lib/vhost/virtio_net.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2016 Intel Corporation
   3  */
   4
   5 #include <stdint.h>
   6 #include <stdbool.h>
   7 #include <linux/virtio_net.h>
   8
   9 #include <rte_mbuf.h>
  10 #include <rte_memcpy.h>
  11 #include <rte_net.h>
  12 #include <rte_ether.h>
  13 #include <rte_ip.h>
  14 #include <rte_vhost.h>
  15 #include <rte_tcp.h>
  16 #include <rte_udp.h>
  17 #include <rte_sctp.h>
  18 #include <rte_arp.h>
  19 #include <rte_spinlock.h>
  20 #include <rte_malloc.h>
  21 #include <rte_vhost_async.h>
  22
  23 #include "iotlb.h"
  24 #include "vhost.h"
  25
  26 #define MAX_BATCH_LEN 256
  27
  28 #define VHOST_ASYNC_BATCH_THRESHOLD 32
  29
  30 static  __rte_always_inline bool
  31 rxvq_is_mergeable(struct virtio_net *dev)
  32 {
  33         return dev->features & (1ULL << VIRTIO_NET_F_MRG_RXBUF);
  34 }
  35
  36 static  __rte_always_inline bool
  37 virtio_net_is_inorder(struct virtio_net *dev)
  38 {
  39         return dev->features & (1ULL << VIRTIO_F_IN_ORDER);
  40 }
  41
  42 static bool
  43 is_valid_virt_queue_idx(uint32_t idx, int is_tx, uint32_t nr_vring)
  44 {
  45         return (is_tx ^ (idx & 1)) == 0 && idx < nr_vring;
  46 }
  47
  48 static inline void
  49 do_data_copy_enqueue(struct virtio_net *dev, struct vhost_virtqueue *vq)
  50 {
  51         struct batch_copy_elem *elem = vq->batch_copy_elems;
  52         uint16_t count = vq->batch_copy_nb_elems;
  53         int i;
  54
  55         for (i = 0; i < count; i++) {
  56                 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
  57                 vhost_log_cache_write_iova(dev, vq, elem[i].log_addr,
  58                                            elem[i].len);
  59                 PRINT_PACKET(dev, (uintptr_t)elem[i].dst, elem[i].len, 0);
  60         }
  61
  62         vq->batch_copy_nb_elems = 0;
  63 }
  64
  65 static inline void
  66 do_data_copy_dequeue(struct vhost_virtqueue *vq)
  67 {
  68         struct batch_copy_elem *elem = vq->batch_copy_elems;
  69         uint16_t count = vq->batch_copy_nb_elems;
  70         int i;
  71
  72         for (i = 0; i < count; i++)
  73                 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
  74
  75         vq->batch_copy_nb_elems = 0;
  76 }
  77
  78 static __rte_always_inline void
  79 do_flush_shadow_used_ring_split(struct virtio_net *dev,
  80                         struct vhost_virtqueue *vq,
  81                         uint16_t to, uint16_t from, uint16_t size)
  82 {
  83         rte_memcpy(&vq->used->ring[to],
  84                         &vq->shadow_used_split[from],
  85                         size * sizeof(struct vring_used_elem));
  86         vhost_log_cache_used_vring(dev, vq,
  87                         offsetof(struct vring_used, ring[to]),
  88                         size * sizeof(struct vring_used_elem));
  89 }
  90
  91 static __rte_always_inline void
  92 flush_shadow_used_ring_split(struct virtio_net *dev, struct vhost_virtqueue *vq)
  93 {
  94         uint16_t used_idx = vq->last_used_idx & (vq->size - 1);
  95
  96         if (used_idx + vq->shadow_used_idx <= vq->size) {
  97                 do_flush_shadow_used_ring_split(dev, vq, used_idx, 0,
  98                                           vq->shadow_used_idx);
  99         } else {
 100                 uint16_t size;
 101
 102                 /* update used ring interval [used_idx, vq->size] */
 103                 size = vq->size - used_idx;
 104                 do_flush_shadow_used_ring_split(dev, vq, used_idx, 0, size);
 105
 106                 /* update the left half used ring interval [0, left_size] */
 107                 do_flush_shadow_used_ring_split(dev, vq, 0, size,
 108                                           vq->shadow_used_idx - size);
 109         }
 110         vq->last_used_idx += vq->shadow_used_idx;
 111
 112         vhost_log_cache_sync(dev, vq);
 113
 114         __atomic_add_fetch(&vq->used->idx, vq->shadow_used_idx,
 115                            __ATOMIC_RELEASE);
 116         vq->shadow_used_idx = 0;
 117         vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
 118                 sizeof(vq->used->idx));
 119 }
 120
 121 static __rte_always_inline void
 122 update_shadow_used_ring_split(struct vhost_virtqueue *vq,
 123                          uint16_t desc_idx, uint32_t len)
 124 {
 125         uint16_t i = vq->shadow_used_idx++;
 126
 127         vq->shadow_used_split[i].id  = desc_idx;
 128         vq->shadow_used_split[i].len = len;
 129 }
 130
 131 static __rte_always_inline void
 132 vhost_flush_enqueue_shadow_packed(struct virtio_net *dev,
 133                                   struct vhost_virtqueue *vq)
 134 {
 135         int i;
 136         uint16_t used_idx = vq->last_used_idx;
 137         uint16_t head_idx = vq->last_used_idx;
 138         uint16_t head_flags = 0;
 139
 140         /* Split loop in two to save memory barriers */
 141         for (i = 0; i < vq->shadow_used_idx; i++) {
 142                 vq->desc_packed[used_idx].id = vq->shadow_used_packed[i].id;
 143                 vq->desc_packed[used_idx].len = vq->shadow_used_packed[i].len;
 144
 145                 used_idx += vq->shadow_used_packed[i].count;
 146                 if (used_idx >= vq->size)
 147                         used_idx -= vq->size;
 148         }
 149
 150         /* The ordering for storing desc flags needs to be enforced. */
 151         rte_atomic_thread_fence(__ATOMIC_RELEASE);
 152
 153         for (i = 0; i < vq->shadow_used_idx; i++) {
 154                 uint16_t flags;
 155
 156                 if (vq->shadow_used_packed[i].len)
 157                         flags = VRING_DESC_F_WRITE;
 158                 else
 159                         flags = 0;
 160
 161                 if (vq->used_wrap_counter) {
 162                         flags |= VRING_DESC_F_USED;
 163                         flags |= VRING_DESC_F_AVAIL;
 164                 } else {
 165                         flags &= ~VRING_DESC_F_USED;
 166                         flags &= ~VRING_DESC_F_AVAIL;
 167                 }
 168
 169                 if (i > 0) {
 170                         vq->desc_packed[vq->last_used_idx].flags = flags;
 171
 172                         vhost_log_cache_used_vring(dev, vq,
 173                                         vq->last_used_idx *
 174                                         sizeof(struct vring_packed_desc),
 175                                         sizeof(struct vring_packed_desc));
 176                 } else {
 177                         head_idx = vq->last_used_idx;
 178                         head_flags = flags;
 179                 }
 180
 181                 vq_inc_last_used_packed(vq, vq->shadow_used_packed[i].count);
 182         }
 183
 184         vq->desc_packed[head_idx].flags = head_flags;
 185
 186         vhost_log_cache_used_vring(dev, vq,
 187                                 head_idx *
 188                                 sizeof(struct vring_packed_desc),
 189                                 sizeof(struct vring_packed_desc));
 190
 191         vq->shadow_used_idx = 0;
 192         vhost_log_cache_sync(dev, vq);
 193 }
 194
 195 static __rte_always_inline void
 196 vhost_flush_dequeue_shadow_packed(struct virtio_net *dev,
 197                                   struct vhost_virtqueue *vq)
 198 {
 199         struct vring_used_elem_packed *used_elem = &vq->shadow_used_packed[0];
 200
 201         vq->desc_packed[vq->shadow_last_used_idx].id = used_elem->id;
 202         /* desc flags is the synchronization point for virtio packed vring */
 203         __atomic_store_n(&vq->desc_packed[vq->shadow_last_used_idx].flags,
 204                          used_elem->flags, __ATOMIC_RELEASE);
 205
 206         vhost_log_cache_used_vring(dev, vq, vq->shadow_last_used_idx *
 207                                    sizeof(struct vring_packed_desc),
 208                                    sizeof(struct vring_packed_desc));
 209         vq->shadow_used_idx = 0;
 210         vhost_log_cache_sync(dev, vq);
 211 }
 212
 213 static __rte_always_inline void
 214 vhost_flush_enqueue_batch_packed(struct virtio_net *dev,
 215                                  struct vhost_virtqueue *vq,
 216                                  uint64_t *lens,
 217                                  uint16_t *ids)
 218 {
 219         uint16_t i;
 220         uint16_t flags;
 221         uint16_t last_used_idx;
 222         struct vring_packed_desc *desc_base;
 223
 224         last_used_idx = vq->last_used_idx;
 225         desc_base = &vq->desc_packed[last_used_idx];
 226
 227         flags = PACKED_DESC_ENQUEUE_USED_FLAG(vq->used_wrap_counter);
 228
 229         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
 230                 desc_base[i].id = ids[i];
