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         rte_smp_wmb();
 175
 176         for (i = 0; i < vq->shadow_used_idx; i++) {
 177                 uint16_t flags;
 178
 179                 if (vq->shadow_used_packed[i].len)
 180                         flags = VRING_DESC_F_WRITE;
 181                 else
 182                         flags = 0;
 183
 184                 if (vq->used_wrap_counter) {
 185                         flags |= VRING_DESC_F_USED;
 186                         flags |= VRING_DESC_F_AVAIL;
 187                 } else {
 188                         flags &= ~VRING_DESC_F_USED;
 189                         flags &= ~VRING_DESC_F_AVAIL;
 190                 }
 191
 192                 if (i > 0) {
 193                         vq->desc_packed[vq->last_used_idx].flags = flags;
 194
 195                         vhost_log_cache_used_vring(dev, vq,
 196                                         vq->last_used_idx *
 197                                         sizeof(struct vring_packed_desc),
 198                                         sizeof(struct vring_packed_desc));
 199                 } else {
 200                         head_idx = vq->last_used_idx;
 201                         head_flags = flags;
 202                 }
 203
 204                 vq_inc_last_used_packed(vq, vq->shadow_used_packed[i].count);
 205         }
 206
 207         vq->desc_packed[head_idx].flags = head_flags;
 208
 209         vhost_log_cache_used_vring(dev, vq,
 210                                 head_idx *
 211                                 sizeof(struct vring_packed_desc),
 212                                 sizeof(struct vring_packed_desc));
 213
 214         vq->shadow_used_idx = 0;
 215         vhost_log_cache_sync(dev, vq);
 216 }
 217
 218 static __rte_always_inline void
 219 vhost_flush_dequeue_shadow_packed(struct virtio_net *dev,
 220                                   struct vhost_virtqueue *vq)
 221 {
 222         struct vring_used_elem_packed *used_elem = &vq->shadow_used_packed[0];
 223
 224         vq->desc_packed[vq->shadow_last_used_idx].id = used_elem->id;
 225         rte_smp_wmb();
 226         vq->desc_packed[vq->shadow_last_used_idx].flags = used_elem->flags;
 227
 228         vhost_log_cache_used_vring(dev, vq, vq->shadow_last_used_idx *
 229                                    sizeof(struct vring_packed_desc),
 230                                    sizeof(struct vring_packed_desc));
 231         vq->shadow_used_idx = 0;
 232         vhost_log_cache_sync(dev, vq);
 233 }
 234
 235 static __rte_always_inline void
 236 vhost_flush_enqueue_batch_packed(struct virtio_net *dev,
 237                                  struct vhost_virtqueue *vq,
 238                                  uint64_t *lens,
 239                                  uint16_t *ids)
 240 {
 241         uint16_t i;
 242         uint16_t flags;
 243
 244         if (vq->shadow_used_idx) {
 245                 do_data_copy_enqueue(dev, vq);
 246                 vhost_flush_enqueue_shadow_packed(dev, vq);
 247         }
 248
 249         flags = PACKED_DESC_ENQUEUE_USED_FLAG(vq->used_wrap_counter);
 250
 251         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
 252                 vq->desc_packed[vq->last_used_idx + i].id = ids[i];
 253                 vq->desc_packed[vq->last_used_idx + i].len = lens[i];
 254         }
 255
 256         rte_smp_wmb();
 257
 258         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
 259                 vq->desc_packed[vq->last_used_idx + i].flags = flags;
 260
 261         vhost_log_cache_used_vring(dev, vq, vq->last_used_idx *
 262                                    sizeof(struct vring_packed_desc),
 263                                    sizeof(struct vring_packed_desc) *
 264                                    PACKED_BATCH_SIZE);
 265         vhost_log_cache_sync(dev, vq);
 266
 267         vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
 268 }
 269
 270 static __rte_always_inline void
 271 vhost_shadow_dequeue_batch_packed_inorder(struct vhost_virtqueue *vq,
 272                                           uint16_t id)
 273 {
 274         vq->shadow_used_packed[0].id = id;
 275
 276         if (!vq->shadow_used_idx) {
 277                 vq->shadow_last_used_idx = vq->last_used_idx;
 278                 vq->shadow_used_packed[0].flags =
 279                         PACKED_DESC_DEQUEUE_USED_FLAG(vq->used_wrap_counter);
 280                 vq->shadow_used_packed[0].len = 0;
 281                 vq->shadow_used_packed[0].count = 1;
 282                 vq->shadow_used_idx++;
 283         }
 284
 285         vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
 286 }
 287
 288 static __rte_always_inline void
 289 vhost_shadow_dequeue_batch_packed(struct virtio_net *dev,
 290                                   struct vhost_virtqueue *vq,
 291                                   uint16_t *ids)
 292 {
 293         uint16_t flags;
 294         uint16_t i;
 295         uint16_t begin;
 296
 297         flags = PACKED_DESC_DEQUEUE_USED_FLAG(vq->used_wrap_counter);
 298
 299         if (!vq->shadow_used_idx) {
 300                 vq->shadow_last_used_idx = vq->last_used_idx;
 301                 vq->shadow_used_packed[0].id  = ids[0];
 302                 vq->shadow_used_packed[0].len = 0;
 303                 vq->shadow_used_packed[0].count = 1;
 304                 vq->shadow_used_packed[0].flags = flags;
 305                 vq->shadow_used_idx++;
 306                 begin = 1;
 307         } else
 308                 begin = 0;
 309
 310         vhost_for_each_try_unroll(i, begin, PACKED_BATCH_SIZE) {
 311                 vq->desc_packed[vq->last_used_idx + i].id = ids[i];
 312                 vq->desc_packed[vq->last_used_idx + i].len = 0;
 313         }
 314
 315         rte_smp_wmb();
 316         vhost_for_each_try_unroll(i, begin, PACKED_BATCH_SIZE)
 317                 vq->desc_packed[vq->last_used_idx + i].flags = flags;
 318
 319         vhost_log_cache_used_vring(dev, vq, vq->last_used_idx *
 320                                    sizeof(struct vring_packed_desc),
 321                                    sizeof(struct vring_packed_desc) *
 322                                    PACKED_BATCH_SIZE);
 323         vhost_log_cache_sync(dev, vq);
 324
 325         vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
 326 }
 327
 328 static __rte_always_inline void
 329 vhost_shadow_dequeue_single_packed(struct vhost_virtqueue *vq,
 330                                    uint16_t buf_id,
 331                                    uint16_t count)
 332 {
 333         uint16_t flags;
 334
 335         flags = vq->desc_packed[vq->last_used_idx].flags;
 336         if (vq->used_wrap_counter) {
 337                 flags |= VRING_DESC_F_USED;
 338                 flags |= VRING_DESC_F_AVAIL;
 339         } else {
 340                 flags &= ~VRING_DESC_F_USED;
 341                 flags &= ~VRING_DESC_F_AVAIL;
 342         }
 343
 344         if (!vq->shadow_used_idx) {
 345                 vq->shadow_last_used_idx = vq->last_used_idx;
 346
 347                 vq->shadow_used_packed[0].id  = buf_id;
 348                 vq->shadow_used_packed[0].len = 0;
 349                 vq->shadow_used_packed[0].flags = flags;
 350                 vq->shadow_used_idx++;
 351         } else {
 352                 vq->desc_packed[vq->last_used_idx].id = buf_id;
 353                 vq->desc_packed[vq->last_used_idx].len = 0;
 354                 vq->desc_packed[vq->last_used_idx].flags = flags;
 355         }
 356
 357         vq_inc_last_used_packed(vq, count);
 358 }
 359
 360 static __rte_always_inline void
 361 vhost_shadow_dequeue_single_packed_inorder(struct vhost_virtqueue *vq,
 362                                            uint16_t buf_id,
 363                                            uint16_t count)
 364 {
 365         uint16_t flags;
 366
 367         vq->shadow_used_packed[0].id = buf_id;
 368
 369         flags = vq->desc_packed[vq->last_used_idx].flags;
 370         if (vq->used_wrap_counter) {
 371                 flags |= VRING_DESC_F_USED;
 372                 flags |= VRING_DESC_F_AVAIL;
 373         } else {
 374                 flags &= ~VRING_DESC_F_USED;
 375                 flags &= ~VRING_DESC_F_AVAIL;
 376         }
 377
 378         if (!vq->shadow_used_idx) {
 379                 vq->shadow_last_used_idx = vq->last_used_idx;
 380                 vq->shadow_used_packed[0].len = 0;
 381                 vq->shadow_used_packed[0].flags = flags;
 382                 vq->shadow_used_idx++;
 383         }
 384
 385         vq_inc_last_used_packed(vq, count);
 386 }
 387
 388 static __rte_always_inline void
 389 vhost_shadow_enqueue_single_packed(struct virtio_net *dev,
 390                                    struct vhost_virtqueue *vq,
 391                                    uint32_t len[],
 392                                    uint16_t id[],
 393                                    uint16_t count[],
 394                                    uint16_t num_buffers)
 395 {
 396         uint16_t i;
 397         for (i = 0; i < num_buffers; i++) {
 398                 /* enqueue shadow flush action aligned with batch num */
 399                 if (!vq->shadow_used_idx)
 400                         vq->shadow_aligned_idx = vq->last_used_idx &
 401                                 PACKED_BATCH_MASK;
 402                 vq->shadow_used_packed[vq->shadow_used_idx].id  = id[i];
 403                 vq->shadow_used_packed[vq->shadow_used_idx].len = len[i];
 404                 vq->shadow_used_packed[vq->shadow_used_idx].count = count[i];
 405                 vq->shadow_aligned_idx += count[i];
 406                 vq->shadow_used_idx++;
 407         }
 408
 409         if (vq->shadow_aligned_idx >= PACKED_BATCH_SIZE) {
 410                 do_data_copy_enqueue(dev, vq);
 411                 vhost_flush_enqueue_shadow_packed(dev, vq);
 412         }
 413 }
 414
 415 /* avoid write operation when necessary, to lessen cache issues */
 416 #define ASSIGN_UNLESS_EQUAL(var, val) do {      \
 417         if ((var) != (val))                     \
 418                 (var) = (val);                  \
 419 } while (0)
 420
 421 static __rte_always_inline void
 422 virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
 423 {
 424         uint64_t csum_l4 = m_buf->ol_flags & PKT_TX_L4_MASK;
 425
 426         if (m_buf->ol_flags & PKT_TX_TCP_SEG)
 427                 csum_l4 |= PKT_TX_TCP_CKSUM;
 428
 429         if (csum_l4) {
 430                 net_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 431                 net_hdr->csum_start = m_buf->l2_len + m_buf->l3_len;
 432
 433                 switch (csum_l4) {
 434                 case PKT_TX_TCP_CKSUM:
 435                         net_hdr->csum_offset = (offsetof(struct rte_tcp_hdr,
 436                                                 cksum));
 437                         break;
 438                 case PKT_TX_UDP_CKSUM:
 439                         net_hdr->csum_offset = (offsetof(struct rte_udp_hdr,
 440                                                 dgram_cksum));
 441                         break;
 442                 case PKT_TX_SCTP_CKSUM:
