1523
1524         return i;
1525 }
1526
1527 uint16_t
1528 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
1529         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1530 {
1531         struct virtio_net *dev;
1532         struct rte_mbuf *rarp_mbuf = NULL;
1533         struct vhost_virtqueue *vq;
1534
1535         dev = get_device(vid);
1536         if (!dev)
1537                 return 0;
1538
1539         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
1540                 RTE_LOG(ERR, VHOST_DATA,
1541                         "(%d) %s: built-in vhost net backend is disabled.\n",
1542                         dev->vid, __func__);
1543                 return 0;
1544         }
1545
1546         if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
1547                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
1548                         dev->vid, __func__, queue_id);
1549                 return 0;
1550         }
1551
1552         vq = dev->virtqueue[queue_id];
1553
1554         if (unlikely(rte_spinlock_trylock(&vq->access_lock) == 0))
1555                 return 0;
1556
1557         if (unlikely(vq->enabled == 0)) {
1558                 count = 0;
1559                 goto out_access_unlock;
1560         }
1561
1562         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1563                 vhost_user_iotlb_rd_lock(vq);
1564
1565         if (unlikely(vq->access_ok == 0))
1566                 if (unlikely(vring_translate(dev, vq) < 0)) {
1567                         count = 0;
1568                         goto out;
1569                 }
1570
1571         /*
1572          * Construct a RARP broadcast packet, and inject it to the "pkts"
1573          * array, to looks like that guest actually send such packet.
1574          *
1575          * Check user_send_rarp() for more information.
1576          *
1577          * broadcast_rarp shares a cacheline in the virtio_net structure
1578          * with some fields that are accessed during enqueue and
1579          * rte_atomic16_cmpset() causes a write if using cmpxchg. This could
1580          * result in false sharing between enqueue and dequeue.
1581          *
1582          * Prevent unnecessary false sharing by reading broadcast_rarp first
1583          * and only performing cmpset if the read indicates it is likely to
1584          * be set.
1585          */
1586         if (unlikely(rte_atomic16_read(&dev->broadcast_rarp) &&
1587                         rte_atomic16_cmpset((volatile uint16_t *)
1588                                 &dev->broadcast_rarp.cnt, 1, 0))) {
1589
1590                 rarp_mbuf = rte_net_make_rarp_packet(mbuf_pool, &dev->mac);
1591                 if (rarp_mbuf == NULL) {
1592                         RTE_LOG(ERR, VHOST_DATA,
1593                                 "Failed to make RARP packet.\n");
1594                         count = 0;
1595                         goto out;
1596                 }
1597                 count -= 1;
1598         }
1599
1600         if (vq_is_packed(dev))
1601                 count = virtio_dev_tx_packed(dev, vq, mbuf_pool, pkts, count);
1602         else
1603                 count = virtio_dev_tx_split(dev, vq, mbuf_pool, pkts, count);
1604
1605 out:
1606         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1607                 vhost_user_iotlb_rd_unlock(vq);
1608
1609 out_access_unlock:
1610         rte_spinlock_unlock(&vq->access_lock);
1611
1612         if (unlikely(rarp_mbuf != NULL)) {
1613                 /*
1614                  * Inject it to the head of "pkts" array, so that switch's mac
1615                  * learning table will get updated first.
1616                  */
1617                 memmove(&pkts[1], pkts, count * sizeof(struct rte_mbuf *));
1618                 pkts[0] = rarp_mbuf;
1619                 count += 1;
1620         }
1621
1622         return count;
1623 }