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 = vq->last_used_idx;
 222         struct vring_packed_desc *desc_base = &vq->desc_packed[last_used_idx];
 223
 224         if (vq->shadow_used_idx) {
 225                 do_data_copy_enqueue(dev, vq);
 226                 vhost_flush_enqueue_shadow_packed(dev, vq);
 227         }
 228
 229         flags = PACKED_DESC_ENQUEUE_USED_FLAG(vq->used_wrap_counter);
 230
 231         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
 232                 desc_base[i].id = ids[i];
 233                 desc_base[i].len = lens[i];
 234         }
 235
 236         rte_atomic_thread_fence(__ATOMIC_RELEASE);
 237
 238         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
 239                 desc_base[i].flags = flags;
 240         }
 241
 242         vhost_log_cache_used_vring(dev, vq, last_used_idx *
 243                                    sizeof(struct vring_packed_desc),
 244                                    sizeof(struct vring_packed_desc) *
 245                                    PACKED_BATCH_SIZE);
 246         vhost_log_cache_sync(dev, vq);
 247
 248         vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
 249 }
 250
 251 static __rte_always_inline void
 252 vhost_shadow_dequeue_batch_packed_inorder(struct vhost_virtqueue *vq,
 253                                           uint16_t id)
 254 {
 255         vq->shadow_used_packed[0].id = id;
 256
 257         if (!vq->shadow_used_idx) {
 258                 vq->shadow_last_used_idx = vq->last_used_idx;
 259                 vq->shadow_used_packed[0].flags =
 260                         PACKED_DESC_DEQUEUE_USED_FLAG(vq->used_wrap_counter);
 261                 vq->shadow_used_packed[0].len = 0;
 262                 vq->shadow_used_packed[0].count = 1;
 263                 vq->shadow_used_idx++;
 264         }
 265
 266         vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
 267 }
 268
 269 static __rte_always_inline void
 270 vhost_shadow_dequeue_batch_packed(struct virtio_net *dev,
 271                                   struct vhost_virtqueue *vq,
 272                                   uint16_t *ids)
 273 {
 274         uint16_t flags;
 275         uint16_t i;
 276         uint16_t begin;
 277
 278         flags = PACKED_DESC_DEQUEUE_USED_FLAG(vq->used_wrap_counter);
 279
 280         if (!vq->shadow_used_idx) {
 281                 vq->shadow_last_used_idx = vq->last_used_idx;
 282                 vq->shadow_used_packed[0].id  = ids[0];
 283                 vq->shadow_used_packed[0].len = 0;
 284                 vq->shadow_used_packed[0].count = 1;
 285                 vq->shadow_used_packed[0].flags = flags;
 286                 vq->shadow_used_idx++;
 287                 begin = 1;
 288         } else
 289                 begin = 0;
 290
 291         vhost_for_each_try_unroll(i, begin, PACKED_BATCH_SIZE) {
 292                 vq->desc_packed[vq->last_used_idx + i].id = ids[i];
 293                 vq->desc_packed[vq->last_used_idx + i].len = 0;
 294         }
 295
 296         rte_atomic_thread_fence(__ATOMIC_RELEASE);
 297         vhost_for_each_try_unroll(i, begin, PACKED_BATCH_SIZE)
 298                 vq->desc_packed[vq->last_used_idx + i].flags = flags;
 299
 300         vhost_log_cache_used_vring(dev, vq, vq->last_used_idx *
 301                                    sizeof(struct vring_packed_desc),
 302                                    sizeof(struct vring_packed_desc) *
 303                                    PACKED_BATCH_SIZE);
 304         vhost_log_cache_sync(dev, vq);
 305
 306         vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
 307 }
 308
 309 static __rte_always_inline void
 310 vhost_shadow_dequeue_single_packed(struct vhost_virtqueue *vq,
 311                                    uint16_t buf_id,
 312                                    uint16_t count)
 313 {
 314         uint16_t flags;
 315
 316         flags = vq->desc_packed[vq->last_used_idx].flags;
 317         if (vq->used_wrap_counter) {
 318                 flags |= VRING_DESC_F_USED;
 319                 flags |= VRING_DESC_F_AVAIL;
 320         } else {
 321                 flags &= ~VRING_DESC_F_USED;
 322                 flags &= ~VRING_DESC_F_AVAIL;
 323         }
 324
 325         if (!vq->shadow_used_idx) {
 326                 vq->shadow_last_used_idx = vq->last_used_idx;
 327
 328                 vq->shadow_used_packed[0].id  = buf_id;
 329                 vq->shadow_used_packed[0].len = 0;
 330                 vq->shadow_used_packed[0].flags = flags;
 331                 vq->shadow_used_idx++;
 332         } else {
 333                 vq->desc_packed[vq->last_used_idx].id = buf_id;
 334                 vq->desc_packed[vq->last_used_idx].len = 0;
 335                 vq->desc_packed[vq->last_used_idx].flags = flags;
 336         }
 337
 338         vq_inc_last_used_packed(vq, count);
 339 }
 340
 341 static __rte_always_inline void
 342 vhost_shadow_dequeue_single_packed_inorder(struct vhost_virtqueue *vq,
 343                                            uint16_t buf_id,
 344                                            uint16_t count)
 345 {
 346         uint16_t flags;
 347
 348         vq->shadow_used_packed[0].id = buf_id;
 349
 350         flags = vq->desc_packed[vq->last_used_idx].flags;
 351         if (vq->used_wrap_counter) {
 352                 flags |= VRING_DESC_F_USED;
 353                 flags |= VRING_DESC_F_AVAIL;
 354         } else {
 355                 flags &= ~VRING_DESC_F_USED;
 356                 flags &= ~VRING_DESC_F_AVAIL;
 357         }
 358
 359         if (!vq->shadow_used_idx) {
 360                 vq->shadow_last_used_idx = vq->last_used_idx;
 361                 vq->shadow_used_packed[0].len = 0;
 362                 vq->shadow_used_packed[0].flags = flags;
 363                 vq->shadow_used_idx++;
 364         }
 365
 366         vq_inc_last_used_packed(vq, count);
 367 }
 368
 369 static __rte_always_inline void
 370 vhost_shadow_enqueue_packed(struct vhost_virtqueue *vq,
 371                                    uint32_t *len,
 372                                    uint16_t *id,
 373                                    uint16_t *count,
 374                                    uint16_t num_buffers)
 375 {
 376         uint16_t i;
 377
 378         for (i = 0; i < num_buffers; i++) {
 379                 /* enqueue shadow flush action aligned with batch num */
 380                 if (!vq->shadow_used_idx)
 381                         vq->shadow_aligned_idx = vq->last_used_idx &
 382                                 PACKED_BATCH_MASK;
 383                 vq->shadow_used_packed[vq->shadow_used_idx].id  = id[i];
 384                 vq->shadow_used_packed[vq->shadow_used_idx].len = len[i];
 385                 vq->shadow_used_packed[vq->shadow_used_idx].count = count[i];
 386                 vq->shadow_aligned_idx += count[i];
 387                 vq->shadow_used_idx++;
 388         }
 389 }
 390
 391 static __rte_always_inline void
 392 vhost_shadow_enqueue_single_packed(struct virtio_net *dev,
 393                                    struct vhost_virtqueue *vq,
 394                                    uint32_t *len,
 395                                    uint16_t *id,
 396                                    uint16_t *count,
 397                                    uint16_t num_buffers)
 398 {
 399         vhost_shadow_enqueue_packed(vq, len, id, count, num_buffers);
 400
 401         if (vq->shadow_aligned_idx >= PACKED_BATCH_SIZE) {
 402                 do_data_copy_enqueue(dev, vq);
 403                 vhost_flush_enqueue_shadow_packed(dev, vq);
 404         }
 405 }
 406
 407 /* avoid write operation when necessary, to lessen cache issues */
 408 #define ASSIGN_UNLESS_EQUAL(var, val) do {      \
 409         if ((var) != (val))                     \
 410                 (var) = (val);                  \
 411 } while (0)
 412
 413 static __rte_always_inline void
 414 virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
 415 {
 416         uint64_t csum_l4 = m_buf->ol_flags & PKT_TX_L4_MASK;
 417
 418         if (m_buf->ol_flags & PKT_TX_TCP_SEG)
 419                 csum_l4 |= PKT_TX_TCP_CKSUM;
 420
 421         if (csum_l4) {
 422                 net_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 423                 net_hdr->csum_start = m_buf->l2_len + m_buf->l3_len;
 424
 425                 switch (csum_l4) {
 426                 case PKT_TX_TCP_CKSUM:
 427                         net_hdr->csum_offset = (offsetof(struct rte_tcp_hdr,
 428                                                 cksum));
 429                         break;
 430                 case PKT_TX_UDP_CKSUM:
 431                         net_hdr->csum_offset = (offsetof(struct rte_udp_hdr,
 432                                                 dgram_cksum));
 433                         break;
 434                 case PKT_TX_SCTP_CKSUM:
 435                         net_hdr->csum_offset = (offsetof(struct rte_sctp_hdr,
 436                                                 cksum));
 437                         break;
 438                 }
 439         } else {
 440                 ASSIGN_UNLESS_EQUAL(net_hdr->csum_start, 0);
 441                 ASSIGN_UNLESS_EQUAL(net_hdr->csum_offset, 0);
 442                 ASSIGN_UNLESS_EQUAL(net_hdr->flags, 0);
 443         }
 444
 445         /* IP cksum verification cannot be bypassed, then calculate here */
 446         if (m_buf->ol_flags & PKT_TX_IP_CKSUM) {
 447                 struct rte_ipv4_hdr *ipv4_hdr;
 448
 449                 ipv4_hdr = rte_pktmbuf_mtod_offset(m_buf, struct rte_ipv4_hdr *,
 450                                                    m_buf->l2_len);
 451                 ipv4_hdr->hdr_checksum = 0;
 452                 ipv4_hdr->hdr_checksum = rte_ipv4_cksum(ipv4_hdr);
 453         }
 454
 455         if (m_buf->ol_flags & PKT_TX_TCP_SEG) {
 456                 if (m_buf->ol_flags & PKT_TX_IPV4)
 457                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 458                 else
 459                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 460                 net_hdr->gso_size = m_buf->tso_segsz;
 461                 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len
 462                                         + m_buf->l4_len;
 463         } else if (m_buf->ol_flags & PKT_TX_UDP_SEG) {
 464                 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
 465                 net_hdr->gso_size = m_buf->tso_segsz;
 466                 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len +
 467                         m_buf->l4_len;
 468         } else {
 469                 ASSIGN_UNLESS_EQUAL(net_hdr->gso_type, 0);
 470                 ASSIGN_UNLESS_EQUAL(net_hdr->gso_size, 0);
 471                 ASSIGN_UNLESS_EQUAL(net_hdr->hdr_len, 0);
 472         }
 473 }
 474
 475 static __rte_always_inline int
 476 map_one_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 477                 struct buf_vector *buf_vec, uint16_t *vec_idx,
 478                 uint64_t desc_iova, uint64_t desc_len, uint8_t perm)
 479 {
 480         uint16_t vec_id = *vec_idx;
 481
 482         while (desc_len) {
 483                 uint64_t desc_addr;
 484                 uint64_t desc_chunck_len = desc_len;
 485
 486                 if (unlikely(vec_id >= BUF_VECTOR_MAX))
 487                         return -1;
 488
 489                 desc_addr = vhost_iova_to_vva(dev, vq,
 490                                 desc_iova,
 491                                 &desc_chunck_len,
 492                                 perm);
 493                 if (unlikely(!desc_addr))
 494                         return -1;
 495
 496                 rte_prefetch0((void *)(uintptr_t)desc_addr);
 497
 498                 buf_vec[vec_id].buf_iova = desc_iova;
 499                 buf_vec[vec_id].buf_addr = desc_addr;
 500                 buf_vec[vec_id].buf_len  = desc_chunck_len;
 501
 502                 desc_len -= desc_chunck_len;
 503                 desc_iova += desc_chunck_len;
 504                 vec_id++;
 505         }
 506         *vec_idx = vec_id;
 507
 508         return 0;
 509 }
 510
 511 static __rte_always_inline int
 512 fill_vec_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 513                          uint32_t avail_idx, uint16_t *vec_idx,
 514                          struct buf_vector *buf_vec, uint16_t *desc_chain_head,
 515                          uint32_t *desc_chain_len, uint8_t perm)
 516 {
 517         uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
 518         uint16_t vec_id = *vec_idx;
 519         uint32_t len    = 0;
 520         uint64_t dlen;
 521         uint32_t nr_descs = vq->size;
 522         uint32_t cnt    = 0;
 523         struct vring_desc *descs = vq->desc;
 524         struct vring_desc *idesc = NULL;
 525
 526         if (unlikely(idx >= vq->size))
 527                 return -1;
 528
 529         *desc_chain_head = idx;
 530
 531         if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
 532                 dlen = vq->desc[idx].len;
 533                 nr_descs = dlen / sizeof(struct vring_desc);
 534                 if (unlikely(nr_descs > vq->size))
 535                         return -1;
 536
 537                 descs = (struct vring_desc *)(uintptr_t)
 538                         vhost_iova_to_vva(dev, vq, vq->desc[idx].addr,
 539                                                 &dlen,
 540                                                 VHOST_ACCESS_RO);
 541                 if (unlikely(!descs))
 542                         return -1;
 543
 544                 if (unlikely(dlen < vq->desc[idx].len)) {
 545                         /*
 546                          * The indirect desc table is not contiguous
 547                          * in process VA space, we have to copy it.