 443                         net_hdr->csum_offset = (offsetof(struct rte_sctp_hdr,
 444                                                 cksum));
 445                         break;
 446                 }
 447         } else {
 448                 ASSIGN_UNLESS_EQUAL(net_hdr->csum_start, 0);
 449                 ASSIGN_UNLESS_EQUAL(net_hdr->csum_offset, 0);
 450                 ASSIGN_UNLESS_EQUAL(net_hdr->flags, 0);
 451         }
 452
 453         /* IP cksum verification cannot be bypassed, then calculate here */
 454         if (m_buf->ol_flags & PKT_TX_IP_CKSUM) {
 455                 struct rte_ipv4_hdr *ipv4_hdr;
 456
 457                 ipv4_hdr = rte_pktmbuf_mtod_offset(m_buf, struct rte_ipv4_hdr *,
 458                                                    m_buf->l2_len);
 459                 ipv4_hdr->hdr_checksum = 0;
 460                 ipv4_hdr->hdr_checksum = rte_ipv4_cksum(ipv4_hdr);
 461         }
 462
 463         if (m_buf->ol_flags & PKT_TX_TCP_SEG) {
 464                 if (m_buf->ol_flags & PKT_TX_IPV4)
 465                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 466                 else
 467                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 468                 net_hdr->gso_size = m_buf->tso_segsz;
 469                 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len
 470                                         + m_buf->l4_len;
 471         } else if (m_buf->ol_flags & PKT_TX_UDP_SEG) {
 472                 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
 473                 net_hdr->gso_size = m_buf->tso_segsz;
 474                 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len +
 475                         m_buf->l4_len;
 476         } else {
 477                 ASSIGN_UNLESS_EQUAL(net_hdr->gso_type, 0);
 478                 ASSIGN_UNLESS_EQUAL(net_hdr->gso_size, 0);
 479                 ASSIGN_UNLESS_EQUAL(net_hdr->hdr_len, 0);
 480         }
 481 }
 482
 483 static __rte_always_inline int
 484 map_one_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 485                 struct buf_vector *buf_vec, uint16_t *vec_idx,
 486                 uint64_t desc_iova, uint64_t desc_len, uint8_t perm)
 487 {
 488         uint16_t vec_id = *vec_idx;
 489
 490         while (desc_len) {
 491                 uint64_t desc_addr;
 492                 uint64_t desc_chunck_len = desc_len;
 493
 494                 if (unlikely(vec_id >= BUF_VECTOR_MAX))
 495                         return -1;
 496
 497                 desc_addr = vhost_iova_to_vva(dev, vq,
 498                                 desc_iova,
 499                                 &desc_chunck_len,
 500                                 perm);
 501                 if (unlikely(!desc_addr))
 502                         return -1;
 503
 504                 rte_prefetch0((void *)(uintptr_t)desc_addr);
 505
 506                 buf_vec[vec_id].buf_iova = desc_iova;
 507                 buf_vec[vec_id].buf_addr = desc_addr;
 508                 buf_vec[vec_id].buf_len  = desc_chunck_len;
 509
 510                 desc_len -= desc_chunck_len;
 511                 desc_iova += desc_chunck_len;
 512                 vec_id++;
 513         }
 514         *vec_idx = vec_id;
 515
 516         return 0;
 517 }
 518
 519 static __rte_always_inline int
 520 fill_vec_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 521                          uint32_t avail_idx, uint16_t *vec_idx,
 522                          struct buf_vector *buf_vec, uint16_t *desc_chain_head,
 523                          uint32_t *desc_chain_len, uint8_t perm)
 524 {
 525         uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
 526         uint16_t vec_id = *vec_idx;
 527         uint32_t len    = 0;
 528         uint64_t dlen;
 529         uint32_t nr_descs = vq->size;
 530         uint32_t cnt    = 0;
 531         struct vring_desc *descs = vq->desc;
 532         struct vring_desc *idesc = NULL;
 533
 534         if (unlikely(idx >= vq->size))
 535                 return -1;
 536
 537         *desc_chain_head = idx;
 538
 539         if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
 540                 dlen = vq->desc[idx].len;
 541                 nr_descs = dlen / sizeof(struct vring_desc);
 542                 if (unlikely(nr_descs > vq->size))
 543                         return -1;
 544
 545                 descs = (struct vring_desc *)(uintptr_t)
 546                         vhost_iova_to_vva(dev, vq, vq->desc[idx].addr,
 547                                                 &dlen,
 548                                                 VHOST_ACCESS_RO);
 549                 if (unlikely(!descs))
 550                         return -1;
 551
 552                 if (unlikely(dlen < vq->desc[idx].len)) {
 553                         /*
 554                          * The indirect desc table is not contiguous
 555                          * in process VA space, we have to copy it.
 556                          */
 557                         idesc = vhost_alloc_copy_ind_table(dev, vq,
 558                                         vq->desc[idx].addr, vq->desc[idx].len);
 559                         if (unlikely(!idesc))
 560                                 return -1;
 561
 562                         descs = idesc;
 563                 }
 564
 565                 idx = 0;
 566         }
 567
 568         while (1) {
 569                 if (unlikely(idx >= nr_descs || cnt++ >= nr_descs)) {
 570                         free_ind_table(idesc);
 571                         return -1;
 572                 }
 573
 574                 len += descs[idx].len;
 575
 576                 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 577                                                 descs[idx].addr, descs[idx].len,
 578                                                 perm))) {
 579                         free_ind_table(idesc);
 580                         return -1;
 581                 }
 582
 583                 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
 584                         break;
 585
 586                 idx = descs[idx].next;
 587         }
 588
 589         *desc_chain_len = len;
 590         *vec_idx = vec_id;
 591
 592         if (unlikely(!!idesc))
 593                 free_ind_table(idesc);
 594
 595         return 0;
 596 }
 597
 598 /*
 599  * Returns -1 on fail, 0 on success
 600  */
 601 static inline int
 602 reserve_avail_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 603                                 uint32_t size, struct buf_vector *buf_vec,
 604                                 uint16_t *num_buffers, uint16_t avail_head,
 605                                 uint16_t *nr_vec)
 606 {
 607         uint16_t cur_idx;
 608         uint16_t vec_idx = 0;
 609         uint16_t max_tries, tries = 0;
 610
 611         uint16_t head_idx = 0;
 612         uint32_t len = 0;
 613
 614         *num_buffers = 0;
 615         cur_idx  = vq->last_avail_idx;
 616
 617         if (rxvq_is_mergeable(dev))
 618                 max_tries = vq->size - 1;
 619         else
 620                 max_tries = 1;
 621
 622         while (size > 0) {
 623                 if (unlikely(cur_idx == avail_head))
 624                         return -1;
 625                 /*
 626                  * if we tried all available ring items, and still
 627                  * can't get enough buf, it means something abnormal
 628                  * happened.
 629                  */
 630                 if (unlikely(++tries > max_tries))
 631                         return -1;
 632
 633                 if (unlikely(fill_vec_buf_split(dev, vq, cur_idx,
 634                                                 &vec_idx, buf_vec,
 635                                                 &head_idx, &len,
 636                                                 VHOST_ACCESS_RW) < 0))
 637                         return -1;
 638                 len = RTE_MIN(len, size);
 639                 update_shadow_used_ring_split(vq, head_idx, len);
 640                 size -= len;
 641
 642                 cur_idx++;
 643                 *num_buffers += 1;
 644         }
 645
 646         *nr_vec = vec_idx;
 647
 648         return 0;
 649 }
 650
 651 static __rte_always_inline int
 652 fill_vec_buf_packed_indirect(struct virtio_net *dev,
 653                         struct vhost_virtqueue *vq,
 654                         struct vring_packed_desc *desc, uint16_t *vec_idx,
 655                         struct buf_vector *buf_vec, uint32_t *len, uint8_t perm)
 656 {
 657         uint16_t i;
 658         uint32_t nr_descs;
 659         uint16_t vec_id = *vec_idx;
 660         uint64_t dlen;
 661         struct vring_packed_desc *descs, *idescs = NULL;
 662
 663         dlen = desc->len;
 664         descs = (struct vring_packed_desc *)(uintptr_t)
 665                 vhost_iova_to_vva(dev, vq, desc->addr, &dlen, VHOST_ACCESS_RO);
 666         if (unlikely(!descs))
 667                 return -1;
 668
 669         if (unlikely(dlen < desc->len)) {
 670                 /*
 671                  * The indirect desc table is not contiguous
 672                  * in process VA space, we have to copy it.
 673                  */
 674                 idescs = vhost_alloc_copy_ind_table(dev,
 675                                 vq, desc->addr, desc->len);
 676                 if (unlikely(!idescs))
 677                         return -1;
 678
 679                 descs = idescs;
 680         }
 681
 682         nr_descs =  desc->len / sizeof(struct vring_packed_desc);
 683         if (unlikely(nr_descs >= vq->size)) {
 684                 free_ind_table(idescs);
 685                 return -1;
 686         }
 687
 688         for (i = 0; i < nr_descs; i++) {
 689                 if (unlikely(vec_id >= BUF_VECTOR_MAX)) {
 690                         free_ind_table(idescs);
 691                         return -1;
 692                 }
 693
 694                 *len += descs[i].len;
 695                 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 696                                                 descs[i].addr, descs[i].len,
 697                                                 perm)))
 698                         return -1;
 699         }
 700         *vec_idx = vec_id;
 701
 702         if (unlikely(!!idescs))
 703                 free_ind_table(idescs);
 704
 705         return 0;
 706 }
 707
 708 static __rte_always_inline int
 709 fill_vec_buf_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
 710                                 uint16_t avail_idx, uint16_t *desc_count,
 711                                 struct buf_vector *buf_vec, uint16_t *vec_idx,
 712                                 uint16_t *buf_id, uint32_t *len, uint8_t perm)
 713 {
 714         bool wrap_counter = vq->avail_wrap_counter;
 715         struct vring_packed_desc *descs = vq->desc_packed;
 716         uint16_t vec_id = *vec_idx;
 717
 718         if (avail_idx < vq->last_avail_idx)
 719                 wrap_counter ^= 1;
 720
 721         /*
 722          * Perform a load-acquire barrier in desc_is_avail to
 723          * enforce the ordering between desc flags and desc
 724          * content.
