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 #include <rte_vhost_async.h>
  21
  22 #include "iotlb.h"
  23 #include "vhost.h"
  24
  25 #define MAX_BATCH_LEN 256
  26
  27 #define VHOST_ASYNC_BATCH_THRESHOLD 32
  28
  29 static  __rte_always_inline bool
  30 rxvq_is_mergeable(struct virtio_net *dev)
  31 {
  32         return dev->features & (1ULL << VIRTIO_NET_F_MRG_RXBUF);
  33 }
  34
  35 static  __rte_always_inline bool
  36 virtio_net_is_inorder(struct virtio_net *dev)
  37 {
  38         return dev->features & (1ULL << VIRTIO_F_IN_ORDER);
  39 }
  40
  41 static bool
  42 is_valid_virt_queue_idx(uint32_t idx, int is_tx, uint32_t nr_vring)
  43 {
  44         return (is_tx ^ (idx & 1)) == 0 && idx < nr_vring;
  45 }
  46
  47 static inline void
  48 do_data_copy_enqueue(struct virtio_net *dev, struct vhost_virtqueue *vq)
  49 {
  50         struct batch_copy_elem *elem = vq->batch_copy_elems;
  51         uint16_t count = vq->batch_copy_nb_elems;
  52         int i;
  53
  54         for (i = 0; i < count; i++) {
  55                 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
  56                 vhost_log_cache_write_iova(dev, vq, elem[i].log_addr,
  57                                            elem[i].len);
  58                 PRINT_PACKET(dev, (uintptr_t)elem[i].dst, elem[i].len, 0);
  59         }
  60
  61         vq->batch_copy_nb_elems = 0;
  62 }
  63
  64 static inline void
  65 do_data_copy_dequeue(struct vhost_virtqueue *vq)
  66 {
  67         struct batch_copy_elem *elem = vq->batch_copy_elems;
  68         uint16_t count = vq->batch_copy_nb_elems;
  69         int i;
  70
  71         for (i = 0; i < count; i++)
  72                 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
  73
  74         vq->batch_copy_nb_elems = 0;
  75 }
  76
  77 static __rte_always_inline void
  78 do_flush_shadow_used_ring_split(struct virtio_net *dev,
  79                         struct vhost_virtqueue *vq,
  80                         uint16_t to, uint16_t from, uint16_t size)
  81 {
  82         rte_memcpy(&vq->used->ring[to],
  83                         &vq->shadow_used_split[from],
  84                         size * sizeof(struct vring_used_elem));
  85         vhost_log_cache_used_vring(dev, vq,
  86                         offsetof(struct vring_used, ring[to]),
  87                         size * sizeof(struct vring_used_elem));
  88 }
  89
  90 static __rte_always_inline void
  91 flush_shadow_used_ring_split(struct virtio_net *dev, struct vhost_virtqueue *vq)
  92 {
  93         uint16_t used_idx = vq->last_used_idx & (vq->size - 1);
  94
  95         if (used_idx + vq->shadow_used_idx <= vq->size) {
  96                 do_flush_shadow_used_ring_split(dev, vq, used_idx, 0,
  97                                           vq->shadow_used_idx);
  98         } else {
  99                 uint16_t size;
 100
 101                 /* update used ring interval [used_idx, vq->size] */
 102                 size = vq->size - used_idx;
 103                 do_flush_shadow_used_ring_split(dev, vq, used_idx, 0, size);
 104
 105                 /* update the left half used ring interval [0, left_size] */
 106                 do_flush_shadow_used_ring_split(dev, vq, 0, size,
 107                                           vq->shadow_used_idx - size);
 108         }
 109         vq->last_used_idx += vq->shadow_used_idx;
 110
 111         vhost_log_cache_sync(dev, vq);
 112
 113         __atomic_add_fetch(&vq->used->idx, vq->shadow_used_idx,
 114                            __ATOMIC_RELEASE);
 115         vq->shadow_used_idx = 0;
 116         vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
 117                 sizeof(vq->used->idx));
 118 }
 119
 120 static __rte_always_inline void
 121 async_flush_shadow_used_ring_split(struct virtio_net *dev,
 122         struct vhost_virtqueue *vq)
 123 {
 124         uint16_t used_idx = vq->last_used_idx & (vq->size - 1);
 125
 126         if (used_idx + vq->shadow_used_idx <= vq->size) {
 127                 do_flush_shadow_used_ring_split(dev, vq, used_idx, 0,
 128                                           vq->shadow_used_idx);
 129         } else {
 130                 uint16_t size;
 131
 132                 /* update used ring interval [used_idx, vq->size] */
 133                 size = vq->size - used_idx;
 134                 do_flush_shadow_used_ring_split(dev, vq, used_idx, 0, size);
 135
 136                 /* update the left half used ring interval [0, left_size] */
 137                 do_flush_shadow_used_ring_split(dev, vq, 0, size,
 138                                           vq->shadow_used_idx - size);
 139         }
 140
 141         vq->last_used_idx += vq->shadow_used_idx;
 142         vq->shadow_used_idx = 0;
 143 }
 144
 145 static __rte_always_inline void
 146 update_shadow_used_ring_split(struct vhost_virtqueue *vq,
 147                          uint16_t desc_idx, uint32_t len)
 148 {
 149         uint16_t i = vq->shadow_used_idx++;
 150
 151         vq->shadow_used_split[i].id  = desc_idx;
 152         vq->shadow_used_split[i].len = len;
 153 }
 154
 155 static __rte_always_inline void
 156 vhost_flush_enqueue_shadow_packed(struct virtio_net *dev,
 157                                   struct vhost_virtqueue *vq)
 158 {
 159         int i;
 160         uint16_t used_idx = vq->last_used_idx;
 161         uint16_t head_idx = vq->last_used_idx;
 162         uint16_t head_flags = 0;
 163
 164         /* Split loop in two to save memory barriers */
 165         for (i = 0; i < vq->shadow_used_idx; i++) {
 166                 vq->desc_packed[used_idx].id = vq->shadow_used_packed[i].id;
 167                 vq->desc_packed[used_idx].len = vq->shadow_used_packed[i].len;
 168
 169                 used_idx += vq->shadow_used_packed[i].count;
 170                 if (used_idx >= vq->size)
 171                         used_idx -= vq->size;
 172         }
 173
 174         /* The ordering for storing desc flags needs to be enforced. */
 175         rte_atomic_thread_fence(__ATOMIC_RELEASE);
 176
 177         for (i = 0; i < vq->shadow_used_idx; i++) {
 178                 uint16_t flags;
 179
 180                 if (vq->shadow_used_packed[i].len)
 181                         flags = VRING_DESC_F_WRITE;
 182                 else
 183                         flags = 0;
 184
 185                 if (vq->used_wrap_counter) {
 186                         flags |= VRING_DESC_F_USED;
 187                         flags |= VRING_DESC_F_AVAIL;
 188                 } else {
 189                         flags &= ~VRING_DESC_F_USED;
 190                         flags &= ~VRING_DESC_F_AVAIL;
 191                 }
 192
 193                 if (i > 0) {
 194                         vq->desc_packed[vq->last_used_idx].flags = flags;
 195
 196                         vhost_log_cache_used_vring(dev, vq,
 197                                         vq->last_used_idx *
 198                                         sizeof(struct vring_packed_desc),
 199                                         sizeof(struct vring_packed_desc));
 200                 } else {
 201                         head_idx = vq->last_used_idx;
 202                         head_flags = flags;
 203                 }
 204
 205                 vq_inc_last_used_packed(vq, vq->shadow_used_packed[i].count);
 206         }
 207
 208         vq->desc_packed[head_idx].flags = head_flags;
 209
 210         vhost_log_cache_used_vring(dev, vq,
 211                                 head_idx *
 212                                 sizeof(struct vring_packed_desc),
 213                                 sizeof(struct vring_packed_desc));
 214
 215         vq->shadow_used_idx = 0;
 216         vhost_log_cache_sync(dev, vq);
 217 }
 218
 219 static __rte_always_inline void
 220 vhost_flush_dequeue_shadow_packed(struct virtio_net *dev,
 221                                   struct vhost_virtqueue *vq)
 222 {
 223         struct vring_used_elem_packed *used_elem = &vq->shadow_used_packed[0];
 224
 225         vq->desc_packed[vq->shadow_last_used_idx].id = used_elem->id;
 226         /* desc flags is the synchronization point for virtio packed vring */
 227         __atomic_store_n(&vq->desc_packed[vq->shadow_last_used_idx].flags,
 228                          used_elem->flags, __ATOMIC_RELEASE);
 229
 230         vhost_log_cache_used_vring(dev, vq, vq->shadow_last_used_idx *
 231                                    sizeof(struct vring_packed_desc),
 232                                    sizeof(struct vring_packed_desc));
 233         vq->shadow_used_idx = 0;
 234         vhost_log_cache_sync(dev, vq);
 235 }
 236
 237 static __rte_always_inline void
 238 vhost_flush_enqueue_batch_packed(struct virtio_net *dev,
 239                                  struct vhost_virtqueue *vq,
 240                                  uint64_t *lens,
 241                                  uint16_t *ids)
 242 {
 243         uint16_t i;
 244         uint16_t flags;
 245
 246         if (vq->shadow_used_idx) {
 247                 do_data_copy_enqueue(dev, vq);
 248                 vhost_flush_enqueue_shadow_packed(dev, vq);
 249         }
 250
 251         flags = PACKED_DESC_ENQUEUE_USED_FLAG(vq->used_wrap_counter);
 252
 253         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
 254                 vq->desc_packed[vq->last_used_idx + i].id = ids[i];
 255                 vq->desc_packed[vq->last_used_idx + i].len = lens[i];
 256         }
 257
 258         rte_atomic_thread_fence(__ATOMIC_RELEASE);
 259
 260         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
 261                 vq->desc_packed[vq->last_used_idx + i].flags = flags;
 262
 263         vhost_log_cache_used_vring(dev, vq, vq->last_used_idx *
 264                                    sizeof(struct vring_packed_desc),
 265                                    sizeof(struct vring_packed_desc) *
 266                                    PACKED_BATCH_SIZE);
 267         vhost_log_cache_sync(dev, vq);
 268
 269         vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
 270 }
 271
 272 static __rte_always_inline void
 273 vhost_shadow_dequeue_batch_packed_inorder(struct vhost_virtqueue *vq,
 274                                           uint16_t id)
 275 {
 276         vq->shadow_used_packed[0].id = id;
 277
 278         if (!vq->shadow_used_idx) {
 279                 vq->shadow_last_used_idx = vq->last_used_idx;
 280                 vq->shadow_used_packed[0].flags =
 281                         PACKED_DESC_DEQUEUE_USED_FLAG(vq->used_wrap_counter);
 282                 vq->shadow_used_packed[0].len = 0;
 283                 vq->shadow_used_packed[0].count = 1;
 284                 vq->shadow_used_idx++;
 285         }
 286
 287         vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
 288 }
 289
 290 static __rte_always_inline void
 291 vhost_shadow_dequeue_batch_packed(struct virtio_net *dev,
 292                                   struct vhost_virtqueue *vq,
 293                                   uint16_t *ids)
 294 {
 295         uint16_t flags;
 296         uint16_t i;
 297         uint16_t begin;
 298
 299         flags = PACKED_DESC_DEQUEUE_USED_FLAG(vq->used_wrap_counter);
 300
 301         if (!vq->shadow_used_idx) {
 302                 vq->shadow_last_used_idx = vq->last_used_idx;
 303                 vq->shadow_used_packed[0].id  = ids[0];
 304                 vq->shadow_used_packed[0].len = 0;
 305                 vq->shadow_used_packed[0].count = 1;
 306                 vq->shadow_used_packed[0].flags = flags;
 307                 vq->shadow_used_idx++;
 308                 begin = 1;
 309         } else
 310                 begin = 0;
 311
 312         vhost_for_each_try_unroll(i, begin, PACKED_BATCH_SIZE) {
 313                 vq->desc_packed[vq->last_used_idx + i].id = ids[i];
 314                 vq->desc_packed[vq->last_used_idx + i].len = 0;
 315         }
 316
 317         rte_atomic_thread_fence(__ATOMIC_RELEASE);
 318         vhost_for_each_try_unroll(i, begin, PACKED_BATCH_SIZE)
 319                 vq->desc_packed[vq->last_used_idx + i].flags = flags;
 320
 321         vhost_log_cache_used_vring(dev, vq, vq->last_used_idx *
 322                                    sizeof(struct vring_packed_desc),
 323                                    sizeof(struct vring_packed_desc) *
 324                                    PACKED_BATCH_SIZE);
