lib/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_net.h>
  12 #include <rte_ether.h>
  13 #include <rte_ip.h>
  14 #include <rte_vhost.h>
  15 #include <rte_tcp.h>
  16 #include <rte_udp.h>
  17 #include <rte_sctp.h>
  18 #include <rte_arp.h>
  19 #include <rte_spinlock.h>
  20 #include <rte_malloc.h>
  21 #include <rte_vhost_async.h>
  22
  23 #include "iotlb.h"
  24 #include "vhost.h"
  25
  26 #define MAX_BATCH_LEN 256
  27
  28 #define VHOST_ASYNC_BATCH_THRESHOLD 32
  29
  30 static  __rte_always_inline bool
  31 rxvq_is_mergeable(struct virtio_net *dev)
  32 {
  33         return dev->features & (1ULL << VIRTIO_NET_F_MRG_RXBUF);
  34 }
  35
  36 static  __rte_always_inline bool
  37 virtio_net_is_inorder(struct virtio_net *dev)
  38 {
  39         return dev->features & (1ULL << VIRTIO_F_IN_ORDER);
  40 }
  41
  42 static bool
  43 is_valid_virt_queue_idx(uint32_t idx, int is_tx, uint32_t nr_vring)
  44 {
  45         return (is_tx ^ (idx & 1)) == 0 && idx < nr_vring;
  46 }
  47
  48 static inline void
  49 do_data_copy_enqueue(struct virtio_net *dev, struct vhost_virtqueue *vq)
  50 {
  51         struct batch_copy_elem *elem = vq->batch_copy_elems;
  52         uint16_t count = vq->batch_copy_nb_elems;
  53         int i;
  54
  55         for (i = 0; i < count; i++) {
  56                 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
  57                 vhost_log_cache_write_iova(dev, vq, elem[i].log_addr,
  58                                            elem[i].len);
  59                 PRINT_PACKET(dev, (uintptr_t)elem[i].dst, elem[i].len, 0);
  60         }
  61
  62         vq->batch_copy_nb_elems = 0;
  63 }
  64
  65 static inline void
  66 do_data_copy_dequeue(struct vhost_virtqueue *vq)
  67 {
  68         struct batch_copy_elem *elem = vq->batch_copy_elems;
  69         uint16_t count = vq->batch_copy_nb_elems;
  70         int i;
  71
  72         for (i = 0; i < count; i++)
  73                 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
  74
  75         vq->batch_copy_nb_elems = 0;
  76 }
  77
  78 static __rte_always_inline void
  79 do_flush_shadow_used_ring_split(struct virtio_net *dev,
  80                         struct vhost_virtqueue *vq,
  81                         uint16_t to, uint16_t from, uint16_t size)
  82 {
  83         rte_memcpy(&vq->used->ring[to],
  84                         &vq->shadow_used_split[from],
  85                         size * sizeof(struct vring_used_elem));
  86         vhost_log_cache_used_vring(dev, vq,
  87                         offsetof(struct vring_used, ring[to]),
  88                         size * sizeof(struct vring_used_elem));
  89 }
  90
  91 static __rte_always_inline void
  92 flush_shadow_used_ring_split(struct virtio_net *dev, struct vhost_virtqueue *vq)
  93 {
  94         uint16_t used_idx = vq->last_used_idx & (vq->size - 1);
  95
  96         if (used_idx + vq->shadow_used_idx <= vq->size) {
  97                 do_flush_shadow_used_ring_split(dev, vq, used_idx, 0,
  98                                           vq->shadow_used_idx);
  99         } else {
 100                 uint16_t size;
 101
 102                 /* update used ring interval [used_idx, vq->size] */
 103                 size = vq->size - used_idx;
 104                 do_flush_shadow_used_ring_split(dev, vq, used_idx, 0, size);
 105
 106                 /* update the left half used ring interval [0, left_size] */
 107                 do_flush_shadow_used_ring_split(dev, vq, 0, size,
 108                                           vq->shadow_used_idx - size);
 109         }
 110         vq->last_used_idx += vq->shadow_used_idx;
 111
 112         vhost_log_cache_sync(dev, vq);
 113
 114         __atomic_add_fetch(&vq->used->idx, vq->shadow_used_idx,
 115                            __ATOMIC_RELEASE);
 116         vq->shadow_used_idx = 0;
 117         vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
 118                 sizeof(vq->used->idx));
 119 }
 120
 121 static __rte_always_inline void
 122 update_shadow_used_ring_split(struct vhost_virtqueue *vq,
 123                          uint16_t desc_idx, uint32_t len)
 124 {
 125         uint16_t i = vq->shadow_used_idx++;
 126
 127         vq->shadow_used_split[i].id  = desc_idx;
 128         vq->shadow_used_split[i].len = len;
 129 }
 130
 131 static __rte_always_inline void
 132 vhost_flush_enqueue_shadow_packed(struct virtio_net *dev,
 133                                   struct vhost_virtqueue *vq)
 134 {
 135         int i;
 136         uint16_t used_idx = vq->last_used_idx;
 137         uint16_t head_idx = vq->last_used_idx;
 138         uint16_t head_flags = 0;
 139
 140         /* Split loop in two to save memory barriers */
 141         for (i = 0; i < vq->shadow_used_idx; i++) {
 142                 vq->desc_packed[used_idx].id = vq->shadow_used_packed[i].id;
 143                 vq->desc_packed[used_idx].len = vq->shadow_used_packed[i].len;
 144
 145                 used_idx += vq->shadow_used_packed[i].count;
 146                 if (used_idx >= vq->size)
 147                         used_idx -= vq->size;
 148         }
 149
 150         /* The ordering for storing desc flags needs to be enforced. */
 151         rte_atomic_thread_fence(__ATOMIC_RELEASE);
 152
 153         for (i = 0; i < vq->shadow_used_idx; i++) {
 154                 uint16_t flags;
 155
 156                 if (vq->shadow_used_packed[i].len)
 157                         flags = VRING_DESC_F_WRITE;
 158                 else
 159                         flags = 0;
 160
 161                 if (vq->used_wrap_counter) {
 162                         flags |= VRING_DESC_F_USED;
 163                         flags |= VRING_DESC_F_AVAIL;
 164                 } else {
 165                         flags &= ~VRING_DESC_F_USED;
 166                         flags &= ~VRING_DESC_F_AVAIL;
 167                 }
 168
 169                 if (i > 0) {
 170                         vq->desc_packed[vq->last_used_idx].flags = flags;
 171
 172                         vhost_log_cache_used_vring(dev, vq,
 173                                         vq->last_used_idx *
 174                                         sizeof(struct vring_packed_desc),
 175                                         sizeof(struct vring_packed_desc));
 176                 } else {
 177                         head_idx = vq->last_used_idx;
 178                         head_flags = flags;
 179                 }
 180
 181                 vq_inc_last_used_packed(vq, vq->shadow_used_packed[i].count);
 182         }
 183
 184         vq->desc_packed[head_idx].flags = head_flags;
 185
 186         vhost_log_cache_used_vring(dev, vq,
 187                                 head_idx *
 188                                 sizeof(struct vring_packed_desc),
 189                                 sizeof(struct vring_packed_desc));
 190
 191         vq->shadow_used_idx = 0;
 192         vhost_log_cache_sync(dev, vq);
 193 }
 194
 195 static __rte_always_inline void
 196 vhost_flush_dequeue_shadow_packed(struct virtio_net *dev,
 197                                   struct vhost_virtqueue *vq)
 198 {
 199         struct vring_used_elem_packed *used_elem = &vq->shadow_used_packed[0];
 200
 201         vq->desc_packed[vq->shadow_last_used_idx].id = used_elem->id;
 202         /* desc flags is the synchronization point for virtio packed vring */
 203         __atomic_store_n(&vq->desc_packed[vq->shadow_last_used_idx].flags,
 204                          used_elem->flags, __ATOMIC_RELEASE);
 205
 206         vhost_log_cache_used_vring(dev, vq, vq->shadow_last_used_idx *
 207                                    sizeof(struct vring_packed_desc),
 208                                    sizeof(struct vring_packed_desc));
 209         vq->shadow_used_idx = 0;
 210         vhost_log_cache_sync(dev, vq);
 211 }
 212
 213 static __rte_always_inline void
 214 vhost_flush_enqueue_batch_packed(struct virtio_net *dev,
 215                                  struct vhost_virtqueue *vq,
 216                                  uint64_t *lens,
 217                                  uint16_t *ids)
 218 {
 219         uint16_t i;
 220         uint16_t flags;
 221         uint16_t last_used_idx;
 222         struct vring_packed_desc *desc_base;
 223
 224         last_used_idx = vq->last_used_idx;
 225         desc_base = &vq->desc_packed[last_used_idx];
 226
 227         flags = PACKED_DESC_ENQUEUE_USED_FLAG(vq->used_wrap_counter);
 228
 229         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
 230                 desc_base[i].id = ids[i];
 231                 desc_base[i].len = lens[i];
 232         }
 233
 234         rte_atomic_thread_fence(__ATOMIC_RELEASE);
 235
 236         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
 237                 desc_base[i].flags = flags;
 238         }
 239
 240         vhost_log_cache_used_vring(dev, vq, last_used_idx *
 241                                    sizeof(struct vring_packed_desc),
 242                                    sizeof(struct vring_packed_desc) *
 243                                    PACKED_BATCH_SIZE);
 244         vhost_log_cache_sync(dev, vq);
 245
 246         vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
 247 }
 248
 249 static __rte_always_inline void
 250 vhost_shadow_dequeue_batch_packed_inorder(struct vhost_virtqueue *vq,
 251                                           uint16_t id)
 252 {
 253         vq->shadow_used_packed[0].id = id;
 254
 255         if (!vq->shadow_used_idx) {
 256                 vq->shadow_last_used_idx = vq->last_used_idx;
 257                 vq->shadow_used_packed[0].flags =
 258                         PACKED_DESC_DEQUEUE_USED_FLAG(vq->used_wrap_counter);
 259                 vq->shadow_used_packed[0].len = 0;
 260                 vq->shadow_used_packed[0].count = 1;
 261                 vq->shadow_used_idx++;
 262         }
 263
 264         vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
 265 }
 266
 267 static __rte_always_inline void
 268 vhost_shadow_dequeue_batch_packed(struct virtio_net *dev,
 269                                   struct vhost_virtqueue *vq,
 270                                   uint16_t *ids)
 271 {
 272         uint16_t flags;
 273         uint16_t i;
 274         uint16_t begin;
 275
 276         flags = PACKED_DESC_DEQUEUE_USED_FLAG(vq->used_wrap_counter);
 277
 278         if (!vq->shadow_used_idx) {
 279                 vq->shadow_last_used_idx = vq->last_used_idx;
 280                 vq->shadow_used_packed[0].id  = ids[0];
 281                 vq->shadow_used_packed[0].len = 0;
 282                 vq->shadow_used_packed[0].count = 1;
 283                 vq->shadow_used_packed[0].flags = flags;
 284                 vq->shadow_used_idx++;
 285                 begin = 1;
 286         } else
 287                 begin = 0;
 288
 289         vhost_for_each_try_unroll(i, begin, PACKED_BATCH_SIZE) {
 290                 vq->desc_packed[vq->last_used_idx + i].id = ids[i];
 291                 vq->desc_packed[vq->last_used_idx + i].len = 0;
 292         }
 293
 294         rte_atomic_thread_fence(__ATOMIC_RELEASE);
 295         vhost_for_each_try_unroll(i, begin, PACKED_BATCH_SIZE)
 296                 vq->desc_packed[vq->last_used_idx + i].flags = flags;
 297
 298         vhost_log_cache_used_vring(dev, vq, vq->last_used_idx *
 299                                    sizeof(struct vring_packed_desc),
 300                                    sizeof(struct vring_packed_desc) *
 301                                    PACKED_BATCH_SIZE);
 302         vhost_log_cache_sync(dev, vq);
 303
 304         vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
 305 }
 306
 307 static __rte_always_inline void
 308 vhost_shadow_dequeue_single_packed(struct vhost_virtqueue *vq,
 309                                    uint16_t buf_id,
 310                                    uint16_t count)
 311 {
 312         uint16_t flags;
 313
 314         flags = vq->desc_packed[vq->last_used_idx].flags;
 315         if (vq->used_wrap_counter) {
 316                 flags |= VRING_DESC_F_USED;
 317                 flags |= VRING_DESC_F_AVAIL;
 318         } else {
 319                 flags &= ~VRING_DESC_F_USED;
 320                 flags &= ~VRING_DESC_F_AVAIL;
 321         }
 322
 323         if (!vq->shadow_used_idx) {
 324                 vq->shadow_last_used_idx = vq->last_used_idx;
 325
 326                 vq->shadow_used_packed[0].id  = buf_id;
 327                 vq->shadow_used_packed[0].len = 0;
 328                 vq->shadow_used_packed[0].flags = flags;
 329                 vq->shadow_used_idx++;
 330         } else {
 331                 vq->desc_packed[vq->last_used_idx].id = buf_id;
 332                 vq->desc_packed[vq->last_used_idx].len = 0;
 333                 vq->desc_packed[vq->last_used_idx].flags = flags;
 334         }
 335
 336         vq_inc_last_used_packed(vq, count);
 337 }
 338
 339 static __rte_always_inline void
 340 vhost_shadow_dequeue_single_packed_inorder(struct vhost_virtqueue *vq,
 341                                            uint16_t buf_id,
 342                                            uint16_t count)
 343 {
 344         uint16_t flags;
 345
 346         vq->shadow_used_packed[0].id = buf_id;
 347
 348         flags = vq->desc_packed[vq->last_used_idx].flags;
 349         if (vq->used_wrap_counter) {
 350                 flags |= VRING_DESC_F_USED;
 351                 flags |= VRING_DESC_F_AVAIL;
 352         } else {
 353                 flags &= ~VRING_DESC_F_USED;
 354                 flags &= ~VRING_DESC_F_AVAIL;
 355         }
 356
 357         if (!vq->shadow_used_idx) {
 358                 vq->shadow_last_used_idx = vq->last_used_idx;
 359                 vq->shadow_used_packed[0].len = 0;
 360                 vq->shadow_used_packed[0].flags = flags;
 361                 vq->shadow_used_idx++;
 362         }
 363
 364         vq_inc_last_used_packed(vq, count);
 365 }
 366
 367 static __rte_always_inline void
 368 vhost_shadow_enqueue_packed(struct vhost_virtqueue *vq,
 369                                    uint32_t *len,
 370                                    uint16_t *id,