 548                          */
 549                         idesc = vhost_alloc_copy_ind_table(dev, vq,
 550                                         vq->desc[idx].addr, vq->desc[idx].len);
 551                         if (unlikely(!idesc))
 552                                 return -1;
 553
 554                         descs = idesc;
 555                 }
 556
 557                 idx = 0;
 558         }
 559
 560         while (1) {
 561                 if (unlikely(idx >= nr_descs || cnt++ >= nr_descs)) {
 562                         free_ind_table(idesc);
 563                         return -1;
 564                 }
 565
 566                 dlen = descs[idx].len;
 567                 len += dlen;
 568
 569                 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 570                                                 descs[idx].addr, dlen,
 571                                                 perm))) {
 572                         free_ind_table(idesc);
 573                         return -1;
 574                 }
 575
 576                 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
 577                         break;
 578
 579                 idx = descs[idx].next;
 580         }
 581
 582         *desc_chain_len = len;
 583         *vec_idx = vec_id;
 584
 585         if (unlikely(!!idesc))
 586                 free_ind_table(idesc);
 587
 588         return 0;
 589 }
 590
 591 /*
 592  * Returns -1 on fail, 0 on success
 593  */
 594 static inline int
 595 reserve_avail_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 596                                 uint32_t size, struct buf_vector *buf_vec,
 597                                 uint16_t *num_buffers, uint16_t avail_head,
 598                                 uint16_t *nr_vec)
 599 {
 600         uint16_t cur_idx;
 601         uint16_t vec_idx = 0;
 602         uint16_t max_tries, tries = 0;
 603
 604         uint16_t head_idx = 0;
 605         uint32_t len = 0;
 606
 607         *num_buffers = 0;
 608         cur_idx  = vq->last_avail_idx;
 609
 610         if (rxvq_is_mergeable(dev))
 611                 max_tries = vq->size - 1;
 612         else
 613                 max_tries = 1;
 614
 615         while (size > 0) {
 616                 if (unlikely(cur_idx == avail_head))
 617                         return -1;
 618                 /*
 619                  * if we tried all available ring items, and still
 620                  * can't get enough buf, it means something abnormal
 621                  * happened.
 622                  */
 623                 if (unlikely(++tries > max_tries))
 624                         return -1;
 625
 626                 if (unlikely(fill_vec_buf_split(dev, vq, cur_idx,
 627                                                 &vec_idx, buf_vec,
 628                                                 &head_idx, &len,
 629                                                 VHOST_ACCESS_RW) < 0))
 630                         return -1;
 631                 len = RTE_MIN(len, size);
 632                 update_shadow_used_ring_split(vq, head_idx, len);
 633                 size -= len;
 634
 635                 cur_idx++;
 636                 *num_buffers += 1;
 637         }
 638
 639         *nr_vec = vec_idx;
 640
 641         return 0;
 642 }
 643
 644 static __rte_always_inline int
 645 fill_vec_buf_packed_indirect(struct virtio_net *dev,
 646                         struct vhost_virtqueue *vq,
 647                         struct vring_packed_desc *desc, uint16_t *vec_idx,
 648                         struct buf_vector *buf_vec, uint32_t *len, uint8_t perm)
 649 {
 650         uint16_t i;
 651         uint32_t nr_descs;
 652         uint16_t vec_id = *vec_idx;
 653         uint64_t dlen;
 654         struct vring_packed_desc *descs, *idescs = NULL;
 655
 656         dlen = desc->len;
 657         descs = (struct vring_packed_desc *)(uintptr_t)
 658                 vhost_iova_to_vva(dev, vq, desc->addr, &dlen, VHOST_ACCESS_RO);
 659         if (unlikely(!descs))
 660                 return -1;
 661
 662         if (unlikely(dlen < desc->len)) {
 663                 /*
 664                  * The indirect desc table is not contiguous
 665                  * in process VA space, we have to copy it.
 666                  */
 667                 idescs = vhost_alloc_copy_ind_table(dev,
 668                                 vq, desc->addr, desc->len);
 669                 if (unlikely(!idescs))
 670                         return -1;
 671
 672                 descs = idescs;
 673         }
 674
 675         nr_descs =  desc->len / sizeof(struct vring_packed_desc);
 676         if (unlikely(nr_descs >= vq->size)) {
 677                 free_ind_table(idescs);
 678                 return -1;
 679         }
 680
 681         for (i = 0; i < nr_descs; i++) {
 682                 if (unlikely(vec_id >= BUF_VECTOR_MAX)) {
 683                         free_ind_table(idescs);
 684                         return -1;
 685                 }
 686
 687                 dlen = descs[i].len;
 688                 *len += dlen;
 689                 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 690                                                 descs[i].addr, dlen,
 691                                                 perm)))
 692                         return -1;
 693         }
 694         *vec_idx = vec_id;
 695
 696         if (unlikely(!!idescs))
 697                 free_ind_table(idescs);
 698
 699         return 0;
 700 }
 701
 702 static __rte_always_inline int
 703 fill_vec_buf_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
 704                                 uint16_t avail_idx, uint16_t *desc_count,
 705                                 struct buf_vector *buf_vec, uint16_t *vec_idx,
 706                                 uint16_t *buf_id, uint32_t *len, uint8_t perm)
 707 {
 708         bool wrap_counter = vq->avail_wrap_counter;
 709         struct vring_packed_desc *descs = vq->desc_packed;
 710         uint16_t vec_id = *vec_idx;
 711         uint64_t dlen;
 712
 713         if (avail_idx < vq->last_avail_idx)
 714                 wrap_counter ^= 1;
 715
 716         /*
 717          * Perform a load-acquire barrier in desc_is_avail to
 718          * enforce the ordering between desc flags and desc
 719          * content.
 720          */
 721         if (unlikely(!desc_is_avail(&descs[avail_idx], wrap_counter)))
 722                 return -1;
 723
 724         *desc_count = 0;
 725         *len = 0;
 726
 727         while (1) {
 728                 if (unlikely(vec_id >= BUF_VECTOR_MAX))
 729                         return -1;
 730
 731                 if (unlikely(*desc_count >= vq->size))
 732                         return -1;
 733
 734                 *desc_count += 1;
 735                 *buf_id = descs[avail_idx].id;
 736
 737                 if (descs[avail_idx].flags & VRING_DESC_F_INDIRECT) {
 738                         if (unlikely(fill_vec_buf_packed_indirect(dev, vq,
 739                                                         &descs[avail_idx],
 740                                                         &vec_id, buf_vec,
 741                                                         len, perm) < 0))
 742                                 return -1;
 743                 } else {
 744                         dlen = descs[avail_idx].len;
 745                         *len += dlen;
 746
 747                         if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 748                                                         descs[avail_idx].addr,
 749                                                         dlen,
 750                                                         perm)))
 751                                 return -1;
 752                 }
 753
 754                 if ((descs[avail_idx].flags & VRING_DESC_F_NEXT) == 0)
 755                         break;
 756
 757                 if (++avail_idx >= vq->size) {
 758                         avail_idx -= vq->size;
 759                         wrap_counter ^= 1;
 760                 }
 761         }
 762
 763         *vec_idx = vec_id;
 764
 765         return 0;
 766 }
 767
 768 static __rte_noinline void
 769 copy_vnet_hdr_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 770                 struct buf_vector *buf_vec,
 771                 struct virtio_net_hdr_mrg_rxbuf *hdr)
 772 {
 773         uint64_t len;
 774         uint64_t remain = dev->vhost_hlen;
 775         uint64_t src = (uint64_t)(uintptr_t)hdr, dst;
 776         uint64_t iova = buf_vec->buf_iova;
 777
 778         while (remain) {
 779                 len = RTE_MIN(remain,
 780                                 buf_vec->buf_len);
 781                 dst = buf_vec->buf_addr;
 782                 rte_memcpy((void *)(uintptr_t)dst,
 783                                 (void *)(uintptr_t)src,
 784                                 len);
 785
 786                 PRINT_PACKET(dev, (uintptr_t)dst,
 787                                 (uint32_t)len, 0);
 788                 vhost_log_cache_write_iova(dev, vq,
 789                                 iova, len);
 790
 791                 remain -= len;
 792                 iova += len;
 793                 src += len;
 794                 buf_vec++;
 795         }
 796 }
 797
 798 static __rte_always_inline int
 799 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 800                             struct rte_mbuf *m, struct buf_vector *buf_vec,
 801                             uint16_t nr_vec, uint16_t num_buffers)
 802 {
 803         uint32_t vec_idx = 0;
 804         uint32_t mbuf_offset, mbuf_avail;
 805         uint32_t buf_offset, buf_avail;
 806         uint64_t buf_addr, buf_iova, buf_len;
 807         uint32_t cpy_len;
 808         uint64_t hdr_addr;
 809         struct rte_mbuf *hdr_mbuf;
 810         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 811         struct virtio_net_hdr_mrg_rxbuf tmp_hdr, *hdr = NULL;
 812         int error = 0;
 813
 814         if (unlikely(m == NULL)) {
 815                 error = -1;
 816                 goto out;
 817         }
 818
 819         buf_addr = buf_vec[vec_idx].buf_addr;
 820         buf_iova = buf_vec[vec_idx].buf_iova;
 821         buf_len = buf_vec[vec_idx].buf_len;
 822
 823         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
 824                 error = -1;
 825                 goto out;
 826         }
 827
 828         hdr_mbuf = m;
 829         hdr_addr = buf_addr;
 830         if (unlikely(buf_len < dev->vhost_hlen)) {
 831                 memset(&tmp_hdr, 0, sizeof(struct virtio_net_hdr_mrg_rxbuf));
 832                 hdr = &tmp_hdr;
 833         } else
 834                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)hdr_addr;
 835
 836         VHOST_LOG_DATA(DEBUG, "(%d) RX: num merge buffers %d\n",
 837                 dev->vid, num_buffers);
 838
 839         if (unlikely(buf_len < dev->vhost_hlen)) {
 840                 buf_offset = dev->vhost_hlen - buf_len;
 841                 vec_idx++;
 842                 buf_addr = buf_vec[vec_idx].buf_addr;
 843                 buf_iova = buf_vec[vec_idx].buf_iova;
 844                 buf_len = buf_vec[vec_idx].buf_len;
 845                 buf_avail = buf_len - buf_offset;
 846         } else {
 847                 buf_offset = dev->vhost_hlen;
 848                 buf_avail = buf_len - dev->vhost_hlen;
 849         }
 850
 851         mbuf_avail  = rte_pktmbuf_data_len(m);
 852         mbuf_offset = 0;
 853         while (mbuf_avail != 0 || m->next != NULL) {
 854                 /* done with current buf, get the next one */
 855                 if (buf_avail == 0) {
 856                         vec_idx++;
 857                         if (unlikely(vec_idx >= nr_vec)) {
 858                                 error = -1;
 859                                 goto out;
 860                         }
 861
 862                         buf_addr = buf_vec[vec_idx].buf_addr;
 863                         buf_iova = buf_vec[vec_idx].buf_iova;
 864                         buf_len = buf_vec[vec_idx].buf_len;
 865
 866                         buf_offset = 0;
 867                         buf_avail  = buf_len;
 868                 }
 869
 870                 /* done with current mbuf, get the next one */
 871                 if (mbuf_avail == 0) {
 872                         m = m->next;
 873
 874                         mbuf_offset = 0;
 875                         mbuf_avail  = rte_pktmbuf_data_len(m);
 876                 }
 877
 878                 if (hdr_addr) {
 879                         virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
 880                         if (rxvq_is_mergeable(dev))
 881                                 ASSIGN_UNLESS_EQUAL(hdr->num_buffers,
 882                                                 num_buffers);
 883
 884                         if (unlikely(hdr == &tmp_hdr)) {
 885                                 copy_vnet_hdr_to_desc(dev, vq, buf_vec, hdr);
 886                         } else {
 887                                 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
 888                                                 dev->vhost_hlen, 0);
 889                                 vhost_log_cache_write_iova(dev, vq,
 890                                                 buf_vec[0].buf_iova,
 891                                                 dev->vhost_hlen);
 892                         }
 893
 894                         hdr_addr = 0;
 895                 }
 896
 897                 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
 898
 899                 if (likely(cpy_len > MAX_BATCH_LEN ||
 900                                         vq->batch_copy_nb_elems >= vq->size)) {
 901                         rte_memcpy((void *)((uintptr_t)(buf_addr + buf_offset)),
 902                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 903                                 cpy_len);
 904                         vhost_log_cache_write_iova(dev, vq,
 905                                                    buf_iova + buf_offset,
 906                                                    cpy_len);
 907                         PRINT_PACKET(dev, (uintptr_t)(buf_addr + buf_offset),
 908                                 cpy_len, 0);
 909                 } else {
 910                         batch_copy[vq->batch_copy_nb_elems].dst =
 911                                 (void *)((uintptr_t)(buf_addr + buf_offset));
 912                         batch_copy[vq->batch_copy_nb_elems].src =
 913                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
 914                         batch_copy[vq->batch_copy_nb_elems].log_addr =
 915                                 buf_iova + buf_offset;
 916                         batch_copy[vq->batch_copy_nb_elems].len = cpy_len;
 917                         vq->batch_copy_nb_elems++;
 918                 }
 919
 920                 mbuf_avail  -= cpy_len;
 921                 mbuf_offset += cpy_len;
 922                 buf_avail  -= cpy_len;
 923                 buf_offset += cpy_len;
 924         }
 925
 926 out:
 927
 928         return error;
 929 }
 930
 931 static __rte_always_inline void
 932 async_fill_vec(struct iovec *v, void *base, size_t len)
 933 {
 934         v->iov_base = base;
 935         v->iov_len = len;
 936 }
 937
 938 static __rte_always_inline void
 939 async_fill_iter(struct rte_vhost_iov_iter *it, size_t count,
 940         struct iovec *vec, unsigned long nr_seg)
 941 {
 942         it->offset = 0;
 943         it->count = count;
 944
 945         if (count) {
 946                 it->iov = vec;
 947                 it->nr_segs = nr_seg;
 948         } else {
 949                 it->iov = 0;
 950                 it->nr_segs = 0;
 951         }
 952 }
 953
 954 static __rte_always_inline void
 955 async_fill_desc(struct rte_vhost_async_desc *desc,
 956         struct rte_vhost_iov_iter *src, struct rte_vhost_iov_iter *dst)
 957 {
 958         desc->src = src;
 959         desc->dst = dst;
 960 }
 961
 962 static __rte_always_inline int
 963 async_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 964                         struct rte_mbuf *m, struct buf_vector *buf_vec,
 965                         uint16_t nr_vec, uint16_t num_buffers,
 966                         struct iovec *src_iovec, struct iovec *dst_iovec,
 967                         struct rte_vhost_iov_iter *src_it,
 968                         struct rte_vhost_iov_iter *dst_it)
 969 {
 970         uint32_t vec_idx = 0;
 971         uint32_t mbuf_offset, mbuf_avail;
 972         uint32_t buf_offset, buf_avail;
 973         uint64_t buf_addr, buf_iova, buf_len;
 974         uint32_t cpy_len, cpy_threshold;
 975         uint64_t hdr_addr;
 976         struct rte_mbuf *hdr_mbuf;
 977         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 978         struct virtio_net_hdr_mrg_rxbuf tmp_hdr, *hdr = NULL;