 231                 desc_base[i].len = lens[i];
 232         }
 233
 234         rte_atomic_thread_fence(__ATOMIC_RELEASE);
 235
 236         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
 237                 desc_base[i].flags = flags;
 238         }
 239
 240         vhost_log_cache_used_vring(dev, vq, last_used_idx *
 241                                    sizeof(struct vring_packed_desc),
 242                                    sizeof(struct vring_packed_desc) *
 243                                    PACKED_BATCH_SIZE);
 244         vhost_log_cache_sync(dev, vq);
 245
 246         vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
 247 }
 248
 249 static __rte_always_inline void
 250 vhost_shadow_dequeue_batch_packed_inorder(struct vhost_virtqueue *vq,
 251                                           uint16_t id)
 252 {
 253         vq->shadow_used_packed[0].id = id;
 254
 255         if (!vq->shadow_used_idx) {
 256                 vq->shadow_last_used_idx = vq->last_used_idx;
 257                 vq->shadow_used_packed[0].flags =
 258                         PACKED_DESC_DEQUEUE_USED_FLAG(vq->used_wrap_counter);
 259                 vq->shadow_used_packed[0].len = 0;
 260                 vq->shadow_used_packed[0].count = 1;
 261                 vq->shadow_used_idx++;
 262         }
 263
 264         vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
 265 }
 266
 267 static __rte_always_inline void
 268 vhost_shadow_dequeue_batch_packed(struct virtio_net *dev,
 269                                   struct vhost_virtqueue *vq,
 270                                   uint16_t *ids)
 271 {
 272         uint16_t flags;
 273         uint16_t i;
 274         uint16_t begin;
 275
 276         flags = PACKED_DESC_DEQUEUE_USED_FLAG(vq->used_wrap_counter);
 277
 278         if (!vq->shadow_used_idx) {
 279                 vq->shadow_last_used_idx = vq->last_used_idx;
 280                 vq->shadow_used_packed[0].id  = ids[0];
 281                 vq->shadow_used_packed[0].len = 0;
 282                 vq->shadow_used_packed[0].count = 1;
 283                 vq->shadow_used_packed[0].flags = flags;
 284                 vq->shadow_used_idx++;
 285                 begin = 1;
 286         } else
 287                 begin = 0;
 288
 289         vhost_for_each_try_unroll(i, begin, PACKED_BATCH_SIZE) {
 290                 vq->desc_packed[vq->last_used_idx + i].id = ids[i];
 291                 vq->desc_packed[vq->last_used_idx + i].len = 0;
 292         }
 293
 294         rte_atomic_thread_fence(__ATOMIC_RELEASE);
 295         vhost_for_each_try_unroll(i, begin, PACKED_BATCH_SIZE)
 296                 vq->desc_packed[vq->last_used_idx + i].flags = flags;
 297
 298         vhost_log_cache_used_vring(dev, vq, vq->last_used_idx *
 299                                    sizeof(struct vring_packed_desc),
 300                                    sizeof(struct vring_packed_desc) *
 301                                    PACKED_BATCH_SIZE);
 302         vhost_log_cache_sync(dev, vq);
 303
 304         vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
 305 }
 306
 307 static __rte_always_inline void
 308 vhost_shadow_dequeue_single_packed(struct vhost_virtqueue *vq,
 309                                    uint16_t buf_id,
 310                                    uint16_t count)
 311 {
 312         uint16_t flags;
 313
 314         flags = vq->desc_packed[vq->last_used_idx].flags;
 315         if (vq->used_wrap_counter) {
 316                 flags |= VRING_DESC_F_USED;
 317                 flags |= VRING_DESC_F_AVAIL;
 318         } else {
 319                 flags &= ~VRING_DESC_F_USED;
 320                 flags &= ~VRING_DESC_F_AVAIL;
 321         }
 322
 323         if (!vq->shadow_used_idx) {
 324                 vq->shadow_last_used_idx = vq->last_used_idx;
 325
 326                 vq->shadow_used_packed[0].id  = buf_id;
 327                 vq->shadow_used_packed[0].len = 0;
 328                 vq->shadow_used_packed[0].flags = flags;
 329                 vq->shadow_used_idx++;
 330         } else {
 331                 vq->desc_packed[vq->last_used_idx].id = buf_id;
 332                 vq->desc_packed[vq->last_used_idx].len = 0;
 333                 vq->desc_packed[vq->last_used_idx].flags = flags;
 334         }
 335
 336         vq_inc_last_used_packed(vq, count);
 337 }
 338
 339 static __rte_always_inline void
 340 vhost_shadow_dequeue_single_packed_inorder(struct vhost_virtqueue *vq,
 341                                            uint16_t buf_id,
 342                                            uint16_t count)
 343 {
 344         uint16_t flags;
 345
 346         vq->shadow_used_packed[0].id = buf_id;
 347
 348         flags = vq->desc_packed[vq->last_used_idx].flags;
 349         if (vq->used_wrap_counter) {
 350                 flags |= VRING_DESC_F_USED;
 351                 flags |= VRING_DESC_F_AVAIL;
 352         } else {
 353                 flags &= ~VRING_DESC_F_USED;
 354                 flags &= ~VRING_DESC_F_AVAIL;
 355         }
 356
 357         if (!vq->shadow_used_idx) {
 358                 vq->shadow_last_used_idx = vq->last_used_idx;
 359                 vq->shadow_used_packed[0].len = 0;
 360                 vq->shadow_used_packed[0].flags = flags;
 361                 vq->shadow_used_idx++;
 362         }
 363
 364         vq_inc_last_used_packed(vq, count);
 365 }
 366
 367 static __rte_always_inline void
 368 vhost_shadow_enqueue_packed(struct vhost_virtqueue *vq,
 369                                    uint32_t *len,
 370                                    uint16_t *id,
 371                                    uint16_t *count,
 372                                    uint16_t num_buffers)
 373 {
 374         uint16_t i;
 375
 376         for (i = 0; i < num_buffers; i++) {
 377                 /* enqueue shadow flush action aligned with batch num */
 378                 if (!vq->shadow_used_idx)
 379                         vq->shadow_aligned_idx = vq->last_used_idx &
 380                                 PACKED_BATCH_MASK;
 381                 vq->shadow_used_packed[vq->shadow_used_idx].id  = id[i];
 382                 vq->shadow_used_packed[vq->shadow_used_idx].len = len[i];
 383                 vq->shadow_used_packed[vq->shadow_used_idx].count = count[i];
 384                 vq->shadow_aligned_idx += count[i];
 385                 vq->shadow_used_idx++;
 386         }
 387 }
 388
 389 static __rte_always_inline void
 390 vhost_shadow_enqueue_single_packed(struct virtio_net *dev,
 391                                    struct vhost_virtqueue *vq,
 392                                    uint32_t *len,
 393                                    uint16_t *id,
 394                                    uint16_t *count,
 395                                    uint16_t num_buffers)
 396 {
 397         vhost_shadow_enqueue_packed(vq, len, id, count, num_buffers);
 398
 399         if (vq->shadow_aligned_idx >= PACKED_BATCH_SIZE) {
 400                 do_data_copy_enqueue(dev, vq);
 401                 vhost_flush_enqueue_shadow_packed(dev, vq);
 402         }
 403 }
 404
 405 /* avoid write operation when necessary, to lessen cache issues */
 406 #define ASSIGN_UNLESS_EQUAL(var, val) do {      \
 407         if ((var) != (val))                     \
 408                 (var) = (val);                  \
 409 } while (0)
 410
 411 static __rte_always_inline void
 412 virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
 413 {
 414         uint64_t csum_l4 = m_buf->ol_flags & RTE_MBUF_F_TX_L4_MASK;
 415
 416         if (m_buf->ol_flags & RTE_MBUF_F_TX_TCP_SEG)
 417                 csum_l4 |= RTE_MBUF_F_TX_TCP_CKSUM;
 418
 419         if (csum_l4) {
 420                 net_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 421                 net_hdr->csum_start = m_buf->l2_len + m_buf->l3_len;
 422
 423                 switch (csum_l4) {
 424                 case RTE_MBUF_F_TX_TCP_CKSUM:
 425                         net_hdr->csum_offset = (offsetof(struct rte_tcp_hdr,
 426                                                 cksum));
 427                         break;
 428                 case RTE_MBUF_F_TX_UDP_CKSUM:
 429                         net_hdr->csum_offset = (offsetof(struct rte_udp_hdr,
 430                                                 dgram_cksum));
 431                         break;
 432                 case RTE_MBUF_F_TX_SCTP_CKSUM:
 433                         net_hdr->csum_offset = (offsetof(struct rte_sctp_hdr,
 434                                                 cksum));
 435                         break;
 436                 }
 437         } else {
 438                 ASSIGN_UNLESS_EQUAL(net_hdr->csum_start, 0);
 439                 ASSIGN_UNLESS_EQUAL(net_hdr->csum_offset, 0);
 440                 ASSIGN_UNLESS_EQUAL(net_hdr->flags, 0);
 441         }