 371                                    uint16_t *count,
 372                                    uint16_t num_buffers)
 373 {
 374         uint16_t i;
 375
 376         for (i = 0; i < num_buffers; i++) {
 377                 /* enqueue shadow flush action aligned with batch num */
 378                 if (!vq->shadow_used_idx)
 379                         vq->shadow_aligned_idx = vq->last_used_idx &
 380                                 PACKED_BATCH_MASK;
 381                 vq->shadow_used_packed[vq->shadow_used_idx].id  = id[i];
 382                 vq->shadow_used_packed[vq->shadow_used_idx].len = len[i];
 383                 vq->shadow_used_packed[vq->shadow_used_idx].count = count[i];
 384                 vq->shadow_aligned_idx += count[i];
 385                 vq->shadow_used_idx++;
 386         }
 387 }
 388
 389 static __rte_always_inline void
 390 vhost_shadow_enqueue_single_packed(struct virtio_net *dev,
 391                                    struct vhost_virtqueue *vq,
 392                                    uint32_t *len,
 393                                    uint16_t *id,
 394                                    uint16_t *count,
 395                                    uint16_t num_buffers)
 396 {
 397         vhost_shadow_enqueue_packed(vq, len, id, count, num_buffers);
 398
 399         if (vq->shadow_aligned_idx >= PACKED_BATCH_SIZE) {
 400                 do_data_copy_enqueue(dev, vq);
 401                 vhost_flush_enqueue_shadow_packed(dev, vq);
 402         }
 403 }
 404
 405 /* avoid write operation when necessary, to lessen cache issues */
 406 #define ASSIGN_UNLESS_EQUAL(var, val) do {      \
 407         if ((var) != (val))                     \
 408                 (var) = (val);                  \
 409 } while (0)
 410
 411 static __rte_always_inline void
 412 virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
 413 {
 414         uint64_t csum_l4 = m_buf->ol_flags & RTE_MBUF_F_TX_L4_MASK;
 415
 416         if (m_buf->ol_flags & RTE_MBUF_F_TX_TCP_SEG)
 417                 csum_l4 |= RTE_MBUF_F_TX_TCP_CKSUM;
 418
 419         if (csum_l4) {
 420                 net_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 421                 net_hdr->csum_start = m_buf->l2_len + m_buf->l3_len;
 422
 423                 switch (csum_l4) {
 424                 case RTE_MBUF_F_TX_TCP_CKSUM:
 425                         net_hdr->csum_offset = (offsetof(struct rte_tcp_hdr,
 426                                                 cksum));
 427                         break;
 428                 case RTE_MBUF_F_TX_UDP_CKSUM:
 429                         net_hdr->csum_offset = (offsetof(struct rte_udp_hdr,
 430                                                 dgram_cksum));
 431                         break;
 432                 case RTE_MBUF_F_TX_SCTP_CKSUM:
 433                         net_hdr->csum_offset = (offsetof(struct rte_sctp_hdr,
 434                                                 cksum));
 435                         break;
 436                 }
 437         } else {
 438                 ASSIGN_UNLESS_EQUAL(net_hdr->csum_start, 0);
 439                 ASSIGN_UNLESS_EQUAL(net_hdr->csum_offset, 0);
 440                 ASSIGN_UNLESS_EQUAL(net_hdr->flags, 0);
 441         }
 442
 443         /* IP cksum verification cannot be bypassed, then calculate here */
 444         if (m_buf->ol_flags & RTE_MBUF_F_TX_IP_CKSUM) {
 445                 struct rte_ipv4_hdr *ipv4_hdr;
 446
 447                 ipv4_hdr = rte_pktmbuf_mtod_offset(m_buf, struct rte_ipv4_hdr *,
 448                                                    m_buf->l2_len);
 449                 ipv4_hdr->hdr_checksum = 0;
 450                 ipv4_hdr->hdr_checksum = rte_ipv4_cksum(ipv4_hdr);
 451         }
 452
 453         if (m_buf->ol_flags & RTE_MBUF_F_TX_TCP_SEG) {
 454                 if (m_buf->ol_flags & RTE_MBUF_F_TX_IPV4)
 455                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 456                 else
 457                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 458                 net_hdr->gso_size = m_buf->tso_segsz;
 459                 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len
 460                                         + m_buf->l4_len;
 461         } else if (m_buf->ol_flags & RTE_MBUF_F_TX_UDP_SEG) {
 462                 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
 463                 net_hdr->gso_size = m_buf->tso_segsz;
 464                 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len +
 465                         m_buf->l4_len;
 466         } else {
 467                 ASSIGN_UNLESS_EQUAL(net_hdr->gso_type, 0);
 468                 ASSIGN_UNLESS_EQUAL(net_hdr->gso_size, 0);
 469                 ASSIGN_UNLESS_EQUAL(net_hdr->hdr_len, 0);
 470         }
 471 }
 472
 473 static __rte_always_inline int
 474 map_one_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 475                 struct buf_vector *buf_vec, uint16_t *vec_idx,
 476                 uint64_t desc_iova, uint64_t desc_len, uint8_t perm)
 477 {
 478         uint16_t vec_id = *vec_idx;
 479
 480         while (desc_len) {
 481                 uint64_t desc_addr;
 482                 uint64_t desc_chunck_len = desc_len;
 483
 484                 if (unlikely(vec_id >= BUF_VECTOR_MAX))
 485                         return -1;
 486
 487                 desc_addr = vhost_iova_to_vva(dev, vq,
 488                                 desc_iova,
 489                                 &desc_chunck_len,
 490                                 perm);
 491                 if (unlikely(!desc_addr))
 492                         return -1;
 493
 494                 rte_prefetch0((void *)(uintptr_t)desc_addr);
 495
 496                 buf_vec[vec_id].buf_iova = desc_iova;
 497                 buf_vec[vec_id].buf_addr = desc_addr;
 498                 buf_vec[vec_id].buf_len  = desc_chunck_len;
 499
 500                 desc_len -= desc_chunck_len;
 501                 desc_iova += desc_chunck_len;
 502                 vec_id++;
 503         }
 504         *vec_idx = vec_id;
 505
 506         return 0;
 507 }
 508
 509 static __rte_always_inline int
 510 fill_vec_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 511                          uint32_t avail_idx, uint16_t *vec_idx,
 512                          struct buf_vector *buf_vec, uint16_t *desc_chain_head,
 513                          uint32_t *desc_chain_len, uint8_t perm)
 514 {
 515         uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
 516         uint16_t vec_id = *vec_idx;
 517         uint32_t len    = 0;
 518         uint64_t dlen;
 519         uint32_t nr_descs = vq->size;
 520         uint32_t cnt    = 0;
 521         struct vring_desc *descs = vq->desc;
 522         struct vring_desc *idesc = NULL;
 523
 524         if (unlikely(idx >= vq->size))
 525                 return -1;
 526
 527         *desc_chain_head = idx;
 528
 529         if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
 530                 dlen = vq->desc[idx].len;
 531                 nr_descs = dlen / sizeof(struct vring_desc);
 532                 if (unlikely(nr_descs > vq->size))
 533                         return -1;
 534
 535                 descs = (struct vring_desc *)(uintptr_t)
 536                         vhost_iova_to_vva(dev, vq, vq->desc[idx].addr,
 537                                                 &dlen,
 538                                                 VHOST_ACCESS_RO);
 539                 if (unlikely(!descs))
 540                         return -1;
 541
 542                 if (unlikely(dlen < vq->desc[idx].len)) {
 543                         /*
 544                          * The indirect desc table is not contiguous
 545                          * in process VA space, we have to copy it.
 546                          */
 547                         idesc = vhost_alloc_copy_ind_table(dev, vq,
 548                                         vq->desc[idx].addr, vq->desc[idx].len);
 549                         if (unlikely(!idesc))
 550                                 return -1;
 551
 552                         descs = idesc;
 553                 }
 554
 555                 idx = 0;
 556         }
 557
 558         while (1) {
 559                 if (unlikely(idx >= nr_descs || cnt++ >= nr_descs)) {
 560                         free_ind_table(idesc);
 561                         return -1;
 562                 }
 563
 564                 dlen = descs[idx].len;
 565                 len += dlen;
 566
 567                 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 568                                                 descs[idx].addr, dlen,
 569                                                 perm))) {
 570                         free_ind_table(idesc);
 571                         return -1;
 572                 }
 573
 574                 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
 575                         break;
 576
 577                 idx = descs[idx].next;
 578         }
 579
 580         *desc_chain_len = len;
 581         *vec_idx = vec_id;
 582
 583         if (unlikely(!!idesc))
 584                 free_ind_table(idesc);
 585
 586         return 0;
 587 }
 588
 589 /*
 590  * Returns -1 on fail, 0 on success
 591  */
 592 static inline int
 593 reserve_avail_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 594                                 uint32_t size, struct buf_vector *buf_vec,
 595                                 uint16_t *num_buffers, uint16_t avail_head,
 596                                 uint16_t *nr_vec)
 597 {
 598         uint16_t cur_idx;
 599         uint16_t vec_idx = 0;
 600         uint16_t max_tries, tries = 0;
 601
 602         uint16_t head_idx = 0;
 603         uint32_t len = 0;
 604
 605         *num_buffers = 0;
 606         cur_idx  = vq->last_avail_idx;
 607
 608         if (rxvq_is_mergeable(dev))
 609                 max_tries = vq->size - 1;
 610         else
 611                 max_tries = 1;
 612
 613         while (size > 0) {
 614                 if (unlikely(cur_idx == avail_head))
 615                         return -1;
 616                 /*
 617                  * if we tried all available ring items, and still
 618                  * can't get enough buf, it means something abnormal
 619                  * happened.
 620                  */
 621                 if (unlikely(++tries > max_tries))
 622                         return -1;
 623
 624                 if (unlikely(fill_vec_buf_split(dev, vq, cur_idx,
 625                                                 &vec_idx, buf_vec,
 626                                                 &head_idx, &len,
 627                                                 VHOST_ACCESS_RW) < 0))
 628                         return -1;
 629                 len = RTE_MIN(len, size);
 630                 update_shadow_used_ring_split(vq, head_idx, len);
 631                 size -= len;
 632
 633                 cur_idx++;
 634                 *num_buffers += 1;
 635         }
 636
 637         *nr_vec = vec_idx;
 638
 639         return 0;
 640 }
 641
 642 static __rte_always_inline int
 643 fill_vec_buf_packed_indirect(struct virtio_net *dev,
 644                         struct vhost_virtqueue *vq,
 645                         struct vring_packed_desc *desc, uint16_t *vec_idx,
 646                         struct buf_vector *buf_vec, uint32_t *len, uint8_t perm)
 647 {
 648         uint16_t i;
 649         uint32_t nr_descs;
 650         uint16_t vec_id = *vec_idx;
 651         uint64_t dlen;
 652         struct vring_packed_desc *descs, *idescs = NULL;
 653
 654         dlen = desc->len;
 655         descs = (struct vring_packed_desc *)(uintptr_t)
 656                 vhost_iova_to_vva(dev, vq, desc->addr, &dlen, VHOST_ACCESS_RO);
 657         if (unlikely(!descs))
 658                 return -1;
 659
 660         if (unlikely(dlen < desc->len)) {
 661                 /*
 662                  * The indirect desc table is not contiguous
 663                  * in process VA space, we have to copy it.
 664                  */
 665                 idescs = vhost_alloc_copy_ind_table(dev,
 666                                 vq, desc->addr, desc->len);
 667                 if (unlikely(!idescs))
 668                         return -1;
 669
 670                 descs = idescs;
 671         }
 672
 673         nr_descs =  desc->len / sizeof(struct vring_packed_desc);
 674         if (unlikely(nr_descs >= vq->size)) {
 675                 free_ind_table(idescs);
 676                 return -1;
 677         }
 678
 679         for (i = 0; i < nr_descs; i++) {
 680                 if (unlikely(vec_id >= BUF_VECTOR_MAX)) {
 681                         free_ind_table(idescs);
 682                         return -1;
 683                 }
 684
 685                 dlen = descs[i].len;
 686                 *len += dlen;
 687                 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 688                                                 descs[i].addr, dlen,
 689                                                 perm)))
 690                         return -1;
 691         }
 692         *vec_idx = vec_id;
 693
 694         if (unlikely(!!idescs))
 695                 free_ind_table(idescs);
 696
 697         return 0;
 698 }
 699
 700 static __rte_always_inline int
 701 fill_vec_buf_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
 702                                 uint16_t avail_idx, uint16_t *desc_count,
 703                                 struct buf_vector *buf_vec, uint16_t *vec_idx,
 704                                 uint16_t *buf_id, uint32_t *len, uint8_t perm)
 705 {
 706         bool wrap_counter = vq->avail_wrap_counter;
 707         struct vring_packed_desc *descs = vq->desc_packed;
 708         uint16_t vec_id = *vec_idx;
 709         uint64_t dlen;
 710
 711         if (avail_idx < vq->last_avail_idx)
 712                 wrap_counter ^= 1;
 713
 714         /*
 715          * Perform a load-acquire barrier in desc_is_avail to
 716          * enforce the ordering between desc flags and desc
 717          * content.