 979         int error = 0;
 980         uint64_t mapped_len;
 981
 982         uint32_t tlen = 0;
 983         int tvec_idx = 0;
 984         void *hpa;
 985
 986         if (unlikely(m == NULL)) {
 987                 error = -1;
 988                 goto out;
 989         }
 990
 991         cpy_threshold = vq->async_threshold;
 992
 993         buf_addr = buf_vec[vec_idx].buf_addr;
 994         buf_iova = buf_vec[vec_idx].buf_iova;
 995         buf_len = buf_vec[vec_idx].buf_len;
 996
 997         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
 998                 error = -1;
 999                 goto out;
1000         }
1001
1002         hdr_mbuf = m;
1003         hdr_addr = buf_addr;
1004         if (unlikely(buf_len < dev->vhost_hlen)) {
1005                 memset(&tmp_hdr, 0, sizeof(struct virtio_net_hdr_mrg_rxbuf));
1006                 hdr = &tmp_hdr;
1007         } else
1008                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)hdr_addr;
1009
1010         VHOST_LOG_DATA(DEBUG, "(%d) RX: num merge buffers %d\n",
1011                 dev->vid, num_buffers);
1012
1013         if (unlikely(buf_len < dev->vhost_hlen)) {
1014                 buf_offset = dev->vhost_hlen - buf_len;
1015                 vec_idx++;
1016                 buf_addr = buf_vec[vec_idx].buf_addr;
1017                 buf_iova = buf_vec[vec_idx].buf_iova;
1018                 buf_len = buf_vec[vec_idx].buf_len;
1019                 buf_avail = buf_len - buf_offset;
1020         } else {
1021                 buf_offset = dev->vhost_hlen;
1022                 buf_avail = buf_len - dev->vhost_hlen;
1023         }
1024
1025         mbuf_avail  = rte_pktmbuf_data_len(m);
1026         mbuf_offset = 0;
1027
1028         while (mbuf_avail != 0 || m->next != NULL) {
1029                 /* done with current buf, get the next one */
1030                 if (buf_avail == 0) {
1031                         vec_idx++;
1032                         if (unlikely(vec_idx >= nr_vec)) {
1033                                 error = -1;
1034                                 goto out;
1035                         }
1036
1037                         buf_addr = buf_vec[vec_idx].buf_addr;
1038                         buf_iova = buf_vec[vec_idx].buf_iova;
1039                         buf_len = buf_vec[vec_idx].buf_len;
1040
1041                         buf_offset = 0;
1042                         buf_avail  = buf_len;
1043                 }
1044
1045                 /* done with current mbuf, get the next one */
1046                 if (mbuf_avail == 0) {
1047                         m = m->next;
1048
1049                         mbuf_offset = 0;
1050                         mbuf_avail  = rte_pktmbuf_data_len(m);
1051                 }
1052
1053                 if (hdr_addr) {
1054                         virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
1055                         if (rxvq_is_mergeable(dev))
1056                                 ASSIGN_UNLESS_EQUAL(hdr->num_buffers,
1057                                                 num_buffers);
1058
1059                         if (unlikely(hdr == &tmp_hdr)) {
1060                                 copy_vnet_hdr_to_desc(dev, vq, buf_vec, hdr);
1061                         } else {
1062                                 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
1063                                                 dev->vhost_hlen, 0);
1064                                 vhost_log_cache_write_iova(dev, vq,
1065                                                 buf_vec[0].buf_iova,
1066                                                 dev->vhost_hlen);
1067                         }
1068
1069                         hdr_addr = 0;
1070                 }
1071
1072                 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
1073
1074                 while (unlikely(cpy_len && cpy_len >= cpy_threshold)) {
1075                         hpa = (void *)(uintptr_t)gpa_to_first_hpa(dev,
1076                                         buf_iova + buf_offset,
1077                                         cpy_len, &mapped_len);
1078
1079                         if (unlikely(!hpa || mapped_len < cpy_threshold))
1080                                 break;
1081
1082                         async_fill_vec(src_iovec + tvec_idx,
1083                                 (void *)(uintptr_t)rte_pktmbuf_iova_offset(m,
1084                                 mbuf_offset), (size_t)mapped_len);
1085
1086                         async_fill_vec(dst_iovec + tvec_idx,
1087                                         hpa, (size_t)mapped_len);
1088
1089                         tlen += (uint32_t)mapped_len;
1090                         cpy_len -= (uint32_t)mapped_len;
1091                         mbuf_avail  -= (uint32_t)mapped_len;
1092                         mbuf_offset += (uint32_t)mapped_len;
1093                         buf_avail  -= (uint32_t)mapped_len;
1094                         buf_offset += (uint32_t)mapped_len;
1095                         tvec_idx++;
1096                 }
1097
1098                 if (likely(cpy_len)) {
1099                         if (unlikely(vq->batch_copy_nb_elems >= vq->size)) {
1100                                 rte_memcpy(
1101                                 (void *)((uintptr_t)(buf_addr + buf_offset)),
1102                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
1103                                 cpy_len);
1104
1105                                 PRINT_PACKET(dev,
1106                                         (uintptr_t)(buf_addr + buf_offset),
1107                                         cpy_len, 0);
1108                         } else {
1109                                 batch_copy[vq->batch_copy_nb_elems].dst =
1110                                 (void *)((uintptr_t)(buf_addr + buf_offset));
1111                                 batch_copy[vq->batch_copy_nb_elems].src =
1112                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
1113                                 batch_copy[vq->batch_copy_nb_elems].log_addr =
1114                                         buf_iova + buf_offset;
1115                                 batch_copy[vq->batch_copy_nb_elems].len =
1116                                         cpy_len;
1117                                 vq->batch_copy_nb_elems++;
1118                         }
1119
1120                         mbuf_avail  -= cpy_len;
1121                         mbuf_offset += cpy_len;
1122                         buf_avail  -= cpy_len;
1123                         buf_offset += cpy_len;
1124                 }
1125
1126         }
1127
1128 out:
1129         if (tlen) {
1130                 async_fill_iter(src_it, tlen, src_iovec, tvec_idx);
1131                 async_fill_iter(dst_it, tlen, dst_iovec, tvec_idx);
1132         } else {
1133                 src_it->count = 0;
1134         }
1135
1136         return error;
1137 }
1138
1139 static __rte_always_inline int
1140 vhost_enqueue_single_packed(struct virtio_net *dev,
1141                             struct vhost_virtqueue *vq,
1142                             struct rte_mbuf *pkt,
1143                             struct buf_vector *buf_vec,
1144                             uint16_t *nr_descs)
1145 {
1146         uint16_t nr_vec = 0;
1147         uint16_t avail_idx = vq->last_avail_idx;
1148         uint16_t max_tries, tries = 0;
1149         uint16_t buf_id = 0;
1150         uint32_t len = 0;
1151         uint16_t desc_count;
1152         uint32_t size = pkt->pkt_len + sizeof(struct virtio_net_hdr_mrg_rxbuf);
1153         uint16_t num_buffers = 0;
1154         uint32_t buffer_len[vq->size];
1155         uint16_t buffer_buf_id[vq->size];
1156         uint16_t buffer_desc_count[vq->size];
1157
1158         if (rxvq_is_mergeable(dev))
1159                 max_tries = vq->size - 1;
1160         else
1161                 max_tries = 1;
1162
1163         while (size > 0) {
1164                 /*
1165                  * if we tried all available ring items, and still
1166                  * can't get enough buf, it means something abnormal
1167                  * happened.
1168                  */
1169                 if (unlikely(++tries > max_tries))
1170                         return -1;
1171
1172                 if (unlikely(fill_vec_buf_packed(dev, vq,
1173                                                 avail_idx, &desc_count,
1174                                                 buf_vec, &nr_vec,
1175                                                 &buf_id, &len,
1176                                                 VHOST_ACCESS_RW) < 0))
1177                         return -1;
1178
1179                 len = RTE_MIN(len, size);
1180                 size -= len;
1181
1182                 buffer_len[num_buffers] = len;
1183                 buffer_buf_id[num_buffers] = buf_id;
1184                 buffer_desc_count[num_buffers] = desc_count;
1185                 num_buffers += 1;
1186
1187                 *nr_descs += desc_count;
1188                 avail_idx += desc_count;
1189                 if (avail_idx >= vq->size)
1190                         avail_idx -= vq->size;
1191         }
1192
1193         if (copy_mbuf_to_desc(dev, vq, pkt, buf_vec, nr_vec, num_buffers) < 0)
1194                 return -1;
1195
1196         vhost_shadow_enqueue_single_packed(dev, vq, buffer_len, buffer_buf_id,
1197                                            buffer_desc_count, num_buffers);
1198
1199         return 0;
1200 }
1201
1202 static __rte_noinline uint32_t
1203 virtio_dev_rx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
1204         struct rte_mbuf **pkts, uint32_t count)
1205 {
1206         uint32_t pkt_idx = 0;
1207         uint16_t num_buffers;
1208         struct buf_vector buf_vec[BUF_VECTOR_MAX];
1209         uint16_t avail_head;
1210
1211         /*
1212          * The ordering between avail index and
1213          * desc reads needs to be enforced.
1214          */
1215         avail_head = __atomic_load_n(&vq->avail->idx, __ATOMIC_ACQUIRE);
1216
1217         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
1218
1219         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
1220                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
1221                 uint16_t nr_vec = 0;
1222
1223                 if (unlikely(reserve_avail_buf_split(dev, vq,
1224                                                 pkt_len, buf_vec, &num_buffers,
1225                                                 avail_head, &nr_vec) < 0)) {
1226                         VHOST_LOG_DATA(DEBUG,
1227                                 "(%d) failed to get enough desc from vring\n",
1228                                 dev->vid);
1229                         vq->shadow_used_idx -= num_buffers;
1230                         break;
1231                 }
1232
1233                 VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
1234                         dev->vid, vq->last_avail_idx,
1235                         vq->last_avail_idx + num_buffers);
1236
1237                 if (copy_mbuf_to_desc(dev, vq, pkts[pkt_idx],
1238                                                 buf_vec, nr_vec,
1239                                                 num_buffers) < 0) {
1240                         vq->shadow_used_idx -= num_buffers;
1241                         break;
1242                 }
1243
1244                 vq->last_avail_idx += num_buffers;
1245         }
1246
1247         do_data_copy_enqueue(dev, vq);
1248
1249         if (likely(vq->shadow_used_idx)) {
1250                 flush_shadow_used_ring_split(dev, vq);
1251                 vhost_vring_call_split(dev, vq);
1252         }
1253
1254         return pkt_idx;
1255 }
1256
1257 static __rte_always_inline int
1258 virtio_dev_rx_batch_packed(struct virtio_net *dev,
1259                            struct vhost_virtqueue *vq,
1260                            struct rte_mbuf **pkts)
1261 {
1262         bool wrap_counter = vq->avail_wrap_counter;
1263         struct vring_packed_desc *descs = vq->desc_packed;
1264         uint16_t avail_idx = vq->last_avail_idx;
1265         uint64_t desc_addrs[PACKED_BATCH_SIZE];
1266         struct virtio_net_hdr_mrg_rxbuf *hdrs[PACKED_BATCH_SIZE];
1267         uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1268         uint64_t lens[PACKED_BATCH_SIZE];
1269         uint16_t ids[PACKED_BATCH_SIZE];
1270         uint16_t i;
1271
1272         if (unlikely(avail_idx & PACKED_BATCH_MASK))
1273                 return -1;
1274
1275         if (unlikely((avail_idx + PACKED_BATCH_SIZE) > vq->size))
1276                 return -1;
1277
1278         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1279                 if (unlikely(pkts[i]->next != NULL))
1280                         return -1;
1281                 if (unlikely(!desc_is_avail(&descs[avail_idx + i],
1282                                             wrap_counter)))
1283                         return -1;
1284         }
1285
1286         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1287                 lens[i] = descs[avail_idx + i].len;
1288
1289         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1290                 if (unlikely(pkts[i]->pkt_len > (lens[i] - buf_offset)))
1291                         return -1;
1292         }
1293
1294         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1295                 desc_addrs[i] = vhost_iova_to_vva(dev, vq,
1296                                                   descs[avail_idx + i].addr,
1297                                                   &lens[i],
1298                                                   VHOST_ACCESS_RW);
1299
1300         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1301                 if (unlikely(!desc_addrs[i]))
1302                         return -1;
1303                 if (unlikely(lens[i] != descs[avail_idx + i].len))
1304                         return -1;
1305         }
1306
1307         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1308                 rte_prefetch0((void *)(uintptr_t)desc_addrs[i]);
1309                 hdrs[i] = (struct virtio_net_hdr_mrg_rxbuf *)
1310                                         (uintptr_t)desc_addrs[i];
1311                 lens[i] = pkts[i]->pkt_len +
1312                         sizeof(struct virtio_net_hdr_mrg_rxbuf);
1313         }
1314
1315         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1316                 virtio_enqueue_offload(pkts[i], &hdrs[i]->hdr);
1317
1318         vq_inc_last_avail_packed(vq, PACKED_BATCH_SIZE);
1319
1320         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1321                 rte_memcpy((void *)(uintptr_t)(desc_addrs[i] + buf_offset),
1322                            rte_pktmbuf_mtod_offset(pkts[i], void *, 0),
1323                            pkts[i]->pkt_len);
1324         }
1325
1326         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1327                 vhost_log_cache_write_iova(dev, vq, descs[avail_idx + i].addr,
1328                                            lens[i]);
1329
1330         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1331                 ids[i] = descs[avail_idx + i].id;
1332
1333         vhost_flush_enqueue_batch_packed(dev, vq, lens, ids);
1334
1335         return 0;
1336 }
1337
1338 static __rte_always_inline int16_t
1339 virtio_dev_rx_single_packed(struct virtio_net *dev,
1340                             struct vhost_virtqueue *vq,
1341                             struct rte_mbuf *pkt)
1342 {
1343         struct buf_vector buf_vec[BUF_VECTOR_MAX];
1344         uint16_t nr_descs = 0;
1345
1346         if (unlikely(vhost_enqueue_single_packed(dev, vq, pkt, buf_vec,
1347                                                  &nr_descs) < 0)) {
1348                 VHOST_LOG_DATA(DEBUG,
1349                                 "(%d) failed to get enough desc from vring\n",
1350                                 dev->vid);
1351                 return -1;
1352         }
1353
1354         VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
1355                         dev->vid, vq->last_avail_idx,
1356                         vq->last_avail_idx + nr_descs);
1357
1358         vq_inc_last_avail_packed(vq, nr_descs);
1359
1360         return 0;
1361 }
1362
1363 static __rte_noinline uint32_t
1364 virtio_dev_rx_packed(struct virtio_net *dev,
1365                      struct vhost_virtqueue *__rte_restrict vq,
1366                      struct rte_mbuf **__rte_restrict pkts,
1367                      uint32_t count)
1368 {
1369         uint32_t pkt_idx = 0;
1370
1371         do {
1372                 rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
1373
1374                 if (count - pkt_idx >= PACKED_BATCH_SIZE) {
1375                         if (!virtio_dev_rx_batch_packed(dev, vq,
1376                                                         &pkts[pkt_idx])) {