 442
 443         /* IP cksum verification cannot be bypassed, then calculate here */
 444         if (m_buf->ol_flags & RTE_MBUF_F_TX_IP_CKSUM) {
 445                 struct rte_ipv4_hdr *ipv4_hdr;
 446
 447                 ipv4_hdr = rte_pktmbuf_mtod_offset(m_buf, struct rte_ipv4_hdr *,
 448                                                    m_buf->l2_len);
 449                 ipv4_hdr->hdr_checksum = 0;
 450                 ipv4_hdr->hdr_checksum = rte_ipv4_cksum(ipv4_hdr);
 451         }
 452
 453         if (m_buf->ol_flags & RTE_MBUF_F_TX_TCP_SEG) {
 454                 if (m_buf->ol_flags & RTE_MBUF_F_TX_IPV4)
 455                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 456                 else
 457                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 458                 net_hdr->gso_size = m_buf->tso_segsz;
 459                 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len
 460                                         + m_buf->l4_len;
 461         } else if (m_buf->ol_flags & RTE_MBUF_F_TX_UDP_SEG) {
 462                 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
 463                 net_hdr->gso_size = m_buf->tso_segsz;
 464                 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len +
 465                         m_buf->l4_len;
 466         } else {
 467                 ASSIGN_UNLESS_EQUAL(net_hdr->gso_type, 0);
 468                 ASSIGN_UNLESS_EQUAL(net_hdr->gso_size, 0);
 469                 ASSIGN_UNLESS_EQUAL(net_hdr->hdr_len, 0);
 470         }
 471 }
 472
 473 static __rte_always_inline int
 474 map_one_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 475                 struct buf_vector *buf_vec, uint16_t *vec_idx,
 476                 uint64_t desc_iova, uint64_t desc_len, uint8_t perm)
 477 {
 478         uint16_t vec_id = *vec_idx;
 479
 480         while (desc_len) {
 481                 uint64_t desc_addr;
 482                 uint64_t desc_chunck_len = desc_len;
 483
 484                 if (unlikely(vec_id >= BUF_VECTOR_MAX))
 485                         return -1;
 486
 487                 desc_addr = vhost_iova_to_vva(dev, vq,
 488                                 desc_iova,
 489                                 &desc_chunck_len,
 490                                 perm);
 491                 if (unlikely(!desc_addr))
 492                         return -1;
 493
 494                 rte_prefetch0((void *)(uintptr_t)desc_addr);
 495
 496                 buf_vec[vec_id].buf_iova = desc_iova;
 497                 buf_vec[vec_id].buf_addr = desc_addr;
 498                 buf_vec[vec_id].buf_len  = desc_chunck_len;
 499
 500                 desc_len -= desc_chunck_len;
 501                 desc_iova += desc_chunck_len;
 502                 vec_id++;
 503         }
 504         *vec_idx = vec_id;
 505
 506         return 0;
 507 }
 508
 509 static __rte_always_inline int
 510 fill_vec_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 511                          uint32_t avail_idx, uint16_t *vec_idx,
 512                          struct buf_vector *buf_vec, uint16_t *desc_chain_head,
 513                          uint32_t *desc_chain_len, uint8_t perm)
 514 {
 515         uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
 516         uint16_t vec_id = *vec_idx;
 517         uint32_t len    = 0;
 518         uint64_t dlen;
 519         uint32_t nr_descs = vq->size;
 520         uint32_t cnt    = 0;
 521         struct vring_desc *descs = vq->desc;
 522         struct vring_desc *idesc = NULL;
 523
 524         if (unlikely(idx >= vq->size))
 525                 return -1;
 526
 527         *desc_chain_head = idx;
 528
 529         if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
 530                 dlen = vq->desc[idx].len;
 531                 nr_descs = dlen / sizeof(struct vring_desc);
 532                 if (unlikely(nr_descs > vq->size))
 533                         return -1;
 534
 535                 descs = (struct vring_desc *)(uintptr_t)
 536                         vhost_iova_to_vva(dev, vq, vq->desc[idx].addr,
 537                                                 &dlen,
 538                                                 VHOST_ACCESS_RO);
 539                 if (unlikely(!descs))
 540                         return -1;
 541
 542                 if (unlikely(dlen < vq->desc[idx].len)) {
 543                         /*
 544                          * The indirect desc table is not contiguous
 545                          * in process VA space, we have to copy it.
 546                          */
 547                         idesc = vhost_alloc_copy_ind_table(dev, vq,
 548                                         vq->desc[idx].addr, vq->desc[idx].len);
 549                         if (unlikely(!idesc))
 550                                 return -1;
 551
 552                         descs = idesc;
 553                 }
 554
 555                 idx = 0;
 556         }
 557
 558         while (1) {
 559                 if (unlikely(idx >= nr_descs || cnt++ >= nr_descs)) {
 560                         free_ind_table(idesc);
 561                         return -1;
 562                 }
 563
 564                 dlen = descs[idx].len;
 565                 len += dlen;
 566
 567                 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 568                                                 descs[idx].addr, dlen,
 569                                                 perm))) {
 570                         free_ind_table(idesc);
 571                         return -1;
 572                 }
 573
 574                 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
 575                         break;
 576
 577                 idx = descs[idx].next;
 578         }
 579
 580         *desc_chain_len = len;
 581         *vec_idx = vec_id;
 582
 583         if (unlikely(!!idesc))
 584                 free_ind_table(idesc);
 585
 586         return 0;
 587 }
 588
 589 /*
 590  * Returns -1 on fail, 0 on success
 591  */
 592 static inline int
 593 reserve_avail_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 594                                 uint32_t size, struct buf_vector *buf_vec,
 595                                 uint16_t *num_buffers, uint16_t avail_head,
 596                                 uint16_t *nr_vec)
 597 {
 598         uint16_t cur_idx;
 599         uint16_t vec_idx = 0;
 600         uint16_t max_tries, tries = 0;
 601
 602         uint16_t head_idx = 0;
 603         uint32_t len = 0;
 604
 605         *num_buffers = 0;
 606         cur_idx  = vq->last_avail_idx;
 607
 608         if (rxvq_is_mergeable(dev))
 609                 max_tries = vq->size - 1;
 610         else
 611                 max_tries = 1;
 612
 613         while (size > 0) {
 614                 if (unlikely(cur_idx == avail_head))
 615                         return -1;
 616                 /*
 617                  * if we tried all available ring items, and still
 618                  * can't get enough buf, it means something abnormal
 619                  * happened.
 620                  */
 621                 if (unlikely(++tries > max_tries))
 622                         return -1;
 623
 624                 if (unlikely(fill_vec_buf_split(dev, vq, cur_idx,
 625                                                 &vec_idx, buf_vec,
 626                                                 &head_idx, &len,
 627                                                 VHOST_ACCESS_RW) < 0))
 628                         return -1;
 629                 len = RTE_MIN(len, size);
 630                 update_shadow_used_ring_split(vq, head_idx, len);
 631                 size -= len;
 632
 633                 cur_idx++;
 634                 *num_buffers += 1;
 635         }
 636
 637         *nr_vec = vec_idx;
 638
 639         return 0;
 640 }
 641
 642 static __rte_always_inline int
 643 fill_vec_buf_packed_indirect(struct virtio_net *dev,
 644                         struct vhost_virtqueue *vq,
 645                         struct vring_packed_desc *desc, uint16_t *vec_idx,
 646                         struct buf_vector *buf_vec, uint32_t *len, uint8_t perm)
 647 {
 648         uint16_t i;
 649         uint32_t nr_descs;
 650         uint16_t vec_id = *vec_idx;
 651         uint64_t dlen;
 652         struct vring_packed_desc *descs, *idescs = NULL;
 653
 654         dlen = desc->len;
 655         descs = (struct vring_packed_desc *)(uintptr_t)
 656                 vhost_iova_to_vva(dev, vq, desc->addr, &dlen, VHOST_ACCESS_RO);
 657         if (unlikely(!descs))
 658                 return -1;
 659
 660         if (unlikely(dlen < desc->len)) {
 661                 /*
 662                  * The indirect desc table is not contiguous
 663                  * in process VA space, we have to copy it.