 725          */
 726         if (unlikely(!desc_is_avail(&descs[avail_idx], wrap_counter)))
 727                 return -1;
 728
 729         *desc_count = 0;
 730         *len = 0;
 731
 732         while (1) {
 733                 if (unlikely(vec_id >= BUF_VECTOR_MAX))
 734                         return -1;
 735
 736                 if (unlikely(*desc_count >= vq->size))
 737                         return -1;
 738
 739                 *desc_count += 1;
 740                 *buf_id = descs[avail_idx].id;
 741
 742                 if (descs[avail_idx].flags & VRING_DESC_F_INDIRECT) {
 743                         if (unlikely(fill_vec_buf_packed_indirect(dev, vq,
 744                                                         &descs[avail_idx],
 745                                                         &vec_id, buf_vec,
 746                                                         len, perm) < 0))
 747                                 return -1;
 748                 } else {
 749                         *len += descs[avail_idx].len;
 750
 751                         if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 752                                                         descs[avail_idx].addr,
 753                                                         descs[avail_idx].len,
 754                                                         perm)))
 755                                 return -1;
 756                 }
 757
 758                 if ((descs[avail_idx].flags & VRING_DESC_F_NEXT) == 0)
 759                         break;
 760
 761                 if (++avail_idx >= vq->size) {
 762                         avail_idx -= vq->size;
 763                         wrap_counter ^= 1;
 764                 }
 765         }
 766
 767         *vec_idx = vec_id;
 768
 769         return 0;
 770 }
 771
 772 static __rte_noinline void
 773 copy_vnet_hdr_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 774                 struct buf_vector *buf_vec,
 775                 struct virtio_net_hdr_mrg_rxbuf *hdr)
 776 {
 777         uint64_t len;
 778         uint64_t remain = dev->vhost_hlen;
 779         uint64_t src = (uint64_t)(uintptr_t)hdr, dst;
 780         uint64_t iova = buf_vec->buf_iova;
 781
 782         while (remain) {
 783                 len = RTE_MIN(remain,
 784                                 buf_vec->buf_len);
 785                 dst = buf_vec->buf_addr;
 786                 rte_memcpy((void *)(uintptr_t)dst,
 787                                 (void *)(uintptr_t)src,
 788                                 len);
 789
 790                 PRINT_PACKET(dev, (uintptr_t)dst,
 791                                 (uint32_t)len, 0);
 792                 vhost_log_cache_write_iova(dev, vq,
 793                                 iova, len);
 794
 795                 remain -= len;
 796                 iova += len;
 797                 src += len;
 798                 buf_vec++;
 799         }
 800 }
 801
 802 static __rte_always_inline int
 803 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 804                             struct rte_mbuf *m, struct buf_vector *buf_vec,
 805                             uint16_t nr_vec, uint16_t num_buffers)
 806 {
 807         uint32_t vec_idx = 0;
 808         uint32_t mbuf_offset, mbuf_avail;
 809         uint32_t buf_offset, buf_avail;
 810         uint64_t buf_addr, buf_iova, buf_len;
 811         uint32_t cpy_len;
 812         uint64_t hdr_addr;
 813         struct rte_mbuf *hdr_mbuf;
 814         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 815         struct virtio_net_hdr_mrg_rxbuf tmp_hdr, *hdr = NULL;
 816         int error = 0;
 817
 818         if (unlikely(m == NULL)) {
 819                 error = -1;
 820                 goto out;
 821         }
 822
 823         buf_addr = buf_vec[vec_idx].buf_addr;
 824         buf_iova = buf_vec[vec_idx].buf_iova;
 825         buf_len = buf_vec[vec_idx].buf_len;
 826
 827         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
 828                 error = -1;
 829                 goto out;
 830         }
 831
 832         hdr_mbuf = m;
 833         hdr_addr = buf_addr;
 834         if (unlikely(buf_len < dev->vhost_hlen))
 835                 hdr = &tmp_hdr;
 836         else
 837                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)hdr_addr;
 838
 839         VHOST_LOG_DATA(DEBUG, "(%d) RX: num merge buffers %d\n",
 840                 dev->vid, num_buffers);
 841
 842         if (unlikely(buf_len < dev->vhost_hlen)) {
 843                 buf_offset = dev->vhost_hlen - buf_len;
 844                 vec_idx++;
 845                 buf_addr = buf_vec[vec_idx].buf_addr;
 846                 buf_iova = buf_vec[vec_idx].buf_iova;
 847                 buf_len = buf_vec[vec_idx].buf_len;
 848                 buf_avail = buf_len - buf_offset;
 849         } else {
 850                 buf_offset = dev->vhost_hlen;
 851                 buf_avail = buf_len - dev->vhost_hlen;
 852         }
 853
 854         mbuf_avail  = rte_pktmbuf_data_len(m);
 855         mbuf_offset = 0;
 856         while (mbuf_avail != 0 || m->next != NULL) {
 857                 /* done with current buf, get the next one */
 858                 if (buf_avail == 0) {
 859                         vec_idx++;
 860                         if (unlikely(vec_idx >= nr_vec)) {
 861                                 error = -1;
 862                                 goto out;
 863                         }
 864
 865                         buf_addr = buf_vec[vec_idx].buf_addr;
 866                         buf_iova = buf_vec[vec_idx].buf_iova;
 867                         buf_len = buf_vec[vec_idx].buf_len;
 868
 869                         buf_offset = 0;
 870                         buf_avail  = buf_len;
 871                 }
 872
 873                 /* done with current mbuf, get the next one */
 874                 if (mbuf_avail == 0) {
 875                         m = m->next;
 876
 877                         mbuf_offset = 0;
 878                         mbuf_avail  = rte_pktmbuf_data_len(m);
 879                 }
 880
 881                 if (hdr_addr) {
 882                         virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
 883                         if (rxvq_is_mergeable(dev))
 884                                 ASSIGN_UNLESS_EQUAL(hdr->num_buffers,
 885                                                 num_buffers);
 886
 887                         if (unlikely(hdr == &tmp_hdr)) {
 888                                 copy_vnet_hdr_to_desc(dev, vq, buf_vec, hdr);
 889                         } else {
 890                                 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
 891                                                 dev->vhost_hlen, 0);
 892                                 vhost_log_cache_write_iova(dev, vq,
 893                                                 buf_vec[0].buf_iova,
 894                                                 dev->vhost_hlen);
 895                         }
 896
 897                         hdr_addr = 0;
 898                 }
 899
 900                 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
 901
 902                 if (likely(cpy_len > MAX_BATCH_LEN ||
 903                                         vq->batch_copy_nb_elems >= vq->size)) {
 904                         rte_memcpy((void *)((uintptr_t)(buf_addr + buf_offset)),
 905                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 906                                 cpy_len);
 907                         vhost_log_cache_write_iova(dev, vq,
 908                                                    buf_iova + buf_offset,
 909                                                    cpy_len);
 910                         PRINT_PACKET(dev, (uintptr_t)(buf_addr + buf_offset),
 911                                 cpy_len, 0);
 912                 } else {
 913                         batch_copy[vq->batch_copy_nb_elems].dst =
 914                                 (void *)((uintptr_t)(buf_addr + buf_offset));
 915                         batch_copy[vq->batch_copy_nb_elems].src =
 916                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
 917                         batch_copy[vq->batch_copy_nb_elems].log_addr =
 918                                 buf_iova + buf_offset;
 919                         batch_copy[vq->batch_copy_nb_elems].len = cpy_len;
 920                         vq->batch_copy_nb_elems++;
 921                 }
 922
 923                 mbuf_avail  -= cpy_len;
 924                 mbuf_offset += cpy_len;
 925                 buf_avail  -= cpy_len;
 926                 buf_offset += cpy_len;
 927         }
 928
 929 out:
 930
 931         return error;
 932 }
 933
 934 static __rte_always_inline void
 935 async_fill_vec(struct iovec *v, void *base, size_t len)
 936 {
 937         v->iov_base = base;
 938         v->iov_len = len;
 939 }
 940
 941 static __rte_always_inline void
 942 async_fill_iter(struct rte_vhost_iov_iter *it, size_t count,
 943         struct iovec *vec, unsigned long nr_seg)
 944 {
 945         it->offset = 0;
 946         it->count = count;
 947
 948         if (count) {
 949                 it->iov = vec;
 950                 it->nr_segs = nr_seg;
 951         } else {
 952                 it->iov = 0;
 953                 it->nr_segs = 0;
 954         }
 955 }
 956
 957 static __rte_always_inline void
 958 async_fill_desc(struct rte_vhost_async_desc *desc,
 959         struct rte_vhost_iov_iter *src, struct rte_vhost_iov_iter *dst)
 960 {
 961         desc->src = src;
 962         desc->dst = dst;
 963 }
 964
 965 static __rte_always_inline int
 966 async_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 967                         struct rte_mbuf *m, struct buf_vector *buf_vec,
 968                         uint16_t nr_vec, uint16_t num_buffers,
 969                         struct iovec *src_iovec, struct iovec *dst_iovec,
 970                         struct rte_vhost_iov_iter *src_it,
 971                         struct rte_vhost_iov_iter *dst_it)
 972 {
 973         uint32_t vec_idx = 0;
 974         uint32_t mbuf_offset, mbuf_avail;
 975         uint32_t buf_offset, buf_avail;
 976         uint64_t buf_addr, buf_iova, buf_len;
 977         uint32_t cpy_len, cpy_threshold;
 978         uint64_t hdr_addr;
 979         struct rte_mbuf *hdr_mbuf;
 980         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 981         struct virtio_net_hdr_mrg_rxbuf tmp_hdr, *hdr = NULL;
 982         int error = 0;
 983         uint64_t mapped_len;
 984
 985         uint32_t tlen = 0;
 986         int tvec_idx = 0;
 987         void *hpa;
 988
 989         if (unlikely(m == NULL)) {
 990                 error = -1;
 991                 goto out;
 992         }
 993
 994         cpy_threshold = vq->async_threshold;
 995
 996         buf_addr = buf_vec[vec_idx].buf_addr;
 997         buf_iova = buf_vec[vec_idx].buf_iova;
 998         buf_len = buf_vec[vec_idx].buf_len;
 999
1000         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
1001                 error = -1;
1002                 goto out;
1003         }
1004
1005         hdr_mbuf = m;
1006         hdr_addr = buf_addr;
1007         if (unlikely(buf_len < dev->vhost_hlen))
1008                 hdr = &tmp_hdr;
1009         else
1010                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)hdr_addr;
1011
1012         VHOST_LOG_DATA(DEBUG, "(%d) RX: num merge buffers %d\n",
1013                 dev->vid, num_buffers);
1014
1015         if (unlikely(buf_len < dev->vhost_hlen)) {
1016                 buf_offset = dev->vhost_hlen - buf_len;
1017                 vec_idx++;
1018                 buf_addr = buf_vec[vec_idx].buf_addr;
1019                 buf_iova = buf_vec[vec_idx].buf_iova;
1020                 buf_len = buf_vec[vec_idx].buf_len;
1021                 buf_avail = buf_len - buf_offset;
1022         } else {
1023                 buf_offset = dev->vhost_hlen;
1024                 buf_avail = buf_len - dev->vhost_hlen;
1025         }
1026
1027         mbuf_avail  = rte_pktmbuf_data_len(m);
1028         mbuf_offset = 0;
1029
1030         while (mbuf_avail != 0 || m->next != NULL) {
1031                 /* done with current buf, get the next one */
1032                 if (buf_avail == 0) {
1033                         vec_idx++;
1034                         if (unlikely(vec_idx >= nr_vec)) {
1035                                 error = -1;
1036                                 goto out;
1037                         }
1038
1039                         buf_addr = buf_vec[vec_idx].buf_addr;