1377                                 pkt_idx += PACKED_BATCH_SIZE;
1378                                 continue;
1379                         }
1380                 }
1381
1382                 if (virtio_dev_rx_single_packed(dev, vq, pkts[pkt_idx]))
1383                         break;
1384                 pkt_idx++;
1385
1386         } while (pkt_idx < count);
1387
1388         if (vq->shadow_used_idx) {
1389                 do_data_copy_enqueue(dev, vq);
1390                 vhost_flush_enqueue_shadow_packed(dev, vq);
1391         }
1392
1393         if (pkt_idx)
1394                 vhost_vring_call_packed(dev, vq);
1395
1396         return pkt_idx;
1397 }
1398
1399 static __rte_always_inline uint32_t
1400 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
1401         struct rte_mbuf **pkts, uint32_t count)
1402 {
1403         struct vhost_virtqueue *vq;
1404         uint32_t nb_tx = 0;
1405
1406         VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
1407         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
1408                 VHOST_LOG_DATA(ERR, "(%d) %s: invalid virtqueue idx %d.\n",
1409                         dev->vid, __func__, queue_id);
1410                 return 0;
1411         }
1412
1413         vq = dev->virtqueue[queue_id];
1414
1415         rte_spinlock_lock(&vq->access_lock);
1416
1417         if (unlikely(!vq->enabled))
1418                 goto out_access_unlock;
1419
1420         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1421                 vhost_user_iotlb_rd_lock(vq);
1422
1423         if (unlikely(!vq->access_ok))
1424                 if (unlikely(vring_translate(dev, vq) < 0))
1425                         goto out;
1426
1427         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
1428         if (count == 0)
1429                 goto out;
1430
1431         if (vq_is_packed(dev))
1432                 nb_tx = virtio_dev_rx_packed(dev, vq, pkts, count);
1433         else
1434                 nb_tx = virtio_dev_rx_split(dev, vq, pkts, count);
1435
1436 out:
1437         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1438                 vhost_user_iotlb_rd_unlock(vq);
1439
1440 out_access_unlock:
1441         rte_spinlock_unlock(&vq->access_lock);
1442
1443         return nb_tx;
1444 }
1445
1446 uint16_t
1447 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
1448         struct rte_mbuf **__rte_restrict pkts, uint16_t count)
1449 {
1450         struct virtio_net *dev = get_device(vid);
1451
1452         if (!dev)
1453                 return 0;
1454
1455         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
1456                 VHOST_LOG_DATA(ERR,
1457                         "(%d) %s: built-in vhost net backend is disabled.\n",
1458                         dev->vid, __func__);
1459                 return 0;
1460         }
1461
1462         return virtio_dev_rx(dev, queue_id, pkts, count);
1463 }
1464
1465 static __rte_always_inline uint16_t
1466 virtio_dev_rx_async_get_info_idx(uint16_t pkts_idx,
1467         uint16_t vq_size, uint16_t n_inflight)
1468 {
1469         return pkts_idx > n_inflight ? (pkts_idx - n_inflight) :
1470                 (vq_size - n_inflight + pkts_idx) & (vq_size - 1);
1471 }
1472
1473 static __rte_always_inline void
1474 store_dma_desc_info_split(struct vring_used_elem *s_ring, struct vring_used_elem *d_ring,
1475                 uint16_t ring_size, uint16_t s_idx, uint16_t d_idx, uint16_t count)
1476 {
1477         size_t elem_size = sizeof(struct vring_used_elem);
1478
1479         if (d_idx + count <= ring_size) {
1480                 rte_memcpy(d_ring + d_idx, s_ring + s_idx, count * elem_size);
1481         } else {
1482                 uint16_t size = ring_size - d_idx;
1483
1484                 rte_memcpy(d_ring + d_idx, s_ring + s_idx, size * elem_size);
1485                 rte_memcpy(d_ring, s_ring + s_idx + size, (count - size) * elem_size);
1486         }
1487 }
1488
1489 static __rte_always_inline void
1490 store_dma_desc_info_packed(struct vring_used_elem_packed *s_ring,
1491                 struct vring_used_elem_packed *d_ring,
1492                 uint16_t ring_size, uint16_t s_idx, uint16_t d_idx, uint16_t count)
1493 {
1494         size_t elem_size = sizeof(struct vring_used_elem_packed);
1495
1496         if (d_idx + count <= ring_size) {
1497                 rte_memcpy(d_ring + d_idx, s_ring + s_idx, count * elem_size);
1498         } else {
1499                 uint16_t size = ring_size - d_idx;
1500
1501                 rte_memcpy(d_ring + d_idx, s_ring + s_idx, size * elem_size);
1502                 rte_memcpy(d_ring, s_ring + s_idx + size, (count - size) * elem_size);
1503         }
1504 }
1505
1506 static __rte_noinline uint32_t
1507 virtio_dev_rx_async_submit_split(struct virtio_net *dev,
1508         struct vhost_virtqueue *vq, uint16_t queue_id,
1509         struct rte_mbuf **pkts, uint32_t count,
1510         struct rte_mbuf **comp_pkts, uint32_t *comp_count)
1511 {
1512         uint32_t pkt_idx = 0, pkt_burst_idx = 0;
1513         uint16_t num_buffers;
1514         struct buf_vector buf_vec[BUF_VECTOR_MAX];
1515         uint16_t avail_head;
1516
1517         struct rte_vhost_iov_iter *it_pool = vq->it_pool;
1518         struct iovec *vec_pool = vq->vec_pool;
1519         struct rte_vhost_async_desc tdes[MAX_PKT_BURST];
1520         struct iovec *src_iovec = vec_pool;
1521         struct iovec *dst_iovec = vec_pool + (VHOST_MAX_ASYNC_VEC >> 1);
1522         uint16_t slot_idx = 0;
1523         uint16_t segs_await = 0;
1524         uint16_t iovec_idx = 0, it_idx = 0;
1525         struct async_inflight_info *pkts_info = vq->async_pkts_info;
1526         uint32_t n_pkts = 0, pkt_err = 0;
1527         uint32_t num_async_pkts = 0, num_done_pkts = 0;
1528         struct {
1529                 uint16_t pkt_idx;
1530                 uint16_t last_avail_idx;
1531         } async_pkts_log[MAX_PKT_BURST];
1532
1533         /*
1534          * The ordering between avail index and desc reads need to be enforced.
1535          */
1536         avail_head = __atomic_load_n(&vq->avail->idx, __ATOMIC_ACQUIRE);
1537
1538         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
1539
1540         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
1541                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
1542                 uint16_t nr_vec = 0;
1543
1544                 if (unlikely(reserve_avail_buf_split(dev, vq,
1545                                                 pkt_len, buf_vec, &num_buffers,
1546                                                 avail_head, &nr_vec) < 0)) {
1547                         VHOST_LOG_DATA(DEBUG,
1548                                 "(%d) failed to get enough desc from vring\n",
1549                                 dev->vid);
1550                         vq->shadow_used_idx -= num_buffers;
1551                         break;
1552                 }
1553
1554                 VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
1555                         dev->vid, vq->last_avail_idx,
1556                         vq->last_avail_idx + num_buffers);
1557
1558                 if (async_mbuf_to_desc(dev, vq, pkts[pkt_idx], buf_vec, nr_vec, num_buffers,
1559                                 &src_iovec[iovec_idx], &dst_iovec[iovec_idx],
1560                                 &it_pool[it_idx], &it_pool[it_idx + 1]) < 0) {
1561                         vq->shadow_used_idx -= num_buffers;
1562                         break;
1563                 }
1564
1565                 slot_idx = (vq->async_pkts_idx + num_async_pkts) &
1566                         (vq->size - 1);
1567                 if (it_pool[it_idx].count) {
1568                         uint16_t from, to;
1569
1570                         async_fill_desc(&tdes[pkt_burst_idx++],
1571                                 &it_pool[it_idx], &it_pool[it_idx + 1]);
1572                         pkts_info[slot_idx].descs = num_buffers;
1573                         pkts_info[slot_idx].mbuf = pkts[pkt_idx];
1574                         async_pkts_log[num_async_pkts].pkt_idx = pkt_idx;
1575                         async_pkts_log[num_async_pkts++].last_avail_idx =
1576                                 vq->last_avail_idx;
1577
1578                         iovec_idx += it_pool[it_idx].nr_segs;
1579                         it_idx += 2;
1580
1581                         segs_await += it_pool[it_idx].nr_segs;
1582
1583                         /**
1584                          * recover shadow used ring and keep DMA-occupied
1585                          * descriptors.
1586                          */
1587                         from = vq->shadow_used_idx - num_buffers;
1588                         to = vq->async_desc_idx_split & (vq->size - 1);
1589
1590                         store_dma_desc_info_split(vq->shadow_used_split,
1591                                         vq->async_descs_split, vq->size, from, to, num_buffers);
1592
1593                         vq->async_desc_idx_split += num_buffers;
1594                         vq->shadow_used_idx -= num_buffers;
1595                 } else
1596                         comp_pkts[num_done_pkts++] = pkts[pkt_idx];
1597
1598                 vq->last_avail_idx += num_buffers;
1599
1600                 /*
1601                  * conditions to trigger async device transfer:
1602                  * - buffered packet number reaches transfer threshold
1603                  * - unused async iov number is less than max vhost vector
1604                  */
1605                 if (unlikely(pkt_burst_idx >= VHOST_ASYNC_BATCH_THRESHOLD ||
1606                         ((VHOST_MAX_ASYNC_VEC >> 1) - segs_await <
1607                         BUF_VECTOR_MAX))) {
1608                         n_pkts = vq->async_ops.transfer_data(dev->vid,
1609                                         queue_id, tdes, 0, pkt_burst_idx);
1610                         iovec_idx = 0;
1611                         it_idx = 0;
1612
1613                         segs_await = 0;
1614                         vq->async_pkts_inflight_n += n_pkts;
1615
1616                         if (unlikely(n_pkts < pkt_burst_idx)) {
1617                                 /*
1618                                  * log error packets number here and do actual
1619                                  * error processing when applications poll
1620                                  * completion
1621                                  */
1622                                 pkt_err = pkt_burst_idx - n_pkts;
1623                                 pkt_burst_idx = 0;
1624                                 break;
1625                         }
1626
1627                         pkt_burst_idx = 0;
1628                 }
1629         }
1630
1631         if (pkt_burst_idx) {
1632                 n_pkts = vq->async_ops.transfer_data(dev->vid,
1633                                 queue_id, tdes, 0, pkt_burst_idx);
1634                 vq->async_pkts_inflight_n += n_pkts;
1635
1636                 if (unlikely(n_pkts < pkt_burst_idx))
1637                         pkt_err = pkt_burst_idx - n_pkts;
1638         }
1639
1640         do_data_copy_enqueue(dev, vq);
1641
1642         if (unlikely(pkt_err)) {
1643                 uint16_t num_descs = 0;
1644
1645                 num_async_pkts -= pkt_err;
1646                 /* calculate the sum of descriptors of DMA-error packets. */
1647                 while (pkt_err-- > 0) {
1648                         num_descs += pkts_info[slot_idx & (vq->size - 1)].descs;
1649                         slot_idx--;
1650                 }
1651                 vq->async_desc_idx_split -= num_descs;
1652                 /* recover shadow used ring and available ring */
1653                 vq->shadow_used_idx -= (vq->last_avail_idx -
1654                                 async_pkts_log[num_async_pkts].last_avail_idx -
1655                                 num_descs);
1656                 vq->last_avail_idx =
1657                         async_pkts_log[num_async_pkts].last_avail_idx;
1658                 pkt_idx = async_pkts_log[num_async_pkts].pkt_idx;
1659                 num_done_pkts = pkt_idx - num_async_pkts;
1660         }
1661
1662         vq->async_pkts_idx += num_async_pkts;
1663         *comp_count = num_done_pkts;
1664
1665         if (likely(vq->shadow_used_idx)) {
1666                 flush_shadow_used_ring_split(dev, vq);
1667                 vhost_vring_call_split(dev, vq);
1668         }
1669
1670         return pkt_idx;
1671 }
1672
1673 static __rte_always_inline void
1674 vhost_update_used_packed(struct vhost_virtqueue *vq,
1675                         struct vring_used_elem_packed *shadow_ring,
1676                         uint16_t count)
1677 {
1678         int i;
1679         uint16_t used_idx = vq->last_used_idx;
1680         uint16_t head_idx = vq->last_used_idx;
1681         uint16_t head_flags = 0;
1682
1683         if (count == 0)
1684                 return;
1685
1686         /* Split loop in two to save memory barriers */
1687         for (i = 0; i < count; i++) {
1688                 vq->desc_packed[used_idx].id = shadow_ring[i].id;
1689                 vq->desc_packed[used_idx].len = shadow_ring[i].len;
1690
1691                 used_idx += shadow_ring[i].count;
1692                 if (used_idx >= vq->size)
1693                         used_idx -= vq->size;
1694         }
1695
1696         /* The ordering for storing desc flags needs to be enforced. */
1697         rte_atomic_thread_fence(__ATOMIC_RELEASE);
1698
1699         for (i = 0; i < count; i++) {
1700                 uint16_t flags;
1701
1702                 if (vq->shadow_used_packed[i].len)
1703                         flags = VRING_DESC_F_WRITE;
1704                 else
1705                         flags = 0;
1706
1707                 if (vq->used_wrap_counter) {
1708                         flags |= VRING_DESC_F_USED;
1709                         flags |= VRING_DESC_F_AVAIL;
1710                 } else {
1711                         flags &= ~VRING_DESC_F_USED;
1712                         flags &= ~VRING_DESC_F_AVAIL;
1713                 }
1714
1715                 if (i > 0) {
1716                         vq->desc_packed[vq->last_used_idx].flags = flags;
1717                 } else {
1718                         head_idx = vq->last_used_idx;
1719                         head_flags = flags;
1720                 }
1721
1722                 vq_inc_last_used_packed(vq, shadow_ring[i].count);
1723         }
1724
1725         vq->desc_packed[head_idx].flags = head_flags;
1726 }
1727
1728 static __rte_always_inline int
1729 virtio_dev_rx_async_batch_packed(struct virtio_net *dev,
1730                            struct vhost_virtqueue *vq,
1731                            struct rte_mbuf **pkts,
1732                            struct rte_mbuf **comp_pkts, uint32_t *pkt_done)
1733 {
1734         uint16_t i;
1735         uint32_t cpy_threshold = vq->async_threshold;
1736
1737         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1738                 if (unlikely(pkts[i]->pkt_len >= cpy_threshold))
1739                         return -1;
1740         }
1741         if (!virtio_dev_rx_batch_packed(dev, vq, pkts)) {
1742                 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1743                         comp_pkts[(*pkt_done)++] = pkts[i];
1744
1745                 return 0;
1746         }
1747
1748         return -1;
1749 }
1750
1751 static __rte_always_inline int
1752 vhost_enqueue_async_single_packed(struct virtio_net *dev,
1753                             struct vhost_virtqueue *vq,
1754                             struct rte_mbuf *pkt,
1755                             struct buf_vector *buf_vec,
1756                             uint16_t *nr_descs,
1757                             uint16_t *nr_buffers,
1758                             struct vring_packed_desc *async_descs,
1759                             struct iovec *src_iovec, struct iovec *dst_iovec,
1760                             struct rte_vhost_iov_iter *src_it,
1761                             struct rte_vhost_iov_iter *dst_it)
1762 {
1763         uint16_t nr_vec = 0;
1764         uint16_t avail_idx = vq->last_avail_idx;
1765         uint16_t max_tries, tries = 0;
1766         uint16_t buf_id = 0;
1767         uint32_t len = 0;
1768         uint16_t desc_count = 0;
1769         uint32_t size = pkt->pkt_len + sizeof(struct virtio_net_hdr_mrg_rxbuf);
1770         uint32_t buffer_len[vq->size];
1771         uint16_t buffer_buf_id[vq->size];
1772         uint16_t buffer_desc_count[vq->size];
1773
1774         if (rxvq_is_mergeable(dev))
1775                 max_tries = vq->size - 1;
1776         else
1777                 max_tries = 1;
1778
1779         while (size > 0) {
1780                 /*
1781                  * if we tried all available ring items, and still
1782                  * can't get enough buf, it means something abnormal
1783                  * happened.