 664                  */
 665                 idescs = vhost_alloc_copy_ind_table(dev,
 666                                 vq, desc->addr, desc->len);
 667                 if (unlikely(!idescs))
 668                         return -1;
 669
 670                 descs = idescs;
 671         }
 672
 673         nr_descs =  desc->len / sizeof(struct vring_packed_desc);
 674         if (unlikely(nr_descs >= vq->size)) {
 675                 free_ind_table(idescs);
 676                 return -1;
 677         }
 678
 679         for (i = 0; i < nr_descs; i++) {
 680                 if (unlikely(vec_id >= BUF_VECTOR_MAX)) {
 681                         free_ind_table(idescs);
 682                         return -1;
 683                 }
 684
 685                 dlen = descs[i].len;
 686                 *len += dlen;
 687                 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 688                                                 descs[i].addr, dlen,
 689                                                 perm)))
 690                         return -1;
 691         }
 692         *vec_idx = vec_id;
 693
 694         if (unlikely(!!idescs))
 695                 free_ind_table(idescs);
 696
 697         return 0;
 698 }
 699
 700 static __rte_always_inline int
 701 fill_vec_buf_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
 702                                 uint16_t avail_idx, uint16_t *desc_count,
 703                                 struct buf_vector *buf_vec, uint16_t *vec_idx,
 704                                 uint16_t *buf_id, uint32_t *len, uint8_t perm)
 705 {
 706         bool wrap_counter = vq->avail_wrap_counter;
 707         struct vring_packed_desc *descs = vq->desc_packed;
 708         uint16_t vec_id = *vec_idx;
 709         uint64_t dlen;
 710
 711         if (avail_idx < vq->last_avail_idx)
 712                 wrap_counter ^= 1;
 713
 714         /*
 715          * Perform a load-acquire barrier in desc_is_avail to
 716          * enforce the ordering between desc flags and desc
 717          * content.
 718          */
 719         if (unlikely(!desc_is_avail(&descs[avail_idx], wrap_counter)))
 720                 return -1;
 721
 722         *desc_count = 0;
 723         *len = 0;
 724
 725         while (1) {
 726                 if (unlikely(vec_id >= BUF_VECTOR_MAX))
 727                         return -1;
 728
 729                 if (unlikely(*desc_count >= vq->size))
 730                         return -1;
 731
 732                 *desc_count += 1;
 733                 *buf_id = descs[avail_idx].id;
 734
 735                 if (descs[avail_idx].flags & VRING_DESC_F_INDIRECT) {
 736                         if (unlikely(fill_vec_buf_packed_indirect(dev, vq,
 737                                                         &descs[avail_idx],
 738                                                         &vec_id, buf_vec,
 739                                                         len, perm) < 0))
 740                                 return -1;
 741                 } else {
 742                         dlen = descs[avail_idx].len;
 743                         *len += dlen;
 744
 745                         if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 746                                                         descs[avail_idx].addr,
 747                                                         dlen,
 748                                                         perm)))
 749                                 return -1;
 750                 }
 751
 752                 if ((descs[avail_idx].flags & VRING_DESC_F_NEXT) == 0)
 753                         break;
 754
 755                 if (++avail_idx >= vq->size) {
 756                         avail_idx -= vq->size;
 757                         wrap_counter ^= 1;
 758                 }
 759         }
 760
 761         *vec_idx = vec_id;
 762
 763         return 0;
 764 }
 765
 766 static __rte_noinline void
 767 copy_vnet_hdr_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 768                 struct buf_vector *buf_vec,
 769                 struct virtio_net_hdr_mrg_rxbuf *hdr)
 770 {
 771         uint64_t len;
 772         uint64_t remain = dev->vhost_hlen;
 773         uint64_t src = (uint64_t)(uintptr_t)hdr, dst;
 774         uint64_t iova = buf_vec->buf_iova;
 775
 776         while (remain) {
 777                 len = RTE_MIN(remain,
 778                                 buf_vec->buf_len);
 779                 dst = buf_vec->buf_addr;
 780                 rte_memcpy((void *)(uintptr_t)dst,
 781                                 (void *)(uintptr_t)src,
 782                                 len);
 783
 784                 PRINT_PACKET(dev, (uintptr_t)dst,
 785                                 (uint32_t)len, 0);
 786                 vhost_log_cache_write_iova(dev, vq,
 787                                 iova, len);
 788
 789                 remain -= len;
 790                 iova += len;
 791                 src += len;
 792                 buf_vec++;
 793         }
 794 }
 795
 796 static __rte_always_inline int
 797 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 798                             struct rte_mbuf *m, struct buf_vector *buf_vec,
 799                             uint16_t nr_vec, uint16_t num_buffers)
 800 {
 801         uint32_t vec_idx = 0;
 802         uint32_t mbuf_offset, mbuf_avail;
 803         uint32_t buf_offset, buf_avail;
 804         uint64_t buf_addr, buf_iova, buf_len;
 805         uint32_t cpy_len;
 806         uint64_t hdr_addr;
 807         struct rte_mbuf *hdr_mbuf;
 808         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 809         struct virtio_net_hdr_mrg_rxbuf tmp_hdr, *hdr = NULL;
 810         int error = 0;
 811
 812         if (unlikely(m == NULL)) {
 813                 error = -1;
 814                 goto out;
 815         }
 816
 817         buf_addr = buf_vec[vec_idx].buf_addr;
 818         buf_iova = buf_vec[vec_idx].buf_iova;
 819         buf_len = buf_vec[vec_idx].buf_len;
 820
 821         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
 822                 error = -1;
 823                 goto out;
 824         }
 825
 826         hdr_mbuf = m;
 827         hdr_addr = buf_addr;
 828         if (unlikely(buf_len < dev->vhost_hlen)) {
 829                 memset(&tmp_hdr, 0, sizeof(struct virtio_net_hdr_mrg_rxbuf));
 830                 hdr = &tmp_hdr;
 831         } else
 832                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)hdr_addr;
 833
 834         VHOST_LOG_DATA(DEBUG, "(%d) RX: num merge buffers %d\n",
 835                 dev->vid, num_buffers);
 836
 837         if (unlikely(buf_len < dev->vhost_hlen)) {
 838                 buf_offset = dev->vhost_hlen - buf_len;
 839                 vec_idx++;
 840                 buf_addr = buf_vec[vec_idx].buf_addr;
 841                 buf_iova = buf_vec[vec_idx].buf_iova;
 842                 buf_len = buf_vec[vec_idx].buf_len;
 843                 buf_avail = buf_len - buf_offset;
 844         } else {
 845                 buf_offset = dev->vhost_hlen;
 846                 buf_avail = buf_len - dev->vhost_hlen;
 847         }
 848
 849         mbuf_avail  = rte_pktmbuf_data_len(m);
 850         mbuf_offset = 0;
 851         while (mbuf_avail != 0 || m->next != NULL) {
 852                 /* done with current buf, get the next one */
 853                 if (buf_avail == 0) {
 854                         vec_idx++;
 855                         if (unlikely(vec_idx >= nr_vec)) {
 856                                 error = -1;
 857                                 goto out;
 858                         }
 859
 860                         buf_addr = buf_vec[vec_idx].buf_addr;
 861                         buf_iova = buf_vec[vec_idx].buf_iova;
 862                         buf_len = buf_vec[vec_idx].buf_len;
 863
 864                         buf_offset = 0;
 865                         buf_avail  = buf_len;
 866                 }
 867
 868                 /* done with current mbuf, get the next one */
 869                 if (mbuf_avail == 0) {
 870                         m = m->next;
 871
 872                         mbuf_offset = 0;
 873                         mbuf_avail  = rte_pktmbuf_data_len(m);
 874                 }
 875
 876                 if (hdr_addr) {
 877                         virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
 878                         if (rxvq_is_mergeable(dev))
 879                                 ASSIGN_UNLESS_EQUAL(hdr->num_buffers,
 880                                                 num_buffers);
 881
 882                         if (unlikely(hdr == &tmp_hdr)) {
 883                                 copy_vnet_hdr_to_desc(dev, vq, buf_vec, hdr);
 884                         } else {
 885                                 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
 886                                                 dev->vhost_hlen, 0);
 887                                 vhost_log_cache_write_iova(dev, vq,
 888                                                 buf_vec[0].buf_iova,
 889                                                 dev->vhost_hlen);
 890                         }
 891
 892                         hdr_addr = 0;
 893                 }
 894
 895                 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
 896
 897                 if (likely(cpy_len > MAX_BATCH_LEN ||
 898                                         vq->batch_copy_nb_elems >= vq->size)) {
 899                         rte_memcpy((void *)((uintptr_t)(buf_addr + buf_offset)),