 718          */
 719         if (unlikely(!desc_is_avail(&descs[avail_idx], wrap_counter)))
 720                 return -1;
 721
 722         *desc_count = 0;
 723         *len = 0;
 724
 725         while (1) {
 726                 if (unlikely(vec_id >= BUF_VECTOR_MAX))
 727                         return -1;
 728
 729                 if (unlikely(*desc_count >= vq->size))
 730                         return -1;
 731
 732                 *desc_count += 1;
 733                 *buf_id = descs[avail_idx].id;
 734
 735                 if (descs[avail_idx].flags & VRING_DESC_F_INDIRECT) {
 736                         if (unlikely(fill_vec_buf_packed_indirect(dev, vq,
 737                                                         &descs[avail_idx],
 738                                                         &vec_id, buf_vec,
 739                                                         len, perm) < 0))
 740                                 return -1;
 741                 } else {
 742                         dlen = descs[avail_idx].len;
 743                         *len += dlen;
 744
 745                         if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 746                                                         descs[avail_idx].addr,
 747                                                         dlen,
 748                                                         perm)))
 749                                 return -1;
 750                 }
 751
 752                 if ((descs[avail_idx].flags & VRING_DESC_F_NEXT) == 0)
 753                         break;
 754
 755                 if (++avail_idx >= vq->size) {
 756                         avail_idx -= vq->size;
 757                         wrap_counter ^= 1;
 758                 }
 759         }
 760
 761         *vec_idx = vec_id;
 762
 763         return 0;
 764 }
 765
 766 static __rte_noinline void
 767 copy_vnet_hdr_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 768                 struct buf_vector *buf_vec,
 769                 struct virtio_net_hdr_mrg_rxbuf *hdr)
 770 {
 771         uint64_t len;
 772         uint64_t remain = dev->vhost_hlen;
 773         uint64_t src = (uint64_t)(uintptr_t)hdr, dst;
 774         uint64_t iova = buf_vec->buf_iova;
 775
 776         while (remain) {
 777                 len = RTE_MIN(remain,
 778                                 buf_vec->buf_len);
 779                 dst = buf_vec->buf_addr;
 780                 rte_memcpy((void *)(uintptr_t)dst,
 781                                 (void *)(uintptr_t)src,
 782                                 len);
 783
 784                 PRINT_PACKET(dev, (uintptr_t)dst,
 785                                 (uint32_t)len, 0);
 786                 vhost_log_cache_write_iova(dev, vq,
 787                                 iova, len);
 788
 789                 remain -= len;
 790                 iova += len;
 791                 src += len;
 792                 buf_vec++;
 793         }
 794 }
 795
 796 static __rte_always_inline int
 797 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 798                             struct rte_mbuf *m, struct buf_vector *buf_vec,
 799                             uint16_t nr_vec, uint16_t num_buffers)
 800 {
 801         uint32_t vec_idx = 0;
 802         uint32_t mbuf_offset, mbuf_avail;
 803         uint32_t buf_offset, buf_avail;
 804         uint64_t buf_addr, buf_iova, buf_len;
 805         uint32_t cpy_len;
 806         uint64_t hdr_addr;
 807         struct rte_mbuf *hdr_mbuf;
 808         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 809         struct virtio_net_hdr_mrg_rxbuf tmp_hdr, *hdr = NULL;
 810         int error = 0;
 811
 812         if (unlikely(m == NULL)) {
 813                 error = -1;
 814                 goto out;
 815         }
 816
 817         buf_addr = buf_vec[vec_idx].buf_addr;
 818         buf_iova = buf_vec[vec_idx].buf_iova;
 819         buf_len = buf_vec[vec_idx].buf_len;
 820
 821         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
 822                 error = -1;
 823                 goto out;
 824         }
 825
 826         hdr_mbuf = m;
 827         hdr_addr = buf_addr;
 828         if (unlikely(buf_len < dev->vhost_hlen)) {
 829                 memset(&tmp_hdr, 0, sizeof(struct virtio_net_hdr_mrg_rxbuf));
 830                 hdr = &tmp_hdr;
 831         } else
 832                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)hdr_addr;
 833
 834         VHOST_LOG_DATA(DEBUG, "(%d) RX: num merge buffers %d\n",
 835                 dev->vid, num_buffers);
 836
 837         if (unlikely(buf_len < dev->vhost_hlen)) {
 838                 buf_offset = dev->vhost_hlen - buf_len;
 839                 vec_idx++;
 840                 buf_addr = buf_vec[vec_idx].buf_addr;
 841                 buf_iova = buf_vec[vec_idx].buf_iova;
 842                 buf_len = buf_vec[vec_idx].buf_len;
 843                 buf_avail = buf_len - buf_offset;
 844         } else {
 845                 buf_offset = dev->vhost_hlen;
 846                 buf_avail = buf_len - dev->vhost_hlen;
 847         }
 848
 849         mbuf_avail  = rte_pktmbuf_data_len(m);
 850         mbuf_offset = 0;
 851         while (mbuf_avail != 0 || m->next != NULL) {
 852                 /* done with current buf, get the next one */
 853                 if (buf_avail == 0) {
 854                         vec_idx++;
 855                         if (unlikely(vec_idx >= nr_vec)) {
 856                                 error = -1;
 857                                 goto out;
 858                         }
 859
 860                         buf_addr = buf_vec[vec_idx].buf_addr;
 861                         buf_iova = buf_vec[vec_idx].buf_iova;
 862                         buf_len = buf_vec[vec_idx].buf_len;
 863
 864                         buf_offset = 0;
 865                         buf_avail  = buf_len;
 866                 }
 867
 868                 /* done with current mbuf, get the next one */
 869                 if (mbuf_avail == 0) {
 870                         m = m->next;
 871
 872                         mbuf_offset = 0;
 873                         mbuf_avail  = rte_pktmbuf_data_len(m);
 874                 }
 875
 876                 if (hdr_addr) {
 877                         virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
 878                         if (rxvq_is_mergeable(dev))
 879                                 ASSIGN_UNLESS_EQUAL(hdr->num_buffers,
 880                                                 num_buffers);
 881
 882                         if (unlikely(hdr == &tmp_hdr)) {
 883                                 copy_vnet_hdr_to_desc(dev, vq, buf_vec, hdr);
 884                         } else {
 885                                 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
 886                                                 dev->vhost_hlen, 0);
 887                                 vhost_log_cache_write_iova(dev, vq,
 888                                                 buf_vec[0].buf_iova,
 889                                                 dev->vhost_hlen);
 890                         }
 891
 892                         hdr_addr = 0;
 893                 }
 894
 895                 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
 896
 897                 if (likely(cpy_len > MAX_BATCH_LEN ||
 898                                         vq->batch_copy_nb_elems >= vq->size)) {
 899                         rte_memcpy((void *)((uintptr_t)(buf_addr + buf_offset)),
 900                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 901                                 cpy_len);
 902                         vhost_log_cache_write_iova(dev, vq,
 903                                                    buf_iova + buf_offset,
 904                                                    cpy_len);
 905                         PRINT_PACKET(dev, (uintptr_t)(buf_addr + buf_offset),
 906                                 cpy_len, 0);
 907                 } else {
 908                         batch_copy[vq->batch_copy_nb_elems].dst =
 909                                 (void *)((uintptr_t)(buf_addr + buf_offset));
 910                         batch_copy[vq->batch_copy_nb_elems].src =
 911                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
 912                         batch_copy[vq->batch_copy_nb_elems].log_addr =
 913                                 buf_iova + buf_offset;
 914                         batch_copy[vq->batch_copy_nb_elems].len = cpy_len;
 915                         vq->batch_copy_nb_elems++;
 916                 }
 917
 918                 mbuf_avail  -= cpy_len;
 919                 mbuf_offset += cpy_len;
 920                 buf_avail  -= cpy_len;
 921                 buf_offset += cpy_len;
 922         }
 923
 924 out:
 925
 926         return error;
 927 }
 928
 929 static __rte_always_inline void
 930 async_fill_vec(struct iovec *v, void *base, size_t len)
 931 {
 932         v->iov_base = base;
 933         v->iov_len = len;
 934 }
 935
 936 static __rte_always_inline void
 937 async_fill_iter(struct rte_vhost_iov_iter *it, size_t count,
 938         struct iovec *vec, unsigned long nr_seg)
 939 {
 940         it->offset = 0;
 941         it->count = count;
 942
 943         if (count) {
 944                 it->iov = vec;
 945                 it->nr_segs = nr_seg;
 946         } else {
 947                 it->iov = 0;
 948                 it->nr_segs = 0;
 949         }
 950 }
 951
 952 static __rte_always_inline void
 953 async_fill_desc(struct rte_vhost_async_desc *desc,
 954         struct rte_vhost_iov_iter *src, struct rte_vhost_iov_iter *dst)
 955 {
 956         desc->src = src;
 957         desc->dst = dst;
 958 }
 959
 960 static __rte_always_inline int
 961 async_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 962                         struct rte_mbuf *m, struct buf_vector *buf_vec,
 963                         uint16_t nr_vec, uint16_t num_buffers,
 964                         struct iovec *src_iovec, struct iovec *dst_iovec,
 965                         struct rte_vhost_iov_iter *src_it,
 966                         struct rte_vhost_iov_iter *dst_it)
 967 {
 968         uint32_t vec_idx = 0;
 969         uint32_t mbuf_offset, mbuf_avail;
 970         uint32_t buf_offset, buf_avail;
 971         uint64_t buf_addr, buf_iova, buf_len;
 972         uint32_t cpy_len, cpy_threshold;
 973         uint64_t hdr_addr;
 974         struct rte_mbuf *hdr_mbuf;
 975         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 976         struct virtio_net_hdr_mrg_rxbuf tmp_hdr, *hdr = NULL;
 977         int error = 0;
 978         uint64_t mapped_len;
 979
 980         uint32_t tlen = 0;
 981         int tvec_idx = 0;
 982         void *hpa;
 983
 984         if (unlikely(m == NULL)) {
 985                 error = -1;
 986                 goto out;
 987         }
 988
 989         cpy_threshold = vq->async_threshold;
 990
 991         buf_addr = buf_vec[vec_idx].buf_addr;
 992         buf_iova = buf_vec[vec_idx].buf_iova;
 993         buf_len = buf_vec[vec_idx].buf_len;
 994
 995         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
 996                 error = -1;
 997                 goto out;
 998         }
 999
1000         hdr_mbuf = m;
1001         hdr_addr = buf_addr;
1002         if (unlikely(buf_len < dev->vhost_hlen)) {
1003                 memset(&tmp_hdr, 0, sizeof(struct virtio_net_hdr_mrg_rxbuf));
1004                 hdr = &tmp_hdr;
1005         } else
1006                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)hdr_addr;
1007
1008         VHOST_LOG_DATA(DEBUG, "(%d) RX: num merge buffers %d\n",
1009                 dev->vid, num_buffers);
1010
1011         if (unlikely(buf_len < dev->vhost_hlen)) {
1012                 buf_offset = dev->vhost_hlen - buf_len;
1013                 vec_idx++;
1014                 buf_addr = buf_vec[vec_idx].buf_addr;
1015                 buf_iova = buf_vec[vec_idx].buf_iova;
1016                 buf_len = buf_vec[vec_idx].buf_len;
1017                 buf_avail = buf_len - buf_offset;
1018         } else {
1019                 buf_offset = dev->vhost_hlen;
1020                 buf_avail = buf_len - dev->vhost_hlen;
1021         }
1022
1023         mbuf_avail  = rte_pktmbuf_data_len(m);
1024         mbuf_offset = 0;
1025
1026         while (mbuf_avail != 0 || m->next != NULL) {
1027                 /* done with current buf, get the next one */
1028                 if (buf_avail == 0) {
1029                         vec_idx++;
1030                         if (unlikely(vec_idx >= nr_vec)) {
1031                                 error = -1;
1032                                 goto out;
1033                         }
1034
1035                         buf_addr = buf_vec[vec_idx].buf_addr;
1036                         buf_iova = buf_vec[vec_idx].buf_iova;
1037                         buf_len = buf_vec[vec_idx].buf_len;
1038
1039                         buf_offset = 0;
1040                         buf_avail  = buf_len;
1041                 }
1042
1043                 /* done with current mbuf, get the next one */
1044                 if (mbuf_avail == 0) {
1045                         m = m->next;
1046
1047                         mbuf_offset = 0;
1048                         mbuf_avail  = rte_pktmbuf_data_len(m);
1049                 }
1050
1051                 if (hdr_addr) {
1052                         virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
1053                         if (rxvq_is_mergeable(dev))
1054                                 ASSIGN_UNLESS_EQUAL(hdr->num_buffers,
1055                                                 num_buffers);
1056
1057                         if (unlikely(hdr == &tmp_hdr)) {
1058                                 copy_vnet_hdr_to_desc(dev, vq, buf_vec, hdr);
1059                         } else {
1060                                 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
1061                                                 dev->vhost_hlen, 0);
1062                                 vhost_log_cache_write_iova(dev, vq,
1063                                                 buf_vec[0].buf_iova,
1064                                                 dev->vhost_hlen);
1065                         }
1066
1067                         hdr_addr = 0;
1068                 }
1069
1070                 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
1071
1072                 while (unlikely(cpy_len && cpy_len >= cpy_threshold)) {
1073                         hpa = (void *)(uintptr_t)gpa_to_first_hpa(dev,
1074                                         buf_iova + buf_offset,
1075                                         cpy_len, &mapped_len);
1076
1077                         if (unlikely(!hpa || mapped_len < cpy_threshold))
1078                                 break;
1079
1080                         async_fill_vec(src_iovec + tvec_idx,
1081                                 (void *)(uintptr_t)rte_pktmbuf_iova_offset(m,
1082                                 mbuf_offset), (size_t)mapped_len);
1083
1084                         async_fill_vec(dst_iovec + tvec_idx,
1085                                         hpa, (size_t)mapped_len);
1086
1087                         tlen += (uint32_t)mapped_len;
1088                         cpy_len -= (uint32_t)mapped_len;
1089                         mbuf_avail  -= (uint32_t)mapped_len;
1090                         mbuf_offset += (uint32_t)mapped_len;
1091                         buf_avail  -= (uint32_t)mapped_len;
1092                         buf_offset += (uint32_t)mapped_len;
1093                         tvec_idx++;
1094                 }
1095
1096                 if (likely(cpy_len)) {
1097                         if (unlikely(vq->batch_copy_nb_elems >= vq->size)) {
1098                                 rte_memcpy(
1099                                 (void *)((uintptr_t)(buf_addr + buf_offset)),
1100                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
1101                                 cpy_len);
1102
1103                                 PRINT_PACKET(dev,
1104                                         (uintptr_t)(buf_addr + buf_offset),
1105                                         cpy_len, 0);
1106                         } else {
1107                                 batch_copy[vq->batch_copy_nb_elems].dst =
1108                                 (void *)((uintptr_t)(buf_addr + buf_offset));
1109                                 batch_copy[vq->batch_copy_nb_elems].src =
1110                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
1111                                 batch_copy[vq->batch_copy_nb_elems].log_addr =
1112                                         buf_iova + buf_offset;
1113                                 batch_copy[vq->batch_copy_nb_elems].len =
1114                                         cpy_len;
1115                                 vq->batch_copy_nb_elems++;
1116                         }
1117
1118                         mbuf_avail  -= cpy_len;
1119                         mbuf_offset += cpy_len;
1120                         buf_avail  -= cpy_len;
1121                         buf_offset += cpy_len;
1122                 }
1123
1124         }
1125
1126 out:
1127         if (tlen) {
1128                 async_fill_iter(src_it, tlen, src_iovec, tvec_idx);
1129                 async_fill_iter(dst_it, tlen, dst_iovec, tvec_idx);
1130         } else {
1131                 src_it->count = 0;
1132         }
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_sync_batch_check(struct virtio_net *dev,
1257                            struct vhost_virtqueue *vq,
1258                            struct rte_mbuf **pkts,
1259                            uint64_t *desc_addrs,
1260                            uint64_t *lens)
1261 {
1262         bool wrap_counter = vq->avail_wrap_counter;
1263         struct vring_packed_desc *descs = vq->desc_packed;
1264         uint16_t avail_idx = vq->last_avail_idx;
1265         uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1266         uint16_t i;
1267
1268         if (unlikely(avail_idx & PACKED_BATCH_MASK))
1269                 return -1;
1270
1271         if (unlikely((avail_idx + PACKED_BATCH_SIZE) > vq->size))
1272                 return -1;
1273
1274         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1275                 if (unlikely(pkts[i]->next != NULL))
1276                         return -1;
1277                 if (unlikely(!desc_is_avail(&descs[avail_idx + i],
1278                                             wrap_counter)))
1279                         return -1;
1280         }
1281
1282         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1283                 lens[i] = descs[avail_idx + i].len;
1284
1285         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1286                 if (unlikely(pkts[i]->pkt_len > (lens[i] - buf_offset)))
1287                         return -1;
1288         }
1289
1290         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1291                 desc_addrs[i] = vhost_iova_to_vva(dev, vq,
1292                                                   descs[avail_idx + i].addr,
1293                                                   &lens[i],
1294                                                   VHOST_ACCESS_RW);
1295
1296         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1297                 if (unlikely(!desc_addrs[i]))
1298                         return -1;
1299                 if (unlikely(lens[i] != descs[avail_idx + i].len))
1300                         return -1;
1301         }
1302
1303         return 0;
1304 }
1305
1306 static __rte_always_inline int
1307 virtio_dev_rx_async_batch_check(struct virtio_net *dev,
1308                            struct vhost_virtqueue *vq,
1309                            struct rte_mbuf **pkts,
1310                            uint64_t *desc_addrs,
1311                            uint64_t *lens)
1312 {
1313         bool wrap_counter = vq->avail_wrap_counter;
1314         struct vring_packed_desc *descs = vq->desc_packed;
1315         uint16_t avail_idx = vq->last_avail_idx;
1316         uint16_t used_idx = vq->last_used_idx;
1317         uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1318         uint32_t cpy_threshold = vq->async_threshold;
1319         uint16_t i;
1320
1321         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1322                 if (unlikely(pkts[i]->data_len >= cpy_threshold))
1323                         return -1;
1324         }
1325
1326         if (unlikely(avail_idx & PACKED_BATCH_MASK))
1327                 return -1;
1328
1329         if (unlikely((avail_idx + PACKED_BATCH_SIZE) > vq->size))
1330                 return -1;
1331
1332         if (unlikely((used_idx + PACKED_BATCH_SIZE) > vq->size))
1333                 return -1;
1334
1335         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1336                 if (unlikely(pkts[i]->next != NULL))
1337                         return -1;
1338                 if (unlikely(!desc_is_avail(&descs[avail_idx + i],
1339                                             wrap_counter)))
1340                         return -1;
1341         }
1342
1343         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1344                 lens[i] = descs[avail_idx + i].len;
1345
1346         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1347                 if (unlikely(pkts[i]->pkt_len > (lens[i] - buf_offset)))
1348                         return -1;
1349         }
1350
1351         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1352                 desc_addrs[i] = vhost_iova_to_vva(dev, vq,
1353                                                   descs[avail_idx + i].addr,
1354                                                   &lens[i],
1355                                                   VHOST_ACCESS_RW);
1356
1357         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1358                 if (unlikely(!desc_addrs[i]))
1359                         return -1;
1360                 if (unlikely(lens[i] != descs[avail_idx + i].len))
1361                         return -1;
1362         }
1363
1364         return 0;