1040                         buf_iova = buf_vec[vec_idx].buf_iova;
1041                         buf_len = buf_vec[vec_idx].buf_len;
1042
1043                         buf_offset = 0;
1044                         buf_avail  = buf_len;
1045                 }
1046
1047                 /* done with current mbuf, get the next one */
1048                 if (mbuf_avail == 0) {
1049                         m = m->next;
1050
1051                         mbuf_offset = 0;
1052                         mbuf_avail  = rte_pktmbuf_data_len(m);
1053                 }
1054
1055                 if (hdr_addr) {
1056                         virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
1057                         if (rxvq_is_mergeable(dev))
1058                                 ASSIGN_UNLESS_EQUAL(hdr->num_buffers,
1059                                                 num_buffers);
1060
1061                         if (unlikely(hdr == &tmp_hdr)) {
1062                                 copy_vnet_hdr_to_desc(dev, vq, buf_vec, hdr);
1063                         } else {
1064                                 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
1065                                                 dev->vhost_hlen, 0);
1066                                 vhost_log_cache_write_iova(dev, vq,
1067                                                 buf_vec[0].buf_iova,
1068                                                 dev->vhost_hlen);
1069                         }
1070
1071                         hdr_addr = 0;
1072                 }
1073
1074                 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
1075
1076                 while (unlikely(cpy_len && cpy_len >= cpy_threshold)) {
1077                         hpa = (void *)(uintptr_t)gpa_to_first_hpa(dev,
1078                                         buf_iova + buf_offset,
1079                                         cpy_len, &mapped_len);
1080
1081                         if (unlikely(!hpa || mapped_len < cpy_threshold))
1082                                 break;
1083
1084                         async_fill_vec(src_iovec + tvec_idx,
1085                                 (void *)(uintptr_t)rte_pktmbuf_iova_offset(m,
1086                                 mbuf_offset), (size_t)mapped_len);
1087
1088                         async_fill_vec(dst_iovec + tvec_idx,
1089                                         hpa, (size_t)mapped_len);
1090
1091                         tlen += (uint32_t)mapped_len;
1092                         cpy_len -= (uint32_t)mapped_len;
1093                         mbuf_avail  -= (uint32_t)mapped_len;
1094                         mbuf_offset += (uint32_t)mapped_len;
1095                         buf_avail  -= (uint32_t)mapped_len;
1096                         buf_offset += (uint32_t)mapped_len;
1097                         tvec_idx++;
1098                 }
1099
1100                 if (likely(cpy_len)) {
1101                         if (unlikely(vq->batch_copy_nb_elems >= vq->size)) {
1102                                 rte_memcpy(
1103                                 (void *)((uintptr_t)(buf_addr + buf_offset)),
1104                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
1105                                 cpy_len);
1106
1107                                 PRINT_PACKET(dev,
1108                                         (uintptr_t)(buf_addr + buf_offset),
1109                                         cpy_len, 0);
1110                         } else {
1111                                 batch_copy[vq->batch_copy_nb_elems].dst =
1112                                 (void *)((uintptr_t)(buf_addr + buf_offset));
1113                                 batch_copy[vq->batch_copy_nb_elems].src =
1114                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
1115                                 batch_copy[vq->batch_copy_nb_elems].log_addr =
1116                                         buf_iova + buf_offset;
1117                                 batch_copy[vq->batch_copy_nb_elems].len =
1118                                         cpy_len;
1119                                 vq->batch_copy_nb_elems++;
1120                         }
1121
1122                         mbuf_avail  -= cpy_len;
1123                         mbuf_offset += cpy_len;
1124                         buf_avail  -= cpy_len;
1125                         buf_offset += cpy_len;
1126                 }
1127
1128         }
1129
1130 out:
1131         async_fill_iter(src_it, tlen, src_iovec, tvec_idx);
1132         async_fill_iter(dst_it, tlen, dst_iovec, tvec_idx);
1133
1134         return error;
1135 }
1136
1137 static __rte_always_inline int
1138 vhost_enqueue_single_packed(struct virtio_net *dev,
1139                             struct vhost_virtqueue *vq,
1140                             struct rte_mbuf *pkt,
1141                             struct buf_vector *buf_vec,
1142                             uint16_t *nr_descs)
1143 {
1144         uint16_t nr_vec = 0;
1145         uint16_t avail_idx = vq->last_avail_idx;
1146         uint16_t max_tries, tries = 0;
1147         uint16_t buf_id = 0;
1148         uint32_t len = 0;
1149         uint16_t desc_count;
1150         uint32_t size = pkt->pkt_len + sizeof(struct virtio_net_hdr_mrg_rxbuf);
1151         uint16_t num_buffers = 0;
1152         uint32_t buffer_len[vq->size];
1153         uint16_t buffer_buf_id[vq->size];
1154         uint16_t buffer_desc_count[vq->size];
1155
1156         if (rxvq_is_mergeable(dev))
1157                 max_tries = vq->size - 1;
1158         else
1159                 max_tries = 1;
1160
1161         while (size > 0) {
1162                 /*
1163                  * if we tried all available ring items, and still
1164                  * can't get enough buf, it means something abnormal
1165                  * happened.
1166                  */
1167                 if (unlikely(++tries > max_tries))
1168                         return -1;
1169
1170                 if (unlikely(fill_vec_buf_packed(dev, vq,
1171                                                 avail_idx, &desc_count,
1172                                                 buf_vec, &nr_vec,
1173                                                 &buf_id, &len,
1174                                                 VHOST_ACCESS_RW) < 0))
1175                         return -1;
1176
1177                 len = RTE_MIN(len, size);
1178                 size -= len;
1179
1180                 buffer_len[num_buffers] = len;
1181                 buffer_buf_id[num_buffers] = buf_id;
1182                 buffer_desc_count[num_buffers] = desc_count;
1183                 num_buffers += 1;
1184
1185                 *nr_descs += desc_count;
1186                 avail_idx += desc_count;
1187                 if (avail_idx >= vq->size)
1188                         avail_idx -= vq->size;
1189         }
1190
1191         if (copy_mbuf_to_desc(dev, vq, pkt, buf_vec, nr_vec, num_buffers) < 0)
1192                 return -1;
1193
1194         vhost_shadow_enqueue_single_packed(dev, vq, buffer_len, buffer_buf_id,
1195                                            buffer_desc_count, num_buffers);
1196
1197         return 0;
1198 }
1199
1200 static __rte_noinline uint32_t
1201 virtio_dev_rx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
1202         struct rte_mbuf **pkts, uint32_t count)
1203 {
1204         uint32_t pkt_idx = 0;
1205         uint16_t num_buffers;
1206         struct buf_vector buf_vec[BUF_VECTOR_MAX];
1207         uint16_t avail_head;
1208
1209         /*
1210          * The ordering between avail index and
1211          * desc reads needs to be enforced.
1212          */
1213         avail_head = __atomic_load_n(&vq->avail->idx, __ATOMIC_ACQUIRE);
1214
1215         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
1216
1217         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
1218                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
1219                 uint16_t nr_vec = 0;
1220
1221                 if (unlikely(reserve_avail_buf_split(dev, vq,
1222                                                 pkt_len, buf_vec, &num_buffers,
1223                                                 avail_head, &nr_vec) < 0)) {
1224                         VHOST_LOG_DATA(DEBUG,
1225                                 "(%d) failed to get enough desc from vring\n",
1226                                 dev->vid);
1227                         vq->shadow_used_idx -= num_buffers;
1228                         break;
1229                 }
1230
1231                 VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
1232                         dev->vid, vq->last_avail_idx,
1233                         vq->last_avail_idx + num_buffers);
1234
1235                 if (copy_mbuf_to_desc(dev, vq, pkts[pkt_idx],
1236                                                 buf_vec, nr_vec,
1237                                                 num_buffers) < 0) {
1238                         vq->shadow_used_idx -= num_buffers;
1239                         break;
1240                 }
1241
1242                 vq->last_avail_idx += num_buffers;
1243         }
1244
1245         do_data_copy_enqueue(dev, vq);
1246
1247         if (likely(vq->shadow_used_idx)) {
1248                 flush_shadow_used_ring_split(dev, vq);
1249                 vhost_vring_call_split(dev, vq);
1250         }
1251
1252         return pkt_idx;
1253 }
1254
1255 static __rte_always_inline int
1256 virtio_dev_rx_batch_packed(struct virtio_net *dev,
1257                            struct vhost_virtqueue *vq,
1258                            struct rte_mbuf **pkts)
1259 {
1260         bool wrap_counter = vq->avail_wrap_counter;
1261         struct vring_packed_desc *descs = vq->desc_packed;
1262         uint16_t avail_idx = vq->last_avail_idx;
1263         uint64_t desc_addrs[PACKED_BATCH_SIZE];
1264         struct virtio_net_hdr_mrg_rxbuf *hdrs[PACKED_BATCH_SIZE];
1265         uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1266         uint64_t lens[PACKED_BATCH_SIZE];
1267         uint16_t ids[PACKED_BATCH_SIZE];
1268         uint16_t i;
1269
1270         if (unlikely(avail_idx & PACKED_BATCH_MASK))
1271                 return -1;
1272
1273         if (unlikely((avail_idx + PACKED_BATCH_SIZE) > vq->size))
1274                 return -1;
1275
1276         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1277                 if (unlikely(pkts[i]->next != NULL))
1278                         return -1;
1279                 if (unlikely(!desc_is_avail(&descs[avail_idx + i],
1280                                             wrap_counter)))
1281                         return -1;
1282         }
1283
1284         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1285                 lens[i] = descs[avail_idx + i].len;
1286
1287         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1288                 if (unlikely(pkts[i]->pkt_len > (lens[i] - buf_offset)))
1289                         return -1;
1290         }
1291
1292         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1293                 desc_addrs[i] = vhost_iova_to_vva(dev, vq,
1294                                                   descs[avail_idx + i].addr,
1295                                                   &lens[i],
1296                                                   VHOST_ACCESS_RW);
1297
1298         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1299                 if (unlikely(!desc_addrs[i]))
1300                         return -1;
1301                 if (unlikely(lens[i] != descs[avail_idx + i].len))
1302                         return -1;
1303         }
1304
1305         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1306                 rte_prefetch0((void *)(uintptr_t)desc_addrs[i]);
1307                 hdrs[i] = (struct virtio_net_hdr_mrg_rxbuf *)
1308                                         (uintptr_t)desc_addrs[i];
1309                 lens[i] = pkts[i]->pkt_len +
1310                         sizeof(struct virtio_net_hdr_mrg_rxbuf);
1311         }
1312
1313         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1314                 virtio_enqueue_offload(pkts[i], &hdrs[i]->hdr);
1315
1316         vq_inc_last_avail_packed(vq, PACKED_BATCH_SIZE);
1317
1318         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1319                 rte_memcpy((void *)(uintptr_t)(desc_addrs[i] + buf_offset),
1320                            rte_pktmbuf_mtod_offset(pkts[i], void *, 0),
1321                            pkts[i]->pkt_len);
1322         }
1323