 900                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 901                                 cpy_len);
 902                         vhost_log_cache_write_iova(dev, vq,
 903                                                    buf_iova + buf_offset,
 904                                                    cpy_len);
 905                         PRINT_PACKET(dev, (uintptr_t)(buf_addr + buf_offset),
 906                                 cpy_len, 0);
 907                 } else {
 908                         batch_copy[vq->batch_copy_nb_elems].dst =
 909                                 (void *)((uintptr_t)(buf_addr + buf_offset));
 910                         batch_copy[vq->batch_copy_nb_elems].src =
 911                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
 912                         batch_copy[vq->batch_copy_nb_elems].log_addr =
 913                                 buf_iova + buf_offset;
 914                         batch_copy[vq->batch_copy_nb_elems].len = cpy_len;
 915                         vq->batch_copy_nb_elems++;
 916                 }
 917
 918                 mbuf_avail  -= cpy_len;
 919                 mbuf_offset += cpy_len;
 920                 buf_avail  -= cpy_len;
 921                 buf_offset += cpy_len;
 922         }
 923
 924 out:
 925
 926         return error;
 927 }
 928
 929 static __rte_always_inline void
 930 async_fill_vec(struct iovec *v, void *base, size_t len)
 931 {
 932         v->iov_base = base;
 933         v->iov_len = len;
 934 }
 935
 936 static __rte_always_inline void
 937 async_fill_iter(struct rte_vhost_iov_iter *it, size_t count,
 938         struct iovec *vec, unsigned long nr_seg)
 939 {
 940         it->offset = 0;
 941         it->count = count;
 942
 943         if (count) {
 944                 it->iov = vec;
 945                 it->nr_segs = nr_seg;
 946         } else {
 947                 it->iov = 0;
 948                 it->nr_segs = 0;
 949         }
 950 }
 951
 952 static __rte_always_inline void
 953 async_fill_desc(struct rte_vhost_async_desc *desc,
 954         struct rte_vhost_iov_iter *src, struct rte_vhost_iov_iter *dst)
 955 {
 956         desc->src = src;
 957         desc->dst = dst;
 958 }
 959
 960 static __rte_always_inline int
 961 async_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 962                         struct rte_mbuf *m, struct buf_vector *buf_vec,
 963                         uint16_t nr_vec, uint16_t num_buffers,
 964                         struct iovec *src_iovec, struct iovec *dst_iovec,
 965                         struct rte_vhost_iov_iter *src_it,
 966                         struct rte_vhost_iov_iter *dst_it)
 967 {
 968         struct rte_mbuf *hdr_mbuf;
 969         struct virtio_net_hdr_mrg_rxbuf tmp_hdr, *hdr = NULL;
 970         uint64_t buf_addr, buf_iova;
 971         uint64_t hdr_addr;
 972         uint64_t mapped_len;
 973         uint32_t vec_idx = 0;
 974         uint32_t mbuf_offset, mbuf_avail;
 975         uint32_t buf_offset, buf_avail;
 976         uint32_t cpy_len, buf_len;
 977         int error = 0;
 978
 979         uint32_t tlen = 0;
 980         int tvec_idx = 0;
 981         void *hpa;
 982
 983         if (unlikely(m == NULL)) {
 984                 error = -1;
 985                 goto out;
 986         }
 987
 988         buf_addr = buf_vec[vec_idx].buf_addr;
 989         buf_iova = buf_vec[vec_idx].buf_iova;
 990         buf_len = buf_vec[vec_idx].buf_len;
 991
 992         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
 993                 error = -1;
 994                 goto out;
 995         }
 996
 997         hdr_mbuf = m;
 998         hdr_addr = buf_addr;
 999         if (unlikely(buf_len < dev->vhost_hlen)) {
1000                 memset(&tmp_hdr, 0, sizeof(struct virtio_net_hdr_mrg_rxbuf));
1001                 hdr = &tmp_hdr;
1002         } else
1003                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)hdr_addr;
1004
1005         VHOST_LOG_DATA(DEBUG, "(%d) RX: num merge buffers %d\n",
1006                 dev->vid, num_buffers);
1007
1008         if (unlikely(buf_len < dev->vhost_hlen)) {
1009                 buf_offset = dev->vhost_hlen - buf_len;
1010                 vec_idx++;
1011                 buf_addr = buf_vec[vec_idx].buf_addr;
1012                 buf_iova = buf_vec[vec_idx].buf_iova;
1013                 buf_len = buf_vec[vec_idx].buf_len;
1014                 buf_avail = buf_len - buf_offset;
1015         } else {
1016                 buf_offset = dev->vhost_hlen;
1017                 buf_avail = buf_len - dev->vhost_hlen;
1018         }
1019
1020         mbuf_avail  = rte_pktmbuf_data_len(m);
1021         mbuf_offset = 0;
1022
1023         while (mbuf_avail != 0 || m->next != NULL) {
1024                 /* done with current buf, get the next one */
1025                 if (buf_avail == 0) {
1026                         vec_idx++;
1027                         if (unlikely(vec_idx >= nr_vec)) {
1028                                 error = -1;
1029                                 goto out;
1030                         }
1031
1032                         buf_addr = buf_vec[vec_idx].buf_addr;
1033                         buf_iova = buf_vec[vec_idx].buf_iova;
1034                         buf_len = buf_vec[vec_idx].buf_len;
1035
1036                         buf_offset = 0;
1037                         buf_avail = buf_len;
1038                 }
1039
1040                 /* done with current mbuf, get the next one */
1041                 if (mbuf_avail == 0) {
1042                         m = m->next;
1043
1044                         mbuf_offset = 0;
1045                         mbuf_avail = rte_pktmbuf_data_len(m);
1046                 }
1047
1048                 if (hdr_addr) {
1049                         virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
1050                         if (rxvq_is_mergeable(dev))
1051                                 ASSIGN_UNLESS_EQUAL(hdr->num_buffers,
1052                                                 num_buffers);
1053
1054                         if (unlikely(hdr == &tmp_hdr)) {
1055                                 copy_vnet_hdr_to_desc(dev, vq, buf_vec, hdr);
1056                         } else {
1057                                 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
1058                                                 dev->vhost_hlen, 0);
1059                                 vhost_log_cache_write_iova(dev, vq,
1060                                                 buf_vec[0].buf_iova,
1061                                                 dev->vhost_hlen);
1062                         }
1063
1064                         hdr_addr = 0;
1065                 }
1066
1067                 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
1068
1069                 while (unlikely(cpy_len)) {
1070                         hpa = (void *)(uintptr_t)gpa_to_first_hpa(dev,
1071                                         buf_iova + buf_offset,
1072                                         cpy_len, &mapped_len);
1073                         if (unlikely(!hpa)) {
1074                                 VHOST_LOG_DATA(ERR, "(%d) %s: failed to get hpa.\n",
1075                                 dev->vid, __func__);
1076                                 error = -1;
1077                                 goto out;
1078                         }
1079
1080                         async_fill_vec(src_iovec + tvec_idx,
1081                                 (void *)(uintptr_t)rte_pktmbuf_iova_offset(m,
1082                                 mbuf_offset), (size_t)mapped_len);
1083                         async_fill_vec(dst_iovec + tvec_idx,
1084                                         hpa, (size_t)mapped_len);
1085
1086                         tlen += (uint32_t)mapped_len;
1087                         cpy_len -= (uint32_t)mapped_len;
1088                         mbuf_avail  -= (uint32_t)mapped_len;
1089                         mbuf_offset += (uint32_t)mapped_len;
1090                         buf_avail  -= (uint32_t)mapped_len;
1091                         buf_offset += (uint32_t)mapped_len;
1092                         tvec_idx++;
1093                 }
1094         }
1095
1096         async_fill_iter(src_it, tlen, src_iovec, tvec_idx);
1097         async_fill_iter(dst_it, tlen, dst_iovec, tvec_idx);
1098 out:
1099         return error;
1100 }
1101
1102 static __rte_always_inline int
1103 vhost_enqueue_single_packed(struct virtio_net *dev,
1104                             struct vhost_virtqueue *vq,
1105                             struct rte_mbuf *pkt,
1106                             struct buf_vector *buf_vec,
1107                             uint16_t *nr_descs)
1108 {
1109         uint16_t nr_vec = 0;
1110         uint16_t avail_idx = vq->last_avail_idx;
1111         uint16_t max_tries, tries = 0;
1112         uint16_t buf_id = 0;
1113         uint32_t len = 0;
1114         uint16_t desc_count;
1115         uint32_t size = pkt->pkt_len + sizeof(struct virtio_net_hdr_mrg_rxbuf);
1116         uint16_t num_buffers = 0;
1117         uint32_t buffer_len[vq->size];
1118         uint16_t buffer_buf_id[vq->size];
1119         uint16_t buffer_desc_count[vq->size];
1120
1121         if (rxvq_is_mergeable(dev))
1122                 max_tries = vq->size - 1;
1123         else
1124                 max_tries = 1;
1125
1126         while (size > 0) {
1127                 /*
1128                  * if we tried all available ring items, and still
1129                  * can't get enough buf, it means something abnormal
1130                  * happened.