1784                  */
1785                 if (unlikely(++tries > max_tries))
1786                         return -1;
1787
1788                 if (unlikely(fill_vec_buf_packed(dev, vq, avail_idx, &desc_count, buf_vec, &nr_vec,
1789                                                 &buf_id, &len, VHOST_ACCESS_RW) < 0))
1790                         return -1;
1791
1792                 len = RTE_MIN(len, size);
1793                 size -= len;
1794
1795                 buffer_len[*nr_buffers] = len;
1796                 buffer_buf_id[*nr_buffers] = buf_id;
1797                 buffer_desc_count[*nr_buffers] = desc_count;
1798                 *nr_buffers += 1;
1799
1800                 *nr_descs += desc_count;
1801                 avail_idx += desc_count;
1802                 if (avail_idx >= vq->size)
1803                         avail_idx -= vq->size;
1804         }
1805
1806         if (async_mbuf_to_desc(dev, vq, pkt, buf_vec, nr_vec, *nr_buffers, src_iovec, dst_iovec,
1807                         src_it, dst_it) < 0)
1808                 return -1;
1809         /* store descriptors for DMA */
1810         if (avail_idx >= *nr_descs) {
1811                 rte_memcpy(async_descs, &vq->desc_packed[vq->last_avail_idx],
1812                         *nr_descs * sizeof(struct vring_packed_desc));
1813         } else {
1814                 uint16_t nr_copy = vq->size - vq->last_avail_idx;
1815
1816                 rte_memcpy(async_descs, &vq->desc_packed[vq->last_avail_idx],
1817                         nr_copy * sizeof(struct vring_packed_desc));
1818                 rte_memcpy(async_descs + nr_copy, vq->desc_packed,
1819                         (*nr_descs - nr_copy) * sizeof(struct vring_packed_desc));
1820         }
1821
1822         vhost_shadow_enqueue_packed(vq, buffer_len, buffer_buf_id, buffer_desc_count, *nr_buffers);
1823
1824         return 0;
1825 }
1826
1827 static __rte_always_inline int16_t
1828 virtio_dev_rx_async_single_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
1829                             struct rte_mbuf *pkt, uint16_t *nr_descs, uint16_t *nr_buffers,
1830                             struct vring_packed_desc *async_descs,
1831                             struct iovec *src_iovec, struct iovec *dst_iovec,
1832                             struct rte_vhost_iov_iter *src_it, struct rte_vhost_iov_iter *dst_it)
1833 {
1834         struct buf_vector buf_vec[BUF_VECTOR_MAX];
1835
1836         if (unlikely(vhost_enqueue_async_single_packed(dev, vq, pkt, buf_vec, nr_descs, nr_buffers,
1837                                                  async_descs, src_iovec, dst_iovec,
1838                                                  src_it, dst_it) < 0)) {
1839                 VHOST_LOG_DATA(DEBUG, "(%d) failed to get enough desc from vring\n", dev->vid);
1840                 return -1;
1841         }
1842
1843         VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
1844                         dev->vid, vq->last_avail_idx, vq->last_avail_idx + *nr_descs);
1845
1846         return 0;
1847 }
1848
1849 static __rte_always_inline void
1850 dma_error_handler_packed(struct vhost_virtqueue *vq, struct vring_packed_desc *async_descs,
1851                         uint16_t async_descs_idx, uint16_t slot_idx, uint32_t nr_err,
1852                         uint32_t *pkt_idx, uint32_t *num_async_pkts, uint32_t *num_done_pkts)
1853 {
1854         uint16_t descs_err = 0;
1855         uint16_t buffers_err = 0;
1856         struct async_inflight_info *pkts_info = vq->async_pkts_info;
1857
1858         *num_async_pkts -= nr_err;
1859         *pkt_idx -= nr_err;
1860         /* calculate the sum of buffers and descs of DMA-error packets. */
1861         while (nr_err-- > 0) {
1862                 descs_err += pkts_info[slot_idx % vq->size].descs;
1863                 buffers_err += pkts_info[slot_idx % vq->size].nr_buffers;
1864                 slot_idx--;
1865         }
1866
1867         vq->async_buffer_idx_packed -= buffers_err;
1868
1869         if (vq->last_avail_idx >= descs_err) {
1870                 vq->last_avail_idx -= descs_err;
1871
1872                 rte_memcpy(&vq->desc_packed[vq->last_avail_idx],
1873                         &async_descs[async_descs_idx - descs_err],
1874                         descs_err * sizeof(struct vring_packed_desc));
1875         } else {
1876                 uint16_t nr_copy;
1877
1878                 vq->last_avail_idx = vq->last_avail_idx + vq->size - descs_err;
1879                 nr_copy = vq->size - vq->last_avail_idx;
1880                 rte_memcpy(&vq->desc_packed[vq->last_avail_idx],
1881                         &async_descs[async_descs_idx - descs_err],
1882                         nr_copy * sizeof(struct vring_packed_desc));
1883                 descs_err -= nr_copy;
1884                 rte_memcpy(&vq->desc_packed[0], &async_descs[async_descs_idx - descs_err],
1885                         descs_err * sizeof(struct vring_packed_desc));
1886                 vq->avail_wrap_counter ^= 1;
1887         }
1888
1889         *num_done_pkts = *pkt_idx - *num_async_pkts;
1890 }
1891
1892 static __rte_noinline uint32_t
1893 virtio_dev_rx_async_submit_packed(struct virtio_net *dev,
1894         struct vhost_virtqueue *vq, uint16_t queue_id,
1895         struct rte_mbuf **pkts, uint32_t count,
1896         struct rte_mbuf **comp_pkts, uint32_t *comp_count)
1897 {
1898         uint32_t pkt_idx = 0, pkt_burst_idx = 0;
1899         uint32_t remained = count;
1900         uint16_t async_descs_idx = 0;
1901         uint16_t num_buffers;
1902         uint16_t num_descs;
1903
1904         struct rte_vhost_iov_iter *it_pool = vq->it_pool;
1905         struct iovec *vec_pool = vq->vec_pool;
1906         struct rte_vhost_async_desc tdes[MAX_PKT_BURST];
1907         struct iovec *src_iovec = vec_pool;
1908         struct iovec *dst_iovec = vec_pool + (VHOST_MAX_ASYNC_VEC >> 1);
1909         uint16_t slot_idx = 0;
1910         uint16_t segs_await = 0;
1911         uint16_t iovec_idx = 0, it_idx = 0;
1912         struct async_inflight_info *pkts_info = vq->async_pkts_info;
1913         uint32_t n_pkts = 0, pkt_err = 0;
1914         uint32_t num_async_pkts = 0, num_done_pkts = 0;
1915         struct vring_packed_desc async_descs[vq->size];
1916
1917         do {
1918                 rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
1919                 if (remained >= PACKED_BATCH_SIZE) {
1920                         if (!virtio_dev_rx_async_batch_packed(dev, vq,
1921                                 &pkts[pkt_idx], comp_pkts, &num_done_pkts)) {
1922                                 pkt_idx += PACKED_BATCH_SIZE;
1923                                 remained -= PACKED_BATCH_SIZE;
1924                                 continue;
1925                         }
1926                 }
1927
1928                 num_buffers = 0;
1929                 num_descs = 0;
1930                 if (unlikely(virtio_dev_rx_async_single_packed(dev, vq, pkts[pkt_idx],
1931                                                 &num_descs, &num_buffers,
1932                                                 &async_descs[async_descs_idx],
1933                                                 &src_iovec[iovec_idx], &dst_iovec[iovec_idx],
1934                                                 &it_pool[it_idx], &it_pool[it_idx + 1]) < 0))
1935                         break;
1936
1937                 VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
1938                         dev->vid, vq->last_avail_idx,
1939                         vq->last_avail_idx + num_descs);
1940
1941                 slot_idx = (vq->async_pkts_idx + num_async_pkts) % vq->size;
1942                 if (it_pool[it_idx].count) {
1943                         uint16_t from, to;
1944
1945                         async_descs_idx += num_descs;
1946                         async_fill_desc(&tdes[pkt_burst_idx++],
1947                                 &it_pool[it_idx], &it_pool[it_idx + 1]);
1948                         pkts_info[slot_idx].descs = num_descs;
1949                         pkts_info[slot_idx].nr_buffers = num_buffers;
1950                         pkts_info[slot_idx].mbuf = pkts[pkt_idx];
1951                         num_async_pkts++;
1952                         iovec_idx += it_pool[it_idx].nr_segs;
1953                         it_idx += 2;
1954
1955                         segs_await += it_pool[it_idx].nr_segs;
1956
1957                         /**
1958                          * recover shadow used ring and keep DMA-occupied
1959                          * descriptors.