1365 }
1366
1367 static __rte_always_inline void
1368 virtio_dev_rx_batch_packed_copy(struct virtio_net *dev,
1369                            struct vhost_virtqueue *vq,
1370                            struct rte_mbuf **pkts,
1371                            uint64_t *desc_addrs,
1372                            uint64_t *lens)
1373 {
1374         uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1375         struct virtio_net_hdr_mrg_rxbuf *hdrs[PACKED_BATCH_SIZE];
1376         struct vring_packed_desc *descs = vq->desc_packed;
1377         uint16_t avail_idx = vq->last_avail_idx;
1378         uint16_t ids[PACKED_BATCH_SIZE];
1379         uint16_t i;
1380
1381         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1382                 rte_prefetch0((void *)(uintptr_t)desc_addrs[i]);
1383                 hdrs[i] = (struct virtio_net_hdr_mrg_rxbuf *)
1384                                         (uintptr_t)desc_addrs[i];
1385                 lens[i] = pkts[i]->pkt_len +
1386                         sizeof(struct virtio_net_hdr_mrg_rxbuf);
1387         }
1388
1389         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1390                 virtio_enqueue_offload(pkts[i], &hdrs[i]->hdr);
1391
1392         vq_inc_last_avail_packed(vq, PACKED_BATCH_SIZE);
1393
1394         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1395                 rte_memcpy((void *)(uintptr_t)(desc_addrs[i] + buf_offset),
1396                            rte_pktmbuf_mtod_offset(pkts[i], void *, 0),
1397                            pkts[i]->pkt_len);
1398         }
1399
1400         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1401                 vhost_log_cache_write_iova(dev, vq, descs[avail_idx + i].addr,
1402                                            lens[i]);
1403
1404         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1405                 ids[i] = descs[avail_idx + i].id;
1406
1407         vhost_flush_enqueue_batch_packed(dev, vq, lens, ids);
1408 }
1409
1410 static __rte_always_inline int
1411 virtio_dev_rx_sync_batch_packed(struct virtio_net *dev,
1412                            struct vhost_virtqueue *vq,
1413                            struct rte_mbuf **pkts)
1414 {
1415         uint64_t desc_addrs[PACKED_BATCH_SIZE];
1416         uint64_t lens[PACKED_BATCH_SIZE];
1417
1418         if (virtio_dev_rx_sync_batch_check(dev, vq, pkts, desc_addrs, lens) == -1)
1419                 return -1;
1420
1421         if (vq->shadow_used_idx) {
1422                 do_data_copy_enqueue(dev, vq);
1423                 vhost_flush_enqueue_shadow_packed(dev, vq);
1424         }
1425
1426         virtio_dev_rx_batch_packed_copy(dev, vq, pkts, desc_addrs, lens);
1427
1428         return 0;
1429 }
1430
1431 static __rte_always_inline int
1432 virtio_dev_rx_async_batch_packed(struct virtio_net *dev,
1433                            struct vhost_virtqueue *vq,
1434                            struct rte_mbuf **pkts,
1435                            struct rte_mbuf **comp_pkts, uint32_t *pkt_done)
1436 {
1437         uint16_t i;
1438         uint64_t desc_addrs[PACKED_BATCH_SIZE];
1439         uint64_t lens[PACKED_BATCH_SIZE];
1440
1441         if (virtio_dev_rx_async_batch_check(dev, vq, pkts, desc_addrs, lens) == -1)
1442                 return -1;
1443
1444         virtio_dev_rx_batch_packed_copy(dev, vq, pkts, desc_addrs, lens);
1445
1446         if (vq->shadow_used_idx) {
1447                 do_data_copy_enqueue(dev, vq);
1448                 vhost_flush_enqueue_shadow_packed(dev, vq);
1449         }
1450
1451         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1452                 comp_pkts[(*pkt_done)++] = pkts[i];
1453
1454         return 0;
1455 }
1456
1457 static __rte_always_inline int16_t
1458 virtio_dev_rx_single_packed(struct virtio_net *dev,
1459                             struct vhost_virtqueue *vq,
1460                             struct rte_mbuf *pkt)
1461 {
1462         struct buf_vector buf_vec[BUF_VECTOR_MAX];
1463         uint16_t nr_descs = 0;
1464
1465         if (unlikely(vhost_enqueue_single_packed(dev, vq, pkt, buf_vec,
1466                                                  &nr_descs) < 0)) {
1467                 VHOST_LOG_DATA(DEBUG,
1468                                 "(%d) failed to get enough desc from vring\n",
1469                                 dev->vid);
1470                 return -1;
1471         }
1472
1473         VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
1474                         dev->vid, vq->last_avail_idx,
1475                         vq->last_avail_idx + nr_descs);
1476
1477         vq_inc_last_avail_packed(vq, nr_descs);
1478
1479         return 0;
1480 }
1481
1482 static __rte_noinline uint32_t
1483 virtio_dev_rx_packed(struct virtio_net *dev,
1484                      struct vhost_virtqueue *__rte_restrict vq,
1485                      struct rte_mbuf **__rte_restrict pkts,
1486                      uint32_t count)
1487 {
1488         uint32_t pkt_idx = 0;
1489
1490         do {
1491                 rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
1492
1493                 if (count - pkt_idx >= PACKED_BATCH_SIZE) {
1494                         if (!virtio_dev_rx_sync_batch_packed(dev, vq,
1495                                                         &pkts[pkt_idx])) {
1496                                 pkt_idx += PACKED_BATCH_SIZE;
1497                                 continue;
1498                         }
1499                 }
1500
1501                 if (virtio_dev_rx_single_packed(dev, vq, pkts[pkt_idx]))
1502                         break;
1503                 pkt_idx++;
1504
1505         } while (pkt_idx < count);
1506
1507         if (vq->shadow_used_idx) {
1508                 do_data_copy_enqueue(dev, vq);
1509                 vhost_flush_enqueue_shadow_packed(dev, vq);
1510         }
1511
1512         if (pkt_idx)
1513                 vhost_vring_call_packed(dev, vq);
1514
1515         return pkt_idx;
1516 }
1517
1518 static __rte_always_inline uint32_t
1519 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
1520         struct rte_mbuf **pkts, uint32_t count)
1521 {
1522         struct vhost_virtqueue *vq;
1523         uint32_t nb_tx = 0;
1524
1525         VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
1526         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
1527                 VHOST_LOG_DATA(ERR, "(%d) %s: invalid virtqueue idx %d.\n",
1528                         dev->vid, __func__, queue_id);
1529                 return 0;
1530         }
1531
1532         vq = dev->virtqueue[queue_id];
1533
1534         rte_spinlock_lock(&vq->access_lock);
1535
1536         if (unlikely(!vq->enabled))
1537                 goto out_access_unlock;
1538
1539         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1540                 vhost_user_iotlb_rd_lock(vq);
1541
1542         if (unlikely(!vq->access_ok))
1543                 if (unlikely(vring_translate(dev, vq) < 0))
1544                         goto out;
1545
1546         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
1547         if (count == 0)
1548                 goto out;
1549
1550         if (vq_is_packed(dev))
1551                 nb_tx = virtio_dev_rx_packed(dev, vq, pkts, count);
1552         else
1553                 nb_tx = virtio_dev_rx_split(dev, vq, pkts, count);
1554
1555 out:
1556         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1557                 vhost_user_iotlb_rd_unlock(vq);
1558
1559 out_access_unlock:
1560         rte_spinlock_unlock(&vq->access_lock);
1561
1562         return nb_tx;
1563 }
1564
1565 uint16_t
1566 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
1567         struct rte_mbuf **__rte_restrict pkts, uint16_t count)
1568 {
1569         struct virtio_net *dev = get_device(vid);
1570
1571         if (!dev)
1572                 return 0;
1573
1574         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
1575                 VHOST_LOG_DATA(ERR,
1576                         "(%d) %s: built-in vhost net backend is disabled.\n",
1577                         dev->vid, __func__);
1578                 return 0;
1579         }
1580
1581         return virtio_dev_rx(dev, queue_id, pkts, count);
1582 }
1583
1584 static __rte_always_inline uint16_t
1585 virtio_dev_rx_async_get_info_idx(uint16_t pkts_idx,
1586         uint16_t vq_size, uint16_t n_inflight)
1587 {
1588         return pkts_idx > n_inflight ? (pkts_idx - n_inflight) :
1589                 (vq_size - n_inflight + pkts_idx) % vq_size;
1590 }
1591
1592 static __rte_always_inline void
1593 store_dma_desc_info_split(struct vring_used_elem *s_ring, struct vring_used_elem *d_ring,
1594                 uint16_t ring_size, uint16_t s_idx, uint16_t d_idx, uint16_t count)
1595 {
1596         size_t elem_size = sizeof(struct vring_used_elem);
1597
1598         if (d_idx + count <= ring_size) {
1599                 rte_memcpy(d_ring + d_idx, s_ring + s_idx, count * elem_size);
1600         } else {
1601                 uint16_t size = ring_size - d_idx;
1602
1603                 rte_memcpy(d_ring + d_idx, s_ring + s_idx, size * elem_size);
1604                 rte_memcpy(d_ring, s_ring + s_idx + size, (count - size) * elem_size);
1605         }
1606 }
1607
1608 static __rte_always_inline void
1609 store_dma_desc_info_packed(struct vring_used_elem_packed *s_ring,
1610                 struct vring_used_elem_packed *d_ring,
1611                 uint16_t ring_size, uint16_t s_idx, uint16_t d_idx, uint16_t count)
1612 {
1613         size_t elem_size = sizeof(struct vring_used_elem_packed);
1614
1615         if (d_idx + count <= ring_size) {
1616                 rte_memcpy(d_ring + d_idx, s_ring + s_idx, count * elem_size);
1617         } else {
1618                 uint16_t size = ring_size - d_idx;
1619
1620                 rte_memcpy(d_ring + d_idx, s_ring + s_idx, size * elem_size);
1621                 rte_memcpy(d_ring, s_ring + s_idx + size, (count - size) * elem_size);
1622         }
1623 }
1624
1625 static __rte_noinline uint32_t
1626 virtio_dev_rx_async_submit_split(struct virtio_net *dev,
1627         struct vhost_virtqueue *vq, uint16_t queue_id,
1628         struct rte_mbuf **pkts, uint32_t count,
1629         struct rte_mbuf **comp_pkts, uint32_t *comp_count)
1630 {
1631         uint32_t pkt_idx = 0, pkt_burst_idx = 0;
1632         uint16_t num_buffers;
1633         struct buf_vector buf_vec[BUF_VECTOR_MAX];
1634         uint16_t avail_head;
1635
1636         struct rte_vhost_iov_iter *it_pool = vq->it_pool;
1637         struct iovec *vec_pool = vq->vec_pool;
1638         struct rte_vhost_async_desc tdes[MAX_PKT_BURST];
1639         struct iovec *src_iovec = vec_pool;
1640         struct iovec *dst_iovec = vec_pool + (VHOST_MAX_ASYNC_VEC >> 1);
1641         uint16_t slot_idx = 0;
1642         uint16_t segs_await = 0;
1643         uint16_t iovec_idx = 0, it_idx = 0;
1644         struct async_inflight_info *pkts_info = vq->async_pkts_info;
1645         uint32_t n_pkts = 0, pkt_err = 0;
1646         uint32_t num_async_pkts = 0, num_done_pkts = 0;
1647         int32_t n_xfer;
1648         struct {
1649                 uint16_t pkt_idx;
1650                 uint16_t last_avail_idx;
1651         } async_pkts_log[MAX_PKT_BURST];
1652
1653         /*
1654          * The ordering between avail index and desc reads need to be enforced.
1655          */
1656         avail_head = __atomic_load_n(&vq->avail->idx, __ATOMIC_ACQUIRE);
1657
1658         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
1659
1660         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
1661                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
1662                 uint16_t nr_vec = 0;
1663
1664                 if (unlikely(reserve_avail_buf_split(dev, vq,
1665                                                 pkt_len, buf_vec, &num_buffers,
1666                                                 avail_head, &nr_vec) < 0)) {
1667                         VHOST_LOG_DATA(DEBUG,
1668                                 "(%d) failed to get enough desc from vring\n",
1669                                 dev->vid);
1670                         vq->shadow_used_idx -= num_buffers;
1671                         break;
1672                 }
1673
1674                 VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
1675                         dev->vid, vq->last_avail_idx,
1676                         vq->last_avail_idx + num_buffers);
1677
1678                 if (async_mbuf_to_desc(dev, vq, pkts[pkt_idx], buf_vec, nr_vec, num_buffers,
1679                                 &src_iovec[iovec_idx], &dst_iovec[iovec_idx],
1680                                 &it_pool[it_idx], &it_pool[it_idx + 1]) < 0) {
1681                         vq->shadow_used_idx -= num_buffers;
1682                         break;
1683                 }
1684
1685                 slot_idx = (vq->async_pkts_idx + num_async_pkts) &
1686                         (vq->size - 1);
1687                 if (it_pool[it_idx].count) {
1688                         uint16_t from, to;
1689
1690                         async_fill_desc(&tdes[pkt_burst_idx++],
1691                                 &it_pool[it_idx], &it_pool[it_idx + 1]);
1692                         pkts_info[slot_idx].descs = num_buffers;
1693                         pkts_info[slot_idx].mbuf = pkts[pkt_idx];
1694                         async_pkts_log[num_async_pkts].pkt_idx = pkt_idx;
1695                         async_pkts_log[num_async_pkts++].last_avail_idx =
1696                                 vq->last_avail_idx;
1697
1698                         iovec_idx += it_pool[it_idx].nr_segs;
1699                         it_idx += 2;
1700
1701                         segs_await += it_pool[it_idx].nr_segs;
1702
1703                         /**
1704                          * recover shadow used ring and keep DMA-occupied
1705                          * descriptors.
1706                          */
1707                         from = vq->shadow_used_idx - num_buffers;
1708                         to = vq->async_desc_idx_split & (vq->size - 1);
1709
1710                         store_dma_desc_info_split(vq->shadow_used_split,
1711                                         vq->async_descs_split, vq->size, from, to, num_buffers);
1712
1713                         vq->async_desc_idx_split += num_buffers;
1714                         vq->shadow_used_idx -= num_buffers;
1715                 } else
1716                         comp_pkts[num_done_pkts++] = pkts[pkt_idx];
1717
1718                 vq->last_avail_idx += num_buffers;
1719
1720                 /*
1721                  * conditions to trigger async device transfer:
1722                  * - buffered packet number reaches transfer threshold
1723                  * - unused async iov number is less than max vhost vector
1724                  */
1725                 if (unlikely(pkt_burst_idx >= VHOST_ASYNC_BATCH_THRESHOLD ||
1726                         ((VHOST_MAX_ASYNC_VEC >> 1) - segs_await <
1727                         BUF_VECTOR_MAX))) {
1728                         n_xfer = vq->async_ops.transfer_data(dev->vid,
1729                                         queue_id, tdes, 0, pkt_burst_idx);
1730                         if (n_xfer >= 0) {
1731                                 n_pkts = n_xfer;
1732                         } else {
1733                                 VHOST_LOG_DATA(ERR,
1734                                         "(%d) %s: failed to transfer data for queue id %d.\n",
1735                                         dev->vid, __func__, queue_id);
1736                                 n_pkts = 0;
1737                         }
1738
1739                         iovec_idx = 0;
1740                         it_idx = 0;
1741
1742                         segs_await = 0;
1743                         vq->async_pkts_inflight_n += n_pkts;
1744
1745                         if (unlikely(n_pkts < pkt_burst_idx)) {
1746                                 /*
1747                                  * log error packets number here and do actual
1748                                  * error processing when applications poll
1749                                  * completion
1750                                  */
1751                                 pkt_err = pkt_burst_idx - n_pkts;
1752                                 pkt_burst_idx = 0;
1753                                 break;
1754                         }
1755
1756                         pkt_burst_idx = 0;
1757                 }
1758         }
1759
1760         if (pkt_burst_idx) {
1761                 n_xfer = vq->async_ops.transfer_data(dev->vid, queue_id, tdes, 0, pkt_burst_idx);
1762                 if (n_xfer >= 0) {
1763                         n_pkts = n_xfer;
1764                 } else {
1765                         VHOST_LOG_DATA(ERR, "(%d) %s: failed to transfer data for queue id %d.\n",
1766                                 dev->vid, __func__, queue_id);
1767                         n_pkts = 0;
1768                 }
1769
1770                 vq->async_pkts_inflight_n += n_pkts;
1771
1772                 if (unlikely(n_pkts < pkt_burst_idx))
1773                         pkt_err = pkt_burst_idx - n_pkts;
1774         }
1775
1776         do_data_copy_enqueue(dev, vq);
1777
1778         if (unlikely(pkt_err)) {
1779                 uint16_t num_descs = 0;
1780
1781                 num_async_pkts -= pkt_err;
1782                 /* calculate the sum of descriptors of DMA-error packets. */
1783                 while (pkt_err-- > 0) {
1784                         num_descs += pkts_info[slot_idx & (vq->size - 1)].descs;
1785                         slot_idx--;
1786                 }
1787                 vq->async_desc_idx_split -= num_descs;
1788                 /* recover shadow used ring and available ring */
1789                 vq->shadow_used_idx -= (vq->last_avail_idx -
1790                                 async_pkts_log[num_async_pkts].last_avail_idx -
1791                                 num_descs);
1792                 vq->last_avail_idx =
1793                         async_pkts_log[num_async_pkts].last_avail_idx;
1794                 pkt_idx = async_pkts_log[num_async_pkts].pkt_idx;
1795                 num_done_pkts = pkt_idx - num_async_pkts;
1796         }
1797
1798         vq->async_pkts_idx += num_async_pkts;
1799         *comp_count = num_done_pkts;
1800
1801         if (likely(vq->shadow_used_idx)) {
1802                 flush_shadow_used_ring_split(dev, vq);
1803                 vhost_vring_call_split(dev, vq);
1804         }
1805
1806         return pkt_idx;
1807 }
1808
1809 static __rte_always_inline void
1810 vhost_update_used_packed(struct vhost_virtqueue *vq,
1811                         struct vring_used_elem_packed *shadow_ring,
1812                         uint16_t count)
1813 {
1814         int i;
1815         uint16_t used_idx = vq->last_used_idx;
1816         uint16_t head_idx = vq->last_used_idx;
1817         uint16_t head_flags = 0;
1818
1819         if (count == 0)
1820                 return;
1821
1822         /* Split loop in two to save memory barriers */
1823         for (i = 0; i < count; i++) {
1824                 vq->desc_packed[used_idx].id = shadow_ring[i].id;
1825                 vq->desc_packed[used_idx].len = shadow_ring[i].len;
1826
1827                 used_idx += shadow_ring[i].count;
1828                 if (used_idx >= vq->size)
1829                         used_idx -= vq->size;
1830         }
1831
1832         /* The ordering for storing desc flags needs to be enforced. */
1833         rte_atomic_thread_fence(__ATOMIC_RELEASE);
1834
1835         for (i = 0; i < count; i++) {
1836                 uint16_t flags;
1837
1838                 if (vq->shadow_used_packed[i].len)
1839                         flags = VRING_DESC_F_WRITE;
1840                 else
1841                         flags = 0;
1842
1843                 if (vq->used_wrap_counter) {
1844                         flags |= VRING_DESC_F_USED;
1845                         flags |= VRING_DESC_F_AVAIL;
1846                 } else {
1847                         flags &= ~VRING_DESC_F_USED;
1848                         flags &= ~VRING_DESC_F_AVAIL;
1849                 }
1850
1851                 if (i > 0) {
1852                         vq->desc_packed[vq->last_used_idx].flags = flags;
1853                 } else {
1854                         head_idx = vq->last_used_idx;
1855                         head_flags = flags;
1856                 }
1857
1858                 vq_inc_last_used_packed(vq, shadow_ring[i].count);
1859         }
1860
1861         vq->desc_packed[head_idx].flags = head_flags;
1862 }
1863
1864 static __rte_always_inline int
1865 vhost_enqueue_async_single_packed(struct virtio_net *dev,
1866                             struct vhost_virtqueue *vq,
1867                             struct rte_mbuf *pkt,
1868                             struct buf_vector *buf_vec,
1869                             uint16_t *nr_descs,
1870                             uint16_t *nr_buffers,
1871                             struct vring_packed_desc *async_descs,
1872                             struct iovec *src_iovec, struct iovec *dst_iovec,
1873                             struct rte_vhost_iov_iter *src_it,
1874                             struct rte_vhost_iov_iter *dst_it)
1875 {
1876         uint16_t nr_vec = 0;
1877         uint16_t avail_idx = vq->last_avail_idx;
1878         uint16_t max_tries, tries = 0;
1879         uint16_t buf_id = 0;
1880         uint32_t len = 0;
1881         uint16_t desc_count = 0;
1882         uint32_t size = pkt->pkt_len + sizeof(struct virtio_net_hdr_mrg_rxbuf);
1883         uint32_t buffer_len[vq->size];
1884         uint16_t buffer_buf_id[vq->size];
1885         uint16_t buffer_desc_count[vq->size];
1886
1887         if (rxvq_is_mergeable(dev))
1888                 max_tries = vq->size - 1;
1889         else
1890                 max_tries = 1;
1891
1892         while (size > 0) {
1893                 /*
1894                  * if we tried all available ring items, and still
1895                  * can't get enough buf, it means something abnormal
1896                  * happened.