1131                  */
1132                 if (unlikely(++tries > max_tries))
1133                         return -1;
1134
1135                 if (unlikely(fill_vec_buf_packed(dev, vq,
1136                                                 avail_idx, &desc_count,
1137                                                 buf_vec, &nr_vec,
1138                                                 &buf_id, &len,
1139                                                 VHOST_ACCESS_RW) < 0))
1140                         return -1;
1141
1142                 len = RTE_MIN(len, size);
1143                 size -= len;
1144
1145                 buffer_len[num_buffers] = len;
1146                 buffer_buf_id[num_buffers] = buf_id;
1147                 buffer_desc_count[num_buffers] = desc_count;
1148                 num_buffers += 1;
1149
1150                 *nr_descs += desc_count;
1151                 avail_idx += desc_count;
1152                 if (avail_idx >= vq->size)
1153                         avail_idx -= vq->size;
1154         }
1155
1156         if (copy_mbuf_to_desc(dev, vq, pkt, buf_vec, nr_vec, num_buffers) < 0)
1157                 return -1;
1158
1159         vhost_shadow_enqueue_single_packed(dev, vq, buffer_len, buffer_buf_id,
1160                                            buffer_desc_count, num_buffers);
1161
1162         return 0;
1163 }
1164
1165 static __rte_noinline uint32_t
1166 virtio_dev_rx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
1167         struct rte_mbuf **pkts, uint32_t count)
1168 {
1169         uint32_t pkt_idx = 0;
1170         uint16_t num_buffers;
1171         struct buf_vector buf_vec[BUF_VECTOR_MAX];
1172         uint16_t avail_head;
1173
1174         /*
1175          * The ordering between avail index and
1176          * desc reads needs to be enforced.
1177          */
1178         avail_head = __atomic_load_n(&vq->avail->idx, __ATOMIC_ACQUIRE);
1179
1180         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
1181
1182         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
1183                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
1184                 uint16_t nr_vec = 0;
1185
1186                 if (unlikely(reserve_avail_buf_split(dev, vq,
1187                                                 pkt_len, buf_vec, &num_buffers,
1188                                                 avail_head, &nr_vec) < 0)) {
1189                         VHOST_LOG_DATA(DEBUG,
1190                                 "(%d) failed to get enough desc from vring\n",
1191                                 dev->vid);
1192                         vq->shadow_used_idx -= num_buffers;
1193                         break;
1194                 }
1195
1196                 VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
1197                         dev->vid, vq->last_avail_idx,
1198                         vq->last_avail_idx + num_buffers);
1199
1200                 if (copy_mbuf_to_desc(dev, vq, pkts[pkt_idx],
1201                                                 buf_vec, nr_vec,
1202                                                 num_buffers) < 0) {
1203                         vq->shadow_used_idx -= num_buffers;
1204                         break;
1205                 }
1206
1207                 vq->last_avail_idx += num_buffers;
1208         }
1209
1210         do_data_copy_enqueue(dev, vq);
1211
1212         if (likely(vq->shadow_used_idx)) {
1213                 flush_shadow_used_ring_split(dev, vq);
1214                 vhost_vring_call_split(dev, vq);
1215         }
1216
1217         return pkt_idx;
1218 }
1219
1220 static __rte_always_inline int
1221 virtio_dev_rx_sync_batch_check(struct virtio_net *dev,
1222                            struct vhost_virtqueue *vq,
1223                            struct rte_mbuf **pkts,
1224                            uint64_t *desc_addrs,
1225                            uint64_t *lens)
1226 {
1227         bool wrap_counter = vq->avail_wrap_counter;
1228         struct vring_packed_desc *descs = vq->desc_packed;
1229         uint16_t avail_idx = vq->last_avail_idx;
1230         uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1231         uint16_t i;
1232
1233         if (unlikely(avail_idx & PACKED_BATCH_MASK))
1234                 return -1;
1235
1236         if (unlikely((avail_idx + PACKED_BATCH_SIZE) > vq->size))
1237                 return -1;
1238
1239         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1240                 if (unlikely(pkts[i]->next != NULL))
1241                         return -1;
1242                 if (unlikely(!desc_is_avail(&descs[avail_idx + i],
1243                                             wrap_counter)))
1244                         return -1;
1245         }
1246
1247         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1248                 lens[i] = descs[avail_idx + i].len;
1249
1250         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1251                 if (unlikely(pkts[i]->pkt_len > (lens[i] - buf_offset)))
1252                         return -1;
1253         }
1254
1255         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1256                 desc_addrs[i] = vhost_iova_to_vva(dev, vq,
1257                                                   descs[avail_idx + i].addr,
1258                                                   &lens[i],
1259                                                   VHOST_ACCESS_RW);
1260
1261         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1262                 if (unlikely(!desc_addrs[i]))
1263                         return -1;
1264                 if (unlikely(lens[i] != descs[avail_idx + i].len))
1265                         return -1;
1266         }
1267
1268         return 0;
1269 }
1270
1271 static __rte_always_inline void
1272 virtio_dev_rx_batch_packed_copy(struct virtio_net *dev,
1273                            struct vhost_virtqueue *vq,
1274                            struct rte_mbuf **pkts,
1275                            uint64_t *desc_addrs,
1276                            uint64_t *lens)
1277 {
1278         uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1279         struct virtio_net_hdr_mrg_rxbuf *hdrs[PACKED_BATCH_SIZE];
1280         struct vring_packed_desc *descs = vq->desc_packed;
1281         uint16_t avail_idx = vq->last_avail_idx;
1282         uint16_t ids[PACKED_BATCH_SIZE];
1283         uint16_t i;
1284
1285         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1286                 rte_prefetch0((void *)(uintptr_t)desc_addrs[i]);
1287                 hdrs[i] = (struct virtio_net_hdr_mrg_rxbuf *)
1288                                         (uintptr_t)desc_addrs[i];
1289                 lens[i] = pkts[i]->pkt_len +
1290                         sizeof(struct virtio_net_hdr_mrg_rxbuf);
1291         }
1292
1293         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1294                 virtio_enqueue_offload(pkts[i], &hdrs[i]->hdr);
1295
1296         vq_inc_last_avail_packed(vq, PACKED_BATCH_SIZE);
1297
1298         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1299                 rte_memcpy((void *)(uintptr_t)(desc_addrs[i] + buf_offset),
1300                            rte_pktmbuf_mtod_offset(pkts[i], void *, 0),
1301                            pkts[i]->pkt_len);
1302         }
1303
1304         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1305                 vhost_log_cache_write_iova(dev, vq, descs[avail_idx + i].addr,
1306                                            lens[i]);
1307
1308         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1309                 ids[i] = descs[avail_idx + i].id;
1310
1311         vhost_flush_enqueue_batch_packed(dev, vq, lens, ids);
1312 }
1313
1314 static __rte_always_inline int
1315 virtio_dev_rx_sync_batch_packed(struct virtio_net *dev,
1316                            struct vhost_virtqueue *vq,
1317                            struct rte_mbuf **pkts)