 325         vhost_log_cache_sync(dev, vq);
 326
 327         vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
 328 }
 329
 330 static __rte_always_inline void
 331 vhost_shadow_dequeue_single_packed(struct vhost_virtqueue *vq,
 332                                    uint16_t buf_id,
 333                                    uint16_t count)
 334 {
 335         uint16_t flags;
 336
 337         flags = vq->desc_packed[vq->last_used_idx].flags;
 338         if (vq->used_wrap_counter) {
 339                 flags |= VRING_DESC_F_USED;
 340                 flags |= VRING_DESC_F_AVAIL;
 341         } else {
 342                 flags &= ~VRING_DESC_F_USED;
 343                 flags &= ~VRING_DESC_F_AVAIL;
 344         }
 345
 346         if (!vq->shadow_used_idx) {
 347                 vq->shadow_last_used_idx = vq->last_used_idx;
 348
 349                 vq->shadow_used_packed[0].id  = buf_id;
 350                 vq->shadow_used_packed[0].len = 0;
 351                 vq->shadow_used_packed[0].flags = flags;
 352                 vq->shadow_used_idx++;
 353         } else {
 354                 vq->desc_packed[vq->last_used_idx].id = buf_id;
 355                 vq->desc_packed[vq->last_used_idx].len = 0;
 356                 vq->desc_packed[vq->last_used_idx].flags = flags;
 357         }
 358
 359         vq_inc_last_used_packed(vq, count);
 360 }
 361
 362 static __rte_always_inline void
 363 vhost_shadow_dequeue_single_packed_inorder(struct vhost_virtqueue *vq,
 364                                            uint16_t buf_id,
 365                                            uint16_t count)
 366 {
 367         uint16_t flags;
 368
 369         vq->shadow_used_packed[0].id = buf_id;
 370
 371         flags = vq->desc_packed[vq->last_used_idx].flags;
 372         if (vq->used_wrap_counter) {
 373                 flags |= VRING_DESC_F_USED;
 374                 flags |= VRING_DESC_F_AVAIL;
 375         } else {
 376                 flags &= ~VRING_DESC_F_USED;
 377                 flags &= ~VRING_DESC_F_AVAIL;
 378         }
 379
 380         if (!vq->shadow_used_idx) {
 381                 vq->shadow_last_used_idx = vq->last_used_idx;
 382                 vq->shadow_used_packed[0].len = 0;
 383                 vq->shadow_used_packed[0].flags = flags;
 384                 vq->shadow_used_idx++;
 385         }
 386
 387         vq_inc_last_used_packed(vq, count);
 388 }
 389
 390 static __rte_always_inline void
 391 vhost_shadow_enqueue_single_packed(struct virtio_net *dev,
 392                                    struct vhost_virtqueue *vq,
 393                                    uint32_t len[],
 394                                    uint16_t id[],
 395                                    uint16_t count[],
 396                                    uint16_t num_buffers)
 397 {
 398         uint16_t i;
 399         for (i = 0; i < num_buffers; i++) {
 400                 /* enqueue shadow flush action aligned with batch num */
 401                 if (!vq->shadow_used_idx)
 402                         vq->shadow_aligned_idx = vq->last_used_idx &
 403                                 PACKED_BATCH_MASK;
 404                 vq->shadow_used_packed[vq->shadow_used_idx].id  = id[i];
 405                 vq->shadow_used_packed[vq->shadow_used_idx].len = len[i];
 406                 vq->shadow_used_packed[vq->shadow_used_idx].count = count[i];
 407                 vq->shadow_aligned_idx += count[i];
 408                 vq->shadow_used_idx++;
 409         }
 410
 411         if (vq->shadow_aligned_idx >= PACKED_BATCH_SIZE) {
 412                 do_data_copy_enqueue(dev, vq);
 413                 vhost_flush_enqueue_shadow_packed(dev, vq);
 414         }
 415 }
 416
 417 /* avoid write operation when necessary, to lessen cache issues */
 418 #define ASSIGN_UNLESS_EQUAL(var, val) do {      \
 419         if ((var) != (val))                     \
 420                 (var) = (val);                  \
 421 } while (0)
 422
 423 static __rte_always_inline void
 424 virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
 425 {
 426         uint64_t csum_l4 = m_buf->ol_flags & PKT_TX_L4_MASK;
 427
 428         if (m_buf->ol_flags & PKT_TX_TCP_SEG)
 429                 csum_l4 |= PKT_TX_TCP_CKSUM;
 430
 431         if (csum_l4) {
 432                 net_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 433                 net_hdr->csum_start = m_buf->l2_len + m_buf->l3_len;
 434
 435                 switch (csum_l4) {
 436                 case PKT_TX_TCP_CKSUM:
 437                         net_hdr->csum_offset = (offsetof(struct rte_tcp_hdr,
 438                                                 cksum));
 439                         break;
 440                 case PKT_TX_UDP_CKSUM:
 441                         net_hdr->csum_offset = (offsetof(struct rte_udp_hdr,
 442                                                 dgram_cksum));
 443                         break;
 444                 case PKT_TX_SCTP_CKSUM:
 445                         net_hdr->csum_offset = (offsetof(struct rte_sctp_hdr,
 446                                                 cksum));
 447                         break;
 448                 }
 449         } else {
 450                 ASSIGN_UNLESS_EQUAL(net_hdr->csum_start, 0);
 451                 ASSIGN_UNLESS_EQUAL(net_hdr->csum_offset, 0);
 452                 ASSIGN_UNLESS_EQUAL(net_hdr->flags, 0);
 453         }
 454
 455         /* IP cksum verification cannot be bypassed, then calculate here */
 456         if (m_buf->ol_flags & PKT_TX_IP_CKSUM) {
 457                 struct rte_ipv4_hdr *ipv4_hdr;
 458
 459                 ipv4_hdr = rte_pktmbuf_mtod_offset(m_buf, struct rte_ipv4_hdr *,
 460                                                    m_buf->l2_len);
 461                 ipv4_hdr->hdr_checksum = 0;
 462                 ipv4_hdr->hdr_checksum = rte_ipv4_cksum(ipv4_hdr);
 463         }
 464
 465         if (m_buf->ol_flags & PKT_TX_TCP_SEG) {
 466                 if (m_buf->ol_flags & PKT_TX_IPV4)
 467                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 468                 else
 469                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 470                 net_hdr->gso_size = m_buf->tso_segsz;
 471                 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len
 472                                         + m_buf->l4_len;
 473         } else if (m_buf->ol_flags & PKT_TX_UDP_SEG) {
 474                 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
 475                 net_hdr->gso_size = m_buf->tso_segsz;
 476                 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len +
 477                         m_buf->l4_len;
 478         } else {
 479                 ASSIGN_UNLESS_EQUAL(net_hdr->gso_type, 0);
 480                 ASSIGN_UNLESS_EQUAL(net_hdr->gso_size, 0);
 481                 ASSIGN_UNLESS_EQUAL(net_hdr->hdr_len, 0);
 482         }
 483 }
 484
 485 static __rte_always_inline int
 486 map_one_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 487                 struct buf_vector *buf_vec, uint16_t *vec_idx,
 488                 uint64_t desc_iova, uint64_t desc_len, uint8_t perm)
 489 {
 490         uint16_t vec_id = *vec_idx;
 491
 492         while (desc_len) {
 493                 uint64_t desc_addr;
 494                 uint64_t desc_chunck_len = desc_len;
 495
 496                 if (unlikely(vec_id >= BUF_VECTOR_MAX))
 497                         return -1;
 498
 499                 desc_addr = vhost_iova_to_vva(dev, vq,
 500                                 desc_iova,
 501                                 &desc_chunck_len,
 502                                 perm);
 503                 if (unlikely(!desc_addr))
 504                         return -1;
 505
 506                 rte_prefetch0((void *)(uintptr_t)desc_addr);
 507
 508                 buf_vec[vec_id].buf_iova = desc_iova;
 509                 buf_vec[vec_id].buf_addr = desc_addr;
 510                 buf_vec[vec_id].buf_len  = desc_chunck_len;
 511
 512                 desc_len -= desc_chunck_len;
 513                 desc_iova += desc_chunck_len;
 514                 vec_id++;
 515         }
 516         *vec_idx = vec_id;
 517
 518         return 0;
 519 }
 520
 521 static __rte_always_inline int
 522 fill_vec_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 523                          uint32_t avail_idx, uint16_t *vec_idx,
 524                          struct buf_vector *buf_vec, uint16_t *desc_chain_head,
 525                          uint32_t *desc_chain_len, uint8_t perm)
 526 {
 527         uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
 528         uint16_t vec_id = *vec_idx;
 529         uint32_t len    = 0;
 530         uint64_t dlen;
 531         uint32_t nr_descs = vq->size;
 532         uint32_t cnt    = 0;
 533         struct vring_desc *descs = vq->desc;
 534         struct vring_desc *idesc = NULL;
 535
 536         if (unlikely(idx >= vq->size))
 537                 return -1;
 538
 539         *desc_chain_head = idx;
 540
 541         if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
 542                 dlen = vq->desc[idx].len;
 543                 nr_descs = dlen / sizeof(struct vring_desc);
 544                 if (unlikely(nr_descs > vq->size))
 545                         return -1;
 546
 547                 descs = (struct vring_desc *)(uintptr_t)
 548                         vhost_iova_to_vva(dev, vq, vq->desc[idx].addr,
 549                                                 &dlen,
 550                                                 VHOST_ACCESS_RO);
 551                 if (unlikely(!descs))
 552                         return -1;
 553
 554                 if (unlikely(dlen < vq->desc[idx].len)) {
 555                         /*
 556                          * The indirect desc table is not contiguous
 557                          * in process VA space, we have to copy it.
 558                          */
 559                         idesc = vhost_alloc_copy_ind_table(dev, vq,
 560                                         vq->desc[idx].addr, vq->desc[idx].len);
 561                         if (unlikely(!idesc))
 562                                 return -1;
 563
 564                         descs = idesc;
 565                 }
 566
 567                 idx = 0;
 568         }
 569
 570         while (1) {
 571                 if (unlikely(idx >= nr_descs || cnt++ >= nr_descs)) {
 572                         free_ind_table(idesc);
 573                         return -1;
 574                 }
 575
 576                 len += descs[idx].len;
 577
 578                 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 579                                                 descs[idx].addr, descs[idx].len,
 580                                                 perm))) {
 581                         free_ind_table(idesc);
 582                         return -1;
 583                 }
 584
 585                 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
 586                         break;
 587
 588                 idx = descs[idx].next;
 589         }
 590
 591         *desc_chain_len = len;
 592         *vec_idx = vec_id;
 593
 594         if (unlikely(!!idesc))
 595                 free_ind_table(idesc);
 596
 597         return 0;
 598 }
 599
 600 /*
 601  * Returns -1 on fail, 0 on success
 602  */
 603 static inline int
 604 reserve_avail_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 605                                 uint32_t size, struct buf_vector *buf_vec,
 606                                 uint16_t *num_buffers, uint16_t avail_head,
 607                                 uint16_t *nr_vec)
 608 {
 609         uint16_t cur_idx;
 610         uint16_t vec_idx = 0;
 611         uint16_t max_tries, tries = 0;
 612
 613         uint16_t head_idx = 0;
 614         uint32_t len = 0;
 615
 616         *num_buffers = 0;
 617         cur_idx  = vq->last_avail_idx;
 618
 619         if (rxvq_is_mergeable(dev))
 620                 max_tries = vq->size - 1;
 621         else
 622                 max_tries = 1;
 623
 624         while (size > 0) {
 625                 if (unlikely(cur_idx == avail_head))
 626                         return -1;
 627                 /*
 628                  * if we tried all available ring items, and still
 629                  * can't get enough buf, it means something abnormal
 630                  * happened.