1897                  */
1898                 if (unlikely(++tries > max_tries))
1899                         return -1;
1900
1901                 if (unlikely(fill_vec_buf_packed(dev, vq, avail_idx, &desc_count, buf_vec, &nr_vec,
1902                                                 &buf_id, &len, VHOST_ACCESS_RW) < 0))
1903                         return -1;
1904
1905                 len = RTE_MIN(len, size);
1906                 size -= len;
1907
1908                 buffer_len[*nr_buffers] = len;
1909                 buffer_buf_id[*nr_buffers] = buf_id;
1910                 buffer_desc_count[*nr_buffers] = desc_count;
1911                 *nr_buffers += 1;
1912
1913                 *nr_descs += desc_count;
1914                 avail_idx += desc_count;
1915                 if (avail_idx >= vq->size)
1916                         avail_idx -= vq->size;
1917         }
1918
1919         if (async_mbuf_to_desc(dev, vq, pkt, buf_vec, nr_vec, *nr_buffers, src_iovec, dst_iovec,
1920                         src_it, dst_it) < 0)
1921                 return -1;
1922         /* store descriptors for DMA */
1923         if (avail_idx >= *nr_descs) {
1924                 rte_memcpy(async_descs, &vq->desc_packed[vq->last_avail_idx],
1925                         *nr_descs * sizeof(struct vring_packed_desc));
1926         } else {
1927                 uint16_t nr_copy = vq->size - vq->last_avail_idx;
1928
1929                 rte_memcpy(async_descs, &vq->desc_packed[vq->last_avail_idx],
1930                         nr_copy * sizeof(struct vring_packed_desc));
1931                 rte_memcpy(async_descs + nr_copy, vq->desc_packed,
1932                         (*nr_descs - nr_copy) * sizeof(struct vring_packed_desc));
1933         }
1934
1935         vhost_shadow_enqueue_packed(vq, buffer_len, buffer_buf_id, buffer_desc_count, *nr_buffers);
1936
1937         return 0;
1938 }
1939
1940 static __rte_always_inline int16_t
1941 virtio_dev_rx_async_single_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
1942                             struct rte_mbuf *pkt, uint16_t *nr_descs, uint16_t *nr_buffers,
1943                             struct vring_packed_desc *async_descs,
1944                             struct iovec *src_iovec, struct iovec *dst_iovec,
1945                             struct rte_vhost_iov_iter *src_it, struct rte_vhost_iov_iter *dst_it)
1946 {
1947         struct buf_vector buf_vec[BUF_VECTOR_MAX];
1948
1949         if (unlikely(vhost_enqueue_async_single_packed(dev, vq, pkt, buf_vec, nr_descs, nr_buffers,
1950                                                  async_descs, src_iovec, dst_iovec,
1951                                                  src_it, dst_it) < 0)) {
1952                 VHOST_LOG_DATA(DEBUG, "(%d) failed to get enough desc from vring\n", dev->vid);
1953                 return -1;
1954         }
1955
1956         VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
1957                         dev->vid, vq->last_avail_idx, vq->last_avail_idx + *nr_descs);
1958
1959         return 0;
1960 }
1961
1962 static __rte_always_inline void
1963 dma_error_handler_packed(struct vhost_virtqueue *vq, struct vring_packed_desc *async_descs,
1964                         uint16_t async_descs_idx, uint16_t slot_idx, uint32_t nr_err,
1965                         uint32_t *pkt_idx, uint32_t *num_async_pkts, uint32_t *num_done_pkts)
1966 {
1967         uint16_t descs_err = 0;
1968         uint16_t buffers_err = 0;
1969         struct async_inflight_info *pkts_info = vq->async_pkts_info;
1970
1971         *num_async_pkts -= nr_err;
1972         *pkt_idx -= nr_err;
1973         /* calculate the sum of buffers and descs of DMA-error packets. */
1974         while (nr_err-- > 0) {
1975                 descs_err += pkts_info[slot_idx % vq->size].descs;
1976                 buffers_err += pkts_info[slot_idx % vq->size].nr_buffers;
1977                 slot_idx--;
1978         }
1979
1980         vq->async_buffer_idx_packed -= buffers_err;
1981
1982         if (vq->last_avail_idx >= descs_err) {
1983                 vq->last_avail_idx -= descs_err;
1984
1985                 rte_memcpy(&vq->desc_packed[vq->last_avail_idx],
1986                         &async_descs[async_descs_idx - descs_err],
1987                         descs_err * sizeof(struct vring_packed_desc));
1988         } else {
1989                 uint16_t nr_copy;
1990
1991                 vq->last_avail_idx = vq->last_avail_idx + vq->size - descs_err;
1992                 nr_copy = vq->size - vq->last_avail_idx;
1993                 rte_memcpy(&vq->desc_packed[vq->last_avail_idx],
1994                         &async_descs[async_descs_idx - descs_err],
1995                         nr_copy * sizeof(struct vring_packed_desc));
1996                 descs_err -= nr_copy;
1997                 rte_memcpy(&vq->desc_packed[0], &async_descs[async_descs_idx - descs_err],
1998                         descs_err * sizeof(struct vring_packed_desc));
1999                 vq->avail_wrap_counter ^= 1;
2000         }
2001
2002         *num_done_pkts = *pkt_idx - *num_async_pkts;
2003 }
2004
2005 static __rte_noinline uint32_t
2006 virtio_dev_rx_async_submit_packed(struct virtio_net *dev,
2007         struct vhost_virtqueue *vq, uint16_t queue_id,
2008         struct rte_mbuf **pkts, uint32_t count,
2009         struct rte_mbuf **comp_pkts, uint32_t *comp_count)
2010 {
2011         uint32_t pkt_idx = 0, pkt_burst_idx = 0;
2012         uint32_t remained = count;
2013         uint16_t async_descs_idx = 0;
2014         uint16_t num_buffers;
2015         uint16_t num_descs;
2016         int32_t n_xfer;
2017
2018         struct rte_vhost_iov_iter *it_pool = vq->it_pool;
2019         struct iovec *vec_pool = vq->vec_pool;
2020         struct rte_vhost_async_desc tdes[MAX_PKT_BURST];
2021         struct iovec *src_iovec = vec_pool;
2022         struct iovec *dst_iovec = vec_pool + (VHOST_MAX_ASYNC_VEC >> 1);
2023         uint16_t slot_idx = 0;
2024         uint16_t segs_await = 0;
2025         uint16_t iovec_idx = 0, it_idx = 0;
2026         struct async_inflight_info *pkts_info = vq->async_pkts_info;
2027         uint32_t n_pkts = 0, pkt_err = 0;
2028         uint32_t num_async_pkts = 0, num_done_pkts = 0;
2029         struct vring_packed_desc async_descs[vq->size];
2030
2031         do {
2032                 rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
2033                 if (remained >= PACKED_BATCH_SIZE) {
2034                         if (!virtio_dev_rx_async_batch_packed(dev, vq,
2035                                 &pkts[pkt_idx], comp_pkts, &num_done_pkts)) {
2036                                 pkt_idx += PACKED_BATCH_SIZE;
2037                                 remained -= PACKED_BATCH_SIZE;
2038                                 continue;
2039                         }
2040                 }
2041
2042                 num_buffers = 0;
2043                 num_descs = 0;
2044                 if (unlikely(virtio_dev_rx_async_single_packed(dev, vq, pkts[pkt_idx],
2045                                                 &num_descs, &num_buffers,
2046                                                 &async_descs[async_descs_idx],
2047                                                 &src_iovec[iovec_idx], &dst_iovec[iovec_idx],
2048                                                 &it_pool[it_idx], &it_pool[it_idx + 1]) < 0))
2049                         break;
2050
2051                 VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
2052                         dev->vid, vq->last_avail_idx,
2053                         vq->last_avail_idx + num_descs);
2054
2055                 slot_idx = (vq->async_pkts_idx + num_async_pkts) % vq->size;
2056                 if (it_pool[it_idx].count) {
2057                         uint16_t from;
2058
2059                         async_descs_idx += num_descs;
2060                         async_fill_desc(&tdes[pkt_burst_idx++],
2061                                 &it_pool[it_idx], &it_pool[it_idx + 1]);
2062                         pkts_info[slot_idx].descs = num_descs;
2063                         pkts_info[slot_idx].nr_buffers = num_buffers;
2064                         pkts_info[slot_idx].mbuf = pkts[pkt_idx];
2065                         num_async_pkts++;
2066                         iovec_idx += it_pool[it_idx].nr_segs;
2067                         it_idx += 2;
2068
2069                         segs_await += it_pool[it_idx].nr_segs;
2070
2071                         /**
2072                          * recover shadow used ring and keep DMA-occupied
2073                          * descriptors.