1324         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1325                 vhost_log_cache_write_iova(dev, vq, descs[avail_idx + i].addr,
1326                                            lens[i]);
1327
1328         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1329                 ids[i] = descs[avail_idx + i].id;
1330
1331         vhost_flush_enqueue_batch_packed(dev, vq, lens, ids);
1332
1333         return 0;
1334 }
1335
1336 static __rte_always_inline int16_t
1337 virtio_dev_rx_single_packed(struct virtio_net *dev,
1338                             struct vhost_virtqueue *vq,
1339                             struct rte_mbuf *pkt)
1340 {
1341         struct buf_vector buf_vec[BUF_VECTOR_MAX];
1342         uint16_t nr_descs = 0;
1343
1344         if (unlikely(vhost_enqueue_single_packed(dev, vq, pkt, buf_vec,
1345                                                  &nr_descs) < 0)) {
1346                 VHOST_LOG_DATA(DEBUG,
1347                                 "(%d) failed to get enough desc from vring\n",
1348                                 dev->vid);
1349                 return -1;
1350         }
1351
1352         VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
1353                         dev->vid, vq->last_avail_idx,
1354                         vq->last_avail_idx + nr_descs);
1355
1356         vq_inc_last_avail_packed(vq, nr_descs);
1357
1358         return 0;
1359 }
1360
1361 static __rte_noinline uint32_t
1362 virtio_dev_rx_packed(struct virtio_net *dev,
1363                      struct vhost_virtqueue *__rte_restrict vq,
1364                      struct rte_mbuf **__rte_restrict pkts,
1365                      uint32_t count)
1366 {
1367         uint32_t pkt_idx = 0;
1368         uint32_t remained = count;
1369
1370         do {
1371                 rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
1372
1373                 if (remained >= PACKED_BATCH_SIZE) {
1374                         if (!virtio_dev_rx_batch_packed(dev, vq,
1375                                                         &pkts[pkt_idx])) {
1376                                 pkt_idx += PACKED_BATCH_SIZE;
1377                                 remained -= PACKED_BATCH_SIZE;
1378                                 continue;
1379                         }
1380                 }
1381
1382                 if (virtio_dev_rx_single_packed(dev, vq, pkts[pkt_idx]))
1383                         break;
1384                 pkt_idx++;
1385                 remained--;
1386
1387         } while (pkt_idx < count);
1388
1389         if (vq->shadow_used_idx) {
1390                 do_data_copy_enqueue(dev, vq);
1391                 vhost_flush_enqueue_shadow_packed(dev, vq);
1392         }
1393
1394         if (pkt_idx)
1395                 vhost_vring_call_packed(dev, vq);
1396
1397         return pkt_idx;
1398 }
1399
1400 static __rte_always_inline uint32_t
1401 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
1402         struct rte_mbuf **pkts, uint32_t count)
1403 {
1404         struct vhost_virtqueue *vq;
1405         uint32_t nb_tx = 0;
1406
1407         VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
1408         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
1409                 VHOST_LOG_DATA(ERR, "(%d) %s: invalid virtqueue idx %d.\n",
1410                         dev->vid, __func__, queue_id);
1411                 return 0;
1412         }
1413
1414         vq = dev->virtqueue[queue_id];
1415
1416         rte_spinlock_lock(&vq->access_lock);
1417
1418         if (unlikely(vq->enabled == 0))
1419                 goto out_access_unlock;
1420
1421         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1422                 vhost_user_iotlb_rd_lock(vq);
1423
1424         if (unlikely(vq->access_ok == 0))
1425                 if (unlikely(vring_translate(dev, vq) < 0))
1426                         goto out;
1427
1428         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
1429         if (count == 0)
1430                 goto out;
1431
1432         if (vq_is_packed(dev))
1433                 nb_tx = virtio_dev_rx_packed(dev, vq, pkts, count);
1434         else
1435                 nb_tx = virtio_dev_rx_split(dev, vq, pkts, count);
1436
1437 out:
1438         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1439                 vhost_user_iotlb_rd_unlock(vq);
1440
1441 out_access_unlock:
1442         rte_spinlock_unlock(&vq->access_lock);
1443
1444         return nb_tx;
1445 }
1446
1447 uint16_t
1448 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
1449         struct rte_mbuf **__rte_restrict pkts, uint16_t count)
1450 {
1451         struct virtio_net *dev = get_device(vid);
1452
1453         if (!dev)
1454                 return 0;
1455
1456         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
1457                 VHOST_LOG_DATA(ERR,
1458                         "(%d) %s: built-in vhost net backend is disabled.\n",
1459                         dev->vid, __func__);
1460                 return 0;
1461         }
1462
1463         return virtio_dev_rx(dev, queue_id, pkts, count);
1464 }
1465
1466 static __rte_always_inline uint16_t
1467 virtio_dev_rx_async_get_info_idx(uint16_t pkts_idx,
1468         uint16_t vq_size, uint16_t n_inflight)
1469 {
1470         return pkts_idx > n_inflight ? (pkts_idx - n_inflight) :
1471                 (vq_size - n_inflight + pkts_idx) & (vq_size - 1);
1472 }
1473
1474 static __rte_noinline uint32_t
1475 virtio_dev_rx_async_submit_split(struct virtio_net *dev,
1476         struct vhost_virtqueue *vq, uint16_t queue_id,
1477         struct rte_mbuf **pkts, uint32_t count)
1478 {
1479         uint32_t pkt_idx = 0, pkt_burst_idx = 0;
1480         uint16_t num_buffers;
1481         struct buf_vector buf_vec[BUF_VECTOR_MAX];
1482         uint16_t avail_head;
1483
1484         struct rte_vhost_iov_iter *it_pool = vq->it_pool;
1485         struct iovec *vec_pool = vq->vec_pool;
1486         struct rte_vhost_async_desc tdes[MAX_PKT_BURST];
1487         struct iovec *src_iovec = vec_pool;
1488         struct iovec *dst_iovec = vec_pool + (VHOST_MAX_ASYNC_VEC >> 1);
1489         struct rte_vhost_iov_iter *src_it = it_pool;
1490         struct rte_vhost_iov_iter *dst_it = it_pool + 1;
1491         uint16_t n_free_slot, slot_idx = 0;
1492         uint16_t pkt_err = 0;
1493         uint16_t segs_await = 0;
1494         struct async_inflight_info *pkts_info = vq->async_pkts_info;
1495         int n_pkts = 0;
1496
1497         avail_head = __atomic_load_n(&vq->avail->idx, __ATOMIC_ACQUIRE);
1498
1499         /*
1500          * The ordering between avail index and
1501          * desc reads needs to be enforced.
1502          */
1503         rte_smp_rmb();
1504
1505         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
1506
1507         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
1508                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
1509                 uint16_t nr_vec = 0;
1510
1511                 if (unlikely(reserve_avail_buf_split(dev, vq,
1512                                                 pkt_len, buf_vec, &num_buffers,
1513                                                 avail_head, &nr_vec) < 0)) {
1514                         VHOST_LOG_DATA(DEBUG,
1515                                 "(%d) failed to get enough desc from vring\n",
1516                                 dev->vid);
1517                         vq->shadow_used_idx -= num_buffers;
1518                         break;
1519                 }
1520
1521                 VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
1522                         dev->vid, vq->last_avail_idx,
1523                         vq->last_avail_idx + num_buffers);
1524
1525                 if (async_mbuf_to_desc(dev, vq, pkts[pkt_idx],
1526                                 buf_vec, nr_vec, num_buffers,
1527                                 src_iovec, dst_iovec, src_it, dst_it) < 0) {
1528                         vq->shadow_used_idx -= num_buffers;
1529                         break;
1530                 }
1531
1532                 slot_idx = (vq->async_pkts_idx + pkt_idx) & (vq->size - 1);
1533                 if (src_it->count) {
1534                         async_fill_desc(&tdes[pkt_burst_idx], src_it, dst_it);
1535                         pkt_burst_idx++;
1536                         pkts_info[slot_idx].descs = num_buffers;
1537                         pkts_info[slot_idx].segs = src_it->nr_segs;
1538                         src_iovec += src_it->nr_segs;
1539                         dst_iovec += dst_it->nr_segs;
1540                         src_it += 2;
1541                         dst_it += 2;
1542                         segs_await += src_it->nr_segs;
1543                 } else {
1544                         pkts_info[slot_idx].info = num_buffers;
1545                         vq->async_pkts_inflight_n++;
1546                 }
1547
1548                 vq->last_avail_idx += num_buffers;
1549
1550                 /*
1551                  * conditions to trigger async device transfer:
1552                  * - buffered packet number reaches transfer threshold
1553                  * - this is the last packet in the burst enqueue
1554                  * - unused async iov number is less than max vhost vector
1555                  */
1556                 if (pkt_burst_idx >= VHOST_ASYNC_BATCH_THRESHOLD ||
1557                         (pkt_idx == count - 1 && pkt_burst_idx) ||
1558                         (VHOST_MAX_ASYNC_VEC / 2 - segs_await <
1559                         BUF_VECTOR_MAX)) {
1560                         n_pkts = vq->async_ops.transfer_data(dev->vid,
1561                                         queue_id, tdes, 0, pkt_burst_idx);
1562                         src_iovec = vec_pool;
1563                         dst_iovec = vec_pool + (VHOST_MAX_ASYNC_VEC >> 1);
1564                         src_it = it_pool;
1565                         dst_it = it_pool + 1;
1566                         segs_await = 0;
1567                         vq->async_pkts_inflight_n += pkt_burst_idx;
1568
1569                         if (unlikely(n_pkts < (int)pkt_burst_idx)) {
1570                                 /*
1571                                  * log error packets number here and do actual
1572                                  * error processing when applications poll
1573                                  * completion
1574                                  */
1575                                 pkt_err = pkt_burst_idx - n_pkts;
1576                                 pkt_burst_idx = 0;
1577                                 break;
1578                         }
1579
1580                         pkt_burst_idx = 0;
1581                 }
1582         }
1583
1584         if (pkt_burst_idx) {
1585                 n_pkts = vq->async_ops.transfer_data(dev->vid,
1586                                 queue_id, tdes, 0, pkt_burst_idx);
1587                 vq->async_pkts_inflight_n += pkt_burst_idx;
1588
1589                 if (unlikely(n_pkts < (int)pkt_burst_idx))
1590                         pkt_err = pkt_burst_idx - n_pkts;
1591         }
1592
1593         do_data_copy_enqueue(dev, vq);
1594
1595         while (unlikely(pkt_err && pkt_idx)) {
1596                 if (pkts_info[slot_idx].segs)
1597                         pkt_err--;
1598                 vq->last_avail_idx -= pkts_info[slot_idx].descs;
1599                 vq->shadow_used_idx -= pkts_info[slot_idx].descs;
1600                 vq->async_pkts_inflight_n--;
1601                 slot_idx = (slot_idx - 1) & (vq->size - 1);
1602                 pkt_idx--;
1603         }
1604
1605         n_free_slot = vq->size - vq->async_pkts_idx;
1606         if (n_free_slot > pkt_idx) {
1607                 rte_memcpy(&vq->async_pkts_pending[vq->async_pkts_idx],
1608                         pkts, pkt_idx * sizeof(uintptr_t));
1609                 vq->async_pkts_idx += pkt_idx;
1610         } else {
1611                 rte_memcpy(&vq->async_pkts_pending[vq->async_pkts_idx],
1612                         pkts, n_free_slot * sizeof(uintptr_t));
1613                 rte_memcpy(&vq->async_pkts_pending[0],
1614                         &pkts[n_free_slot],
1615                         (pkt_idx - n_free_slot) * sizeof(uintptr_t));
1616                 vq->async_pkts_idx = pkt_idx - n_free_slot;
1617         }
1618
1619         if (likely(vq->shadow_used_idx))
1620                 async_flush_shadow_used_ring_split(dev, vq);
1621
1622         return pkt_idx;
1623 }
1624
1625 uint16_t rte_vhost_poll_enqueue_completed(int vid, uint16_t queue_id,
1626                 struct rte_mbuf **pkts, uint16_t count)
1627 {