 631                  */
 632                 if (unlikely(++tries > max_tries))
 633                         return -1;
 634
 635                 if (unlikely(fill_vec_buf_split(dev, vq, cur_idx,
 636                                                 &vec_idx, buf_vec,
 637                                                 &head_idx, &len,
 638                                                 VHOST_ACCESS_RW) < 0))
 639                         return -1;
 640                 len = RTE_MIN(len, size);
 641                 update_shadow_used_ring_split(vq, head_idx, len);
 642                 size -= len;
 643
 644                 cur_idx++;
 645                 *num_buffers += 1;
 646         }
 647
 648         *nr_vec = vec_idx;
 649
 650         return 0;
 651 }
 652
 653 static __rte_always_inline int
 654 fill_vec_buf_packed_indirect(struct virtio_net *dev,
 655                         struct vhost_virtqueue *vq,
 656                         struct vring_packed_desc *desc, uint16_t *vec_idx,
 657                         struct buf_vector *buf_vec, uint32_t *len, uint8_t perm)
 658 {
 659         uint16_t i;
 660         uint32_t nr_descs;
 661         uint16_t vec_id = *vec_idx;
 662         uint64_t dlen;
 663         struct vring_packed_desc *descs, *idescs = NULL;
 664
 665         dlen = desc->len;
 666         descs = (struct vring_packed_desc *)(uintptr_t)
 667                 vhost_iova_to_vva(dev, vq, desc->addr, &dlen, VHOST_ACCESS_RO);
 668         if (unlikely(!descs))
 669                 return -1;
 670
 671         if (unlikely(dlen < desc->len)) {
 672                 /*
 673                  * The indirect desc table is not contiguous
 674                  * in process VA space, we have to copy it.
 675                  */
 676                 idescs = vhost_alloc_copy_ind_table(dev,
 677                                 vq, desc->addr, desc->len);
 678                 if (unlikely(!idescs))
 679                         return -1;
 680
 681                 descs = idescs;
 682         }
 683
 684         nr_descs =  desc->len / sizeof(struct vring_packed_desc);
 685         if (unlikely(nr_descs >= vq->size)) {
 686                 free_ind_table(idescs);
 687                 return -1;
 688         }
 689
 690         for (i = 0; i < nr_descs; i++) {
 691                 if (unlikely(vec_id >= BUF_VECTOR_MAX)) {
 692                         free_ind_table(idescs);
 693                         return -1;
 694                 }
 695
 696                 *len += descs[i].len;
 697                 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 698                                                 descs[i].addr, descs[i].len,
 699                                                 perm)))
 700                         return -1;
 701         }
 702         *vec_idx = vec_id;
 703
 704         if (unlikely(!!idescs))
 705                 free_ind_table(idescs);
 706
 707         return 0;
 708 }
 709
 710 static __rte_always_inline int
 711 fill_vec_buf_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
 712                                 uint16_t avail_idx, uint16_t *desc_count,
 713                                 struct buf_vector *buf_vec, uint16_t *vec_idx,
 714                                 uint16_t *buf_id, uint32_t *len, uint8_t perm)
 715 {
 716         bool wrap_counter = vq->avail_wrap_counter;
 717         struct vring_packed_desc *descs = vq->desc_packed;
 718         uint16_t vec_id = *vec_idx;
 719
 720         if (avail_idx < vq->last_avail_idx)
 721                 wrap_counter ^= 1;
 722
 723         /*
 724          * Perform a load-acquire barrier in desc_is_avail to
 725          * enforce the ordering between desc flags and desc
 726          * content.
 727          */
 728         if (unlikely(!desc_is_avail(&descs[avail_idx], wrap_counter)))
 729                 return -1;
 730
 731         *desc_count = 0;
 732         *len = 0;
 733
 734         while (1) {
 735                 if (unlikely(vec_id >= BUF_VECTOR_MAX))
 736                         return -1;
 737
 738                 if (unlikely(*desc_count >= vq->size))
 739                         return -1;
 740
 741                 *desc_count += 1;
 742                 *buf_id = descs[avail_idx].id;
 743
 744                 if (descs[avail_idx].flags & VRING_DESC_F_INDIRECT) {
 745                         if (unlikely(fill_vec_buf_packed_indirect(dev, vq,
 746                                                         &descs[avail_idx],
 747                                                         &vec_id, buf_vec,
 748                                                         len, perm) < 0))
 749                                 return -1;
 750                 } else {
 751                         *len += descs[avail_idx].len;
 752
 753                         if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 754                                                         descs[avail_idx].addr,
 755                                                         descs[avail_idx].len,
 756                                                         perm)))
 757                                 return -1;
 758                 }
 759
 760                 if ((descs[avail_idx].flags & VRING_DESC_F_NEXT) == 0)
 761                         break;
 762
 763                 if (++avail_idx >= vq->size) {
 764                         avail_idx -= vq->size;
 765                         wrap_counter ^= 1;
 766                 }
 767         }
 768
 769         *vec_idx = vec_id;
 770
 771         return 0;
 772 }
 773
 774 static __rte_noinline void
 775 copy_vnet_hdr_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 776                 struct buf_vector *buf_vec,
 777                 struct virtio_net_hdr_mrg_rxbuf *hdr)
 778 {
 779         uint64_t len;
 780         uint64_t remain = dev->vhost_hlen;
 781         uint64_t src = (uint64_t)(uintptr_t)hdr, dst;
 782         uint64_t iova = buf_vec->buf_iova;
 783
 784         while (remain) {
 785                 len = RTE_MIN(remain,
 786                                 buf_vec->buf_len);
 787                 dst = buf_vec->buf_addr;
 788                 rte_memcpy((void *)(uintptr_t)dst,
 789                                 (void *)(uintptr_t)src,
 790                                 len);
 791
 792                 PRINT_PACKET(dev, (uintptr_t)dst,
 793                                 (uint32_t)len, 0);
 794                 vhost_log_cache_write_iova(dev, vq,
 795                                 iova, len);
 796
 797                 remain -= len;
 798                 iova += len;
 799                 src += len;
 800                 buf_vec++;
 801         }
 802 }
 803
 804 static __rte_always_inline int
 805 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 806                             struct rte_mbuf *m, struct buf_vector *buf_vec,
 807                             uint16_t nr_vec, uint16_t num_buffers)
 808 {
 809         uint32_t vec_idx = 0;
 810         uint32_t mbuf_offset, mbuf_avail;
 811         uint32_t buf_offset, buf_avail;
 812         uint64_t buf_addr, buf_iova, buf_len;
 813         uint32_t cpy_len;
 814         uint64_t hdr_addr;
 815         struct rte_mbuf *hdr_mbuf;
 816         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 817         struct virtio_net_hdr_mrg_rxbuf tmp_hdr, *hdr = NULL;
 818         int error = 0;
 819
 820         if (unlikely(m == NULL)) {
 821                 error = -1;
 822                 goto out;
 823         }
 824
 825         buf_addr = buf_vec[vec_idx].buf_addr;
 826         buf_iova = buf_vec[vec_idx].buf_iova;
 827         buf_len = buf_vec[vec_idx].buf_len;
 828
 829         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
 830                 error = -1;
 831                 goto out;
 832         }
 833
 834         hdr_mbuf = m;
 835         hdr_addr = buf_addr;
 836         if (unlikely(buf_len < dev->vhost_hlen))
 837                 hdr = &tmp_hdr;
 838         else
 839                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)hdr_addr;
 840
 841         VHOST_LOG_DATA(DEBUG, "(%d) RX: num merge buffers %d\n",
 842                 dev->vid, num_buffers);
 843
 844         if (unlikely(buf_len < dev->vhost_hlen)) {
 845                 buf_offset = dev->vhost_hlen - buf_len;
 846                 vec_idx++;
 847                 buf_addr = buf_vec[vec_idx].buf_addr;
 848                 buf_iova = buf_vec[vec_idx].buf_iova;
 849                 buf_len = buf_vec[vec_idx].buf_len;
 850                 buf_avail = buf_len - buf_offset;
 851         } else {
 852                 buf_offset = dev->vhost_hlen;
 853                 buf_avail = buf_len - dev->vhost_hlen;
 854         }
 855
 856         mbuf_avail  = rte_pktmbuf_data_len(m);
 857         mbuf_offset = 0;
 858         while (mbuf_avail != 0 || m->next != NULL) {
 859                 /* done with current buf, get the next one */
 860                 if (buf_avail == 0) {
 861                         vec_idx++;
 862                         if (unlikely(vec_idx >= nr_vec)) {
 863                                 error = -1;
 864                                 goto out;
 865                         }
 866
 867                         buf_addr = buf_vec[vec_idx].buf_addr;
 868                         buf_iova = buf_vec[vec_idx].buf_iova;
 869                         buf_len = buf_vec[vec_idx].buf_len;
 870
 871                         buf_offset = 0;
 872                         buf_avail  = buf_len;
 873                 }
 874
 875                 /* done with current mbuf, get the next one */
 876                 if (mbuf_avail == 0) {
 877                         m = m->next;
 878
 879                         mbuf_offset = 0;
 880                         mbuf_avail  = rte_pktmbuf_data_len(m);
 881                 }
 882
 883                 if (hdr_addr) {
 884                         virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
 885                         if (rxvq_is_mergeable(dev))
 886                                 ASSIGN_UNLESS_EQUAL(hdr->num_buffers,
 887                                                 num_buffers);
 888
 889                         if (unlikely(hdr == &tmp_hdr)) {
 890                                 copy_vnet_hdr_to_desc(dev, vq, buf_vec, hdr);
 891                         } else {
 892                                 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
 893                                                 dev->vhost_hlen, 0);
 894                                 vhost_log_cache_write_iova(dev, vq,
 895                                                 buf_vec[0].buf_iova,
 896                                                 dev->vhost_hlen);
 897                         }
 898
 899                         hdr_addr = 0;
 900                 }
 901
 902                 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
 903
 904                 if (likely(cpy_len > MAX_BATCH_LEN ||
 905                                         vq->batch_copy_nb_elems >= vq->size)) {
 906                         rte_memcpy((void *)((uintptr_t)(buf_addr + buf_offset)),
 907                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 908                                 cpy_len);
 909                         vhost_log_cache_write_iova(dev, vq,
 910                                                    buf_iova + buf_offset,
 911                                                    cpy_len);
 912                         PRINT_PACKET(dev, (uintptr_t)(buf_addr + buf_offset),
 913                                 cpy_len, 0);
 914                 } else {
 915                         batch_copy[vq->batch_copy_nb_elems].dst =
 916                                 (void *)((uintptr_t)(buf_addr + buf_offset));
 917                         batch_copy[vq->batch_copy_nb_elems].src =
 918                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
 919                         batch_copy[vq->batch_copy_nb_elems].log_addr =