2074                          */
2075                         from = vq->shadow_used_idx - num_buffers;
2076                         store_dma_desc_info_packed(vq->shadow_used_packed,
2077                                         vq->async_buffers_packed, vq->size, from,
2078                                         vq->async_buffer_idx_packed, num_buffers);
2079
2080                         vq->async_buffer_idx_packed += num_buffers;
2081                         if (vq->async_buffer_idx_packed >= vq->size)
2082                                 vq->async_buffer_idx_packed -= vq->size;
2083                         vq->shadow_used_idx -= num_buffers;
2084                 } else {
2085                         comp_pkts[num_done_pkts++] = pkts[pkt_idx];
2086                 }
2087
2088                 pkt_idx++;
2089                 remained--;
2090                 vq_inc_last_avail_packed(vq, num_descs);
2091
2092                 /*
2093                  * conditions to trigger async device transfer:
2094                  * - buffered packet number reaches transfer threshold
2095                  * - unused async iov number is less than max vhost vector
2096                  */
2097                 if (unlikely(pkt_burst_idx >= VHOST_ASYNC_BATCH_THRESHOLD ||
2098                         ((VHOST_MAX_ASYNC_VEC >> 1) - segs_await < BUF_VECTOR_MAX))) {
2099                         n_xfer = vq->async_ops.transfer_data(dev->vid,
2100                                         queue_id, tdes, 0, pkt_burst_idx);
2101                         if (n_xfer >= 0) {
2102                                 n_pkts = n_xfer;
2103                         } else {
2104                                 VHOST_LOG_DATA(ERR,
2105                                         "(%d) %s: failed to transfer data for queue id %d.\n",
2106                                         dev->vid, __func__, queue_id);
2107                                 n_pkts = 0;
2108                         }
2109
2110                         iovec_idx = 0;
2111                         it_idx = 0;
2112                         segs_await = 0;
2113                         vq->async_pkts_inflight_n += n_pkts;
2114
2115                         if (unlikely(n_pkts < pkt_burst_idx)) {
2116                                 /*
2117                                  * log error packets number here and do actual
2118                                  * error processing when applications poll
2119                                  * completion
2120                                  */
2121                                 pkt_err = pkt_burst_idx - n_pkts;
2122                                 pkt_burst_idx = 0;
2123                                 break;
2124                         }
2125
2126                         pkt_burst_idx = 0;
2127                 }
2128         } while (pkt_idx < count);
2129
2130         if (pkt_burst_idx) {
2131                 n_xfer = vq->async_ops.transfer_data(dev->vid, queue_id, tdes, 0, pkt_burst_idx);
2132                 if (n_xfer >= 0) {
2133                         n_pkts = n_xfer;
2134                 } else {
2135                         VHOST_LOG_DATA(ERR, "(%d) %s: failed to transfer data for queue id %d.\n",
2136                                 dev->vid, __func__, queue_id);
2137                         n_pkts = 0;
2138                 }
2139
2140                 vq->async_pkts_inflight_n += n_pkts;
2141
2142                 if (unlikely(n_pkts < pkt_burst_idx))
2143                         pkt_err = pkt_burst_idx - n_pkts;
2144         }
2145
2146         do_data_copy_enqueue(dev, vq);
2147
2148         if (unlikely(pkt_err))
2149                 dma_error_handler_packed(vq, async_descs, async_descs_idx, slot_idx, pkt_err,
2150                                         &pkt_idx, &num_async_pkts, &num_done_pkts);
2151         vq->async_pkts_idx += num_async_pkts;
2152         if (vq->async_pkts_idx >= vq->size)
2153                 vq->async_pkts_idx -= vq->size;
2154         *comp_count = num_done_pkts;
2155
2156         if (likely(vq->shadow_used_idx)) {
2157                 vhost_flush_enqueue_shadow_packed(dev, vq);
2158                 vhost_vring_call_packed(dev, vq);
2159         }
2160
2161         return pkt_idx;
2162 }
2163
2164 static __rte_always_inline void
2165 write_back_completed_descs_split(struct vhost_virtqueue *vq, uint16_t n_descs)
2166 {
2167         uint16_t nr_left = n_descs;
2168         uint16_t nr_copy;
2169         uint16_t to, from;
2170
2171         do {
2172                 from = vq->last_async_desc_idx_split & (vq->size - 1);
2173                 nr_copy = nr_left + from <= vq->size ? nr_left : vq->size - from;
2174                 to = vq->last_used_idx & (vq->size - 1);
2175
2176                 if (to + nr_copy <= vq->size) {
2177                         rte_memcpy(&vq->used->ring[to], &vq->async_descs_split[from],
2178                                         nr_copy * sizeof(struct vring_used_elem));
2179                 } else {
2180                         uint16_t size = vq->size - to;
2181
2182                         rte_memcpy(&vq->used->ring[to], &vq->async_descs_split[from],
2183                                         size * sizeof(struct vring_used_elem));
2184                         rte_memcpy(&vq->used->ring[0], &vq->async_descs_split[from + size],
2185                                         (nr_copy - size) * sizeof(struct vring_used_elem));
2186                 }
2187
2188                 vq->last_async_desc_idx_split += nr_copy;
2189                 vq->last_used_idx += nr_copy;
2190                 nr_left -= nr_copy;
2191         } while (nr_left > 0);
2192 }
2193
2194 static __rte_always_inline void
2195 write_back_completed_descs_packed(struct vhost_virtqueue *vq,
2196                                 uint16_t n_buffers)
2197 {
2198         uint16_t nr_left = n_buffers;
2199         uint16_t from, to;
2200
2201         do {
2202                 from = vq->last_async_buffer_idx_packed;
2203                 to = (from + nr_left) % vq->size;
2204                 if (to > from) {
2205                         vhost_update_used_packed(vq, vq->async_buffers_packed + from, to - from);
2206                         vq->last_async_buffer_idx_packed += nr_left;
2207                         nr_left = 0;
2208                 } else {
2209                         vhost_update_used_packed(vq, vq->async_buffers_packed + from,
2210                                 vq->size - from);
2211                         vq->last_async_buffer_idx_packed = 0;
2212                         nr_left -= vq->size - from;
2213                 }
2214         } while (nr_left > 0);
2215 }
2216
2217 static __rte_always_inline uint16_t
2218 vhost_poll_enqueue_completed(struct virtio_net *dev, uint16_t queue_id,
2219                 struct rte_mbuf **pkts, uint16_t count)
2220 {
2221         struct vhost_virtqueue *vq;
2222         uint16_t n_pkts_cpl = 0, n_pkts_put = 0, n_descs = 0, n_buffers = 0;
2223         uint16_t start_idx, pkts_idx, vq_size;
2224         struct async_inflight_info *pkts_info;
2225         uint16_t from, i;
2226         int32_t n_cpl;
2227
2228         vq = dev->virtqueue[queue_id];
2229
2230         pkts_idx = vq->async_pkts_idx % vq->size;
2231         pkts_info = vq->async_pkts_info;
2232         vq_size = vq->size;
2233         start_idx = virtio_dev_rx_async_get_info_idx(pkts_idx,
2234                 vq_size, vq->async_pkts_inflight_n);
2235
2236         if (count > vq->async_last_pkts_n) {
2237                 n_cpl = vq->async_ops.check_completed_copies(dev->vid,
2238                         queue_id, 0, count - vq->async_last_pkts_n);
2239                 if (n_cpl >= 0) {
2240                         n_pkts_cpl = n_cpl;
2241                 } else {
2242                         VHOST_LOG_DATA(ERR,
2243                                 "(%d) %s: failed to check completed copies for queue id %d.\n",
2244                                 dev->vid, __func__, queue_id);
2245                         n_pkts_cpl = 0;
2246                 }
2247         }
2248         n_pkts_cpl += vq->async_last_pkts_n;
2249
2250         n_pkts_put = RTE_MIN(count, n_pkts_cpl);
2251         if (unlikely(n_pkts_put == 0)) {
2252                 vq->async_last_pkts_n = n_pkts_cpl;
2253                 return 0;
2254         }
2255
2256         if (vq_is_packed(dev)) {
2257                 for (i = 0; i < n_pkts_put; i++) {
2258                         from = (start_idx + i) % vq_size;
2259                         n_buffers += pkts_info[from].nr_buffers;
2260                         pkts[i] = pkts_info[from].mbuf;
2261                 }
2262         } else {
2263                 for (i = 0; i < n_pkts_put; i++) {
2264                         from = (start_idx + i) & (vq_size - 1);
2265                         n_descs += pkts_info[from].descs;
2266                         pkts[i] = pkts_info[from].mbuf;
2267                 }
2268         }
2269
2270         vq->async_last_pkts_n = n_pkts_cpl - n_pkts_put;
2271         vq->async_pkts_inflight_n -= n_pkts_put;
2272
2273         if (likely(vq->enabled && vq->access_ok)) {
2274                 if (vq_is_packed(dev)) {
2275                         write_back_completed_descs_packed(vq, n_buffers);
2276
2277                         vhost_vring_call_packed(dev, vq);
2278                 } else {
2279                         write_back_completed_descs_split(vq, n_descs);
2280
2281                         __atomic_add_fetch(&vq->used->idx, n_descs,
2282                                         __ATOMIC_RELEASE);
2283                         vhost_vring_call_split(dev, vq);
2284                 }
2285         } else {
2286                 if (vq_is_packed(dev)) {
2287                         vq->last_async_buffer_idx_packed += n_buffers;
2288                         if (vq->last_async_buffer_idx_packed >= vq->size)
2289                                 vq->last_async_buffer_idx_packed -= vq->size;
2290                 } else {
2291                         vq->last_async_desc_idx_split += n_descs;
2292                 }
2293         }
2294
2295         return n_pkts_put;
2296 }
2297
2298 uint16_t
2299 rte_vhost_poll_enqueue_completed(int vid, uint16_t queue_id,
2300                 struct rte_mbuf **pkts, uint16_t count)
2301 {
2302         struct virtio_net *dev = get_device(vid);
2303         struct vhost_virtqueue *vq;
2304         uint16_t n_pkts_cpl = 0;
2305
2306         if (!dev)
2307                 return 0;
2308
2309         VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
2310         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
2311                 VHOST_LOG_DATA(ERR, "(%d) %s: invalid virtqueue idx %d.\n",
2312                         dev->vid, __func__, queue_id);
2313                 return 0;
2314         }
2315
2316         vq = dev->virtqueue[queue_id];
2317
2318         if (unlikely(!vq->async_registered)) {
2319                 VHOST_LOG_DATA(ERR, "(%d) %s: async not registered for queue id %d.\n",
2320                         dev->vid, __func__, queue_id);
2321                 return 0;
2322         }
2323
2324         rte_spinlock_lock(&vq->access_lock);
2325
2326         n_pkts_cpl = vhost_poll_enqueue_completed(dev, queue_id, pkts, count);
2327
2328         rte_spinlock_unlock(&vq->access_lock);
2329
2330         return n_pkts_cpl;
2331 }
2332
2333 uint16_t
2334 rte_vhost_clear_queue_thread_unsafe(int vid, uint16_t queue_id,
2335                 struct rte_mbuf **pkts, uint16_t count)
2336 {
2337         struct virtio_net *dev = get_device(vid);
2338         struct vhost_virtqueue *vq;
2339         uint16_t n_pkts_cpl = 0;
2340
2341         if (!dev)
2342                 return 0;
2343
2344         VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
2345         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
2346                 VHOST_LOG_DATA(ERR, "(%d) %s: invalid virtqueue idx %d.\n",
2347                         dev->vid, __func__, queue_id);
2348                 return 0;
2349         }
2350
2351         vq = dev->virtqueue[queue_id];
2352
2353         if (unlikely(!vq->async_registered)) {
2354                 VHOST_LOG_DATA(ERR, "(%d) %s: async not registered for queue id %d.\n",
2355                         dev->vid, __func__, queue_id);
2356                 return 0;
2357         }
2358
2359         n_pkts_cpl = vhost_poll_enqueue_completed(dev, queue_id, pkts, count);
2360
2361         return n_pkts_cpl;
2362 }
2363
2364 static __rte_always_inline uint32_t
2365 virtio_dev_rx_async_submit(struct virtio_net *dev, uint16_t queue_id,
2366         struct rte_mbuf **pkts, uint32_t count,
2367         struct rte_mbuf **comp_pkts, uint32_t *comp_count)
2368 {
2369         struct vhost_virtqueue *vq;
2370         uint32_t nb_tx = 0;
2371
2372         VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
2373         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
2374                 VHOST_LOG_DATA(ERR, "(%d) %s: invalid virtqueue idx %d.\n",
2375                         dev->vid, __func__, queue_id);
2376                 return 0;
2377         }
2378
2379         vq = dev->virtqueue[queue_id];
2380
2381         rte_spinlock_lock(&vq->access_lock);
2382
2383         if (unlikely(!vq->enabled || !vq->async_registered))
2384                 goto out_access_unlock;
2385
2386         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
2387                 vhost_user_iotlb_rd_lock(vq);
2388
2389         if (unlikely(!vq->access_ok))
2390                 if (unlikely(vring_translate(dev, vq) < 0))
2391                         goto out;
2392
2393         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
2394         if (count == 0)
2395                 goto out;
2396
2397         if (vq_is_packed(dev))
2398                 nb_tx = virtio_dev_rx_async_submit_packed(dev,
2399                                 vq, queue_id, pkts, count, comp_pkts,
2400                                 comp_count);
2401         else
2402                 nb_tx = virtio_dev_rx_async_submit_split(dev,
2403                                 vq, queue_id, pkts, count, comp_pkts,
2404                                 comp_count);
2405
2406 out:
2407         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
2408                 vhost_user_iotlb_rd_unlock(vq);
2409
2410 out_access_unlock:
2411         rte_spinlock_unlock(&vq->access_lock);
2412
2413         return nb_tx;
2414 }
2415
2416 uint16_t
2417 rte_vhost_submit_enqueue_burst(int vid, uint16_t queue_id,
2418                 struct rte_mbuf **pkts, uint16_t count,
2419                 struct rte_mbuf **comp_pkts, uint32_t *comp_count)
2420 {
2421         struct virtio_net *dev = get_device(vid);
2422
2423         *comp_count = 0;
2424         if (!dev)
2425                 return 0;
2426
2427         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
2428                 VHOST_LOG_DATA(ERR,
2429                         "(%d) %s: built-in vhost net backend is disabled.\n",
2430                         dev->vid, __func__);
2431                 return 0;
2432         }
2433
2434         return virtio_dev_rx_async_submit(dev, queue_id, pkts, count, comp_pkts,
2435                         comp_count);
2436 }
2437
2438 static inline bool
2439 virtio_net_with_host_offload(struct virtio_net *dev)
2440 {
2441         if (dev->features &
2442                         ((1ULL << VIRTIO_NET_F_CSUM) |
2443                          (1ULL << VIRTIO_NET_F_HOST_ECN) |
2444                          (1ULL << VIRTIO_NET_F_HOST_TSO4) |
2445                          (1ULL << VIRTIO_NET_F_HOST_TSO6) |
2446                          (1ULL << VIRTIO_NET_F_HOST_UFO)))
2447                 return true;
2448
2449         return false;
2450 }
2451
2452 static int
2453 parse_headers(struct rte_mbuf *m, uint8_t *l4_proto)
2454 {
2455         struct rte_ipv4_hdr *ipv4_hdr;
2456         struct rte_ipv6_hdr *ipv6_hdr;
2457         struct rte_ether_hdr *eth_hdr;
2458         uint16_t ethertype;
2459         uint16_t data_len = rte_pktmbuf_data_len(m);
2460
2461         if (data_len < sizeof(struct rte_ether_hdr))
2462                 return -EINVAL;
2463
2464         eth_hdr = rte_pktmbuf_mtod(m, struct rte_ether_hdr *);
2465
2466         m->l2_len = sizeof(struct rte_ether_hdr);
2467         ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
2468
2469         if (ethertype == RTE_ETHER_TYPE_VLAN) {
2470                 if (data_len < sizeof(struct rte_ether_hdr) +
2471                                 sizeof(struct rte_vlan_hdr))
2472                         goto error;
2473
2474                 struct rte_vlan_hdr *vlan_hdr =
2475                         (struct rte_vlan_hdr *)(eth_hdr + 1);
2476
2477                 m->l2_len += sizeof(struct rte_vlan_hdr);
2478                 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
2479         }
2480
2481         switch (ethertype) {
2482         case RTE_ETHER_TYPE_IPV4:
2483                 if (data_len < m->l2_len + sizeof(struct rte_ipv4_hdr))
2484                         goto error;
2485                 ipv4_hdr = rte_pktmbuf_mtod_offset(m, struct rte_ipv4_hdr *,
2486                                 m->l2_len);
2487                 m->l3_len = rte_ipv4_hdr_len(ipv4_hdr);
2488                 if (data_len < m->l2_len + m->l3_len)
2489                         goto error;
2490                 m->ol_flags |= RTE_MBUF_F_TX_IPV4;
2491                 *l4_proto = ipv4_hdr->next_proto_id;
2492                 break;
2493         case RTE_ETHER_TYPE_IPV6:
2494                 if (data_len < m->l2_len + sizeof(struct rte_ipv6_hdr))
2495                         goto error;
2496                 ipv6_hdr = rte_pktmbuf_mtod_offset(m, struct rte_ipv6_hdr *,
2497                                 m->l2_len);
2498                 m->l3_len = sizeof(struct rte_ipv6_hdr);
2499                 m->ol_flags |= RTE_MBUF_F_TX_IPV6;
2500                 *l4_proto = ipv6_hdr->proto;
2501                 break;
2502         default:
2503                 /* a valid L3 header is needed for further L4 parsing */
2504                 goto error;
2505         }
2506
2507         /* both CSUM and GSO need a valid L4 header */
2508         switch (*l4_proto) {
2509         case IPPROTO_TCP:
2510                 if (data_len < m->l2_len + m->l3_len +
2511                                 sizeof(struct rte_tcp_hdr))
2512                         goto error;
2513                 break;
2514         case IPPROTO_UDP:
2515                 if (data_len < m->l2_len + m->l3_len +
2516                                 sizeof(struct rte_udp_hdr))
2517                         goto error;
2518                 break;
2519         case IPPROTO_SCTP:
2520                 if (data_len < m->l2_len + m->l3_len +
2521                                 sizeof(struct rte_sctp_hdr))
2522                         goto error;
2523                 break;
2524         default:
2525                 goto error;
2526         }
2527
2528         return 0;
2529
2530 error:
2531         m->l2_len = 0;
2532         m->l3_len = 0;
2533         m->ol_flags = 0;
2534         return -EINVAL;
2535 }
2536
2537 static __rte_always_inline void
2538 vhost_dequeue_offload_legacy(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
2539 {
2540         uint8_t l4_proto = 0;
2541         struct rte_tcp_hdr *tcp_hdr = NULL;
2542         uint16_t tcp_len;
2543         uint16_t data_len = rte_pktmbuf_data_len(m);
2544
2545         if (parse_headers(m, &l4_proto) < 0)
2546                 return;
2547
2548         if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
2549                 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
2550                         switch (hdr->csum_offset) {
2551                         case (offsetof(struct rte_tcp_hdr, cksum)):
2552                                 if (l4_proto != IPPROTO_TCP)
2553                                         goto error;
2554                                 m->ol_flags |= RTE_MBUF_F_TX_TCP_CKSUM;
2555                                 break;
2556                         case (offsetof(struct rte_udp_hdr, dgram_cksum)):
2557                                 if (l4_proto != IPPROTO_UDP)
2558                                         goto error;
2559                                 m->ol_flags |= RTE_MBUF_F_TX_UDP_CKSUM;
2560                                 break;
2561                         case (offsetof(struct rte_sctp_hdr, cksum)):
2562                                 if (l4_proto != IPPROTO_SCTP)
2563                                         goto error;
2564                                 m->ol_flags |= RTE_MBUF_F_TX_SCTP_CKSUM;
2565                                 break;
2566                         default:
2567                                 goto error;
2568                         }
2569                 } else {
2570                         goto error;
2571                 }
2572         }
2573
2574         if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
2575                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
2576                 case VIRTIO_NET_HDR_GSO_TCPV4:
2577                 case VIRTIO_NET_HDR_GSO_TCPV6:
2578                         if (l4_proto != IPPROTO_TCP)
2579                                 goto error;
2580                         tcp_hdr = rte_pktmbuf_mtod_offset(m,
2581                                         struct rte_tcp_hdr *,
2582                                         m->l2_len + m->l3_len);
2583                         tcp_len = (tcp_hdr->data_off & 0xf0) >> 2;
2584                         if (data_len < m->l2_len + m->l3_len + tcp_len)
2585                                 goto error;
2586                         m->ol_flags |= RTE_MBUF_F_TX_TCP_SEG;
2587                         m->tso_segsz = hdr->gso_size;
2588                         m->l4_len = tcp_len;
2589                         break;
2590                 case VIRTIO_NET_HDR_GSO_UDP:
2591                         if (l4_proto != IPPROTO_UDP)
2592                                 goto error;
2593                         m->ol_flags |= RTE_MBUF_F_TX_UDP_SEG;
2594                         m->tso_segsz = hdr->gso_size;
2595                         m->l4_len = sizeof(struct rte_udp_hdr);
2596                         break;
2597                 default:
2598                         VHOST_LOG_DATA(WARNING,
2599                                 "unsupported gso type %u.\n", hdr->gso_type);
2600                         goto error;
2601                 }
2602         }
2603         return;
2604
2605 error:
2606         m->l2_len = 0;
2607         m->l3_len = 0;
2608         m->ol_flags = 0;
2609 }
2610
2611 static __rte_always_inline void
2612 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m,
2613         bool legacy_ol_flags)
2614 {
2615         struct rte_net_hdr_lens hdr_lens;
2616         int l4_supported = 0;
2617         uint32_t ptype;
2618
2619         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
2620                 return;
2621
2622         if (legacy_ol_flags) {
2623                 vhost_dequeue_offload_legacy(hdr, m);
2624                 return;
2625         }
2626
2627         m->ol_flags |= RTE_MBUF_F_RX_IP_CKSUM_UNKNOWN;
2628
2629         ptype = rte_net_get_ptype(m, &hdr_lens, RTE_PTYPE_ALL_MASK);
2630         m->packet_type = ptype;
2631         if ((ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_TCP ||
2632             (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_UDP ||
2633             (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_SCTP)
2634                 l4_supported = 1;
2635
2636         /* According to Virtio 1.1 spec, the device only needs to look at
2637          * VIRTIO_NET_HDR_F_NEEDS_CSUM in the packet transmission path.