1628         struct virtio_net *dev = get_device(vid);
1629         struct vhost_virtqueue *vq;
1630         uint16_t n_pkts_cpl = 0, n_pkts_put = 0, n_descs = 0;
1631         uint16_t start_idx, pkts_idx, vq_size;
1632         uint16_t n_inflight;
1633         struct async_inflight_info *pkts_info;
1634
1635         if (!dev)
1636                 return 0;
1637
1638         VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
1639         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
1640                 VHOST_LOG_DATA(ERR, "(%d) %s: invalid virtqueue idx %d.\n",
1641                         dev->vid, __func__, queue_id);
1642                 return 0;
1643         }
1644
1645         vq = dev->virtqueue[queue_id];
1646
1647         if (unlikely(!vq->async_registered)) {
1648                 VHOST_LOG_DATA(ERR, "(%d) %s: async not registered for queue id %d.\n",
1649                         dev->vid, __func__, queue_id);
1650                 return 0;
1651         }
1652
1653         rte_spinlock_lock(&vq->access_lock);
1654
1655         n_inflight = vq->async_pkts_inflight_n;
1656         pkts_idx = vq->async_pkts_idx;
1657         pkts_info = vq->async_pkts_info;
1658         vq_size = vq->size;
1659         start_idx = virtio_dev_rx_async_get_info_idx(pkts_idx,
1660                 vq_size, vq->async_pkts_inflight_n);
1661
1662         if (count > vq->async_last_pkts_n)
1663                 n_pkts_cpl = vq->async_ops.check_completed_copies(vid,
1664                         queue_id, 0, count - vq->async_last_pkts_n);
1665         n_pkts_cpl += vq->async_last_pkts_n;
1666
1667         rte_smp_wmb();
1668
1669         while (likely((n_pkts_put < count) && n_inflight)) {
1670                 uint16_t info_idx = (start_idx + n_pkts_put) & (vq_size - 1);
1671                 if (n_pkts_cpl && pkts_info[info_idx].segs)
1672                         n_pkts_cpl--;
1673                 else if (!n_pkts_cpl && pkts_info[info_idx].segs)
1674                         break;
1675                 n_pkts_put++;
1676                 n_inflight--;
1677                 n_descs += pkts_info[info_idx].descs;
1678         }
1679
1680         vq->async_last_pkts_n = n_pkts_cpl;
1681
1682         if (n_pkts_put) {
1683                 vq->async_pkts_inflight_n = n_inflight;
1684                 if (likely(vq->enabled && vq->access_ok)) {
1685                         __atomic_add_fetch(&vq->used->idx,
1686                                         n_descs, __ATOMIC_RELEASE);
1687                         vhost_vring_call_split(dev, vq);
1688                 }
1689
1690                 if (start_idx + n_pkts_put <= vq_size) {
1691                         rte_memcpy(pkts, &vq->async_pkts_pending[start_idx],
1692                                 n_pkts_put * sizeof(uintptr_t));
1693                 } else {
1694                         rte_memcpy(pkts, &vq->async_pkts_pending[start_idx],
1695                                 (vq_size - start_idx) * sizeof(uintptr_t));
1696                         rte_memcpy(&pkts[vq_size - start_idx],
1697                                 vq->async_pkts_pending,
1698                                 (n_pkts_put + start_idx - vq_size) *
1699                                 sizeof(uintptr_t));
1700                 }
1701         }
1702
1703         rte_spinlock_unlock(&vq->access_lock);
1704
1705         return n_pkts_put;
1706 }
1707
1708 static __rte_always_inline uint32_t
1709 virtio_dev_rx_async_submit(struct virtio_net *dev, uint16_t queue_id,
1710         struct rte_mbuf **pkts, uint32_t count)
1711 {
1712         struct vhost_virtqueue *vq;
1713         uint32_t nb_tx = 0;
1714
1715         VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
1716         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
1717                 VHOST_LOG_DATA(ERR, "(%d) %s: invalid virtqueue idx %d.\n",
1718                         dev->vid, __func__, queue_id);
1719                 return 0;
1720         }
1721
1722         vq = dev->virtqueue[queue_id];
1723
1724         rte_spinlock_lock(&vq->access_lock);
1725
1726         if (unlikely(vq->enabled == 0 || !vq->async_registered))
1727                 goto out_access_unlock;
1728
1729         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1730                 vhost_user_iotlb_rd_lock(vq);
1731
1732         if (unlikely(vq->access_ok == 0))
1733                 if (unlikely(vring_translate(dev, vq) < 0))
1734                         goto out;
1735
1736         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
1737         if (count == 0)
1738                 goto out;
1739
1740         /* TODO: packed queue not implemented */
1741         if (vq_is_packed(dev))
1742                 nb_tx = 0;
1743         else
1744                 nb_tx = virtio_dev_rx_async_submit_split(dev,
1745                                 vq, queue_id, pkts, count);
1746
1747 out:
1748         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1749                 vhost_user_iotlb_rd_unlock(vq);
1750
1751 out_access_unlock:
1752         rte_spinlock_unlock(&vq->access_lock);
1753
1754         return nb_tx;
1755 }
1756
1757 uint16_t
1758 rte_vhost_submit_enqueue_burst(int vid, uint16_t queue_id,
1759                 struct rte_mbuf **pkts, uint16_t count)
1760 {
1761         struct virtio_net *dev = get_device(vid);
1762
1763         if (!dev)
1764                 return 0;
1765
1766         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
1767                 VHOST_LOG_DATA(ERR,
1768                         "(%d) %s: built-in vhost net backend is disabled.\n",
1769                         dev->vid, __func__);
1770                 return 0;
1771         }
1772
1773         return virtio_dev_rx_async_submit(dev, queue_id, pkts, count);
1774 }
1775
1776 static inline bool
1777 virtio_net_with_host_offload(struct virtio_net *dev)
1778 {
1779         if (dev->features &
1780                         ((1ULL << VIRTIO_NET_F_CSUM) |
1781                          (1ULL << VIRTIO_NET_F_HOST_ECN) |
1782                          (1ULL << VIRTIO_NET_F_HOST_TSO4) |
1783                          (1ULL << VIRTIO_NET_F_HOST_TSO6) |
1784                          (1ULL << VIRTIO_NET_F_HOST_UFO)))
1785                 return true;
1786
1787         return false;
1788 }
1789
1790 static void
1791 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
1792 {
1793         struct rte_ipv4_hdr *ipv4_hdr;
1794         struct rte_ipv6_hdr *ipv6_hdr;
1795         void *l3_hdr = NULL;
1796         struct rte_ether_hdr *eth_hdr;
1797         uint16_t ethertype;
1798
1799         eth_hdr = rte_pktmbuf_mtod(m, struct rte_ether_hdr *);
1800
1801         m->l2_len = sizeof(struct rte_ether_hdr);
1802         ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
1803
1804         if (ethertype == RTE_ETHER_TYPE_VLAN) {
1805                 struct rte_vlan_hdr *vlan_hdr =
1806                         (struct rte_vlan_hdr *)(eth_hdr + 1);
1807
1808                 m->l2_len += sizeof(struct rte_vlan_hdr);
1809                 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
1810         }
1811
1812         l3_hdr = (char *)eth_hdr + m->l2_len;
1813
1814         switch (ethertype) {
1815         case RTE_ETHER_TYPE_IPV4:
1816                 ipv4_hdr = l3_hdr;
1817                 *l4_proto = ipv4_hdr->next_proto_id;
1818                 m->l3_len = rte_ipv4_hdr_len(ipv4_hdr);
1819                 *l4_hdr = (char *)l3_hdr + m->l3_len;
1820                 m->ol_flags |= PKT_TX_IPV4;
1821                 break;
1822         case RTE_ETHER_TYPE_IPV6:
1823                 ipv6_hdr = l3_hdr;
1824                 *l4_proto = ipv6_hdr->proto;
1825                 m->l3_len = sizeof(struct rte_ipv6_hdr);
1826                 *l4_hdr = (char *)l3_hdr + m->l3_len;
1827                 m->ol_flags |= PKT_TX_IPV6;
1828                 break;
1829         default:
1830                 m->l3_len = 0;
1831                 *l4_proto = 0;
1832                 *l4_hdr = NULL;
1833                 break;
1834         }
1835 }
1836
1837 static __rte_always_inline void
1838 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
1839 {
1840         uint16_t l4_proto = 0;
1841         void *l4_hdr = NULL;
1842         struct rte_tcp_hdr *tcp_hdr = NULL;
1843
1844         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
1845                 return;
1846
1847         parse_ethernet(m, &l4_proto, &l4_hdr);
1848         if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
1849                 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
1850                         switch (hdr->csum_offset) {
1851                         case (offsetof(struct rte_tcp_hdr, cksum)):
1852                                 if (l4_proto == IPPROTO_TCP)
1853                                         m->ol_flags |= PKT_TX_TCP_CKSUM;
1854                                 break;
1855                         case (offsetof(struct rte_udp_hdr, dgram_cksum)):
1856                                 if (l4_proto == IPPROTO_UDP)
1857                                         m->ol_flags |= PKT_TX_UDP_CKSUM;
1858                                 break;
1859                         case (offsetof(struct rte_sctp_hdr, cksum)):
1860                                 if (l4_proto == IPPROTO_SCTP)
1861                                         m->ol_flags |= PKT_TX_SCTP_CKSUM;
1862                                 break;
1863                         default:
1864                                 break;
1865                         }
1866                 }
1867         }
1868
1869         if (l4_hdr && hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
1870                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
1871                 case VIRTIO_NET_HDR_GSO_TCPV4:
1872                 case VIRTIO_NET_HDR_GSO_TCPV6:
1873                         tcp_hdr = l4_hdr;
1874                         m->ol_flags |= PKT_TX_TCP_SEG;
1875                         m->tso_segsz = hdr->gso_size;
1876                         m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
1877                         break;
1878                 case VIRTIO_NET_HDR_GSO_UDP:
1879                         m->ol_flags |= PKT_TX_UDP_SEG;
1880                         m->tso_segsz = hdr->gso_size;
1881                         m->l4_len = sizeof(struct rte_udp_hdr);
1882                         break;
1883                 default:
1884                         VHOST_LOG_DATA(WARNING,
1885                                 "unsupported gso type %u.\n", hdr->gso_type);
1886                         break;
1887                 }
1888         }
1889 }
1890
1891 static __rte_noinline void
1892 copy_vnet_hdr_from_desc(struct virtio_net_hdr *hdr,
1893                 struct buf_vector *buf_vec)
1894 {
1895         uint64_t len;
1896         uint64_t remain = sizeof(struct virtio_net_hdr);
1897         uint64_t src;
1898         uint64_t dst = (uint64_t)(uintptr_t)hdr;
1899
1900         while (remain) {
1901                 len = RTE_MIN(remain, buf_vec->buf_len);
1902                 src = buf_vec->buf_addr;
1903                 rte_memcpy((void *)(uintptr_t)dst,
1904                                 (void *)(uintptr_t)src, len);
1905
1906                 remain -= len;
1907                 dst += len;
1908                 buf_vec++;
1909         }
1910 }
1911
1912 static __rte_always_inline int
1913 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
1914                   struct buf_vector *buf_vec, uint16_t nr_vec,
1915                   struct rte_mbuf *m, struct rte_mempool *mbuf_pool)
1916 {
1917         uint32_t buf_avail, buf_offset;
1918         uint64_t buf_addr, buf_len;
1919         uint32_t mbuf_avail, mbuf_offset;
1920         uint32_t cpy_len;
1921         struct rte_mbuf *cur = m, *prev = m;
1922         struct virtio_net_hdr tmp_hdr;
1923         struct virtio_net_hdr *hdr = NULL;
1924         /* A counter to avoid desc dead loop chain */
1925         uint16_t vec_idx = 0;
1926         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
1927         int error = 0;
1928
1929         buf_addr = buf_vec[vec_idx].buf_addr;
1930         buf_len = buf_vec[vec_idx].buf_len;
1931
1932         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
1933                 error = -1;
1934                 goto out;
1935         }
1936
1937         if (virtio_net_with_host_offload(dev)) {
1938                 if (unlikely(buf_len < sizeof(struct virtio_net_hdr))) {
1939                         /*
1940                          * No luck, the virtio-net header doesn't fit
1941                          * in a contiguous virtual area.