 920                                 buf_iova + buf_offset;
 921                         batch_copy[vq->batch_copy_nb_elems].len = cpy_len;
 922                         vq->batch_copy_nb_elems++;
 923                 }
 924
 925                 mbuf_avail  -= cpy_len;
 926                 mbuf_offset += cpy_len;
 927                 buf_avail  -= cpy_len;
 928                 buf_offset += cpy_len;
 929         }
 930
 931 out:
 932
 933         return error;
 934 }
 935
 936 static __rte_always_inline void
 937 async_fill_vec(struct iovec *v, void *base, size_t len)
 938 {
 939         v->iov_base = base;
 940         v->iov_len = len;
 941 }
 942
 943 static __rte_always_inline void
 944 async_fill_iter(struct rte_vhost_iov_iter *it, size_t count,
 945         struct iovec *vec, unsigned long nr_seg)
 946 {
 947         it->offset = 0;
 948         it->count = count;
 949
 950         if (count) {
 951                 it->iov = vec;
 952                 it->nr_segs = nr_seg;
 953         } else {
 954                 it->iov = 0;
 955                 it->nr_segs = 0;
 956         }
 957 }
 958
 959 static __rte_always_inline void
 960 async_fill_desc(struct rte_vhost_async_desc *desc,
 961         struct rte_vhost_iov_iter *src, struct rte_vhost_iov_iter *dst)
 962 {
 963         desc->src = src;
 964         desc->dst = dst;
 965 }
 966
 967 static __rte_always_inline int
 968 async_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 969                         struct rte_mbuf *m, struct buf_vector *buf_vec,
 970                         uint16_t nr_vec, uint16_t num_buffers,
 971                         struct iovec *src_iovec, struct iovec *dst_iovec,
 972                         struct rte_vhost_iov_iter *src_it,
 973                         struct rte_vhost_iov_iter *dst_it)
 974 {
 975         uint32_t vec_idx = 0;
 976         uint32_t mbuf_offset, mbuf_avail;
 977         uint32_t buf_offset, buf_avail;
 978         uint64_t buf_addr, buf_iova, buf_len;
 979         uint32_t cpy_len, cpy_threshold;
 980         uint64_t hdr_addr;
 981         struct rte_mbuf *hdr_mbuf;
 982         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 983         struct virtio_net_hdr_mrg_rxbuf tmp_hdr, *hdr = NULL;
 984         int error = 0;
 985         uint64_t mapped_len;
 986
 987         uint32_t tlen = 0;
 988         int tvec_idx = 0;
 989         void *hpa;
 990
 991         if (unlikely(m == NULL)) {
 992                 error = -1;
 993                 goto out;
 994         }
 995
 996         cpy_threshold = vq->async_threshold;
 997
 998         buf_addr = buf_vec[vec_idx].buf_addr;
 999         buf_iova = buf_vec[vec_idx].buf_iova;
1000         buf_len = buf_vec[vec_idx].buf_len;
1001
1002         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
1003                 error = -1;
1004                 goto out;
1005         }
1006
1007         hdr_mbuf = m;
1008         hdr_addr = buf_addr;
1009         if (unlikely(buf_len < dev->vhost_hlen))
1010                 hdr = &tmp_hdr;
1011         else
1012                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)hdr_addr;
1013
1014         VHOST_LOG_DATA(DEBUG, "(%d) RX: num merge buffers %d\n",
1015                 dev->vid, num_buffers);
1016
1017         if (unlikely(buf_len < dev->vhost_hlen)) {
1018                 buf_offset = dev->vhost_hlen - buf_len;
1019                 vec_idx++;
1020                 buf_addr = buf_vec[vec_idx].buf_addr;
1021                 buf_iova = buf_vec[vec_idx].buf_iova;
1022                 buf_len = buf_vec[vec_idx].buf_len;
1023                 buf_avail = buf_len - buf_offset;
1024         } else {
1025                 buf_offset = dev->vhost_hlen;
1026                 buf_avail = buf_len - dev->vhost_hlen;
1027         }
1028
1029         mbuf_avail  = rte_pktmbuf_data_len(m);
1030         mbuf_offset = 0;
1031
1032         while (mbuf_avail != 0 || m->next != NULL) {
1033                 /* done with current buf, get the next one */
1034                 if (buf_avail == 0) {
1035                         vec_idx++;
1036                         if (unlikely(vec_idx >= nr_vec)) {
1037                                 error = -1;
1038                                 goto out;
1039                         }
1040
1041                         buf_addr = buf_vec[vec_idx].buf_addr;
1042                         buf_iova = buf_vec[vec_idx].buf_iova;
1043                         buf_len = buf_vec[vec_idx].buf_len;
1044
1045                         buf_offset = 0;
1046                         buf_avail  = buf_len;
1047                 }
1048
1049                 /* done with current mbuf, get the next one */
1050                 if (mbuf_avail == 0) {
1051                         m = m->next;
1052
1053                         mbuf_offset = 0;
1054                         mbuf_avail  = rte_pktmbuf_data_len(m);
1055                 }
1056
1057                 if (hdr_addr) {
1058                         virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
1059                         if (rxvq_is_mergeable(dev))
1060                                 ASSIGN_UNLESS_EQUAL(hdr->num_buffers,
1061                                                 num_buffers);
1062
1063                         if (unlikely(hdr == &tmp_hdr)) {
1064                                 copy_vnet_hdr_to_desc(dev, vq, buf_vec, hdr);
1065                         } else {
1066                                 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
1067                                                 dev->vhost_hlen, 0);
1068                                 vhost_log_cache_write_iova(dev, vq,
1069                                                 buf_vec[0].buf_iova,
1070                                                 dev->vhost_hlen);
1071                         }
1072
1073                         hdr_addr = 0;
1074                 }
1075
1076                 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
1077
1078                 while (unlikely(cpy_len && cpy_len >= cpy_threshold)) {
1079                         hpa = (void *)(uintptr_t)gpa_to_first_hpa(dev,
1080                                         buf_iova + buf_offset,
1081                                         cpy_len, &mapped_len);
1082
1083                         if (unlikely(!hpa || mapped_len < cpy_threshold))
1084                                 break;
1085
1086                         async_fill_vec(src_iovec + tvec_idx,
1087                                 (void *)(uintptr_t)rte_pktmbuf_iova_offset(m,
1088                                 mbuf_offset), (size_t)mapped_len);
1089
1090                         async_fill_vec(dst_iovec + tvec_idx,
1091                                         hpa, (size_t)mapped_len);
1092
1093                         tlen += (uint32_t)mapped_len;
1094                         cpy_len -= (uint32_t)mapped_len;
1095                         mbuf_avail  -= (uint32_t)mapped_len;
1096                         mbuf_offset += (uint32_t)mapped_len;
1097                         buf_avail  -= (uint32_t)mapped_len;
1098                         buf_offset += (uint32_t)mapped_len;
1099                         tvec_idx++;
1100                 }
1101
1102                 if (likely(cpy_len)) {
1103                         if (unlikely(vq->batch_copy_nb_elems >= vq->size)) {
1104                                 rte_memcpy(
1105                                 (void *)((uintptr_t)(buf_addr + buf_offset)),
1106                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
1107                                 cpy_len);
1108
1109                                 PRINT_PACKET(dev,
1110                                         (uintptr_t)(buf_addr + buf_offset),
1111                                         cpy_len, 0);
1112                         } else {
1113                                 batch_copy[vq->batch_copy_nb_elems].dst =
1114                                 (void *)((uintptr_t)(buf_addr + buf_offset));
1115                                 batch_copy[vq->batch_copy_nb_elems].src =
1116                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
1117                                 batch_copy[vq->batch_copy_nb_elems].log_addr =
1118                                         buf_iova + buf_offset;
1119                                 batch_copy[vq->batch_copy_nb_elems].len =
1120                                         cpy_len;
1121                                 vq->batch_copy_nb_elems++;
1122                         }
1123
1124                         mbuf_avail  -= cpy_len;
1125                         mbuf_offset += cpy_len;
1126                         buf_avail  -= cpy_len;
1127                         buf_offset += cpy_len;
1128                 }
1129
1130         }
1131
1132 out:
1133         async_fill_iter(src_it, tlen, src_iovec, tvec_idx);
1134         async_fill_iter(dst_it, tlen, dst_iovec, tvec_idx);
1135
1136         return error;
1137 }
1138
1139 static __rte_always_inline int
1140 vhost_enqueue_single_packed(struct virtio_net *dev,
1141                             struct vhost_virtqueue *vq,
1142                             struct rte_mbuf *pkt,
1143                             struct buf_vector *buf_vec,
1144                             uint16_t *nr_descs)
1145 {
1146         uint16_t nr_vec = 0;
1147         uint16_t avail_idx = vq->last_avail_idx;
1148         uint16_t max_tries, tries = 0;
1149         uint16_t buf_id = 0;
1150         uint32_t len = 0;
1151         uint16_t desc_count;
1152         uint32_t size = pkt->pkt_len + sizeof(struct virtio_net_hdr_mrg_rxbuf);
1153         uint16_t num_buffers = 0;
1154         uint32_t buffer_len[vq->size];
1155         uint16_t buffer_buf_id[vq->size];
1156         uint16_t buffer_desc_count[vq->size];
1157
1158         if (rxvq_is_mergeable(dev))
1159                 max_tries = vq->size - 1;
1160         else
1161                 max_tries = 1;
1162
1163         while (size > 0) {
1164                 /*
1165                  * if we tried all available ring items, and still
1166                  * can't get enough buf, it means something abnormal
1167                  * happened.
1168                  */
1169                 if (unlikely(++tries > max_tries))
1170                         return -1;
1171
1172                 if (unlikely(fill_vec_buf_packed(dev, vq,
1173                                                 avail_idx, &desc_count,
1174                                                 buf_vec, &nr_vec,
1175                                                 &buf_id, &len,
1176                                                 VHOST_ACCESS_RW) < 0))
1177                         return -1;
1178
1179                 len = RTE_MIN(len, size);
1180                 size -= len;
1181
1182                 buffer_len[num_buffers] = len;
1183                 buffer_buf_id[num_buffers] = buf_id;
1184                 buffer_desc_count[num_buffers] = desc_count;
1185                 num_buffers += 1;
1186
1187                 *nr_descs += desc_count;
1188                 avail_idx += desc_count;
1189                 if (avail_idx >= vq->size)
1190                         avail_idx -= vq->size;
1191         }
1192
1193         if (copy_mbuf_to_desc(dev, vq, pkt, buf_vec, nr_vec, num_buffers) < 0)
1194                 return -1;
1195
1196         vhost_shadow_enqueue_single_packed(dev, vq, buffer_len, buffer_buf_id,
1197                                            buffer_desc_count, num_buffers);
1198
1199         return 0;
1200 }
1201
1202 static __rte_noinline uint32_t
1203 virtio_dev_rx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
1204         struct rte_mbuf **pkts, uint32_t count)
1205 {
1206         uint32_t pkt_idx = 0;
1207         uint16_t num_buffers;
1208         struct buf_vector buf_vec[BUF_VECTOR_MAX];
1209         uint16_t avail_head;
1210
1211         /*
1212          * The ordering between avail index and
1213          * desc reads needs to be enforced.