2638          * This differs from the processing incoming packets path where the
2639          * driver could rely on VIRTIO_NET_HDR_F_DATA_VALID flag set by the
2640          * device.
2641          *
2642          * 5.1.6.2.1 Driver Requirements: Packet Transmission
2643          * The driver MUST NOT set the VIRTIO_NET_HDR_F_DATA_VALID and
2644          * VIRTIO_NET_HDR_F_RSC_INFO bits in flags.
2645          *
2646          * 5.1.6.2.2 Device Requirements: Packet Transmission
2647          * The device MUST ignore flag bits that it does not recognize.
2648          */
2649         if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
2650                 uint32_t hdrlen;
2651
2652                 hdrlen = hdr_lens.l2_len + hdr_lens.l3_len + hdr_lens.l4_len;
2653                 if (hdr->csum_start <= hdrlen && l4_supported != 0) {
2654                         m->ol_flags |= RTE_MBUF_F_RX_L4_CKSUM_NONE;
2655                 } else {
2656                         /* Unknown proto or tunnel, do sw cksum. We can assume
2657                          * the cksum field is in the first segment since the
2658                          * buffers we provided to the host are large enough.
2659                          * In case of SCTP, this will be wrong since it's a CRC
2660                          * but there's nothing we can do.
2661                          */
2662                         uint16_t csum = 0, off;
2663
2664                         if (rte_raw_cksum_mbuf(m, hdr->csum_start,
2665                                         rte_pktmbuf_pkt_len(m) - hdr->csum_start, &csum) < 0)
2666                                 return;
2667                         if (likely(csum != 0xffff))
2668                                 csum = ~csum;
2669                         off = hdr->csum_offset + hdr->csum_start;
2670                         if (rte_pktmbuf_data_len(m) >= off + 1)
2671                                 *rte_pktmbuf_mtod_offset(m, uint16_t *, off) = csum;
2672                 }
2673         }
2674
2675         if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
2676                 if (hdr->gso_size == 0)
2677                         return;
2678
2679                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
2680                 case VIRTIO_NET_HDR_GSO_TCPV4:
2681                 case VIRTIO_NET_HDR_GSO_TCPV6:
2682                         if ((ptype & RTE_PTYPE_L4_MASK) != RTE_PTYPE_L4_TCP)
2683                                 break;
2684                         m->ol_flags |= RTE_MBUF_F_RX_LRO | RTE_MBUF_F_RX_L4_CKSUM_NONE;
2685                         m->tso_segsz = hdr->gso_size;
2686                         break;
2687                 case VIRTIO_NET_HDR_GSO_UDP:
2688                         if ((ptype & RTE_PTYPE_L4_MASK) != RTE_PTYPE_L4_UDP)
2689                                 break;
2690                         m->ol_flags |= RTE_MBUF_F_RX_LRO | RTE_MBUF_F_RX_L4_CKSUM_NONE;
2691                         m->tso_segsz = hdr->gso_size;
2692                         break;
2693                 default:
2694                         break;
2695                 }
2696         }
2697 }
2698
2699 static __rte_noinline void
2700 copy_vnet_hdr_from_desc(struct virtio_net_hdr *hdr,
2701                 struct buf_vector *buf_vec)
2702 {
2703         uint64_t len;
2704         uint64_t remain = sizeof(struct virtio_net_hdr);
2705         uint64_t src;
2706         uint64_t dst = (uint64_t)(uintptr_t)hdr;
2707
2708         while (remain) {
2709                 len = RTE_MIN(remain, buf_vec->buf_len);
2710                 src = buf_vec->buf_addr;
2711                 rte_memcpy((void *)(uintptr_t)dst,
2712                                 (void *)(uintptr_t)src, len);
2713
2714                 remain -= len;
2715                 dst += len;
2716                 buf_vec++;
2717         }
2718 }
2719
2720 static __rte_always_inline int
2721 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
2722                   struct buf_vector *buf_vec, uint16_t nr_vec,
2723                   struct rte_mbuf *m, struct rte_mempool *mbuf_pool,
2724                   bool legacy_ol_flags)
2725 {
2726         uint32_t buf_avail, buf_offset;
2727         uint64_t buf_addr, buf_len;
2728         uint32_t mbuf_avail, mbuf_offset;
2729         uint32_t cpy_len;
2730         struct rte_mbuf *cur = m, *prev = m;
2731         struct virtio_net_hdr tmp_hdr;
2732         struct virtio_net_hdr *hdr = NULL;
2733         /* A counter to avoid desc dead loop chain */
2734         uint16_t vec_idx = 0;
2735         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
2736         int error = 0;
2737
2738         buf_addr = buf_vec[vec_idx].buf_addr;
2739         buf_len = buf_vec[vec_idx].buf_len;
2740
2741         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
2742                 error = -1;
2743                 goto out;
2744         }
2745
2746         if (virtio_net_with_host_offload(dev)) {
2747                 if (unlikely(buf_len < sizeof(struct virtio_net_hdr))) {
2748                         /*
2749                          * No luck, the virtio-net header doesn't fit
2750                          * in a contiguous virtual area.
2751                          */
2752                         copy_vnet_hdr_from_desc(&tmp_hdr, buf_vec);
2753                         hdr = &tmp_hdr;
2754                 } else {
2755                         hdr = (struct virtio_net_hdr *)((uintptr_t)buf_addr);
2756                 }
2757         }
2758
2759         /*
2760          * A virtio driver normally uses at least 2 desc buffers
2761          * for Tx: the first for storing the header, and others
2762          * for storing the data.
2763          */
2764         if (unlikely(buf_len < dev->vhost_hlen)) {
2765                 buf_offset = dev->vhost_hlen - buf_len;
2766                 vec_idx++;
2767                 buf_addr = buf_vec[vec_idx].buf_addr;
2768                 buf_len = buf_vec[vec_idx].buf_len;
2769                 buf_avail  = buf_len - buf_offset;
2770         } else if (buf_len == dev->vhost_hlen) {
2771                 if (unlikely(++vec_idx >= nr_vec))
2772                         goto out;
2773                 buf_addr = buf_vec[vec_idx].buf_addr;
2774                 buf_len = buf_vec[vec_idx].buf_len;
2775
2776                 buf_offset = 0;
2777                 buf_avail = buf_len;
2778         } else {
2779                 buf_offset = dev->vhost_hlen;
2780                 buf_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
2781         }
2782
2783         PRINT_PACKET(dev,
2784                         (uintptr_t)(buf_addr + buf_offset),
2785                         (uint32_t)buf_avail, 0);
2786
2787         mbuf_offset = 0;
2788         mbuf_avail  = m->buf_len - RTE_PKTMBUF_HEADROOM;
2789         while (1) {
2790                 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
2791
2792                 if (likely(cpy_len > MAX_BATCH_LEN ||
2793                                         vq->batch_copy_nb_elems >= vq->size ||
2794                                         (hdr && cur == m))) {
2795                         rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
2796                                                 mbuf_offset),
2797                                         (void *)((uintptr_t)(buf_addr +
2798                                                         buf_offset)), cpy_len);
2799                 } else {
2800                         batch_copy[vq->batch_copy_nb_elems].dst =
2801                                 rte_pktmbuf_mtod_offset(cur, void *,
2802                                                 mbuf_offset);
2803                         batch_copy[vq->batch_copy_nb_elems].src =
2804                                 (void *)((uintptr_t)(buf_addr + buf_offset));
2805                         batch_copy[vq->batch_copy_nb_elems].len = cpy_len;
2806                         vq->batch_copy_nb_elems++;
2807                 }
2808
2809                 mbuf_avail  -= cpy_len;
2810                 mbuf_offset += cpy_len;
2811                 buf_avail -= cpy_len;
2812                 buf_offset += cpy_len;
2813
2814                 /* This buf reaches to its end, get the next one */
2815                 if (buf_avail == 0) {
2816                         if (++vec_idx >= nr_vec)
2817                                 break;
2818
2819                         buf_addr = buf_vec[vec_idx].buf_addr;
2820                         buf_len = buf_vec[vec_idx].buf_len;
2821
2822                         buf_offset = 0;
2823                         buf_avail  = buf_len;
2824
2825                         PRINT_PACKET(dev, (uintptr_t)buf_addr,
2826                                         (uint32_t)buf_avail, 0);
2827                 }
2828
2829                 /*
2830                  * This mbuf reaches to its end, get a new one
2831                  * to hold more data.
2832                  */
2833                 if (mbuf_avail == 0) {
2834                         cur = rte_pktmbuf_alloc(mbuf_pool);
2835                         if (unlikely(cur == NULL)) {
2836                                 VHOST_LOG_DATA(ERR, "Failed to "
2837                                         "allocate memory for mbuf.\n");
2838                                 error = -1;
2839                                 goto out;
2840                         }
2841
2842                         prev->next = cur;
2843                         prev->data_len = mbuf_offset;
2844                         m->nb_segs += 1;
2845                         m->pkt_len += mbuf_offset;
2846                         prev = cur;
2847
2848                         mbuf_offset = 0;
2849                         mbuf_avail  = cur->buf_len - RTE_PKTMBUF_HEADROOM;
2850                 }
2851         }
2852
2853         prev->data_len = mbuf_offset;
2854         m->pkt_len    += mbuf_offset;
2855
2856         if (hdr)
2857                 vhost_dequeue_offload(hdr, m, legacy_ol_flags);
2858
2859 out:
2860
2861         return error;
2862 }
2863
2864 static void
2865 virtio_dev_extbuf_free(void *addr __rte_unused, void *opaque)
2866 {
2867         rte_free(opaque);
2868 }
2869
2870 static int
2871 virtio_dev_extbuf_alloc(struct rte_mbuf *pkt, uint32_t size)
2872 {
2873         struct rte_mbuf_ext_shared_info *shinfo = NULL;
2874         uint32_t total_len = RTE_PKTMBUF_HEADROOM + size;
2875         uint16_t buf_len;
2876         rte_iova_t iova;
2877         void *buf;
2878
2879         total_len += sizeof(*shinfo) + sizeof(uintptr_t);
2880         total_len = RTE_ALIGN_CEIL(total_len, sizeof(uintptr_t));
2881
2882         if (unlikely(total_len > UINT16_MAX))
2883                 return -ENOSPC;
2884
2885         buf_len = total_len;
2886         buf = rte_malloc(NULL, buf_len, RTE_CACHE_LINE_SIZE);
2887         if (unlikely(buf == NULL))
2888                 return -ENOMEM;
2889
2890         /* Initialize shinfo */
2891         shinfo = rte_pktmbuf_ext_shinfo_init_helper(buf, &buf_len,
2892                                                 virtio_dev_extbuf_free, buf);
2893         if (unlikely(shinfo == NULL)) {
2894                 rte_free(buf);
2895                 VHOST_LOG_DATA(ERR, "Failed to init shinfo\n");
2896                 return -1;
2897         }
2898
2899         iova = rte_malloc_virt2iova(buf);
2900         rte_pktmbuf_attach_extbuf(pkt, buf, iova, buf_len, shinfo);
2901         rte_pktmbuf_reset_headroom(pkt);
2902
2903         return 0;
2904 }
2905
2906 /*
2907  * Prepare a host supported pktmbuf.
2908  */
2909 static __rte_always_inline int
2910 virtio_dev_pktmbuf_prep(struct virtio_net *dev, struct rte_mbuf *pkt,
2911                          uint32_t data_len)
2912 {
2913         if (rte_pktmbuf_tailroom(pkt) >= data_len)
2914                 return 0;
2915
2916         /* attach an external buffer if supported */
2917         if (dev->extbuf && !virtio_dev_extbuf_alloc(pkt, data_len))
2918                 return 0;
2919
2920         /* check if chained buffers are allowed */
2921         if (!dev->linearbuf)
2922                 return 0;
2923
2924         return -1;
2925 }
2926
2927 __rte_always_inline
2928 static uint16_t
2929 virtio_dev_tx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
2930         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count,
2931         bool legacy_ol_flags)
2932 {
2933         uint16_t i;
2934         uint16_t free_entries;
2935         uint16_t dropped = 0;
2936         static bool allocerr_warned;
2937
2938         /*
2939          * The ordering between avail index and
2940          * desc reads needs to be enforced.
2941          */
2942         free_entries = __atomic_load_n(&vq->avail->idx, __ATOMIC_ACQUIRE) -
2943                         vq->last_avail_idx;
2944         if (free_entries == 0)
2945                 return 0;
2946
2947         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
2948
2949         VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
2950
2951         count = RTE_MIN(count, MAX_PKT_BURST);
2952         count = RTE_MIN(count, free_entries);
2953         VHOST_LOG_DATA(DEBUG, "(%d) about to dequeue %u buffers\n",
2954                         dev->vid, count);
2955
2956         if (rte_pktmbuf_alloc_bulk(mbuf_pool, pkts, count))
2957                 return 0;
2958
2959         for (i = 0; i < count; i++) {
2960                 struct buf_vector buf_vec[BUF_VECTOR_MAX];
2961                 uint16_t head_idx;
2962                 uint32_t buf_len;
2963                 uint16_t nr_vec = 0;
2964                 int err;
2965
2966                 if (unlikely(fill_vec_buf_split(dev, vq,
2967                                                 vq->last_avail_idx + i,
2968                                                 &nr_vec, buf_vec,
2969                                                 &head_idx, &buf_len,
2970                                                 VHOST_ACCESS_RO) < 0))
2971                         break;
2972
2973                 update_shadow_used_ring_split(vq, head_idx, 0);
2974
2975                 err = virtio_dev_pktmbuf_prep(dev, pkts[i], buf_len);
2976                 if (unlikely(err)) {
2977                         /*
2978                          * mbuf allocation fails for jumbo packets when external
2979                          * buffer allocation is not allowed and linear buffer
2980                          * is required. Drop this packet.