1942                          */
1943                         copy_vnet_hdr_from_desc(&tmp_hdr, buf_vec);
1944                         hdr = &tmp_hdr;
1945                 } else {
1946                         hdr = (struct virtio_net_hdr *)((uintptr_t)buf_addr);
1947                 }
1948         }
1949
1950         /*
1951          * A virtio driver normally uses at least 2 desc buffers
1952          * for Tx: the first for storing the header, and others
1953          * for storing the data.
1954          */
1955         if (unlikely(buf_len < dev->vhost_hlen)) {
1956                 buf_offset = dev->vhost_hlen - buf_len;
1957                 vec_idx++;
1958                 buf_addr = buf_vec[vec_idx].buf_addr;
1959                 buf_len = buf_vec[vec_idx].buf_len;
1960                 buf_avail  = buf_len - buf_offset;
1961         } else if (buf_len == dev->vhost_hlen) {
1962                 if (unlikely(++vec_idx >= nr_vec))
1963                         goto out;
1964                 buf_addr = buf_vec[vec_idx].buf_addr;
1965                 buf_len = buf_vec[vec_idx].buf_len;
1966
1967                 buf_offset = 0;
1968                 buf_avail = buf_len;
1969         } else {
1970                 buf_offset = dev->vhost_hlen;
1971                 buf_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
1972         }
1973
1974         PRINT_PACKET(dev,
1975                         (uintptr_t)(buf_addr + buf_offset),
1976                         (uint32_t)buf_avail, 0);
1977
1978         mbuf_offset = 0;
1979         mbuf_avail  = m->buf_len - RTE_PKTMBUF_HEADROOM;
1980         while (1) {
1981                 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
1982
1983                 if (likely(cpy_len > MAX_BATCH_LEN ||
1984                                         vq->batch_copy_nb_elems >= vq->size ||
1985                                         (hdr && cur == m))) {
1986                         rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
1987                                                 mbuf_offset),
1988                                         (void *)((uintptr_t)(buf_addr +
1989                                                         buf_offset)), cpy_len);
1990                 } else {
1991                         batch_copy[vq->batch_copy_nb_elems].dst =
1992                                 rte_pktmbuf_mtod_offset(cur, void *,
1993                                                 mbuf_offset);
1994                         batch_copy[vq->batch_copy_nb_elems].src =
1995                                 (void *)((uintptr_t)(buf_addr + buf_offset));
1996                         batch_copy[vq->batch_copy_nb_elems].len = cpy_len;
1997                         vq->batch_copy_nb_elems++;
1998                 }
1999
2000                 mbuf_avail  -= cpy_len;
2001                 mbuf_offset += cpy_len;
2002                 buf_avail -= cpy_len;
2003                 buf_offset += cpy_len;
2004
2005                 /* This buf reaches to its end, get the next one */
2006                 if (buf_avail == 0) {
2007                         if (++vec_idx >= nr_vec)
2008                                 break;
2009
2010                         buf_addr = buf_vec[vec_idx].buf_addr;
2011                         buf_len = buf_vec[vec_idx].buf_len;
2012
2013                         buf_offset = 0;
2014                         buf_avail  = buf_len;
2015
2016                         PRINT_PACKET(dev, (uintptr_t)buf_addr,
2017                                         (uint32_t)buf_avail, 0);
2018                 }
2019
2020                 /*
2021                  * This mbuf reaches to its end, get a new one
2022                  * to hold more data.
2023                  */
2024                 if (mbuf_avail == 0) {
2025                         cur = rte_pktmbuf_alloc(mbuf_pool);
2026                         if (unlikely(cur == NULL)) {
2027                                 VHOST_LOG_DATA(ERR, "Failed to "
2028                                         "allocate memory for mbuf.\n");
2029                                 error = -1;
2030                                 goto out;
2031                         }
2032
2033                         prev->next = cur;
2034                         prev->data_len = mbuf_offset;
2035                         m->nb_segs += 1;
2036                         m->pkt_len += mbuf_offset;
2037                         prev = cur;
2038
2039                         mbuf_offset = 0;
2040                         mbuf_avail  = cur->buf_len - RTE_PKTMBUF_HEADROOM;
2041                 }
2042         }
2043
2044         prev->data_len = mbuf_offset;
2045         m->pkt_len    += mbuf_offset;
2046
2047         if (hdr)
2048                 vhost_dequeue_offload(hdr, m);
2049
2050 out:
2051
2052         return error;
2053 }
2054
2055 static void
2056 virtio_dev_extbuf_free(void *addr __rte_unused, void *opaque)
2057 {
2058         rte_free(opaque);
2059 }
2060
2061 static int
2062 virtio_dev_extbuf_alloc(struct rte_mbuf *pkt, uint32_t size)
2063 {
2064         struct rte_mbuf_ext_shared_info *shinfo = NULL;
2065         uint32_t total_len = RTE_PKTMBUF_HEADROOM + size;
2066         uint16_t buf_len;
2067         rte_iova_t iova;
2068         void *buf;
2069
2070         total_len += sizeof(*shinfo) + sizeof(uintptr_t);
2071         total_len = RTE_ALIGN_CEIL(total_len, sizeof(uintptr_t));
2072
2073         if (unlikely(total_len > UINT16_MAX))
2074                 return -ENOSPC;
2075
2076         buf_len = total_len;
2077         buf = rte_malloc(NULL, buf_len, RTE_CACHE_LINE_SIZE);
2078         if (unlikely(buf == NULL))
2079                 return -ENOMEM;
2080
2081         /* Initialize shinfo */
2082         shinfo = rte_pktmbuf_ext_shinfo_init_helper(buf, &buf_len,
2083                                                 virtio_dev_extbuf_free, buf);
2084         if (unlikely(shinfo == NULL)) {
2085                 rte_free(buf);
2086                 VHOST_LOG_DATA(ERR, "Failed to init shinfo\n");
2087                 return -1;
2088         }
2089
2090         iova = rte_malloc_virt2iova(buf);
2091         rte_pktmbuf_attach_extbuf(pkt, buf, iova, buf_len, shinfo);
2092         rte_pktmbuf_reset_headroom(pkt);
2093
2094         return 0;
2095 }
2096
2097 /*
2098  * Allocate a host supported pktmbuf.
2099  */
2100 static __rte_always_inline struct rte_mbuf *
2101 virtio_dev_pktmbuf_alloc(struct virtio_net *dev, struct rte_mempool *mp,
2102                          uint32_t data_len)
2103 {
2104         struct rte_mbuf *pkt = rte_pktmbuf_alloc(mp);
2105
2106         if (unlikely(pkt == NULL)) {
2107                 VHOST_LOG_DATA(ERR,
2108                         "Failed to allocate memory for mbuf.\n");
2109                 return NULL;
2110         }
2111
2112         if (rte_pktmbuf_tailroom(pkt) >= data_len)
2113                 return pkt;
2114
2115         /* attach an external buffer if supported */
2116         if (dev->extbuf && !virtio_dev_extbuf_alloc(pkt, data_len))
2117                 return pkt;
2118
2119         /* check if chained buffers are allowed */
2120         if (!dev->linearbuf)
2121                 return pkt;
2122
2123         /* Data doesn't fit into the buffer and the host supports
2124          * only linear buffers
2125          */
2126         rte_pktmbuf_free(pkt);
2127
2128         return NULL;
2129 }
2130
2131 static __rte_noinline uint16_t
2132 virtio_dev_tx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
2133         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
2134 {
2135         uint16_t i;
2136         uint16_t free_entries;
2137         uint16_t dropped = 0;
2138         static bool allocerr_warned;
2139
2140         /*
2141          * The ordering between avail index and
2142          * desc reads needs to be enforced.
2143          */
2144         free_entries = __atomic_load_n(&vq->avail->idx, __ATOMIC_ACQUIRE) -
2145                         vq->last_avail_idx;
2146         if (free_entries == 0)
2147                 return 0;
2148
2149         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
2150
2151         VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
2152
2153         count = RTE_MIN(count, MAX_PKT_BURST);
2154         count = RTE_MIN(count, free_entries);
2155         VHOST_LOG_DATA(DEBUG, "(%d) about to dequeue %u buffers\n",
2156                         dev->vid, count);
2157
2158         for (i = 0; i < count; i++) {
2159                 struct buf_vector buf_vec[BUF_VECTOR_MAX];
2160                 uint16_t head_idx;
2161                 uint32_t buf_len;
2162                 uint16_t nr_vec = 0;
2163                 int err;
2164
2165                 if (unlikely(fill_vec_buf_split(dev, vq,
2166                                                 vq->last_avail_idx + i,
2167                                                 &nr_vec, buf_vec,
2168                                                 &head_idx, &buf_len,
2169                                                 VHOST_ACCESS_RO) < 0))
2170                         break;
2171
2172                 update_shadow_used_ring_split(vq, head_idx, 0);
2173
2174                 pkts[i] = virtio_dev_pktmbuf_alloc(dev, mbuf_pool, buf_len);
2175                 if (unlikely(pkts[i] == NULL)) {
2176                         /*
2177                          * mbuf allocation fails for jumbo packets when external
2178                          * buffer allocation is not allowed and linear buffer
2179                          * is required. Drop this packet.