1214          */
1215         avail_head = __atomic_load_n(&vq->avail->idx, __ATOMIC_ACQUIRE);
1216
1217         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
1218
1219         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
1220                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
1221                 uint16_t nr_vec = 0;
1222
1223                 if (unlikely(reserve_avail_buf_split(dev, vq,
1224                                                 pkt_len, buf_vec, &num_buffers,
1225                                                 avail_head, &nr_vec) < 0)) {
1226                         VHOST_LOG_DATA(DEBUG,
1227                                 "(%d) failed to get enough desc from vring\n",
1228                                 dev->vid);
1229                         vq->shadow_used_idx -= num_buffers;
1230                         break;
1231                 }
1232
1233                 VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
1234                         dev->vid, vq->last_avail_idx,
1235                         vq->last_avail_idx + num_buffers);
1236
1237                 if (copy_mbuf_to_desc(dev, vq, pkts[pkt_idx],
1238                                                 buf_vec, nr_vec,
1239                                                 num_buffers) < 0) {
1240                         vq->shadow_used_idx -= num_buffers;
1241                         break;
1242                 }
1243
1244                 vq->last_avail_idx += num_buffers;
1245         }
1246
1247         do_data_copy_enqueue(dev, vq);
1248
1249         if (likely(vq->shadow_used_idx)) {
1250                 flush_shadow_used_ring_split(dev, vq);
1251                 vhost_vring_call_split(dev, vq);
1252         }
1253
1254         return pkt_idx;
1255 }
1256
1257 static __rte_always_inline int
1258 virtio_dev_rx_batch_packed(struct virtio_net *dev,
1259                            struct vhost_virtqueue *vq,
1260                            struct rte_mbuf **pkts)
1261 {
1262         bool wrap_counter = vq->avail_wrap_counter;
1263         struct vring_packed_desc *descs = vq->desc_packed;
1264         uint16_t avail_idx = vq->last_avail_idx;
1265         uint64_t desc_addrs[PACKED_BATCH_SIZE];
1266         struct virtio_net_hdr_mrg_rxbuf *hdrs[PACKED_BATCH_SIZE];
1267         uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1268         uint64_t lens[PACKED_BATCH_SIZE];
1269         uint16_t ids[PACKED_BATCH_SIZE];
1270         uint16_t i;
1271
1272         if (unlikely(avail_idx & PACKED_BATCH_MASK))
1273                 return -1;
1274
1275         if (unlikely((avail_idx + PACKED_BATCH_SIZE) > vq->size))
1276                 return -1;
1277
1278         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1279                 if (unlikely(pkts[i]->next != NULL))
1280                         return -1;
1281                 if (unlikely(!desc_is_avail(&descs[avail_idx + i],
1282                                             wrap_counter)))
1283                         return -1;
1284         }
1285
1286         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1287                 lens[i] = descs[avail_idx + i].len;
1288
1289         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1290                 if (unlikely(pkts[i]->pkt_len > (lens[i] - buf_offset)))
1291                         return -1;
1292         }
1293
1294         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1295                 desc_addrs[i] = vhost_iova_to_vva(dev, vq,
1296                                                   descs[avail_idx + i].addr,
1297                                                   &lens[i],
1298                                                   VHOST_ACCESS_RW);
1299
1300         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1301                 if (unlikely(!desc_addrs[i]))
1302                         return -1;
1303                 if (unlikely(lens[i] != descs[avail_idx + i].len))
1304                         return -1;
1305         }
1306
1307         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1308                 rte_prefetch0((void *)(uintptr_t)desc_addrs[i]);
1309                 hdrs[i] = (struct virtio_net_hdr_mrg_rxbuf *)
1310                                         (uintptr_t)desc_addrs[i];
1311                 lens[i] = pkts[i]->pkt_len +
1312                         sizeof(struct virtio_net_hdr_mrg_rxbuf);
1313         }
1314
1315         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1316                 virtio_enqueue_offload(pkts[i], &hdrs[i]->hdr);
1317
1318         vq_inc_last_avail_packed(vq, PACKED_BATCH_SIZE);
1319
1320         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1321                 rte_memcpy((void *)(uintptr_t)(desc_addrs[i] + buf_offset),
1322                            rte_pktmbuf_mtod_offset(pkts[i], void *, 0),
1323                            pkts[i]->pkt_len);
1324         }
1325
1326         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1327                 vhost_log_cache_write_iova(dev, vq, descs[avail_idx + i].addr,
1328                                            lens[i]);
1329
1330         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1331                 ids[i] = descs[avail_idx + i].id;
1332
1333         vhost_flush_enqueue_batch_packed(dev, vq, lens, ids);
1334
1335         return 0;
1336 }
1337
1338 static __rte_always_inline int16_t
1339 virtio_dev_rx_single_packed(struct virtio_net *dev,
1340                             struct vhost_virtqueue *vq,
1341                             struct rte_mbuf *pkt)
1342 {
1343         struct buf_vector buf_vec[BUF_VECTOR_MAX];
1344         uint16_t nr_descs = 0;
1345
1346         if (unlikely(vhost_enqueue_single_packed(dev, vq, pkt, buf_vec,
1347                                                  &nr_descs) < 0)) {
1348                 VHOST_LOG_DATA(DEBUG,
1349                                 "(%d) failed to get enough desc from vring\n",
1350                                 dev->vid);
1351                 return -1;
1352         }
1353
1354         VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
1355                         dev->vid, vq->last_avail_idx,
1356                         vq->last_avail_idx + nr_descs);
1357
1358         vq_inc_last_avail_packed(vq, nr_descs);
1359
1360         return 0;
1361 }
1362
1363 static __rte_noinline uint32_t
1364 virtio_dev_rx_packed(struct virtio_net *dev,
1365                      struct vhost_virtqueue *__rte_restrict vq,
1366                      struct rte_mbuf **__rte_restrict pkts,
1367                      uint32_t count)
1368 {
1369         uint32_t pkt_idx = 0;
1370         uint32_t remained = count;
1371
1372         do {
1373                 rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
1374
1375                 if (remained >= PACKED_BATCH_SIZE) {
1376                         if (!virtio_dev_rx_batch_packed(dev, vq,
1377                                                         &pkts[pkt_idx])) {
1378                                 pkt_idx += PACKED_BATCH_SIZE;
1379                                 remained -= PACKED_BATCH_SIZE;
1380                                 continue;
1381                         }
1382                 }
1383
1384                 if (virtio_dev_rx_single_packed(dev, vq, pkts[pkt_idx]))
1385                         break;
1386                 pkt_idx++;
1387                 remained--;
1388
1389         } while (pkt_idx < count);
1390
1391         if (vq->shadow_used_idx) {
1392                 do_data_copy_enqueue(dev, vq);
1393                 vhost_flush_enqueue_shadow_packed(dev, vq);
1394         }
1395
1396         if (pkt_idx)
1397                 vhost_vring_call_packed(dev, vq);
1398
1399         return pkt_idx;
1400 }
1401
1402 static __rte_always_inline uint32_t
1403 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
1404         struct rte_mbuf **pkts, uint32_t count)
1405 {
1406         struct vhost_virtqueue *vq;
1407         uint32_t nb_tx = 0;
1408
1409         VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
1410         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
1411                 VHOST_LOG_DATA(ERR, "(%d) %s: invalid virtqueue idx %d.\n",
1412                         dev->vid, __func__, queue_id);
1413                 return 0;
1414         }
1415
1416         vq = dev->virtqueue[queue_id];
1417
1418         rte_spinlock_lock(&vq->access_lock);
1419
1420         if (unlikely(vq->enabled == 0))
1421                 goto out_access_unlock;
1422
1423         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1424                 vhost_user_iotlb_rd_lock(vq);
1425
1426         if (unlikely(vq->access_ok == 0))
1427                 if (unlikely(vring_translate(dev, vq) < 0))
1428                         goto out;
1429
1430         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
1431         if (count == 0)
1432                 goto out;
1433
1434         if (vq_is_packed(dev))
1435                 nb_tx = virtio_dev_rx_packed(dev, vq, pkts, count);
1436         else
1437                 nb_tx = virtio_dev_rx_split(dev, vq, pkts, count);
1438
1439 out:
1440         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1441                 vhost_user_iotlb_rd_unlock(vq);
1442
1443 out_access_unlock:
1444         rte_spinlock_unlock(&vq->access_lock);
1445
1446         return nb_tx;
1447 }
1448
1449 uint16_t
1450 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
1451         struct rte_mbuf **__rte_restrict pkts, uint16_t count)
1452 {
1453         struct virtio_net *dev = get_device(vid);
1454
1455         if (!dev)
1456                 return 0;
1457
1458         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
1459                 VHOST_LOG_DATA(ERR,
1460                         "(%d) %s: built-in vhost net backend is disabled.\n",
1461                         dev->vid, __func__);
1462                 return 0;
1463         }
1464
1465         return virtio_dev_rx(dev, queue_id, pkts, count);
1466 }
1467
1468 static __rte_always_inline uint16_t
1469 virtio_dev_rx_async_get_info_idx(uint16_t pkts_idx,
1470         uint16_t vq_size, uint16_t n_inflight)
1471 {
1472         return pkts_idx > n_inflight ? (pkts_idx - n_inflight) :
1473                 (vq_size - n_inflight + pkts_idx) & (vq_size - 1);
1474 }
1475
1476 static __rte_noinline uint32_t
1477 virtio_dev_rx_async_submit_split(struct virtio_net *dev,
1478         struct vhost_virtqueue *vq, uint16_t queue_id,
1479         struct rte_mbuf **pkts, uint32_t count)
1480 {
1481         uint32_t pkt_idx = 0, pkt_burst_idx = 0;
1482         uint16_t num_buffers;
1483         struct buf_vector buf_vec[BUF_VECTOR_MAX];
1484         uint16_t avail_head;
1485
1486         struct rte_vhost_iov_iter *it_pool = vq->it_pool;
1487         struct iovec *vec_pool = vq->vec_pool;
1488         struct rte_vhost_async_desc tdes[MAX_PKT_BURST];
1489         struct iovec *src_iovec = vec_pool;
1490         struct iovec *dst_iovec = vec_pool + (VHOST_MAX_ASYNC_VEC >> 1);
1491         struct rte_vhost_iov_iter *src_it = it_pool;
1492         struct rte_vhost_iov_iter *dst_it = it_pool + 1;
1493         uint16_t n_free_slot, slot_idx = 0;
1494         uint16_t pkt_err = 0;
1495         uint16_t segs_await = 0;
1496         struct async_inflight_info *pkts_info = vq->async_pkts_info;
1497         int n_pkts = 0;
1498
1499         /*
1500          * The ordering between avail index and desc reads need to be enforced.
1501          */
1502         avail_head = __atomic_load_n(&vq->avail->idx, __ATOMIC_ACQUIRE);
1503
1504         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
1505
1506         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
1507                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
1508                 uint16_t nr_vec = 0;
1509
1510                 if (unlikely(reserve_avail_buf_split(dev, vq,
1511                                                 pkt_len, buf_vec, &num_buffers,
1512                                                 avail_head, &nr_vec) < 0)) {
1513                         VHOST_LOG_DATA(DEBUG,
1514                                 "(%d) failed to get enough desc from vring\n",
1515                                 dev->vid);
1516                         vq->shadow_used_idx -= num_buffers;
1517                         break;
1518                 }
1519
1520                 VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
1521                         dev->vid, vq->last_avail_idx,
1522                         vq->last_avail_idx + num_buffers);
1523
1524                 if (async_mbuf_to_desc(dev, vq, pkts[pkt_idx],
1525                                 buf_vec, nr_vec, num_buffers,
1526                                 src_iovec, dst_iovec, src_it, dst_it) < 0) {
1527                         vq->shadow_used_idx -= num_buffers;
1528                         break;
1529                 }
1530
1531                 slot_idx = (vq->async_pkts_idx + pkt_idx) & (vq->size - 1);
1532                 if (src_it->count) {
1533                         async_fill_desc(&tdes[pkt_burst_idx], src_it, dst_it);
1534                         pkt_burst_idx++;
1535                         pkts_info[slot_idx].descs = num_buffers;
1536                         pkts_info[slot_idx].segs = src_it->nr_segs;
1537                         src_iovec += src_it->nr_segs;
1538                         dst_iovec += dst_it->nr_segs;
1539                         src_it += 2;
1540                         dst_it += 2;
1541                         segs_await += src_it->nr_segs;
1542                 } else {
1543                         pkts_info[slot_idx].info = num_buffers;
1544                         vq->async_pkts_inflight_n++;
1545                 }
1546
1547                 vq->last_avail_idx += num_buffers;
1548
1549                 /*
1550                  * conditions to trigger async device transfer:
1551                  * - buffered packet number reaches transfer threshold
1552                  * - this is the last packet in the burst enqueue
1553                  * - unused async iov number is less than max vhost vector
1554                  */
1555                 if (pkt_burst_idx >= VHOST_ASYNC_BATCH_THRESHOLD ||
1556                         (pkt_idx == count - 1 && pkt_burst_idx) ||
1557                         (VHOST_MAX_ASYNC_VEC / 2 - segs_await <
1558                         BUF_VECTOR_MAX)) {
1559                         n_pkts = vq->async_ops.transfer_data(dev->vid,
1560                                         queue_id, tdes, 0, pkt_burst_idx);
1561                         src_iovec = vec_pool;
1562                         dst_iovec = vec_pool + (VHOST_MAX_ASYNC_VEC >> 1);
1563                         src_it = it_pool;
1564                         dst_it = it_pool + 1;
1565                         segs_await = 0;
1566                         vq->async_pkts_inflight_n += pkt_burst_idx;
1567
1568                         if (unlikely(n_pkts < (int)pkt_burst_idx)) {
1569                                 /*