1318 {
1319         uint64_t desc_addrs[PACKED_BATCH_SIZE];
1320         uint64_t lens[PACKED_BATCH_SIZE];
1321
1322         if (virtio_dev_rx_sync_batch_check(dev, vq, pkts, desc_addrs, lens) == -1)
1323                 return -1;
1324
1325         if (vq->shadow_used_idx) {
1326                 do_data_copy_enqueue(dev, vq);
1327                 vhost_flush_enqueue_shadow_packed(dev, vq);
1328         }
1329
1330         virtio_dev_rx_batch_packed_copy(dev, vq, pkts, desc_addrs, lens);
1331
1332         return 0;
1333 }
1334
1335 static __rte_always_inline int16_t
1336 virtio_dev_rx_single_packed(struct virtio_net *dev,
1337                             struct vhost_virtqueue *vq,
1338                             struct rte_mbuf *pkt)
1339 {
1340         struct buf_vector buf_vec[BUF_VECTOR_MAX];
1341         uint16_t nr_descs = 0;
1342
1343         if (unlikely(vhost_enqueue_single_packed(dev, vq, pkt, buf_vec,
1344                                                  &nr_descs) < 0)) {
1345                 VHOST_LOG_DATA(DEBUG,
1346                                 "(%d) failed to get enough desc from vring\n",
1347                                 dev->vid);
1348                 return -1;
1349         }
1350
1351         VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
1352                         dev->vid, vq->last_avail_idx,
1353                         vq->last_avail_idx + nr_descs);
1354
1355         vq_inc_last_avail_packed(vq, nr_descs);
1356
1357         return 0;
1358 }
1359
1360 static __rte_noinline uint32_t
1361 virtio_dev_rx_packed(struct virtio_net *dev,
1362                      struct vhost_virtqueue *__rte_restrict vq,
1363                      struct rte_mbuf **__rte_restrict pkts,
1364                      uint32_t count)
1365 {
1366         uint32_t pkt_idx = 0;
1367
1368         do {
1369                 rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
1370
1371                 if (count - pkt_idx >= PACKED_BATCH_SIZE) {
1372                         if (!virtio_dev_rx_sync_batch_packed(dev, vq,
1373                                                         &pkts[pkt_idx])) {
1374                                 pkt_idx += PACKED_BATCH_SIZE;
1375                                 continue;
1376                         }
1377                 }
1378
1379                 if (virtio_dev_rx_single_packed(dev, vq, pkts[pkt_idx]))
1380                         break;
1381                 pkt_idx++;
1382
1383         } while (pkt_idx < count);
1384
1385         if (vq->shadow_used_idx) {
1386                 do_data_copy_enqueue(dev, vq);
1387                 vhost_flush_enqueue_shadow_packed(dev, vq);
1388         }
1389
1390         if (pkt_idx)
1391                 vhost_vring_call_packed(dev, vq);
1392
1393         return pkt_idx;
1394 }
1395
1396 static __rte_always_inline uint32_t
1397 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
1398         struct rte_mbuf **pkts, uint32_t count)
1399 {
1400         struct vhost_virtqueue *vq;
1401         uint32_t nb_tx = 0;
1402
1403         VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
1404         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
1405                 VHOST_LOG_DATA(ERR, "(%d) %s: invalid virtqueue idx %d.\n",
1406                         dev->vid, __func__, queue_id);
1407                 return 0;
1408         }
1409
1410         vq = dev->virtqueue[queue_id];
1411
1412         rte_spinlock_lock(&vq->access_lock);
1413
1414         if (unlikely(!vq->enabled))
1415                 goto out_access_unlock;
1416
1417         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1418                 vhost_user_iotlb_rd_lock(vq);
1419
1420         if (unlikely(!vq->access_ok))
1421                 if (unlikely(vring_translate(dev, vq) < 0))
1422                         goto out;
1423
1424         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
1425         if (count == 0)
1426                 goto out;
1427
1428         if (vq_is_packed(dev))
1429                 nb_tx = virtio_dev_rx_packed(dev, vq, pkts, count);
1430         else
1431                 nb_tx = virtio_dev_rx_split(dev, vq, pkts, count);
1432
1433 out:
1434         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1435                 vhost_user_iotlb_rd_unlock(vq);
1436
1437 out_access_unlock:
1438         rte_spinlock_unlock(&vq->access_lock);
1439
1440         return nb_tx;
1441 }
1442
1443 uint16_t
1444 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
1445         struct rte_mbuf **__rte_restrict pkts, uint16_t count)
1446 {
1447         struct virtio_net *dev = get_device(vid);
1448
1449         if (!dev)
1450                 return 0;
1451
1452         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
1453                 VHOST_LOG_DATA(ERR,
1454                         "(%d) %s: built-in vhost net backend is disabled.\n",
1455                         dev->vid, __func__);
1456                 return 0;
1457         }
1458
1459         return virtio_dev_rx(dev, queue_id, pkts, count);
1460 }
1461
1462 static __rte_always_inline uint16_t
1463 virtio_dev_rx_async_get_info_idx(uint16_t pkts_idx,
1464         uint16_t vq_size, uint16_t n_inflight)
1465 {
1466         return pkts_idx > n_inflight ? (pkts_idx - n_inflight) :
1467                 (vq_size - n_inflight + pkts_idx) % vq_size;
1468 }
1469
1470 static __rte_always_inline void
1471 store_dma_desc_info_split(struct vring_used_elem *s_ring, struct vring_used_elem *d_ring,
1472                 uint16_t ring_size, uint16_t s_idx, uint16_t d_idx, uint16_t count)
1473 {
1474         size_t elem_size = sizeof(struct vring_used_elem);
1475
1476         if (d_idx + count <= ring_size) {
1477                 rte_memcpy(d_ring + d_idx, s_ring + s_idx, count * elem_size);
1478         } else {
1479                 uint16_t size = ring_size - d_idx;
1480
1481                 rte_memcpy(d_ring + d_idx, s_ring + s_idx, size * elem_size);
1482                 rte_memcpy(d_ring, s_ring + s_idx + size, (count - size) * elem_size);
1483         }
1484 }
1485
1486 static __rte_always_inline void
1487 store_dma_desc_info_packed(struct vring_used_elem_packed *s_ring,
1488                 struct vring_used_elem_packed *d_ring,
1489                 uint16_t ring_size, uint16_t s_idx, uint16_t d_idx, uint16_t count)
1490 {
1491         size_t elem_size = sizeof(struct vring_used_elem_packed);
1492
1493         if (d_idx + count <= ring_size) {
1494                 rte_memcpy(d_ring + d_idx, s_ring + s_idx, count * elem_size);
1495         } else {
1496                 uint16_t size = ring_size - d_idx;
1497
1498                 rte_memcpy(d_ring + d_idx, s_ring + s_idx, size * elem_size);
1499                 rte_memcpy(d_ring, s_ring + s_idx + size, (count - size) * elem_size);
1500         }
1501 }
1502
1503 static __rte_noinline uint32_t
1504 virtio_dev_rx_async_submit_split(struct virtio_net *dev,
1505         struct vhost_virtqueue *vq, uint16_t queue_id,
1506         struct rte_mbuf **pkts, uint32_t count)
1507 {
1508         struct buf_vector buf_vec[BUF_VECTOR_MAX];
1509         uint32_t pkt_idx = 0, pkt_burst_idx = 0;
1510         uint16_t num_buffers;
1511         uint16_t avail_head;
1512
1513         struct vhost_async *async = vq->async;
1514         struct rte_vhost_iov_iter *it_pool = async->it_pool;
1515         struct iovec *vec_pool = async->vec_pool;
1516         struct rte_vhost_async_desc tdes[MAX_PKT_BURST];
1517         struct iovec *src_iovec = vec_pool;
1518         struct iovec *dst_iovec = vec_pool + (VHOST_MAX_ASYNC_VEC >> 1);
1519         struct async_inflight_info *pkts_info = async->pkts_info;
1520         uint32_t n_pkts = 0, pkt_err = 0;
1521         int32_t n_xfer;
1522         uint16_t segs_await = 0;
1523         uint16_t iovec_idx = 0, it_idx = 0, slot_idx = 0;
1524
1525         /*
1526          * The ordering between avail index and desc reads need to be enforced.