2180                          */
2181                         if (!allocerr_warned) {
2182                                 VHOST_LOG_DATA(ERR,
2183                                         "Failed mbuf alloc of size %d from %s on %s.\n",
2184                                         buf_len, mbuf_pool->name, dev->ifname);
2185                                 allocerr_warned = true;
2186                         }
2187                         dropped += 1;
2188                         i++;
2189                         break;
2190                 }
2191
2192                 err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts[i],
2193                                 mbuf_pool);
2194                 if (unlikely(err)) {
2195                         rte_pktmbuf_free(pkts[i]);
2196                         if (!allocerr_warned) {
2197                                 VHOST_LOG_DATA(ERR,
2198                                         "Failed to copy desc to mbuf on %s.\n",
2199                                         dev->ifname);
2200                                 allocerr_warned = true;
2201                         }
2202                         dropped += 1;
2203                         i++;
2204                         break;
2205                 }
2206         }
2207
2208         vq->last_avail_idx += i;
2209
2210         do_data_copy_dequeue(vq);
2211         if (unlikely(i < count))
2212                 vq->shadow_used_idx = i;
2213         if (likely(vq->shadow_used_idx)) {
2214                 flush_shadow_used_ring_split(dev, vq);
2215                 vhost_vring_call_split(dev, vq);
2216         }
2217
2218         return (i - dropped);
2219 }
2220
2221 static __rte_always_inline int
2222 vhost_reserve_avail_batch_packed(struct virtio_net *dev,
2223                                  struct vhost_virtqueue *vq,
2224                                  struct rte_mempool *mbuf_pool,
2225                                  struct rte_mbuf **pkts,
2226                                  uint16_t avail_idx,
2227                                  uintptr_t *desc_addrs,
2228                                  uint16_t *ids)
2229 {
2230         bool wrap = vq->avail_wrap_counter;
2231         struct vring_packed_desc *descs = vq->desc_packed;
2232         struct virtio_net_hdr *hdr;
2233         uint64_t lens[PACKED_BATCH_SIZE];
2234         uint64_t buf_lens[PACKED_BATCH_SIZE];
2235         uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2236         uint16_t flags, i;
2237
2238         if (unlikely(avail_idx & PACKED_BATCH_MASK))
2239                 return -1;
2240         if (unlikely((avail_idx + PACKED_BATCH_SIZE) > vq->size))
2241                 return -1;
2242
2243         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2244                 flags = descs[avail_idx + i].flags;
2245                 if (unlikely((wrap != !!(flags & VRING_DESC_F_AVAIL)) ||
2246                              (wrap == !!(flags & VRING_DESC_F_USED))  ||
2247                              (flags & PACKED_DESC_SINGLE_DEQUEUE_FLAG)))
2248                         return -1;
2249         }
2250
2251         rte_smp_rmb();
2252
2253         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2254                 lens[i] = descs[avail_idx + i].len;
2255
2256         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2257                 desc_addrs[i] = vhost_iova_to_vva(dev, vq,
2258                                                   descs[avail_idx + i].addr,
2259                                                   &lens[i], VHOST_ACCESS_RW);
2260         }
2261
2262         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2263                 if (unlikely(!desc_addrs[i]))
2264                         return -1;
2265                 if (unlikely((lens[i] != descs[avail_idx + i].len)))
2266                         return -1;
2267         }
2268
2269         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2270                 pkts[i] = virtio_dev_pktmbuf_alloc(dev, mbuf_pool, lens[i]);
2271                 if (!pkts[i])
2272                         goto free_buf;
2273         }
2274
2275         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2276                 buf_lens[i] = pkts[i]->buf_len - pkts[i]->data_off;
2277
2278         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2279                 if (unlikely(buf_lens[i] < (lens[i] - buf_offset)))
2280                         goto free_buf;
2281         }
2282
2283         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2284                 pkts[i]->pkt_len = descs[avail_idx + i].len - buf_offset;
2285                 pkts[i]->data_len = pkts[i]->pkt_len;
2286                 ids[i] = descs[avail_idx + i].id;
2287         }
2288
2289         if (virtio_net_with_host_offload(dev)) {
2290                 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2291                         hdr = (struct virtio_net_hdr *)(desc_addrs[i]);
2292                         vhost_dequeue_offload(hdr, pkts[i]);
2293                 }
2294         }
2295
2296         return 0;
2297
2298 free_buf:
2299         for (i = 0; i < PACKED_BATCH_SIZE; i++)
2300                 rte_pktmbuf_free(pkts[i]);
2301
2302         return -1;
2303 }
2304
2305 static __rte_always_inline int
2306 virtio_dev_tx_batch_packed(struct virtio_net *dev,
2307                            struct vhost_virtqueue *vq,
2308                            struct rte_mempool *mbuf_pool,
2309                            struct rte_mbuf **pkts)
2310 {
2311         uint16_t avail_idx = vq->last_avail_idx;
2312         uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2313         uintptr_t desc_addrs[PACKED_BATCH_SIZE];
2314         uint16_t ids[PACKED_BATCH_SIZE];
2315         uint16_t i;
2316
2317         if (vhost_reserve_avail_batch_packed(dev, vq, mbuf_pool, pkts,
2318                                              avail_idx, desc_addrs, ids))
2319                 return -1;
2320
2321         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2322                 rte_prefetch0((void *)(uintptr_t)desc_addrs[i]);
2323
2324         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2325                 rte_memcpy(rte_pktmbuf_mtod_offset(pkts[i], void *, 0),
2326                            (void *)(uintptr_t)(desc_addrs[i] + buf_offset),
2327                            pkts[i]->pkt_len);
2328
2329         if (virtio_net_is_inorder(dev))
2330                 vhost_shadow_dequeue_batch_packed_inorder(vq,
2331                         ids[PACKED_BATCH_SIZE - 1]);
2332         else
2333                 vhost_shadow_dequeue_batch_packed(dev, vq, ids);
2334
2335         vq_inc_last_avail_packed(vq, PACKED_BATCH_SIZE);
2336
2337         return 0;
2338 }
2339
2340 static __rte_always_inline int
2341 vhost_dequeue_single_packed(struct virtio_net *dev,
2342                             struct vhost_virtqueue *vq,
2343                             struct rte_mempool *mbuf_pool,
2344                             struct rte_mbuf **pkts,
2345                             uint16_t *buf_id,
2346                             uint16_t *desc_count)
2347 {
2348         struct buf_vector buf_vec[BUF_VECTOR_MAX];
2349         uint32_t buf_len;
2350         uint16_t nr_vec = 0;
2351         int err;
2352         static bool allocerr_warned;
2353
2354         if (unlikely(fill_vec_buf_packed(dev, vq,
2355                                          vq->last_avail_idx, desc_count,
2356                                          buf_vec, &nr_vec,
2357                                          buf_id, &buf_len,
2358                                          VHOST_ACCESS_RO) < 0))
2359                 return -1;
2360
2361         *pkts = virtio_dev_pktmbuf_alloc(dev, mbuf_pool, buf_len);
2362         if (unlikely(*pkts == NULL)) {
2363                 if (!allocerr_warned) {
2364                         VHOST_LOG_DATA(ERR,
2365                                 "Failed mbuf alloc of size %d from %s on %s.\n",
2366                                 buf_len, mbuf_pool->name, dev->ifname);
2367                         allocerr_warned = true;
2368                 }
2369                 return -1;
2370         }
2371
2372         err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, *pkts,
2373                                 mbuf_pool);
2374         if (unlikely(err)) {
2375                 if (!allocerr_warned) {
2376                         VHOST_LOG_DATA(ERR,
2377                                 "Failed to copy desc to mbuf on %s.\n",
2378                                 dev->ifname);
2379                         allocerr_warned = true;
2380                 }
2381                 rte_pktmbuf_free(*pkts);
2382                 return -1;
2383         }
2384
2385         return 0;
2386 }
2387
2388 static __rte_always_inline int
2389 virtio_dev_tx_single_packed(struct virtio_net *dev,
2390                             struct vhost_virtqueue *vq,
2391                             struct rte_mempool *mbuf_pool,
2392                             struct rte_mbuf **pkts)
2393 {
2394
2395         uint16_t buf_id, desc_count = 0;
2396         int ret;
2397
2398         ret = vhost_dequeue_single_packed(dev, vq, mbuf_pool, pkts, &buf_id,
2399                                         &desc_count);
2400
2401         if (likely(desc_count > 0)) {
2402                 if (virtio_net_is_inorder(dev))
2403                         vhost_shadow_dequeue_single_packed_inorder(vq, buf_id,
2404                                                                    desc_count);
2405                 else
2406                         vhost_shadow_dequeue_single_packed(vq, buf_id,
2407                                         desc_count);
2408
2409                 vq_inc_last_avail_packed(vq, desc_count);
2410         }
2411
2412         return ret;
2413 }
2414
2415 static __rte_noinline uint16_t
2416 virtio_dev_tx_packed(struct virtio_net *dev,
2417                      struct vhost_virtqueue *__rte_restrict vq,
2418                      struct rte_mempool *mbuf_pool,
2419                      struct rte_mbuf **__rte_restrict pkts,
2420                      uint32_t count)
2421 {
2422         uint32_t pkt_idx = 0;
2423         uint32_t remained = count;
2424
2425         do {
2426                 rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
2427
2428                 if (remained >= PACKED_BATCH_SIZE) {
2429                         if (!virtio_dev_tx_batch_packed(dev, vq, mbuf_pool,
2430                                                         &pkts[pkt_idx])) {
2431                                 pkt_idx += PACKED_BATCH_SIZE;
2432                                 remained -= PACKED_BATCH_SIZE;
2433                                 continue;
2434                         }
2435                 }
2436
2437                 if (virtio_dev_tx_single_packed(dev, vq, mbuf_pool,
2438                                                 &pkts[pkt_idx]))
2439                         break;
2440                 pkt_idx++;
2441                 remained--;
2442
2443         } while (remained);
2444
2445         if (vq->shadow_used_idx) {
2446                 do_data_copy_dequeue(vq);
2447
2448                 vhost_flush_dequeue_shadow_packed(dev, vq);
2449                 vhost_vring_call_packed(dev, vq);
2450         }
2451
2452         return pkt_idx;
2453 }
2454
2455 uint16_t
2456 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
2457         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
2458 {
2459         struct virtio_net *dev;
2460         struct rte_mbuf *rarp_mbuf = NULL;
2461         struct vhost_virtqueue *vq;
2462         int16_t success = 1;
2463
2464         dev = get_device(vid);
2465         if (!dev)
2466                 return 0;
2467
2468         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
2469                 VHOST_LOG_DATA(ERR,
2470                         "(%d) %s: built-in vhost net backend is disabled.\n",
2471                         dev->vid, __func__);
2472                 return 0;
2473         }
2474
2475         if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
2476                 VHOST_LOG_DATA(ERR,
2477                         "(%d) %s: invalid virtqueue idx %d.\n",
2478                         dev->vid, __func__, queue_id);
2479                 return 0;
2480         }
2481
2482         vq = dev->virtqueue[queue_id];
2483
2484         if (unlikely(rte_spinlock_trylock(&vq->access_lock) == 0))
2485                 return 0;
2486
2487         if (unlikely(vq->enabled == 0)) {
2488                 count = 0;
2489                 goto out_access_unlock;
2490         }
2491
2492         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
2493                 vhost_user_iotlb_rd_lock(vq);
2494
2495         if (unlikely(vq->access_ok == 0))
2496                 if (unlikely(vring_translate(dev, vq) < 0)) {
2497                         count = 0;
2498                         goto out;
2499                 }
2500
2501         /*
2502          * Construct a RARP broadcast packet, and inject it to the "pkts"
2503          * array, to looks like that guest actually send such packet.
2504          *
2505          * Check user_send_rarp() for more information.
2506          *
2507          * broadcast_rarp shares a cacheline in the virtio_net structure
2508          * with some fields that are accessed during enqueue and
2509          * __atomic_compare_exchange_n causes a write if performed compare
2510          * and exchange. This could result in false sharing between enqueue
2511          * and dequeue.
2512          *
2513          * Prevent unnecessary false sharing by reading broadcast_rarp first
2514          * and only performing compare and exchange if the read indicates it
2515          * is likely to be set.
2516          */
2517         if (unlikely(__atomic_load_n(&dev->broadcast_rarp, __ATOMIC_ACQUIRE) &&
2518                         __atomic_compare_exchange_n(&dev->broadcast_rarp,
2519                         &success, 0, 0, __ATOMIC_RELEASE, __ATOMIC_RELAXED))) {
2520
2521                 rarp_mbuf = rte_net_make_rarp_packet(mbuf_pool, &dev->mac);
2522                 if (rarp_mbuf == NULL) {
2523                         VHOST_LOG_DATA(ERR, "Failed to make RARP packet.\n");
2524                         count = 0;
2525                         goto out;
2526                 }
2527                 count -= 1;
2528         }
2529
2530         if (vq_is_packed(dev))
2531                 count = virtio_dev_tx_packed(dev, vq, mbuf_pool, pkts, count);
2532         else
2533                 count = virtio_dev_tx_split(dev, vq, mbuf_pool, pkts, count);
2534
2535 out:
2536         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
2537                 vhost_user_iotlb_rd_unlock(vq);
2538
2539 out_access_unlock:
2540         rte_spinlock_unlock(&vq->access_lock);
2541
2542         if (unlikely(rarp_mbuf != NULL)) {
2543                 /*
2544                  * Inject it to the head of "pkts" array, so that switch's mac
2545                  * learning table will get updated first.
2546                  */
2547                 memmove(&pkts[1], pkts, count * sizeof(struct rte_mbuf *));
2548                 pkts[0] = rarp_mbuf;
2549                 count += 1;
2550         }
2551
2552         return count;
2553 }