1570                                  * log error packets number here and do actual
1571                                  * error processing when applications poll
1572                                  * completion
1573                                  */
1574                                 pkt_err = pkt_burst_idx - n_pkts;
1575                                 pkt_burst_idx = 0;
1576                                 break;
1577                         }
1578
1579                         pkt_burst_idx = 0;
1580                 }
1581         }
1582
1583         if (pkt_burst_idx) {
1584                 n_pkts = vq->async_ops.transfer_data(dev->vid,
1585                                 queue_id, tdes, 0, pkt_burst_idx);
1586                 vq->async_pkts_inflight_n += pkt_burst_idx;
1587
1588                 if (unlikely(n_pkts < (int)pkt_burst_idx))
1589                         pkt_err = pkt_burst_idx - n_pkts;
1590         }
1591
1592         do_data_copy_enqueue(dev, vq);
1593
1594         while (unlikely(pkt_err && pkt_idx)) {
1595                 if (pkts_info[slot_idx].segs)
1596                         pkt_err--;
1597                 vq->last_avail_idx -= pkts_info[slot_idx].descs;
1598                 vq->shadow_used_idx -= pkts_info[slot_idx].descs;
1599                 vq->async_pkts_inflight_n--;
1600                 slot_idx = (slot_idx - 1) & (vq->size - 1);
1601                 pkt_idx--;
1602         }
1603
1604         n_free_slot = vq->size - vq->async_pkts_idx;
1605         if (n_free_slot > pkt_idx) {
1606                 rte_memcpy(&vq->async_pkts_pending[vq->async_pkts_idx],
1607                         pkts, pkt_idx * sizeof(uintptr_t));
1608                 vq->async_pkts_idx += pkt_idx;
1609         } else {
1610                 rte_memcpy(&vq->async_pkts_pending[vq->async_pkts_idx],
1611                         pkts, n_free_slot * sizeof(uintptr_t));
1612                 rte_memcpy(&vq->async_pkts_pending[0],
1613                         &pkts[n_free_slot],
1614                         (pkt_idx - n_free_slot) * sizeof(uintptr_t));
1615                 vq->async_pkts_idx = pkt_idx - n_free_slot;
1616         }
1617
1618         if (likely(vq->shadow_used_idx))
1619                 async_flush_shadow_used_ring_split(dev, vq);
1620
1621         return pkt_idx;
1622 }
1623
1624 uint16_t rte_vhost_poll_enqueue_completed(int vid, uint16_t queue_id,
1625                 struct rte_mbuf **pkts, uint16_t count)
1626 {
1627         struct virtio_net *dev = get_device(vid);
1628         struct vhost_virtqueue *vq;
1629         uint16_t n_pkts_cpl = 0, n_pkts_put = 0, n_descs = 0;
1630         uint16_t start_idx, pkts_idx, vq_size;
1631         uint16_t n_inflight;
1632         struct async_inflight_info *pkts_info;
1633
1634         if (!dev)
1635                 return 0;
1636
1637         VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
1638         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
1639                 VHOST_LOG_DATA(ERR, "(%d) %s: invalid virtqueue idx %d.\n",
1640                         dev->vid, __func__, queue_id);
1641                 return 0;
1642         }
1643
1644         vq = dev->virtqueue[queue_id];
1645
1646         if (unlikely(!vq->async_registered)) {
1647                 VHOST_LOG_DATA(ERR, "(%d) %s: async not registered for queue id %d.\n",
1648                         dev->vid, __func__, queue_id);
1649                 return 0;
1650         }
1651
1652         rte_spinlock_lock(&vq->access_lock);
1653
1654         n_inflight = vq->async_pkts_inflight_n;
1655         pkts_idx = vq->async_pkts_idx;
1656         pkts_info = vq->async_pkts_info;
1657         vq_size = vq->size;
1658         start_idx = virtio_dev_rx_async_get_info_idx(pkts_idx,
1659                 vq_size, vq->async_pkts_inflight_n);
1660
1661         if (count > vq->async_last_pkts_n)
1662                 n_pkts_cpl = vq->async_ops.check_completed_copies(vid,
1663                         queue_id, 0, count - vq->async_last_pkts_n);
1664         n_pkts_cpl += vq->async_last_pkts_n;
1665
1666         rte_smp_wmb();
1667
1668         while (likely((n_pkts_put < count) && n_inflight)) {
1669                 uint16_t info_idx = (start_idx + n_pkts_put) & (vq_size - 1);
1670                 if (n_pkts_cpl && pkts_info[info_idx].segs)
1671                         n_pkts_cpl--;
1672                 else if (!n_pkts_cpl && pkts_info[info_idx].segs)
1673                         break;
1674                 n_pkts_put++;
1675                 n_inflight--;
1676                 n_descs += pkts_info[info_idx].descs;
1677         }
1678
1679         vq->async_last_pkts_n = n_pkts_cpl;
1680
1681         if (n_pkts_put) {
1682                 vq->async_pkts_inflight_n = n_inflight;
1683                 if (likely(vq->enabled && vq->access_ok)) {
1684                         __atomic_add_fetch(&vq->used->idx,
1685                                         n_descs, __ATOMIC_RELEASE);
1686                         vhost_vring_call_split(dev, vq);
1687                 }
1688
1689                 if (start_idx + n_pkts_put <= vq_size) {
1690                         rte_memcpy(pkts, &vq->async_pkts_pending[start_idx],
1691                                 n_pkts_put * sizeof(uintptr_t));
1692                 } else {
1693                         rte_memcpy(pkts, &vq->async_pkts_pending[start_idx],
1694                                 (vq_size - start_idx) * sizeof(uintptr_t));
1695                         rte_memcpy(&pkts[vq_size - start_idx],
1696                                 vq->async_pkts_pending,
1697                                 (n_pkts_put + start_idx - vq_size) *
1698                                 sizeof(uintptr_t));
1699                 }
1700         }
1701
1702         rte_spinlock_unlock(&vq->access_lock);
1703
1704         return n_pkts_put;
1705 }
1706
1707 static __rte_always_inline uint32_t
1708 virtio_dev_rx_async_submit(struct virtio_net *dev, uint16_t queue_id,
1709         struct rte_mbuf **pkts, uint32_t count)
1710 {
1711         struct vhost_virtqueue *vq;
1712         uint32_t nb_tx = 0;
1713
1714         VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
1715         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
1716                 VHOST_LOG_DATA(ERR, "(%d) %s: invalid virtqueue idx %d.\n",
1717                         dev->vid, __func__, queue_id);
1718                 return 0;
1719         }
1720
1721         vq = dev->virtqueue[queue_id];
1722
1723         rte_spinlock_lock(&vq->access_lock);
1724
1725         if (unlikely(vq->enabled == 0 || !vq->async_registered))
1726                 goto out_access_unlock;
1727
1728         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1729                 vhost_user_iotlb_rd_lock(vq);
1730
1731         if (unlikely(vq->access_ok == 0))
1732                 if (unlikely(vring_translate(dev, vq) < 0))
1733                         goto out;
1734
1735         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
1736         if (count == 0)
1737                 goto out;
1738
1739         /* TODO: packed queue not implemented */
1740         if (vq_is_packed(dev))
1741                 nb_tx = 0;
1742         else
1743                 nb_tx = virtio_dev_rx_async_submit_split(dev,
1744                                 vq, queue_id, pkts, count);
1745
1746 out:
1747         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1748                 vhost_user_iotlb_rd_unlock(vq);
1749
1750 out_access_unlock:
1751         rte_spinlock_unlock(&vq->access_lock);
1752
1753         return nb_tx;
1754 }
1755
1756 uint16_t
1757 rte_vhost_submit_enqueue_burst(int vid, uint16_t queue_id,
1758                 struct rte_mbuf **pkts, uint16_t count)
1759 {
1760         struct virtio_net *dev = get_device(vid);
1761
1762         if (!dev)
1763                 return 0;
1764
1765         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
1766                 VHOST_LOG_DATA(ERR,
1767                         "(%d) %s: built-in vhost net backend is disabled.\n",
1768                         dev->vid, __func__);
1769                 return 0;
1770         }
1771
1772         return virtio_dev_rx_async_submit(dev, queue_id, pkts, count);
1773 }
1774
1775 static inline bool
1776 virtio_net_with_host_offload(struct virtio_net *dev)
1777 {
1778         if (dev->features &
1779                         ((1ULL << VIRTIO_NET_F_CSUM) |
1780                          (1ULL << VIRTIO_NET_F_HOST_ECN) |
1781                          (1ULL << VIRTIO_NET_F_HOST_TSO4) |
1782                          (1ULL << VIRTIO_NET_F_HOST_TSO6) |
1783                          (1ULL << VIRTIO_NET_F_HOST_UFO)))
1784                 return true;
1785
1786         return false;
1787 }
1788
1789 static void
1790 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
1791 {
1792         struct rte_ipv4_hdr *ipv4_hdr;
1793         struct rte_ipv6_hdr *ipv6_hdr;
1794         void *l3_hdr = NULL;
1795         struct rte_ether_hdr *eth_hdr;
1796         uint16_t ethertype;
1797
1798         eth_hdr = rte_pktmbuf_mtod(m, struct rte_ether_hdr *);
1799
1800         m->l2_len = sizeof(struct rte_ether_hdr);
1801         ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
1802
1803         if (ethertype == RTE_ETHER_TYPE_VLAN) {
1804                 struct rte_vlan_hdr *vlan_hdr =
1805                         (struct rte_vlan_hdr *)(eth_hdr + 1);
1806
1807                 m->l2_len += sizeof(struct rte_vlan_hdr);
1808                 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
1809         }
1810
1811         l3_hdr = (char *)eth_hdr + m->l2_len;
1812
1813         switch (ethertype) {
1814         case RTE_ETHER_TYPE_IPV4:
1815                 ipv4_hdr = l3_hdr;
1816                 *l4_proto = ipv4_hdr->next_proto_id;
1817                 m->l3_len = rte_ipv4_hdr_len(ipv4_hdr);
1818                 *l4_hdr = (char *)l3_hdr + m->l3_len;
1819                 m->ol_flags |= PKT_TX_IPV4;
1820                 break;
1821         case RTE_ETHER_TYPE_IPV6:
1822                 ipv6_hdr = l3_hdr;
1823                 *l4_proto = ipv6_hdr->proto;
1824                 m->l3_len = sizeof(struct rte_ipv6_hdr);
1825                 *l4_hdr = (char *)l3_hdr + m->l3_len;
1826                 m->ol_flags |= PKT_TX_IPV6;
1827                 break;
1828         default:
1829                 m->l3_len = 0;
1830                 *l4_proto = 0;
1831                 *l4_hdr = NULL;
1832                 break;
1833         }
1834 }
1835
1836 static __rte_always_inline void
1837 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
1838 {
1839         uint16_t l4_proto = 0;
1840         void *l4_hdr = NULL;
1841         struct rte_tcp_hdr *tcp_hdr = NULL;
1842
1843         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
1844                 return;
1845
1846         parse_ethernet(m, &l4_proto, &l4_hdr);
1847         if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
1848                 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
1849                         switch (hdr->csum_offset) {
1850                         case (offsetof(struct rte_tcp_hdr, cksum)):
1851                                 if (l4_proto == IPPROTO_TCP)
1852                                         m->ol_flags |= PKT_TX_TCP_CKSUM;
1853                                 break;
1854                         case (offsetof(struct rte_udp_hdr, dgram_cksum)):
1855                                 if (l4_proto == IPPROTO_UDP)
1856                                         m->ol_flags |= PKT_TX_UDP_CKSUM;
1857                                 break;
1858                         case (offsetof(struct rte_sctp_hdr, cksum)):
1859                                 if (l4_proto == IPPROTO_SCTP)
1860                                         m->ol_flags |= PKT_TX_SCTP_CKSUM;
1861                                 break;
1862                         default:
1863                                 break;
1864                         }
1865                 }
1866         }
1867
1868         if (l4_hdr && hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
1869                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
1870                 case VIRTIO_NET_HDR_GSO_TCPV4:
1871                 case VIRTIO_NET_HDR_GSO_TCPV6:
1872                         tcp_hdr = l4_hdr;
1873                         m->ol_flags |= PKT_TX_TCP_SEG;
1874                         m->tso_segsz = hdr->gso_size;
1875                         m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
1876                         break;
1877                 case VIRTIO_NET_HDR_GSO_UDP:
1878                         m->ol_flags |= PKT_TX_UDP_SEG;
1879                         m->tso_segsz = hdr->gso_size;
1880                         m->l4_len = sizeof(struct rte_udp_hdr);
1881                         break;
1882                 default:
1883                         VHOST_LOG_DATA(WARNING,
1884                                 "unsupported gso type %u.\n", hdr->gso_type);
1885                         break;
1886                 }
1887         }
1888 }
1889
1890 static __rte_noinline void
1891 copy_vnet_hdr_from_desc(struct virtio_net_hdr *hdr,
1892                 struct buf_vector *buf_vec)
1893 {
1894         uint64_t len;
1895         uint64_t remain = sizeof(struct virtio_net_hdr);
1896         uint64_t src;
1897         uint64_t dst = (uint64_t)(uintptr_t)hdr;
1898
1899         while (remain) {
1900                 len = RTE_MIN(remain, buf_vec->buf_len);
1901                 src = buf_vec->buf_addr;
1902                 rte_memcpy((void *)(uintptr_t)dst,
1903                                 (void *)(uintptr_t)src, len);
1904
1905                 remain -= len;
1906                 dst += len;
1907                 buf_vec++;
1908         }
1909 }
1910
1911 static __rte_always_inline int
1912 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
1913                   struct buf_vector *buf_vec, uint16_t nr_vec,
1914                   struct rte_mbuf *m, struct rte_mempool *mbuf_pool)
1915 {
1916         uint32_t buf_avail, buf_offset;
1917         uint64_t buf_addr, buf_len;
1918         uint32_t mbuf_avail, mbuf_offset;
1919         uint32_t cpy_len;
1920         struct rte_mbuf *cur = m, *prev = m;
1921         struct virtio_net_hdr tmp_hdr;
1922         struct virtio_net_hdr *hdr = NULL;
1923         /* A counter to avoid desc dead loop chain */
1924         uint16_t vec_idx = 0;
1925         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
1926         int error = 0;
1927
1928         buf_addr = buf_vec[vec_idx].buf_addr;
1929         buf_len = buf_vec[vec_idx].buf_len;
1930
1931         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
1932                 error = -1;
1933                 goto out;
1934         }
1935
1936         if (virtio_net_with_host_offload(dev)) {
1937                 if (unlikely(buf_len < sizeof(struct virtio_net_hdr))) {
1938                         /*
1939                          * No luck, the virtio-net header doesn't fit
1940                          * in a contiguous virtual area.