2981                          */
2982                         if (!allocerr_warned) {
2983                                 VHOST_LOG_DATA(ERR,
2984                                         "Failed mbuf alloc of size %d from %s on %s.\n",
2985                                         buf_len, mbuf_pool->name, dev->ifname);
2986                                 allocerr_warned = true;
2987                         }
2988                         dropped += 1;
2989                         i++;
2990                         break;
2991                 }
2992
2993                 err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts[i],
2994                                 mbuf_pool, legacy_ol_flags);
2995                 if (unlikely(err)) {
2996                         if (!allocerr_warned) {
2997                                 VHOST_LOG_DATA(ERR,
2998                                         "Failed to copy desc to mbuf on %s.\n",
2999                                         dev->ifname);
3000                                 allocerr_warned = true;
3001                         }
3002                         dropped += 1;
3003                         i++;
3004                         break;
3005                 }
3006         }
3007
3008         if (dropped)
3009                 rte_pktmbuf_free_bulk(&pkts[i - 1], count - i + 1);
3010
3011         vq->last_avail_idx += i;
3012
3013         do_data_copy_dequeue(vq);
3014         if (unlikely(i < count))
3015                 vq->shadow_used_idx = i;
3016         if (likely(vq->shadow_used_idx)) {
3017                 flush_shadow_used_ring_split(dev, vq);
3018                 vhost_vring_call_split(dev, vq);
3019         }
3020
3021         return (i - dropped);
3022 }
3023
3024 __rte_noinline
3025 static uint16_t
3026 virtio_dev_tx_split_legacy(struct virtio_net *dev,
3027         struct vhost_virtqueue *vq, struct rte_mempool *mbuf_pool,
3028         struct rte_mbuf **pkts, uint16_t count)
3029 {
3030         return virtio_dev_tx_split(dev, vq, mbuf_pool, pkts, count, true);
3031 }
3032
3033 __rte_noinline
3034 static uint16_t
3035 virtio_dev_tx_split_compliant(struct virtio_net *dev,
3036         struct vhost_virtqueue *vq, struct rte_mempool *mbuf_pool,
3037         struct rte_mbuf **pkts, uint16_t count)
3038 {
3039         return virtio_dev_tx_split(dev, vq, mbuf_pool, pkts, count, false);
3040 }
3041
3042 static __rte_always_inline int
3043 vhost_reserve_avail_batch_packed(struct virtio_net *dev,
3044                                  struct vhost_virtqueue *vq,
3045                                  struct rte_mbuf **pkts,
3046                                  uint16_t avail_idx,
3047                                  uintptr_t *desc_addrs,
3048                                  uint16_t *ids)
3049 {
3050         bool wrap = vq->avail_wrap_counter;
3051         struct vring_packed_desc *descs = vq->desc_packed;
3052         uint64_t lens[PACKED_BATCH_SIZE];
3053         uint64_t buf_lens[PACKED_BATCH_SIZE];
3054         uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
3055         uint16_t flags, i;
3056
3057         if (unlikely(avail_idx & PACKED_BATCH_MASK))
3058                 return -1;
3059         if (unlikely((avail_idx + PACKED_BATCH_SIZE) > vq->size))
3060                 return -1;
3061
3062         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
3063                 flags = descs[avail_idx + i].flags;
3064                 if (unlikely((wrap != !!(flags & VRING_DESC_F_AVAIL)) ||
3065                              (wrap == !!(flags & VRING_DESC_F_USED))  ||
3066                              (flags & PACKED_DESC_SINGLE_DEQUEUE_FLAG)))
3067                         return -1;
3068         }
3069
3070         rte_atomic_thread_fence(__ATOMIC_ACQUIRE);
3071
3072         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
3073                 lens[i] = descs[avail_idx + i].len;
3074
3075         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
3076                 desc_addrs[i] = vhost_iova_to_vva(dev, vq,
3077                                                   descs[avail_idx + i].addr,
3078                                                   &lens[i], VHOST_ACCESS_RW);
3079         }
3080
3081         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
3082                 if (unlikely(!desc_addrs[i]))
3083                         return -1;
3084                 if (unlikely((lens[i] != descs[avail_idx + i].len)))
3085                         return -1;
3086         }
3087
3088         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
3089                 if (virtio_dev_pktmbuf_prep(dev, pkts[i], lens[i]))
3090                         goto err;
3091         }
3092
3093         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
3094                 buf_lens[i] = pkts[i]->buf_len - pkts[i]->data_off;
3095
3096         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
3097                 if (unlikely(buf_lens[i] < (lens[i] - buf_offset)))
3098                         goto err;
3099         }
3100
3101         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
3102                 pkts[i]->pkt_len = lens[i] - buf_offset;
3103                 pkts[i]->data_len = pkts[i]->pkt_len;
3104                 ids[i] = descs[avail_idx + i].id;
3105         }
3106
3107         return 0;
3108
3109 err:
3110         return -1;
3111 }
3112
3113 static __rte_always_inline int
3114 virtio_dev_tx_batch_packed(struct virtio_net *dev,
3115                            struct vhost_virtqueue *vq,
3116                            struct rte_mbuf **pkts,
3117                            bool legacy_ol_flags)
3118 {
3119         uint16_t avail_idx = vq->last_avail_idx;
3120         uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
3121         struct virtio_net_hdr *hdr;
3122         uintptr_t desc_addrs[PACKED_BATCH_SIZE];
3123         uint16_t ids[PACKED_BATCH_SIZE];
3124         uint16_t i;
3125
3126         if (vhost_reserve_avail_batch_packed(dev, vq, pkts, avail_idx,
3127                                              desc_addrs, ids))
3128                 return -1;
3129
3130         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
3131                 rte_prefetch0((void *)(uintptr_t)desc_addrs[i]);
3132
3133         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
3134                 rte_memcpy(rte_pktmbuf_mtod_offset(pkts[i], void *, 0),
3135                            (void *)(uintptr_t)(desc_addrs[i] + buf_offset),
3136                            pkts[i]->pkt_len);
3137
3138         if (virtio_net_with_host_offload(dev)) {
3139                 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
3140                         hdr = (struct virtio_net_hdr *)(desc_addrs[i]);
3141                         vhost_dequeue_offload(hdr, pkts[i], legacy_ol_flags);
3142                 }
3143         }
3144
3145         if (virtio_net_is_inorder(dev))
3146                 vhost_shadow_dequeue_batch_packed_inorder(vq,
3147                         ids[PACKED_BATCH_SIZE - 1]);
3148         else
3149                 vhost_shadow_dequeue_batch_packed(dev, vq, ids);
3150
3151         vq_inc_last_avail_packed(vq, PACKED_BATCH_SIZE);
3152
3153         return 0;
3154 }
3155
3156 static __rte_always_inline int
3157 vhost_dequeue_single_packed(struct virtio_net *dev,
3158                             struct vhost_virtqueue *vq,
3159                             struct rte_mempool *mbuf_pool,
3160                             struct rte_mbuf *pkts,
3161                             uint16_t *buf_id,
3162                             uint16_t *desc_count,
3163                             bool legacy_ol_flags)
3164 {
3165         struct buf_vector buf_vec[BUF_VECTOR_MAX];
3166         uint32_t buf_len;
3167         uint16_t nr_vec = 0;
3168         int err;
3169         static bool allocerr_warned;
3170
3171         if (unlikely(fill_vec_buf_packed(dev, vq,
3172                                          vq->last_avail_idx, desc_count,
3173                                          buf_vec, &nr_vec,
3174                                          buf_id, &buf_len,
3175                                          VHOST_ACCESS_RO) < 0))
3176                 return -1;
3177
3178         if (unlikely(virtio_dev_pktmbuf_prep(dev, pkts, buf_len))) {
3179                 if (!allocerr_warned) {
3180                         VHOST_LOG_DATA(ERR,
3181                                 "Failed mbuf alloc of size %d from %s on %s.\n",
3182                                 buf_len, mbuf_pool->name, dev->ifname);
3183                         allocerr_warned = true;
3184                 }
3185                 return -1;
3186         }
3187
3188         err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts,
3189                                 mbuf_pool, legacy_ol_flags);
3190         if (unlikely(err)) {
3191                 if (!allocerr_warned) {
3192                         VHOST_LOG_DATA(ERR,
3193                                 "Failed to copy desc to mbuf on %s.\n",
3194                                 dev->ifname);
3195                         allocerr_warned = true;
3196                 }
3197                 return -1;
3198         }
3199
3200         return 0;
3201 }
3202
3203 static __rte_always_inline int
3204 virtio_dev_tx_single_packed(struct virtio_net *dev,
3205                             struct vhost_virtqueue *vq,
3206                             struct rte_mempool *mbuf_pool,
3207                             struct rte_mbuf *pkts,
3208                             bool legacy_ol_flags)
3209 {
3210
3211         uint16_t buf_id, desc_count = 0;
3212         int ret;
3213
3214         ret = vhost_dequeue_single_packed(dev, vq, mbuf_pool, pkts, &buf_id,
3215                                         &desc_count, legacy_ol_flags);
3216
3217         if (likely(desc_count > 0)) {
3218                 if (virtio_net_is_inorder(dev))
3219                         vhost_shadow_dequeue_single_packed_inorder(vq, buf_id,
3220                                                                    desc_count);
3221                 else
3222                         vhost_shadow_dequeue_single_packed(vq, buf_id,
3223                                         desc_count);
3224
3225                 vq_inc_last_avail_packed(vq, desc_count);
3226         }
3227
3228         return ret;
3229 }
3230
3231 __rte_always_inline
3232 static uint16_t
3233 virtio_dev_tx_packed(struct virtio_net *dev,
3234                      struct vhost_virtqueue *__rte_restrict vq,
3235                      struct rte_mempool *mbuf_pool,
3236                      struct rte_mbuf **__rte_restrict pkts,
3237                      uint32_t count,
3238                      bool legacy_ol_flags)
3239 {
3240         uint32_t pkt_idx = 0;
3241
3242         if (rte_pktmbuf_alloc_bulk(mbuf_pool, pkts, count))
3243                 return 0;
3244
3245         do {
3246                 rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
3247
3248                 if (count - pkt_idx >= PACKED_BATCH_SIZE) {
3249                         if (!virtio_dev_tx_batch_packed(dev, vq,
3250                                                         &pkts[pkt_idx],
3251                                                         legacy_ol_flags)) {
3252                                 pkt_idx += PACKED_BATCH_SIZE;
3253                                 continue;
3254                         }
3255                 }
3256
3257                 if (virtio_dev_tx_single_packed(dev, vq, mbuf_pool,
3258                                                 pkts[pkt_idx],
3259                                                 legacy_ol_flags))
3260                         break;
3261                 pkt_idx++;
3262         } while (pkt_idx < count);
3263
3264         if (pkt_idx != count)
3265                 rte_pktmbuf_free_bulk(&pkts[pkt_idx], count - pkt_idx);
3266
3267         if (vq->shadow_used_idx) {
3268                 do_data_copy_dequeue(vq);
3269
3270                 vhost_flush_dequeue_shadow_packed(dev, vq);
3271                 vhost_vring_call_packed(dev, vq);
3272         }
3273
3274         return pkt_idx;
3275 }
3276
3277 __rte_noinline
3278 static uint16_t
3279 virtio_dev_tx_packed_legacy(struct virtio_net *dev,
3280         struct vhost_virtqueue *__rte_restrict vq, struct rte_mempool *mbuf_pool,
3281         struct rte_mbuf **__rte_restrict pkts, uint32_t count)
3282 {
3283         return virtio_dev_tx_packed(dev, vq, mbuf_pool, pkts, count, true);
3284 }
3285
3286 __rte_noinline
3287 static uint16_t
3288 virtio_dev_tx_packed_compliant(struct virtio_net *dev,
3289         struct vhost_virtqueue *__rte_restrict vq, struct rte_mempool *mbuf_pool,
3290         struct rte_mbuf **__rte_restrict pkts, uint32_t count)
3291 {
3292         return virtio_dev_tx_packed(dev, vq, mbuf_pool, pkts, count, false);
3293 }
3294
3295 uint16_t
3296 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
3297         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
3298 {
3299         struct virtio_net *dev;
3300         struct rte_mbuf *rarp_mbuf = NULL;
3301         struct vhost_virtqueue *vq;
3302         int16_t success = 1;
3303
3304         dev = get_device(vid);
3305         if (!dev)
3306                 return 0;
3307
3308         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
3309                 VHOST_LOG_DATA(ERR,
3310                         "(%d) %s: built-in vhost net backend is disabled.\n",
3311                         dev->vid, __func__);
3312                 return 0;
3313         }
3314
3315         if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
3316                 VHOST_LOG_DATA(ERR,
3317                         "(%d) %s: invalid virtqueue idx %d.\n",
3318                         dev->vid, __func__, queue_id);
3319                 return 0;
3320         }
3321
3322         vq = dev->virtqueue[queue_id];
3323
3324         if (unlikely(rte_spinlock_trylock(&vq->access_lock) == 0))
3325                 return 0;
3326
3327         if (unlikely(!vq->enabled)) {
3328                 count = 0;
3329                 goto out_access_unlock;
3330         }
3331
3332         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
3333                 vhost_user_iotlb_rd_lock(vq);
3334
3335         if (unlikely(!vq->access_ok))
3336                 if (unlikely(vring_translate(dev, vq) < 0)) {
3337                         count = 0;
3338                         goto out;
3339                 }
3340
3341         /*
3342          * Construct a RARP broadcast packet, and inject it to the "pkts"
3343          * array, to looks like that guest actually send such packet.
3344          *
3345          * Check user_send_rarp() for more information.
3346          *
3347          * broadcast_rarp shares a cacheline in the virtio_net structure
3348          * with some fields that are accessed during enqueue and
3349          * __atomic_compare_exchange_n causes a write if performed compare
3350          * and exchange. This could result in false sharing between enqueue
3351          * and dequeue.
3352          *
3353          * Prevent unnecessary false sharing by reading broadcast_rarp first
3354          * and only performing compare and exchange if the read indicates it
3355          * is likely to be set.
3356          */
3357         if (unlikely(__atomic_load_n(&dev->broadcast_rarp, __ATOMIC_ACQUIRE) &&
3358                         __atomic_compare_exchange_n(&dev->broadcast_rarp,
3359                         &success, 0, 0, __ATOMIC_RELEASE, __ATOMIC_RELAXED))) {
3360
3361                 rarp_mbuf = rte_net_make_rarp_packet(mbuf_pool, &dev->mac);
3362                 if (rarp_mbuf == NULL) {
3363                         VHOST_LOG_DATA(ERR, "Failed to make RARP packet.\n");
3364                         count = 0;
3365                         goto out;
3366                 }
3367                 count -= 1;
3368         }
3369
3370         if (vq_is_packed(dev)) {
3371                 if (dev->flags & VIRTIO_DEV_LEGACY_OL_FLAGS)
3372                         count = virtio_dev_tx_packed_legacy(dev, vq, mbuf_pool, pkts, count);
3373                 else
3374                         count = virtio_dev_tx_packed_compliant(dev, vq, mbuf_pool, pkts, count);
3375         } else {
3376                 if (dev->flags & VIRTIO_DEV_LEGACY_OL_FLAGS)
3377                         count = virtio_dev_tx_split_legacy(dev, vq, mbuf_pool, pkts, count);
3378                 else
3379                         count = virtio_dev_tx_split_compliant(dev, vq, mbuf_pool, pkts, count);
3380         }
3381
3382 out:
3383         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
3384                 vhost_user_iotlb_rd_unlock(vq);
3385
3386 out_access_unlock:
3387         rte_spinlock_unlock(&vq->access_lock);
3388
3389         if (unlikely(rarp_mbuf != NULL)) {
3390                 /*
3391                  * Inject it to the head of "pkts" array, so that switch's mac
3392                  * learning table will get updated first.
3393                  */
3394                 memmove(&pkts[1], pkts, count * sizeof(struct rte_mbuf *));
3395                 pkts[0] = rarp_mbuf;
3396                 count += 1;
3397         }
3398
3399         return count;
3400 }