1960                          */
1961                         from = vq->shadow_used_idx - num_buffers;
1962                         to = vq->async_buffer_idx_packed % vq->size;
1963                         store_dma_desc_info_packed(vq->shadow_used_packed,
1964                                         vq->async_buffers_packed, vq->size, from, to, num_buffers);
1965
1966                         vq->async_buffer_idx_packed += num_buffers;
1967                         vq->shadow_used_idx -= num_buffers;
1968                 } else {
1969                         comp_pkts[num_done_pkts++] = pkts[pkt_idx];
1970                 }
1971
1972                 pkt_idx++;
1973                 remained--;
1974                 vq_inc_last_avail_packed(vq, num_descs);
1975
1976                 /*
1977                  * conditions to trigger async device transfer:
1978                  * - buffered packet number reaches transfer threshold
1979                  * - unused async iov number is less than max vhost vector
1980                  */
1981                 if (unlikely(pkt_burst_idx >= VHOST_ASYNC_BATCH_THRESHOLD ||
1982                         ((VHOST_MAX_ASYNC_VEC >> 1) - segs_await < BUF_VECTOR_MAX))) {
1983                         n_pkts = vq->async_ops.transfer_data(dev->vid, queue_id,
1984                                 tdes, 0, pkt_burst_idx);
1985                         iovec_idx = 0;
1986                         it_idx = 0;
1987                         segs_await = 0;
1988                         vq->async_pkts_inflight_n += n_pkts;
1989
1990                         if (unlikely(n_pkts < pkt_burst_idx)) {
1991                                 /*
1992                                  * log error packets number here and do actual
1993                                  * error processing when applications poll
1994                                  * completion
1995                                  */
1996                                 pkt_err = pkt_burst_idx - n_pkts;
1997                                 pkt_burst_idx = 0;
1998                                 break;
1999                         }
2000
2001                         pkt_burst_idx = 0;
2002                 }
2003         } while (pkt_idx < count);
2004
2005         if (pkt_burst_idx) {
2006                 n_pkts = vq->async_ops.transfer_data(dev->vid, queue_id, tdes, 0, pkt_burst_idx);
2007                 vq->async_pkts_inflight_n += n_pkts;
2008
2009                 if (unlikely(n_pkts < pkt_burst_idx))
2010                         pkt_err = pkt_burst_idx - n_pkts;
2011         }
2012
2013         do_data_copy_enqueue(dev, vq);
2014
2015         if (unlikely(pkt_err))
2016                 dma_error_handler_packed(vq, async_descs, async_descs_idx, slot_idx, pkt_err,
2017                                         &pkt_idx, &num_async_pkts, &num_done_pkts);
2018         vq->async_pkts_idx += num_async_pkts;
2019         *comp_count = num_done_pkts;
2020
2021         if (likely(vq->shadow_used_idx)) {
2022                 vhost_flush_enqueue_shadow_packed(dev, vq);
2023                 vhost_vring_call_packed(dev, vq);
2024         }
2025
2026         return pkt_idx;
2027 }
2028
2029 static __rte_always_inline void
2030 write_back_completed_descs_split(struct vhost_virtqueue *vq, uint16_t n_descs)
2031 {
2032         uint16_t nr_left = n_descs;
2033         uint16_t nr_copy;
2034         uint16_t to, from;
2035
2036         do {
2037                 from = vq->last_async_desc_idx_split & (vq->size - 1);
2038                 nr_copy = nr_left + from <= vq->size ? nr_left : vq->size - from;
2039                 to = vq->last_used_idx & (vq->size - 1);
2040
2041                 if (to + nr_copy <= vq->size) {
2042                         rte_memcpy(&vq->used->ring[to], &vq->async_descs_split[from],
2043                                         nr_copy * sizeof(struct vring_used_elem));
2044                 } else {
2045                         uint16_t size = vq->size - to;
2046
2047                         rte_memcpy(&vq->used->ring[to], &vq->async_descs_split[from],
2048                                         size * sizeof(struct vring_used_elem));
2049                         rte_memcpy(&vq->used->ring[0], &vq->async_descs_split[from + size],
2050                                         (nr_copy - size) * sizeof(struct vring_used_elem));
2051                 }
2052
2053                 vq->last_async_desc_idx_split += nr_copy;
2054                 vq->last_used_idx += nr_copy;
2055                 nr_left -= nr_copy;
2056         } while (nr_left > 0);
2057 }
2058
2059 static __rte_always_inline void
2060 write_back_completed_descs_packed(struct vhost_virtqueue *vq,
2061                                 uint16_t n_buffers)
2062 {
2063         uint16_t nr_left = n_buffers;
2064         uint16_t from, to;
2065
2066         do {
2067                 from = vq->last_async_buffer_idx_packed % vq->size;
2068                 to = (from + nr_left) % vq->size;
2069                 if (to > from) {
2070                         vhost_update_used_packed(vq, vq->async_buffers_packed + from, to - from);
2071                         vq->last_async_buffer_idx_packed += nr_left;
2072                         nr_left = 0;
2073                 } else {
2074                         vhost_update_used_packed(vq, vq->async_buffers_packed + from,
2075                                 vq->size - from);
2076                         vq->last_async_buffer_idx_packed += vq->size - from;
2077                         nr_left -= vq->size - from;
2078                 }
2079         } while (nr_left > 0);
2080 }
2081
2082 uint16_t rte_vhost_poll_enqueue_completed(int vid, uint16_t queue_id,
2083                 struct rte_mbuf **pkts, uint16_t count)
2084 {
2085         struct virtio_net *dev = get_device(vid);
2086         struct vhost_virtqueue *vq;
2087         uint16_t n_pkts_cpl = 0, n_pkts_put = 0, n_descs = 0, n_buffers = 0;
2088         uint16_t start_idx, pkts_idx, vq_size;
2089         struct async_inflight_info *pkts_info;
2090         uint16_t from, i;
2091
2092         if (!dev)
2093                 return 0;
2094
2095         VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
2096         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
2097                 VHOST_LOG_DATA(ERR, "(%d) %s: invalid virtqueue idx %d.\n",
2098                         dev->vid, __func__, queue_id);
2099                 return 0;
2100         }
2101
2102         vq = dev->virtqueue[queue_id];
2103
2104         if (unlikely(!vq->async_registered)) {
2105                 VHOST_LOG_DATA(ERR, "(%d) %s: async not registered for queue id %d.\n",
2106                         dev->vid, __func__, queue_id);
2107                 return 0;
2108         }
2109
2110         rte_spinlock_lock(&vq->access_lock);
2111
2112         pkts_idx = vq->async_pkts_idx % vq->size;
2113         pkts_info = vq->async_pkts_info;
2114         vq_size = vq->size;
2115         start_idx = virtio_dev_rx_async_get_info_idx(pkts_idx,
2116                 vq_size, vq->async_pkts_inflight_n);
2117
2118         if (count > vq->async_last_pkts_n)
2119                 n_pkts_cpl = vq->async_ops.check_completed_copies(vid,
2120                         queue_id, 0, count - vq->async_last_pkts_n);
2121         n_pkts_cpl += vq->async_last_pkts_n;
2122
2123         n_pkts_put = RTE_MIN(count, n_pkts_cpl);
2124         if (unlikely(n_pkts_put == 0)) {
2125                 vq->async_last_pkts_n = n_pkts_cpl;
2126                 goto done;
2127         }
2128
2129         if (vq_is_packed(dev)) {
2130                 for (i = 0; i < n_pkts_put; i++) {
2131                         from = (start_idx + i) & (vq_size - 1);
2132                         n_buffers += pkts_info[from].nr_buffers;
2133                         pkts[i] = pkts_info[from].mbuf;
2134                 }
2135         } else {
2136                 for (i = 0; i < n_pkts_put; i++) {
2137                         from = (start_idx + i) & (vq_size - 1);
2138                         n_descs += pkts_info[from].descs;
2139                         pkts[i] = pkts_info[from].mbuf;
2140                 }
2141         }
2142
2143         vq->async_last_pkts_n = n_pkts_cpl - n_pkts_put;
2144         vq->async_pkts_inflight_n -= n_pkts_put;
2145
2146         if (likely(vq->enabled && vq->access_ok)) {
2147                 if (vq_is_packed(dev)) {
2148                         write_back_completed_descs_packed(vq, n_buffers);
2149
2150                         vhost_vring_call_packed(dev, vq);
2151                 } else {
2152                         write_back_completed_descs_split(vq, n_descs);
2153
2154                         __atomic_add_fetch(&vq->used->idx, n_descs,
2155                                         __ATOMIC_RELEASE);
2156                         vhost_vring_call_split(dev, vq);
2157                 }
2158         } else {
2159                 if (vq_is_packed(dev))
2160                         vq->last_async_buffer_idx_packed += n_buffers;
2161                 else
2162                         vq->last_async_desc_idx_split += n_descs;
2163         }
2164
2165 done:
2166         rte_spinlock_unlock(&vq->access_lock);
2167
2168         return n_pkts_put;
2169 }
2170
2171 static __rte_always_inline uint32_t
2172 virtio_dev_rx_async_submit(struct virtio_net *dev, uint16_t queue_id,
2173         struct rte_mbuf **pkts, uint32_t count,
2174         struct rte_mbuf **comp_pkts, uint32_t *comp_count)
2175 {
2176         struct vhost_virtqueue *vq;
2177         uint32_t nb_tx = 0;
2178
2179         VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
2180         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
2181                 VHOST_LOG_DATA(ERR, "(%d) %s: invalid virtqueue idx %d.\n",
2182                         dev->vid, __func__, queue_id);
2183                 return 0;
2184         }
2185
2186         vq = dev->virtqueue[queue_id];
2187
2188         rte_spinlock_lock(&vq->access_lock);
2189
2190         if (unlikely(!vq->enabled || !vq->async_registered))
2191                 goto out_access_unlock;
2192
2193         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
2194                 vhost_user_iotlb_rd_lock(vq);
2195
2196         if (unlikely(!vq->access_ok))
2197                 if (unlikely(vring_translate(dev, vq) < 0))
2198                         goto out;
2199
2200         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
2201         if (count == 0)
2202                 goto out;
2203
2204         if (vq_is_packed(dev))
2205                 nb_tx = virtio_dev_rx_async_submit_packed(dev,
2206                                 vq, queue_id, pkts, count, comp_pkts,
2207                                 comp_count);
2208         else
2209                 nb_tx = virtio_dev_rx_async_submit_split(dev,
2210                                 vq, queue_id, pkts, count, comp_pkts,
2211                                 comp_count);
2212
2213 out:
2214         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
2215                 vhost_user_iotlb_rd_unlock(vq);
2216
2217 out_access_unlock:
2218         rte_spinlock_unlock(&vq->access_lock);
2219
2220         return nb_tx;
2221 }
2222
2223 uint16_t
2224 rte_vhost_submit_enqueue_burst(int vid, uint16_t queue_id,
2225                 struct rte_mbuf **pkts, uint16_t count,
2226                 struct rte_mbuf **comp_pkts, uint32_t *comp_count)
2227 {
2228         struct virtio_net *dev = get_device(vid);
2229
2230         *comp_count = 0;
2231         if (!dev)
2232                 return 0;
2233
2234         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
2235                 VHOST_LOG_DATA(ERR,
2236                         "(%d) %s: built-in vhost net backend is disabled.\n",
2237                         dev->vid, __func__);
2238                 return 0;
2239         }
2240
2241         return virtio_dev_rx_async_submit(dev, queue_id, pkts, count, comp_pkts,
2242                         comp_count);
2243 }
2244
2245 static inline bool
2246 virtio_net_with_host_offload(struct virtio_net *dev)
2247 {
2248         if (dev->features &
2249                         ((1ULL << VIRTIO_NET_F_CSUM) |
2250                          (1ULL << VIRTIO_NET_F_HOST_ECN) |
2251                          (1ULL << VIRTIO_NET_F_HOST_TSO4) |
2252                          (1ULL << VIRTIO_NET_F_HOST_TSO6) |
2253                          (1ULL << VIRTIO_NET_F_HOST_UFO)))
2254                 return true;
2255
2256         return false;
2257 }
2258
2259 static void
2260 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
2261 {
2262         struct rte_ipv4_hdr *ipv4_hdr;
2263         struct rte_ipv6_hdr *ipv6_hdr;
2264         void *l3_hdr = NULL;
2265         struct rte_ether_hdr *eth_hdr;
2266         uint16_t ethertype;
2267
2268         eth_hdr = rte_pktmbuf_mtod(m, struct rte_ether_hdr *);
2269
2270         m->l2_len = sizeof(struct rte_ether_hdr);
2271         ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
2272
2273         if (ethertype == RTE_ETHER_TYPE_VLAN) {
2274                 struct rte_vlan_hdr *vlan_hdr =
2275                         (struct rte_vlan_hdr *)(eth_hdr + 1);
2276
2277                 m->l2_len += sizeof(struct rte_vlan_hdr);
2278                 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
2279         }
2280
2281         l3_hdr = (char *)eth_hdr + m->l2_len;
2282
2283         switch (ethertype) {
2284         case RTE_ETHER_TYPE_IPV4:
2285                 ipv4_hdr = l3_hdr;
2286                 *l4_proto = ipv4_hdr->next_proto_id;
2287                 m->l3_len = rte_ipv4_hdr_len(ipv4_hdr);
2288                 *l4_hdr = (char *)l3_hdr + m->l3_len;
2289                 m->ol_flags |= PKT_TX_IPV4;
2290                 break;
2291         case RTE_ETHER_TYPE_IPV6:
2292                 ipv6_hdr = l3_hdr;
2293                 *l4_proto = ipv6_hdr->proto;
2294                 m->l3_len = sizeof(struct rte_ipv6_hdr);
2295                 *l4_hdr = (char *)l3_hdr + m->l3_len;
2296                 m->ol_flags |= PKT_TX_IPV6;
2297                 break;
2298         default:
2299                 m->l3_len = 0;
2300                 *l4_proto = 0;
2301                 *l4_hdr = NULL;
2302                 break;
2303         }
2304 }
2305
2306 static __rte_always_inline void
2307 vhost_dequeue_offload_legacy(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
2308 {
2309         uint16_t l4_proto = 0;
2310         void *l4_hdr = NULL;
2311         struct rte_tcp_hdr *tcp_hdr = NULL;
2312
2313         parse_ethernet(m, &l4_proto, &l4_hdr);
2314         if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
2315                 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
2316                         switch (hdr->csum_offset) {
2317                         case (offsetof(struct rte_tcp_hdr, cksum)):
2318                                 if (l4_proto == IPPROTO_TCP)
2319                                         m->ol_flags |= PKT_TX_TCP_CKSUM;
2320                                 break;
2321                         case (offsetof(struct rte_udp_hdr, dgram_cksum)):
2322                                 if (l4_proto == IPPROTO_UDP)
2323                                         m->ol_flags |= PKT_TX_UDP_CKSUM;
2324                                 break;
2325                         case (offsetof(struct rte_sctp_hdr, cksum)):
2326                                 if (l4_proto == IPPROTO_SCTP)
2327                                         m->ol_flags |= PKT_TX_SCTP_CKSUM;
2328                                 break;
2329                         default:
2330                                 break;
2331                         }
2332                 }
2333         }
2334
2335         if (l4_hdr && hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
2336                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
2337                 case VIRTIO_NET_HDR_GSO_TCPV4:
2338                 case VIRTIO_NET_HDR_GSO_TCPV6:
2339                         tcp_hdr = l4_hdr;
2340                         m->ol_flags |= PKT_TX_TCP_SEG;
2341                         m->tso_segsz = hdr->gso_size;
2342                         m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
2343                         break;
2344                 case VIRTIO_NET_HDR_GSO_UDP:
2345                         m->ol_flags |= PKT_TX_UDP_SEG;
2346                         m->tso_segsz = hdr->gso_size;
2347                         m->l4_len = sizeof(struct rte_udp_hdr);
2348                         break;
2349                 default:
2350                         VHOST_LOG_DATA(WARNING,
2351                                 "unsupported gso type %u.\n", hdr->gso_type);
2352                         break;
2353                 }
2354         }
2355 }
2356
2357 static __rte_always_inline void
2358 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m,
2359         bool legacy_ol_flags)
2360 {
2361         struct rte_net_hdr_lens hdr_lens;
2362         int l4_supported = 0;
2363         uint32_t ptype;
2364
2365         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
2366                 return;
2367
2368         if (legacy_ol_flags) {
2369                 vhost_dequeue_offload_legacy(hdr, m);
2370                 return;
2371         }
2372
2373         m->ol_flags |= PKT_RX_IP_CKSUM_UNKNOWN;
2374
2375         ptype = rte_net_get_ptype(m, &hdr_lens, RTE_PTYPE_ALL_MASK);
2376         m->packet_type = ptype;
2377         if ((ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_TCP ||
2378             (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_UDP ||
2379             (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_SCTP)
2380                 l4_supported = 1;
2381
2382         /* According to Virtio 1.1 spec, the device only needs to look at
2383          * VIRTIO_NET_HDR_F_NEEDS_CSUM in the packet transmission path.
2384          * This differs from the processing incoming packets path where the
2385          * driver could rely on VIRTIO_NET_HDR_F_DATA_VALID flag set by the
2386          * device.
2387          *
2388          * 5.1.6.2.1 Driver Requirements: Packet Transmission
2389          * The driver MUST NOT set the VIRTIO_NET_HDR_F_DATA_VALID and
2390          * VIRTIO_NET_HDR_F_RSC_INFO bits in flags.
2391          *
2392          * 5.1.6.2.2 Device Requirements: Packet Transmission
2393          * The device MUST ignore flag bits that it does not recognize.
2394          */
2395         if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
2396                 uint32_t hdrlen;
2397
2398                 hdrlen = hdr_lens.l2_len + hdr_lens.l3_len + hdr_lens.l4_len;
2399                 if (hdr->csum_start <= hdrlen && l4_supported != 0) {
2400                         m->ol_flags |= PKT_RX_L4_CKSUM_NONE;
2401                 } else {
2402                         /* Unknown proto or tunnel, do sw cksum. We can assume
2403                          * the cksum field is in the first segment since the
2404                          * buffers we provided to the host are large enough.
2405                          * In case of SCTP, this will be wrong since it's a CRC
2406                          * but there's nothing we can do.