1941                          */
1942                         copy_vnet_hdr_from_desc(&tmp_hdr, buf_vec);
1943                         hdr = &tmp_hdr;
1944                 } else {
1945                         hdr = (struct virtio_net_hdr *)((uintptr_t)buf_addr);
1946                 }
1947         }
1948
1949         /*
1950          * A virtio driver normally uses at least 2 desc buffers
1951          * for Tx: the first for storing the header, and others
1952          * for storing the data.
1953          */
1954         if (unlikely(buf_len < dev->vhost_hlen)) {
1955                 buf_offset = dev->vhost_hlen - buf_len;
1956                 vec_idx++;
1957                 buf_addr = buf_vec[vec_idx].buf_addr;
1958                 buf_len = buf_vec[vec_idx].buf_len;
1959                 buf_avail  = buf_len - buf_offset;
1960         } else if (buf_len == dev->vhost_hlen) {
1961                 if (unlikely(++vec_idx >= nr_vec))
1962                         goto out;
1963                 buf_addr = buf_vec[vec_idx].buf_addr;
1964                 buf_len = buf_vec[vec_idx].buf_len;
1965
1966                 buf_offset = 0;
1967                 buf_avail = buf_len;
1968         } else {
1969                 buf_offset = dev->vhost_hlen;
1970                 buf_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
1971         }
1972
1973         PRINT_PACKET(dev,
1974                         (uintptr_t)(buf_addr + buf_offset),
1975                         (uint32_t)buf_avail, 0);
1976
1977         mbuf_offset = 0;
1978         mbuf_avail  = m->buf_len - RTE_PKTMBUF_HEADROOM;
1979         while (1) {
1980                 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
1981
1982                 if (likely(cpy_len > MAX_BATCH_LEN ||
1983                                         vq->batch_copy_nb_elems >= vq->size ||
1984                                         (hdr && cur == m))) {
1985                         rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
1986                                                 mbuf_offset),
1987                                         (void *)((uintptr_t)(buf_addr +
1988                                                         buf_offset)), cpy_len);
1989                 } else {
1990                         batch_copy[vq->batch_copy_nb_elems].dst =
1991                                 rte_pktmbuf_mtod_offset(cur, void *,
1992                                                 mbuf_offset);
1993                         batch_copy[vq->batch_copy_nb_elems].src =
1994                                 (void *)((uintptr_t)(buf_addr + buf_offset));
1995                         batch_copy[vq->batch_copy_nb_elems].len = cpy_len;
1996                         vq->batch_copy_nb_elems++;
1997                 }
1998
1999                 mbuf_avail  -= cpy_len;
2000                 mbuf_offset += cpy_len;
2001                 buf_avail -= cpy_len;
2002                 buf_offset += cpy_len;
2003
2004                 /* This buf reaches to its end, get the next one */
2005                 if (buf_avail == 0) {
2006                         if (++vec_idx >= nr_vec)
2007                                 break;
2008
2009                         buf_addr = buf_vec[vec_idx].buf_addr;
2010                         buf_len = buf_vec[vec_idx].buf_len;
2011
2012                         buf_offset = 0;
2013                         buf_avail  = buf_len;
2014
2015                         PRINT_PACKET(dev, (uintptr_t)buf_addr,
2016                                         (uint32_t)buf_avail, 0);
2017                 }
2018
2019                 /*
2020                  * This mbuf reaches to its end, get a new one
2021                  * to hold more data.
2022                  */
2023                 if (mbuf_avail == 0) {
2024                         cur = rte_pktmbuf_alloc(mbuf_pool);
2025                         if (unlikely(cur == NULL)) {
2026                                 VHOST_LOG_DATA(ERR, "Failed to "
2027                                         "allocate memory for mbuf.\n");
2028                                 error = -1;
2029                                 goto out;
2030                         }
2031
2032                         prev->next = cur;
2033                         prev->data_len = mbuf_offset;
2034                         m->nb_segs += 1;
2035                         m->pkt_len += mbuf_offset;
2036                         prev = cur;
2037
2038                         mbuf_offset = 0;
2039                         mbuf_avail  = cur->buf_len - RTE_PKTMBUF_HEADROOM;
2040                 }
2041         }
2042
2043         prev->data_len = mbuf_offset;
2044         m->pkt_len    += mbuf_offset;
2045
2046         if (hdr)
2047                 vhost_dequeue_offload(hdr, m);
2048
2049 out:
2050
2051         return error;
2052 }
2053
2054 static void
2055 virtio_dev_extbuf_free(void *addr __rte_unused, void *opaque)
2056 {
2057         rte_free(opaque);
2058 }
2059
2060 static int
2061 virtio_dev_extbuf_alloc(struct rte_mbuf *pkt, uint32_t size)
2062 {
2063         struct rte_mbuf_ext_shared_info *shinfo = NULL;
2064         uint32_t total_len = RTE_PKTMBUF_HEADROOM + size;
2065         uint16_t buf_len;
2066         rte_iova_t iova;
2067         void *buf;
2068
2069         total_len += sizeof(*shinfo) + sizeof(uintptr_t);
2070         total_len = RTE_ALIGN_CEIL(total_len, sizeof(uintptr_t));
2071
2072         if (unlikely(total_len > UINT16_MAX))
2073                 return -ENOSPC;
2074
2075         buf_len = total_len;
2076         buf = rte_malloc(NULL, buf_len, RTE_CACHE_LINE_SIZE);
2077         if (unlikely(buf == NULL))
2078                 return -ENOMEM;
2079
2080         /* Initialize shinfo */
2081         shinfo = rte_pktmbuf_ext_shinfo_init_helper(buf, &buf_len,
2082                                                 virtio_dev_extbuf_free, buf);
2083         if (unlikely(shinfo == NULL)) {
2084                 rte_free(buf);
2085                 VHOST_LOG_DATA(ERR, "Failed to init shinfo\n");
2086                 return -1;
2087         }
2088
2089         iova = rte_malloc_virt2iova(buf);
2090         rte_pktmbuf_attach_extbuf(pkt, buf, iova, buf_len, shinfo);
2091         rte_pktmbuf_reset_headroom(pkt);
2092
2093         return 0;
2094 }
2095
2096 /*
2097  * Allocate a host supported pktmbuf.
2098  */
2099 static __rte_always_inline struct rte_mbuf *
2100 virtio_dev_pktmbuf_alloc(struct virtio_net *dev, struct rte_mempool *mp,
2101                          uint32_t data_len)
2102 {
2103         struct rte_mbuf *pkt = rte_pktmbuf_alloc(mp);
2104
2105         if (unlikely(pkt == NULL)) {
2106                 VHOST_LOG_DATA(ERR,
2107                         "Failed to allocate memory for mbuf.\n");
2108                 return NULL;
2109         }
2110
2111         if (rte_pktmbuf_tailroom(pkt) >= data_len)
2112                 return pkt;
2113
2114         /* attach an external buffer if supported */
2115         if (dev->extbuf && !virtio_dev_extbuf_alloc(pkt, data_len))
2116                 return pkt;
2117
2118         /* check if chained buffers are allowed */
2119         if (!dev->linearbuf)
2120                 return pkt;
2121
2122         /* Data doesn't fit into the buffer and the host supports
2123          * only linear buffers
2124          */
2125         rte_pktmbuf_free(pkt);
2126
2127         return NULL;
2128 }
2129
2130 static __rte_noinline uint16_t
2131 virtio_dev_tx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
2132         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
2133 {
2134         uint16_t i;
2135         uint16_t free_entries;
2136         uint16_t dropped = 0;
2137         static bool allocerr_warned;
2138
2139         /*
2140          * The ordering between avail index and
2141          * desc reads needs to be enforced.
2142          */
2143         free_entries = __atomic_load_n(&vq->avail->idx, __ATOMIC_ACQUIRE) -
2144                         vq->last_avail_idx;
2145         if (free_entries == 0)
2146                 return 0;
2147
2148         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
2149
2150         VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
2151
2152         count = RTE_MIN(count, MAX_PKT_BURST);
2153         count = RTE_MIN(count, free_entries);
2154         VHOST_LOG_DATA(DEBUG, "(%d) about to dequeue %u buffers\n",
2155                         dev->vid, count);
2156
2157         for (i = 0; i < count; i++) {
2158                 struct buf_vector buf_vec[BUF_VECTOR_MAX];
2159                 uint16_t head_idx;
2160                 uint32_t buf_len;
2161                 uint16_t nr_vec = 0;
2162                 int err;
2163
2164                 if (unlikely(fill_vec_buf_split(dev, vq,
2165                                                 vq->last_avail_idx + i,
2166                                                 &nr_vec, buf_vec,
2167                                                 &head_idx, &buf_len,
2168                                                 VHOST_ACCESS_RO) < 0))
2169                         break;
2170
2171                 update_shadow_used_ring_split(vq, head_idx, 0);
2172
2173                 pkts[i] = virtio_dev_pktmbuf_alloc(dev, mbuf_pool, buf_len);
2174                 if (unlikely(pkts[i] == NULL)) {
2175                         /*
2176                          * mbuf allocation fails for jumbo packets when external
2177                          * buffer allocation is not allowed and linear buffer
2178                          * is required. Drop this packet.