2407                          */
2408                         uint16_t csum = 0, off;
2409
2410                         if (rte_raw_cksum_mbuf(m, hdr->csum_start,
2411                                         rte_pktmbuf_pkt_len(m) - hdr->csum_start, &csum) < 0)
2412                                 return;
2413                         if (likely(csum != 0xffff))
2414                                 csum = ~csum;
2415                         off = hdr->csum_offset + hdr->csum_start;
2416                         if (rte_pktmbuf_data_len(m) >= off + 1)
2417                                 *rte_pktmbuf_mtod_offset(m, uint16_t *, off) = csum;
2418                 }
2419         }
2420
2421         if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
2422                 if (hdr->gso_size == 0)
2423                         return;
2424
2425                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
2426                 case VIRTIO_NET_HDR_GSO_TCPV4:
2427                 case VIRTIO_NET_HDR_GSO_TCPV6:
2428                         if ((ptype & RTE_PTYPE_L4_MASK) != RTE_PTYPE_L4_TCP)
2429                                 break;
2430                         m->ol_flags |= PKT_RX_LRO | PKT_RX_L4_CKSUM_NONE;
2431                         m->tso_segsz = hdr->gso_size;
2432                         break;
2433                 case VIRTIO_NET_HDR_GSO_UDP:
2434                         if ((ptype & RTE_PTYPE_L4_MASK) != RTE_PTYPE_L4_UDP)
2435                                 break;
2436                         m->ol_flags |= PKT_RX_LRO | PKT_RX_L4_CKSUM_NONE;
2437                         m->tso_segsz = hdr->gso_size;
2438                         break;
2439                 default:
2440                         break;
2441                 }
2442         }
2443 }
2444
2445 static __rte_noinline void
2446 copy_vnet_hdr_from_desc(struct virtio_net_hdr *hdr,
2447                 struct buf_vector *buf_vec)
2448 {
2449         uint64_t len;
2450         uint64_t remain = sizeof(struct virtio_net_hdr);
2451         uint64_t src;
2452         uint64_t dst = (uint64_t)(uintptr_t)hdr;
2453
2454         while (remain) {
2455                 len = RTE_MIN(remain, buf_vec->buf_len);
2456                 src = buf_vec->buf_addr;
2457                 rte_memcpy((void *)(uintptr_t)dst,
2458                                 (void *)(uintptr_t)src, len);
2459
2460                 remain -= len;
2461                 dst += len;
2462                 buf_vec++;
2463         }
2464 }
2465
2466 static __rte_always_inline int
2467 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
2468                   struct buf_vector *buf_vec, uint16_t nr_vec,
2469                   struct rte_mbuf *m, struct rte_mempool *mbuf_pool,
2470                   bool legacy_ol_flags)
2471 {
2472         uint32_t buf_avail, buf_offset;
2473         uint64_t buf_addr, buf_len;
2474         uint32_t mbuf_avail, mbuf_offset;
2475         uint32_t cpy_len;
2476         struct rte_mbuf *cur = m, *prev = m;
2477         struct virtio_net_hdr tmp_hdr;
2478         struct virtio_net_hdr *hdr = NULL;
2479         /* A counter to avoid desc dead loop chain */
2480         uint16_t vec_idx = 0;
2481         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
2482         int error = 0;
2483
2484         buf_addr = buf_vec[vec_idx].buf_addr;
2485         buf_len = buf_vec[vec_idx].buf_len;
2486
2487         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
2488                 error = -1;
2489                 goto out;
2490         }
2491
2492         if (virtio_net_with_host_offload(dev)) {
2493                 if (unlikely(buf_len < sizeof(struct virtio_net_hdr))) {
2494                         /*
2495                          * No luck, the virtio-net header doesn't fit
2496                          * in a contiguous virtual area.
2497                          */
2498                         copy_vnet_hdr_from_desc(&tmp_hdr, buf_vec);
2499                         hdr = &tmp_hdr;
2500                 } else {
2501                         hdr = (struct virtio_net_hdr *)((uintptr_t)buf_addr);
2502                 }
2503         }
2504
2505         /*
2506          * A virtio driver normally uses at least 2 desc buffers
2507          * for Tx: the first for storing the header, and others
2508          * for storing the data.
2509          */
2510         if (unlikely(buf_len < dev->vhost_hlen)) {
2511                 buf_offset = dev->vhost_hlen - buf_len;
2512                 vec_idx++;
2513                 buf_addr = buf_vec[vec_idx].buf_addr;
2514                 buf_len = buf_vec[vec_idx].buf_len;
2515                 buf_avail  = buf_len - buf_offset;
2516         } else if (buf_len == dev->vhost_hlen) {
2517                 if (unlikely(++vec_idx >= nr_vec))
2518                         goto out;
2519                 buf_addr = buf_vec[vec_idx].buf_addr;
2520                 buf_len = buf_vec[vec_idx].buf_len;
2521
2522                 buf_offset = 0;
2523                 buf_avail = buf_len;
2524         } else {
2525                 buf_offset = dev->vhost_hlen;
2526                 buf_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
2527         }
2528
2529         PRINT_PACKET(dev,
2530                         (uintptr_t)(buf_addr + buf_offset),
2531                         (uint32_t)buf_avail, 0);
2532
2533         mbuf_offset = 0;
2534         mbuf_avail  = m->buf_len - RTE_PKTMBUF_HEADROOM;
2535         while (1) {
2536                 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
2537
2538                 if (likely(cpy_len > MAX_BATCH_LEN ||
2539                                         vq->batch_copy_nb_elems >= vq->size ||
2540                                         (hdr && cur == m))) {
2541                         rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
2542                                                 mbuf_offset),
2543                                         (void *)((uintptr_t)(buf_addr +
2544                                                         buf_offset)), cpy_len);
2545                 } else {
2546                         batch_copy[vq->batch_copy_nb_elems].dst =
2547                                 rte_pktmbuf_mtod_offset(cur, void *,
2548                                                 mbuf_offset);
2549                         batch_copy[vq->batch_copy_nb_elems].src =
2550                                 (void *)((uintptr_t)(buf_addr + buf_offset));
2551                         batch_copy[vq->batch_copy_nb_elems].len = cpy_len;
2552                         vq->batch_copy_nb_elems++;
2553                 }
2554
2555                 mbuf_avail  -= cpy_len;
2556                 mbuf_offset += cpy_len;
2557                 buf_avail -= cpy_len;
2558                 buf_offset += cpy_len;
2559
2560                 /* This buf reaches to its end, get the next one */
2561                 if (buf_avail == 0) {
2562                         if (++vec_idx >= nr_vec)
2563                                 break;
2564
2565                         buf_addr = buf_vec[vec_idx].buf_addr;
2566                         buf_len = buf_vec[vec_idx].buf_len;
2567
2568                         buf_offset = 0;
2569                         buf_avail  = buf_len;
2570
2571                         PRINT_PACKET(dev, (uintptr_t)buf_addr,
2572                                         (uint32_t)buf_avail, 0);
2573                 }
2574
2575                 /*
2576                  * This mbuf reaches to its end, get a new one
2577                  * to hold more data.
2578                  */
2579                 if (mbuf_avail == 0) {
2580                         cur = rte_pktmbuf_alloc(mbuf_pool);
2581                         if (unlikely(cur == NULL)) {
2582                                 VHOST_LOG_DATA(ERR, "Failed to "
2583                                         "allocate memory for mbuf.\n");
2584                                 error = -1;
2585                                 goto out;
2586                         }
2587
2588                         prev->next = cur;
2589                         prev->data_len = mbuf_offset;
2590                         m->nb_segs += 1;
2591                         m->pkt_len += mbuf_offset;
2592                         prev = cur;
2593
2594                         mbuf_offset = 0;
2595                         mbuf_avail  = cur->buf_len - RTE_PKTMBUF_HEADROOM;
2596                 }
2597         }
2598
2599         prev->data_len = mbuf_offset;
2600         m->pkt_len    += mbuf_offset;
2601
2602         if (hdr)
2603                 vhost_dequeue_offload(hdr, m, legacy_ol_flags);
2604
2605 out:
2606
2607         return error;
2608 }
2609
2610 static void
2611 virtio_dev_extbuf_free(void *addr __rte_unused, void *opaque)
2612 {
2613         rte_free(opaque);
2614 }
2615
2616 static int
2617 virtio_dev_extbuf_alloc(struct rte_mbuf *pkt, uint32_t size)
2618 {
2619         struct rte_mbuf_ext_shared_info *shinfo = NULL;
2620         uint32_t total_len = RTE_PKTMBUF_HEADROOM + size;
2621         uint16_t buf_len;
2622         rte_iova_t iova;
2623         void *buf;
2624
2625         total_len += sizeof(*shinfo) + sizeof(uintptr_t);
2626         total_len = RTE_ALIGN_CEIL(total_len, sizeof(uintptr_t));
2627
2628         if (unlikely(total_len > UINT16_MAX))
2629                 return -ENOSPC;
2630
2631         buf_len = total_len;
2632         buf = rte_malloc(NULL, buf_len, RTE_CACHE_LINE_SIZE);
2633         if (unlikely(buf == NULL))
2634                 return -ENOMEM;
2635
2636         /* Initialize shinfo */
2637         shinfo = rte_pktmbuf_ext_shinfo_init_helper(buf, &buf_len,
2638                                                 virtio_dev_extbuf_free, buf);
2639         if (unlikely(shinfo == NULL)) {
2640                 rte_free(buf);
2641                 VHOST_LOG_DATA(ERR, "Failed to init shinfo\n");
2642                 return -1;
2643         }
2644
2645         iova = rte_malloc_virt2iova(buf);
2646         rte_pktmbuf_attach_extbuf(pkt, buf, iova, buf_len, shinfo);
2647         rte_pktmbuf_reset_headroom(pkt);
2648
2649         return 0;
2650 }
2651
2652 static __rte_always_inline int
2653 virtio_dev_pktmbuf_prep(struct virtio_net *dev, struct rte_mbuf *pkt,
2654                          uint32_t data_len)
2655 {
2656         if (rte_pktmbuf_tailroom(pkt) >= data_len)
2657                 return 0;
2658
2659         /* attach an external buffer if supported */
2660         if (dev->extbuf && !virtio_dev_extbuf_alloc(pkt, data_len))
2661                 return 0;
2662
2663         /* check if chained buffers are allowed */
2664         if (!dev->linearbuf)
2665                 return 0;
2666
2667         return -1;
2668 }
2669
2670 /*
2671  * Allocate a host supported pktmbuf.
2672  */
2673 static __rte_always_inline struct rte_mbuf *
2674 virtio_dev_pktmbuf_alloc(struct virtio_net *dev, struct rte_mempool *mp,
2675                          uint32_t data_len)
2676 {
2677         struct rte_mbuf *pkt = rte_pktmbuf_alloc(mp);
2678
2679         if (unlikely(pkt == NULL)) {
2680                 VHOST_LOG_DATA(ERR,
2681                         "Failed to allocate memory for mbuf.\n");
2682                 return NULL;
2683         }
2684
2685         if (virtio_dev_pktmbuf_prep(dev, pkt, data_len)) {
2686                 /* Data doesn't fit into the buffer and the host supports
2687                  * only linear buffers
2688                  */
2689                 rte_pktmbuf_free(pkt);
2690                 return NULL;
2691         }
2692
2693         return pkt;
2694 }
2695
2696 __rte_always_inline
2697 static uint16_t
2698 virtio_dev_tx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
2699         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count,
2700         bool legacy_ol_flags)
2701 {
2702         uint16_t i;
2703         uint16_t free_entries;
2704         uint16_t dropped = 0;
2705         static bool allocerr_warned;
2706
2707         /*
2708          * The ordering between avail index and
2709          * desc reads needs to be enforced.
2710          */
2711         free_entries = __atomic_load_n(&vq->avail->idx, __ATOMIC_ACQUIRE) -
2712                         vq->last_avail_idx;
2713         if (free_entries == 0)
2714                 return 0;
2715
2716         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
2717
2718         VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
2719
2720         count = RTE_MIN(count, MAX_PKT_BURST);
2721         count = RTE_MIN(count, free_entries);
2722         VHOST_LOG_DATA(DEBUG, "(%d) about to dequeue %u buffers\n",
2723                         dev->vid, count);
2724
2725         for (i = 0; i < count; i++) {
2726                 struct buf_vector buf_vec[BUF_VECTOR_MAX];
2727                 uint16_t head_idx;
2728                 uint32_t buf_len;
2729                 uint16_t nr_vec = 0;
2730                 int err;
2731
2732                 if (unlikely(fill_vec_buf_split(dev, vq,
2733                                                 vq->last_avail_idx + i,
2734                                                 &nr_vec, buf_vec,
2735                                                 &head_idx, &buf_len,
2736                                                 VHOST_ACCESS_RO) < 0))
2737                         break;
2738
2739                 update_shadow_used_ring_split(vq, head_idx, 0);
2740
2741                 pkts[i] = virtio_dev_pktmbuf_alloc(dev, mbuf_pool, buf_len);
2742                 if (unlikely(pkts[i] == NULL)) {
2743                         /*
2744                          * mbuf allocation fails for jumbo packets when external
2745                          * buffer allocation is not allowed and linear buffer
2746                          * is required. Drop this packet.
2747                          */
2748                         if (!allocerr_warned) {
2749                                 VHOST_LOG_DATA(ERR,
2750                                         "Failed mbuf alloc of size %d from %s on %s.\n",
2751                                         buf_len, mbuf_pool->name, dev->ifname);
2752                                 allocerr_warned = true;
2753                         }
2754                         dropped += 1;
2755                         i++;
2756                         break;
2757                 }
2758
2759                 err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts[i],
2760                                 mbuf_pool, legacy_ol_flags);
2761                 if (unlikely(err)) {
2762                         rte_pktmbuf_free(pkts[i]);
2763                         if (!allocerr_warned) {
2764                                 VHOST_LOG_DATA(ERR,
2765                                         "Failed to copy desc to mbuf on %s.\n",
2766                                         dev->ifname);
2767                                 allocerr_warned = true;
2768                         }
2769                         dropped += 1;
2770                         i++;
2771                         break;
2772                 }
2773         }
2774
2775         vq->last_avail_idx += i;
2776
2777         do_data_copy_dequeue(vq);
2778         if (unlikely(i < count))
2779                 vq->shadow_used_idx = i;
2780         if (likely(vq->shadow_used_idx)) {
2781                 flush_shadow_used_ring_split(dev, vq);
2782                 vhost_vring_call_split(dev, vq);
2783         }
2784
2785         return (i - dropped);
2786 }
2787
2788 __rte_noinline
2789 static uint16_t
2790 virtio_dev_tx_split_legacy(struct virtio_net *dev,
2791         struct vhost_virtqueue *vq, struct rte_mempool *mbuf_pool,
2792         struct rte_mbuf **pkts, uint16_t count)
2793 {
2794         return virtio_dev_tx_split(dev, vq, mbuf_pool, pkts, count, true);
2795 }
2796
2797 __rte_noinline
2798 static uint16_t
2799 virtio_dev_tx_split_compliant(struct virtio_net *dev,
2800         struct vhost_virtqueue *vq, struct rte_mempool *mbuf_pool,
2801         struct rte_mbuf **pkts, uint16_t count)
2802 {
2803         return virtio_dev_tx_split(dev, vq, mbuf_pool, pkts, count, false);
2804 }
2805
2806 static __rte_always_inline int
2807 vhost_reserve_avail_batch_packed(struct virtio_net *dev,
2808                                  struct vhost_virtqueue *vq,
2809                                  struct rte_mbuf **pkts,
2810                                  uint16_t avail_idx,
2811                                  uintptr_t *desc_addrs,
2812                                  uint16_t *ids)
2813 {
2814         bool wrap = vq->avail_wrap_counter;
2815         struct vring_packed_desc *descs = vq->desc_packed;
2816         uint64_t lens[PACKED_BATCH_SIZE];
2817         uint64_t buf_lens[PACKED_BATCH_SIZE];
2818         uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2819         uint16_t flags, i;
2820
2821         if (unlikely(avail_idx & PACKED_BATCH_MASK))
2822                 return -1;
2823         if (unlikely((avail_idx + PACKED_BATCH_SIZE) > vq->size))
2824                 return -1;
2825
2826         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2827                 flags = descs[avail_idx + i].flags;
2828                 if (unlikely((wrap != !!(flags & VRING_DESC_F_AVAIL)) ||
2829                              (wrap == !!(flags & VRING_DESC_F_USED))  ||
2830                              (flags & PACKED_DESC_SINGLE_DEQUEUE_FLAG)))
2831                         return -1;
2832         }
2833
2834         rte_atomic_thread_fence(__ATOMIC_ACQUIRE);
2835
2836         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2837                 lens[i] = descs[avail_idx + i].len;
2838
2839         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2840                 desc_addrs[i] = vhost_iova_to_vva(dev, vq,
2841                                                   descs[avail_idx + i].addr,
2842                                                   &lens[i], VHOST_ACCESS_RW);
2843         }
2844
2845         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2846                 if (unlikely(!desc_addrs[i]))
2847                         return -1;
2848                 if (unlikely((lens[i] != descs[avail_idx + i].len)))
2849                         return -1;
2850         }
2851
2852         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2853                 if (virtio_dev_pktmbuf_prep(dev, pkts[i], lens[i]))
2854                         goto err;
2855         }
2856
2857         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2858                 buf_lens[i] = pkts[i]->buf_len - pkts[i]->data_off;
2859
2860         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2861                 if (unlikely(buf_lens[i] < (lens[i] - buf_offset)))
2862                         goto err;
2863         }
2864
2865         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2866                 pkts[i]->pkt_len = lens[i] - buf_offset;
2867                 pkts[i]->data_len = pkts[i]->pkt_len;
2868                 ids[i] = descs[avail_idx + i].id;
2869         }
2870
2871         return 0;
2872
2873 err:
2874         return -1;
2875 }
2876
2877 static __rte_always_inline int
2878 virtio_dev_tx_batch_packed(struct virtio_net *dev,
2879                            struct vhost_virtqueue *vq,
2880                            struct rte_mbuf **pkts,
2881                            bool legacy_ol_flags)
2882 {
2883         uint16_t avail_idx = vq->last_avail_idx;
2884         uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2885         struct virtio_net_hdr *hdr;
2886         uintptr_t desc_addrs[PACKED_BATCH_SIZE];
2887         uint16_t ids[PACKED_BATCH_SIZE];
2888         uint16_t i;
2889
2890         if (vhost_reserve_avail_batch_packed(dev, vq, pkts, avail_idx,
2891                                              desc_addrs, ids))
2892                 return -1;
2893
2894         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2895                 rte_prefetch0((void *)(uintptr_t)desc_addrs[i]);
2896
2897         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2898                 rte_memcpy(rte_pktmbuf_mtod_offset(pkts[i], void *, 0),
2899                            (void *)(uintptr_t)(desc_addrs[i] + buf_offset),
2900                            pkts[i]->pkt_len);
2901
2902         if (virtio_net_with_host_offload(dev)) {
2903                 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2904                         hdr = (struct virtio_net_hdr *)(desc_addrs[i]);
2905                         vhost_dequeue_offload(hdr, pkts[i], legacy_ol_flags);
2906                 }
2907         }
2908
2909         if (virtio_net_is_inorder(dev))
2910                 vhost_shadow_dequeue_batch_packed_inorder(vq,
2911                         ids[PACKED_BATCH_SIZE - 1]);
2912         else
2913                 vhost_shadow_dequeue_batch_packed(dev, vq, ids);
2914
2915         vq_inc_last_avail_packed(vq, PACKED_BATCH_SIZE);
2916
2917         return 0;
2918 }
2919
2920 static __rte_always_inline int
2921 vhost_dequeue_single_packed(struct virtio_net *dev,
2922                             struct vhost_virtqueue *vq,
2923                             struct rte_mempool *mbuf_pool,
2924                             struct rte_mbuf *pkts,
2925                             uint16_t *buf_id,
2926                             uint16_t *desc_count,
2927                             bool legacy_ol_flags)
2928 {
2929         struct buf_vector buf_vec[BUF_VECTOR_MAX];
2930         uint32_t buf_len;
2931         uint16_t nr_vec = 0;
2932         int err;
2933         static bool allocerr_warned;
2934
2935         if (unlikely(fill_vec_buf_packed(dev, vq,
2936                                          vq->last_avail_idx, desc_count,
2937                                          buf_vec, &nr_vec,
2938                                          buf_id, &buf_len,
2939                                          VHOST_ACCESS_RO) < 0))
2940                 return -1;
2941
2942         if (unlikely(virtio_dev_pktmbuf_prep(dev, pkts, buf_len))) {
2943                 if (!allocerr_warned) {
2944                         VHOST_LOG_DATA(ERR,
2945                                 "Failed mbuf alloc of size %d from %s on %s.\n",
2946                                 buf_len, mbuf_pool->name, dev->ifname);
2947                         allocerr_warned = true;
2948                 }
2949                 return -1;
2950         }
2951
2952         err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts,
2953                                 mbuf_pool, legacy_ol_flags);
2954         if (unlikely(err)) {
2955                 if (!allocerr_warned) {
2956                         VHOST_LOG_DATA(ERR,
2957                                 "Failed to copy desc to mbuf on %s.\n",
2958                                 dev->ifname);
2959                         allocerr_warned = true;
2960                 }
2961                 return -1;
2962         }
2963
2964         return 0;
2965 }
2966
2967 static __rte_always_inline int
2968 virtio_dev_tx_single_packed(struct virtio_net *dev,
2969                             struct vhost_virtqueue *vq,
2970                             struct rte_mempool *mbuf_pool,
2971                             struct rte_mbuf *pkts,
2972                             bool legacy_ol_flags)
2973 {
2974
2975         uint16_t buf_id, desc_count = 0;
2976         int ret;
2977
2978         ret = vhost_dequeue_single_packed(dev, vq, mbuf_pool, pkts, &buf_id,
2979                                         &desc_count, legacy_ol_flags);
2980
2981         if (likely(desc_count > 0)) {
2982                 if (virtio_net_is_inorder(dev))
2983                         vhost_shadow_dequeue_single_packed_inorder(vq, buf_id,
2984                                                                    desc_count);
2985                 else
2986                         vhost_shadow_dequeue_single_packed(vq, buf_id,
2987                                         desc_count);
2988
2989                 vq_inc_last_avail_packed(vq, desc_count);
2990         }
2991
2992         return ret;
2993 }
2994
2995 __rte_always_inline
2996 static uint16_t
2997 virtio_dev_tx_packed(struct virtio_net *dev,
2998                      struct vhost_virtqueue *__rte_restrict vq,
2999                      struct rte_mempool *mbuf_pool,
3000                      struct rte_mbuf **__rte_restrict pkts,
3001                      uint32_t count,
3002                      bool legacy_ol_flags)
3003 {
3004         uint32_t pkt_idx = 0;
3005
3006         if (rte_pktmbuf_alloc_bulk(mbuf_pool, pkts, count))
3007                 return 0;
3008
3009         do {
3010                 rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
3011
3012                 if (count - pkt_idx >= PACKED_BATCH_SIZE) {
3013                         if (!virtio_dev_tx_batch_packed(dev, vq,
3014                                                         &pkts[pkt_idx],
3015                                                         legacy_ol_flags)) {
3016                                 pkt_idx += PACKED_BATCH_SIZE;
3017                                 continue;
3018                         }
3019                 }
3020
3021                 if (virtio_dev_tx_single_packed(dev, vq, mbuf_pool,
3022                                                 pkts[pkt_idx],
3023                                                 legacy_ol_flags))
3024                         break;
3025                 pkt_idx++;
3026         } while (pkt_idx < count);
3027
3028         if (pkt_idx != count)
3029                 rte_pktmbuf_free_bulk(&pkts[pkt_idx], count - pkt_idx);
3030
3031         if (vq->shadow_used_idx) {
3032                 do_data_copy_dequeue(vq);
3033
3034                 vhost_flush_dequeue_shadow_packed(dev, vq);
3035                 vhost_vring_call_packed(dev, vq);
3036         }
3037
3038         return pkt_idx;
3039 }
3040
3041 __rte_noinline
3042 static uint16_t
3043 virtio_dev_tx_packed_legacy(struct virtio_net *dev,
3044         struct vhost_virtqueue *__rte_restrict vq, struct rte_mempool *mbuf_pool,
3045         struct rte_mbuf **__rte_restrict pkts, uint32_t count)
3046 {
3047         return virtio_dev_tx_packed(dev, vq, mbuf_pool, pkts, count, true);
3048 }
3049
3050 __rte_noinline
3051 static uint16_t
3052 virtio_dev_tx_packed_compliant(struct virtio_net *dev,
3053         struct vhost_virtqueue *__rte_restrict vq, struct rte_mempool *mbuf_pool,
3054         struct rte_mbuf **__rte_restrict pkts, uint32_t count)
3055 {
3056         return virtio_dev_tx_packed(dev, vq, mbuf_pool, pkts, count, false);
3057 }
3058
3059 uint16_t
3060 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
3061         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
3062 {
3063         struct virtio_net *dev;
3064         struct rte_mbuf *rarp_mbuf = NULL;
3065         struct vhost_virtqueue *vq;
3066         int16_t success = 1;
3067
3068         dev = get_device(vid);
3069         if (!dev)
3070                 return 0;
3071
3072         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
3073                 VHOST_LOG_DATA(ERR,
3074                         "(%d) %s: built-in vhost net backend is disabled.\n",
3075                         dev->vid, __func__);
3076                 return 0;
3077         }
3078
3079         if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
3080                 VHOST_LOG_DATA(ERR,
3081                         "(%d) %s: invalid virtqueue idx %d.\n",
3082                         dev->vid, __func__, queue_id);
3083                 return 0;
3084         }
3085
3086         vq = dev->virtqueue[queue_id];
3087
3088         if (unlikely(rte_spinlock_trylock(&vq->access_lock) == 0))
3089                 return 0;
3090
3091         if (unlikely(!vq->enabled)) {
3092                 count = 0;
3093                 goto out_access_unlock;
3094         }
3095
3096         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
3097                 vhost_user_iotlb_rd_lock(vq);
3098
3099         if (unlikely(!vq->access_ok))
3100                 if (unlikely(vring_translate(dev, vq) < 0)) {
3101                         count = 0;
3102                         goto out;
3103                 }
3104
3105         /*
3106          * Construct a RARP broadcast packet, and inject it to the "pkts"
3107          * array, to looks like that guest actually send such packet.
3108          *
3109          * Check user_send_rarp() for more information.
3110          *
3111          * broadcast_rarp shares a cacheline in the virtio_net structure
3112          * with some fields that are accessed during enqueue and
3113          * __atomic_compare_exchange_n causes a write if performed compare
3114          * and exchange. This could result in false sharing between enqueue
3115          * and dequeue.
3116          *
3117          * Prevent unnecessary false sharing by reading broadcast_rarp first
3118          * and only performing compare and exchange if the read indicates it
3119          * is likely to be set.
3120          */
3121         if (unlikely(__atomic_load_n(&dev->broadcast_rarp, __ATOMIC_ACQUIRE) &&
3122                         __atomic_compare_exchange_n(&dev->broadcast_rarp,
3123                         &success, 0, 0, __ATOMIC_RELEASE, __ATOMIC_RELAXED))) {
3124
3125                 rarp_mbuf = rte_net_make_rarp_packet(mbuf_pool, &dev->mac);
3126                 if (rarp_mbuf == NULL) {
3127                         VHOST_LOG_DATA(ERR, "Failed to make RARP packet.\n");
3128                         count = 0;
3129                         goto out;
3130                 }
3131                 count -= 1;
3132         }
3133
3134         if (vq_is_packed(dev)) {
3135                 if (dev->flags & VIRTIO_DEV_LEGACY_OL_FLAGS)
3136                         count = virtio_dev_tx_packed_legacy(dev, vq, mbuf_pool, pkts, count);
3137                 else
3138                         count = virtio_dev_tx_packed_compliant(dev, vq, mbuf_pool, pkts, count);
3139         } else {
3140                 if (dev->flags & VIRTIO_DEV_LEGACY_OL_FLAGS)
3141                         count = virtio_dev_tx_split_legacy(dev, vq, mbuf_pool, pkts, count);
3142                 else
3143                         count = virtio_dev_tx_split_compliant(dev, vq, mbuf_pool, pkts, count);
3144         }
3145
3146 out:
3147         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
3148                 vhost_user_iotlb_rd_unlock(vq);
3149
3150 out_access_unlock:
3151         rte_spinlock_unlock(&vq->access_lock);
3152
3153         if (unlikely(rarp_mbuf != NULL)) {
3154                 /*
3155                  * Inject it to the head of "pkts" array, so that switch's mac
3156                  * learning table will get updated first.
3157                  */
3158                 memmove(&pkts[1], pkts, count * sizeof(struct rte_mbuf *));
3159                 pkts[0] = rarp_mbuf;
3160                 count += 1;
3161         }
3162
3163         return count;
3164 }