2179                          */
2180                         if (!allocerr_warned) {
2181                                 VHOST_LOG_DATA(ERR,
2182                                         "Failed mbuf alloc of size %d from %s on %s.\n",
2183                                         buf_len, mbuf_pool->name, dev->ifname);
2184                                 allocerr_warned = true;
2185                         }
2186                         dropped += 1;
2187                         i++;
2188                         break;
2189                 }
2190
2191                 err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts[i],
2192                                 mbuf_pool);
2193                 if (unlikely(err)) {
2194                         rte_pktmbuf_free(pkts[i]);
2195                         if (!allocerr_warned) {
2196                                 VHOST_LOG_DATA(ERR,
2197                                         "Failed to copy desc to mbuf on %s.\n",
2198                                         dev->ifname);
2199                                 allocerr_warned = true;
2200                         }
2201                         dropped += 1;
2202                         i++;
2203                         break;
2204                 }
2205         }
2206
2207         vq->last_avail_idx += i;
2208
2209         do_data_copy_dequeue(vq);
2210         if (unlikely(i < count))
2211                 vq->shadow_used_idx = i;
2212         if (likely(vq->shadow_used_idx)) {
2213                 flush_shadow_used_ring_split(dev, vq);
2214                 vhost_vring_call_split(dev, vq);
2215         }
2216
2217         return (i - dropped);
2218 }
2219
2220 static __rte_always_inline int
2221 vhost_reserve_avail_batch_packed(struct virtio_net *dev,
2222                                  struct vhost_virtqueue *vq,
2223                                  struct rte_mempool *mbuf_pool,
2224                                  struct rte_mbuf **pkts,
2225                                  uint16_t avail_idx,
2226                                  uintptr_t *desc_addrs,
2227                                  uint16_t *ids)
2228 {
2229         bool wrap = vq->avail_wrap_counter;
2230         struct vring_packed_desc *descs = vq->desc_packed;
2231         struct virtio_net_hdr *hdr;
2232         uint64_t lens[PACKED_BATCH_SIZE];
2233         uint64_t buf_lens[PACKED_BATCH_SIZE];
2234         uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2235         uint16_t flags, i;
2236
2237         if (unlikely(avail_idx & PACKED_BATCH_MASK))
2238                 return -1;
2239         if (unlikely((avail_idx + PACKED_BATCH_SIZE) > vq->size))
2240                 return -1;
2241
2242         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2243                 flags = descs[avail_idx + i].flags;
2244                 if (unlikely((wrap != !!(flags & VRING_DESC_F_AVAIL)) ||
2245                              (wrap == !!(flags & VRING_DESC_F_USED))  ||
2246                              (flags & PACKED_DESC_SINGLE_DEQUEUE_FLAG)))
2247                         return -1;
2248         }
2249
2250         rte_atomic_thread_fence(__ATOMIC_ACQUIRE);
2251
2252         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2253                 lens[i] = descs[avail_idx + i].len;
2254
2255         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2256                 desc_addrs[i] = vhost_iova_to_vva(dev, vq,
2257                                                   descs[avail_idx + i].addr,
2258                                                   &lens[i], VHOST_ACCESS_RW);
2259         }
2260
2261         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2262                 if (unlikely(!desc_addrs[i]))
2263                         return -1;
2264                 if (unlikely((lens[i] != descs[avail_idx + i].len)))
2265                         return -1;
2266         }
2267
2268         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2269                 pkts[i] = virtio_dev_pktmbuf_alloc(dev, mbuf_pool, lens[i]);
2270                 if (!pkts[i])
2271                         goto free_buf;
2272         }
2273
2274         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2275                 buf_lens[i] = pkts[i]->buf_len - pkts[i]->data_off;
2276
2277         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2278                 if (unlikely(buf_lens[i] < (lens[i] - buf_offset)))
2279                         goto free_buf;
2280         }
2281
2282         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2283                 pkts[i]->pkt_len = descs[avail_idx + i].len - buf_offset;
2284                 pkts[i]->data_len = pkts[i]->pkt_len;
2285                 ids[i] = descs[avail_idx + i].id;
2286         }
2287
2288         if (virtio_net_with_host_offload(dev)) {
2289                 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2290                         hdr = (struct virtio_net_hdr *)(desc_addrs[i]);
2291                         vhost_dequeue_offload(hdr, pkts[i]);
2292                 }
2293         }
2294
2295         return 0;
2296
2297 free_buf:
2298         for (i = 0; i < PACKED_BATCH_SIZE; i++)
2299                 rte_pktmbuf_free(pkts[i]);
2300
2301         return -1;
2302 }
2303
2304 static __rte_always_inline int
2305 virtio_dev_tx_batch_packed(struct virtio_net *dev,
2306                            struct vhost_virtqueue *vq,
2307                            struct rte_mempool *mbuf_pool,
2308                            struct rte_mbuf **pkts)
2309 {
2310         uint16_t avail_idx = vq->last_avail_idx;
2311         uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2312         uintptr_t desc_addrs[PACKED_BATCH_SIZE];
2313         uint16_t ids[PACKED_BATCH_SIZE];
2314         uint16_t i;
2315
2316         if (vhost_reserve_avail_batch_packed(dev, vq, mbuf_pool, pkts,
2317                                              avail_idx, desc_addrs, ids))
2318                 return -1;
2319
2320         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2321                 rte_prefetch0((void *)(uintptr_t)desc_addrs[i]);
2322
2323         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2324                 rte_memcpy(rte_pktmbuf_mtod_offset(pkts[i], void *, 0),
2325                            (void *)(uintptr_t)(desc_addrs[i] + buf_offset),
2326                            pkts[i]->pkt_len);
2327
2328         if (virtio_net_is_inorder(dev))
2329                 vhost_shadow_dequeue_batch_packed_inorder(vq,
2330                         ids[PACKED_BATCH_SIZE - 1]);
2331         else
2332                 vhost_shadow_dequeue_batch_packed(dev, vq, ids);
2333
2334         vq_inc_last_avail_packed(vq, PACKED_BATCH_SIZE);
2335
2336         return 0;
2337 }
2338
2339 static __rte_always_inline int
2340 vhost_dequeue_single_packed(struct virtio_net *dev,
2341                             struct vhost_virtqueue *vq,
2342                             struct rte_mempool *mbuf_pool,
2343                             struct rte_mbuf **pkts,
2344                             uint16_t *buf_id,
2345                             uint16_t *desc_count)
2346 {
2347         struct buf_vector buf_vec[BUF_VECTOR_MAX];
2348         uint32_t buf_len;
2349         uint16_t nr_vec = 0;
2350         int err;
2351         static bool allocerr_warned;
2352
2353         if (unlikely(fill_vec_buf_packed(dev, vq,
2354                                          vq->last_avail_idx, desc_count,
2355                                          buf_vec, &nr_vec,
2356                                          buf_id, &buf_len,
2357                                          VHOST_ACCESS_RO) < 0))
2358                 return -1;
2359
2360         *pkts = virtio_dev_pktmbuf_alloc(dev, mbuf_pool, buf_len);
2361         if (unlikely(*pkts == NULL)) {
2362                 if (!allocerr_warned) {
2363                         VHOST_LOG_DATA(ERR,
2364                                 "Failed mbuf alloc of size %d from %s on %s.\n",
2365                                 buf_len, mbuf_pool->name, dev->ifname);
2366                         allocerr_warned = true;
2367                 }
2368                 return -1;
2369         }
2370
2371         err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, *pkts,
2372                                 mbuf_pool);
2373         if (unlikely(err)) {
2374                 if (!allocerr_warned) {
2375                         VHOST_LOG_DATA(ERR,
2376                                 "Failed to copy desc to mbuf on %s.\n",
2377                                 dev->ifname);
2378                         allocerr_warned = true;
2379                 }
2380                 rte_pktmbuf_free(*pkts);
2381                 return -1;
2382         }
2383
2384         return 0;
2385 }
2386
2387 static __rte_always_inline int
2388 virtio_dev_tx_single_packed(struct virtio_net *dev,
2389                             struct vhost_virtqueue *vq,
2390                             struct rte_mempool *mbuf_pool,
2391                             struct rte_mbuf **pkts)
2392 {
2393
2394         uint16_t buf_id, desc_count = 0;
2395         int ret;
2396
2397         ret = vhost_dequeue_single_packed(dev, vq, mbuf_pool, pkts, &buf_id,
2398                                         &desc_count);
2399
2400         if (likely(desc_count > 0)) {
2401                 if (virtio_net_is_inorder(dev))
2402                         vhost_shadow_dequeue_single_packed_inorder(vq, buf_id,
2403                                                                    desc_count);
2404                 else
2405                         vhost_shadow_dequeue_single_packed(vq, buf_id,
2406                                         desc_count);
2407
2408                 vq_inc_last_avail_packed(vq, desc_count);
2409         }
2410
2411         return ret;
2412 }
2413
2414 static __rte_noinline uint16_t
2415 virtio_dev_tx_packed(struct virtio_net *dev,
2416                      struct vhost_virtqueue *__rte_restrict vq,
2417                      struct rte_mempool *mbuf_pool,
2418                      struct rte_mbuf **__rte_restrict pkts,
2419                      uint32_t count)
2420 {
2421         uint32_t pkt_idx = 0;
2422         uint32_t remained = count;
2423
2424         do {
2425                 rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
2426
2427                 if (remained >= PACKED_BATCH_SIZE) {
2428                         if (!virtio_dev_tx_batch_packed(dev, vq, mbuf_pool,
2429                                                         &pkts[pkt_idx])) {
2430                                 pkt_idx += PACKED_BATCH_SIZE;
2431                                 remained -= PACKED_BATCH_SIZE;
2432                                 continue;
2433                         }
2434                 }
2435
2436                 if (virtio_dev_tx_single_packed(dev, vq, mbuf_pool,
2437                                                 &pkts[pkt_idx]))
2438                         break;
2439                 pkt_idx++;
2440                 remained--;
2441
2442         } while (remained);
2443
2444         if (vq->shadow_used_idx) {
2445                 do_data_copy_dequeue(vq);
2446
2447                 vhost_flush_dequeue_shadow_packed(dev, vq);
2448                 vhost_vring_call_packed(dev, vq);
2449         }
2450
2451         return pkt_idx;
2452 }
2453
2454 uint16_t
2455 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
2456         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
2457 {
2458         struct virtio_net *dev;
2459         struct rte_mbuf *rarp_mbuf = NULL;
2460         struct vhost_virtqueue *vq;
2461         int16_t success = 1;
2462
2463         dev = get_device(vid);
2464         if (!dev)
2465                 return 0;
2466
2467         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
2468                 VHOST_LOG_DATA(ERR,
2469                         "(%d) %s: built-in vhost net backend is disabled.\n",
2470                         dev->vid, __func__);
2471                 return 0;
2472         }
2473
2474         if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
2475                 VHOST_LOG_DATA(ERR,
2476                         "(%d) %s: invalid virtqueue idx %d.\n",
2477                         dev->vid, __func__, queue_id);
2478                 return 0;
2479         }
2480
2481         vq = dev->virtqueue[queue_id];
2482
2483         if (unlikely(rte_spinlock_trylock(&vq->access_lock) == 0))
2484                 return 0;
2485
2486         if (unlikely(vq->enabled == 0)) {
2487                 count = 0;
2488                 goto out_access_unlock;
2489         }
2490
2491         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
2492                 vhost_user_iotlb_rd_lock(vq);
2493
2494         if (unlikely(vq->access_ok == 0))
2495                 if (unlikely(vring_translate(dev, vq) < 0)) {
2496                         count = 0;
2497                         goto out;
2498                 }
2499
2500         /*
2501          * Construct a RARP broadcast packet, and inject it to the "pkts"
2502          * array, to looks like that guest actually send such packet.
2503          *
2504          * Check user_send_rarp() for more information.
2505          *
2506          * broadcast_rarp shares a cacheline in the virtio_net structure
2507          * with some fields that are accessed during enqueue and
2508          * __atomic_compare_exchange_n causes a write if performed compare
2509          * and exchange. This could result in false sharing between enqueue
2510          * and dequeue.
2511          *
2512          * Prevent unnecessary false sharing by reading broadcast_rarp first
2513          * and only performing compare and exchange if the read indicates it
2514          * is likely to be set.
2515          */
2516         if (unlikely(__atomic_load_n(&dev->broadcast_rarp, __ATOMIC_ACQUIRE) &&
2517                         __atomic_compare_exchange_n(&dev->broadcast_rarp,
2518                         &success, 0, 0, __ATOMIC_RELEASE, __ATOMIC_RELAXED))) {
2519
2520                 rarp_mbuf = rte_net_make_rarp_packet(mbuf_pool, &dev->mac);
2521                 if (rarp_mbuf == NULL) {
2522                         VHOST_LOG_DATA(ERR, "Failed to make RARP packet.\n");
2523                         count = 0;
2524                         goto out;
2525                 }
2526                 count -= 1;
2527         }
2528
2529         if (vq_is_packed(dev))
2530                 count = virtio_dev_tx_packed(dev, vq, mbuf_pool, pkts, count);
2531         else
2532                 count = virtio_dev_tx_split(dev, vq, mbuf_pool, pkts, count);
2533
2534 out:
2535         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
2536                 vhost_user_iotlb_rd_unlock(vq);
2537
2538 out_access_unlock:
2539         rte_spinlock_unlock(&vq->access_lock);
2540
2541         if (unlikely(rarp_mbuf != NULL)) {
2542                 /*
2543                  * Inject it to the head of "pkts" array, so that switch's mac
2544                  * learning table will get updated first.
2545                  */
2546                 memmove(&pkts[1], pkts, count * sizeof(struct rte_mbuf *));
2547                 pkts[0] = rarp_mbuf;
2548                 count += 1;
2549         }
2550
2551         return count;
2552 }