1527          */
1528         avail_head = __atomic_load_n(&vq->avail->idx, __ATOMIC_ACQUIRE);
1529
1530         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
1531
1532         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
1533                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
1534                 uint16_t nr_vec = 0;
1535
1536                 if (unlikely(reserve_avail_buf_split(dev, vq,
1537                                                 pkt_len, buf_vec, &num_buffers,
1538                                                 avail_head, &nr_vec) < 0)) {
1539                         VHOST_LOG_DATA(DEBUG,
1540                                 "(%d) failed to get enough desc from vring\n",
1541                                 dev->vid);
1542                         vq->shadow_used_idx -= num_buffers;
1543                         break;
1544                 }
1545
1546                 VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
1547                         dev->vid, vq->last_avail_idx,
1548                         vq->last_avail_idx + num_buffers);
1549
1550                 if (async_mbuf_to_desc(dev, vq, pkts[pkt_idx], buf_vec, nr_vec, num_buffers,
1551                                 &src_iovec[iovec_idx], &dst_iovec[iovec_idx],
1552                                 &it_pool[it_idx], &it_pool[it_idx + 1]) < 0) {
1553                         vq->shadow_used_idx -= num_buffers;
1554                         break;
1555                 }
1556
1557                 async_fill_desc(&tdes[pkt_burst_idx++], &it_pool[it_idx],
1558                                 &it_pool[it_idx + 1]);
1559
1560                 slot_idx = (async->pkts_idx + pkt_idx) & (vq->size - 1);
1561                 pkts_info[slot_idx].descs = num_buffers;
1562                 pkts_info[slot_idx].mbuf = pkts[pkt_idx];
1563
1564                 iovec_idx += it_pool[it_idx].nr_segs;
1565                 segs_await += it_pool[it_idx].nr_segs;
1566                 it_idx += 2;
1567
1568                 vq->last_avail_idx += num_buffers;
1569
1570                 /*
1571                  * conditions to trigger async device transfer:
1572                  * - buffered packet number reaches transfer threshold
1573                  * - unused async iov number is less than max vhost vector
1574                  */
1575                 if (unlikely(pkt_burst_idx >= VHOST_ASYNC_BATCH_THRESHOLD ||
1576                         ((VHOST_MAX_ASYNC_VEC >> 1) - segs_await <
1577                         BUF_VECTOR_MAX))) {
1578                         n_xfer = async->ops.transfer_data(dev->vid,
1579                                         queue_id, tdes, 0, pkt_burst_idx);
1580                         if (likely(n_xfer >= 0)) {
1581                                 n_pkts = n_xfer;
1582                         } else {
1583                                 VHOST_LOG_DATA(ERR,
1584                                         "(%d) %s: failed to transfer data for queue id %d.\n",
1585                                         dev->vid, __func__, queue_id);
1586                                 n_pkts = 0;
1587                         }
1588
1589                         iovec_idx = 0;
1590                         it_idx = 0;
1591                         segs_await = 0;
1592
1593                         if (unlikely(n_pkts < pkt_burst_idx)) {
1594                                 /*
1595                                  * log error packets number here and do actual
1596                                  * error processing when applications poll
1597                                  * completion
1598                                  */
1599                                 pkt_err = pkt_burst_idx - n_pkts;
1600                                 pkt_idx++;
1601                                 pkt_burst_idx = 0;
1602                                 break;
1603                         }
1604
1605                         pkt_burst_idx = 0;
1606                 }
1607         }
1608
1609         if (pkt_burst_idx) {
1610                 n_xfer = async->ops.transfer_data(dev->vid, queue_id, tdes, 0, pkt_burst_idx);
1611                 if (likely(n_xfer >= 0)) {
1612                         n_pkts = n_xfer;
1613                 } else {
1614                         VHOST_LOG_DATA(ERR, "(%d) %s: failed to transfer data for queue id %d.\n",
1615                                 dev->vid, __func__, queue_id);
1616                         n_pkts = 0;
1617                 }
1618
1619                 if (unlikely(n_pkts < pkt_burst_idx))
1620                         pkt_err = pkt_burst_idx - n_pkts;
1621         }
1622
1623         if (unlikely(pkt_err)) {
1624                 uint16_t num_descs = 0;
1625
1626                 /* update number of completed packets */
1627                 pkt_idx -= pkt_err;
1628
1629                 /* calculate the sum of descriptors to revert */
1630                 while (pkt_err-- > 0) {
1631                         num_descs += pkts_info[slot_idx & (vq->size - 1)].descs;
1632                         slot_idx--;
1633                 }
1634
1635                 /* recover shadow used ring and available ring */
1636                 vq->shadow_used_idx -= num_descs;
1637                 vq->last_avail_idx -= num_descs;
1638         }
1639
1640         /* keep used descriptors */
1641         if (likely(vq->shadow_used_idx)) {
1642                 uint16_t to = async->desc_idx_split & (vq->size - 1);
1643
1644                 store_dma_desc_info_split(vq->shadow_used_split,
1645                                 async->descs_split, vq->size, 0, to,
1646                                 vq->shadow_used_idx);
1647
1648                 async->desc_idx_split += vq->shadow_used_idx;
1649                 async->pkts_idx += pkt_idx;
1650                 async->pkts_inflight_n += pkt_idx;
1651                 vq->shadow_used_idx = 0;
1652         }
1653
1654         return pkt_idx;
1655 }
1656
1657 static __rte_always_inline void
1658 vhost_update_used_packed(struct vhost_virtqueue *vq,
1659                         struct vring_used_elem_packed *shadow_ring,
1660                         uint16_t count)
1661 {
1662         int i;
1663         uint16_t used_idx = vq->last_used_idx;
1664         uint16_t head_idx = vq->last_used_idx;
1665         uint16_t head_flags = 0;
1666
1667         if (count == 0)
1668                 return;
1669
1670         /* Split loop in two to save memory barriers */
1671         for (i = 0; i < count; i++) {
1672                 vq->desc_packed[used_idx].id = shadow_ring[i].id;
1673                 vq->desc_packed[used_idx].len = shadow_ring[i].len;
1674
1675                 used_idx += shadow_ring[i].count;
1676                 if (used_idx >= vq->size)
1677                         used_idx -= vq->size;
1678         }
1679
1680         /* The ordering for storing desc flags needs to be enforced. */
1681         rte_atomic_thread_fence(__ATOMIC_RELEASE);
1682
1683         for (i = 0; i < count; i++) {
1684                 uint16_t flags;
1685
1686                 if (vq->shadow_used_packed[i].len)
1687                         flags = VRING_DESC_F_WRITE;
1688                 else
1689                         flags = 0;
1690
1691                 if (vq->used_wrap_counter) {
1692                         flags |= VRING_DESC_F_USED;
1693                         flags |= VRING_DESC_F_AVAIL;
1694                 } else {
1695                         flags &= ~VRING_DESC_F_USED;
1696                         flags &= ~VRING_DESC_F_AVAIL;
1697                 }
1698
1699                 if (i > 0) {
1700                         vq->desc_packed[vq->last_used_idx].flags = flags;
1701                 } else {
1702                         head_idx = vq->last_used_idx;
1703                         head_flags = flags;
1704                 }
1705
1706                 vq_inc_last_used_packed(vq, shadow_ring[i].count);
1707         }
1708
1709         vq->desc_packed[head_idx].flags = head_flags;
1710 }
1711
1712 static __rte_always_inline int
1713 vhost_enqueue_async_packed(struct virtio_net *dev,
1714                             struct vhost_virtqueue *vq,
1715                             struct rte_mbuf *pkt,
1716                             struct buf_vector *buf_vec,
1717                             uint16_t *nr_descs,
1718                             uint16_t *nr_buffers,
1719                             struct iovec *src_iovec, struct iovec *dst_iovec,
1720                             struct rte_vhost_iov_iter *src_it,
1721                             struct rte_vhost_iov_iter *dst_it)
1722 {
1723         uint16_t nr_vec = 0;
1724         uint16_t avail_idx = vq->last_avail_idx;
1725         uint16_t max_tries, tries = 0;
1726         uint16_t buf_id = 0;
1727         uint32_t len = 0;
1728         uint16_t desc_count = 0;
1729         uint32_t size = pkt->pkt_len + sizeof(struct virtio_net_hdr_mrg_rxbuf);
1730         uint32_t buffer_len[vq->size];
1731         uint16_t buffer_buf_id[vq->size];
1732         uint16_t buffer_desc_count[vq->size];
1733
1734         if (rxvq_is_mergeable(dev))
1735                 max_tries = vq->size - 1;
1736         else
1737                 max_tries = 1;
1738
1739         while (size > 0) {
1740                 /*
1741                  * if we tried all available ring items, and still
1742                  * can't get enough buf, it means something abnormal
1743                  * happened.
1744                  */
1745                 if (unlikely(++tries > max_tries))
1746                         return -1;
1747
1748                 if (unlikely(fill_vec_buf_packed(dev, vq, avail_idx, &desc_count, buf_vec, &nr_vec,
1749                                                 &buf_id, &len, VHOST_ACCESS_RW) < 0))
1750                         return -1;
1751
1752                 len = RTE_MIN(len, size);
1753                 size -= len;
1754
1755                 buffer_len[*nr_buffers] = len;
1756                 buffer_buf_id[*nr_buffers] = buf_id;
1757                 buffer_desc_count[*nr_buffers] = desc_count;
1758                 *nr_buffers += 1;
1759                 *nr_descs += desc_count;
1760                 avail_idx += desc_count;
1761                 if (avail_idx >= vq->size)
1762                         avail_idx -= vq->size;
1763         }
1764
1765         if (unlikely(async_mbuf_to_desc(dev, vq, pkt, buf_vec, nr_vec,
1766                                         *nr_buffers, src_iovec, dst_iovec,
1767                                         src_it, dst_it) < 0))
1768                 return -1;
1769
1770         vhost_shadow_enqueue_packed(vq, buffer_len, buffer_buf_id, buffer_desc_count, *nr_buffers);
1771
1772         return 0;
1773 }
1774
1775 static __rte_always_inline int16_t
1776 virtio_dev_rx_async_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
1777                             struct rte_mbuf *pkt, uint16_t *nr_descs, uint16_t *nr_buffers,
1778                             struct iovec *src_iovec, struct iovec *dst_iovec,
1779                             struct rte_vhost_iov_iter *src_it, struct rte_vhost_iov_iter *dst_it)
1780 {
1781         struct buf_vector buf_vec[BUF_VECTOR_MAX];
1782
1783         if (unlikely(vhost_enqueue_async_packed(dev, vq, pkt, buf_vec, nr_descs, nr_buffers,
1784                                                  src_iovec, dst_iovec,
1785                                                  src_it, dst_it) < 0)) {
1786                 VHOST_LOG_DATA(DEBUG, "(%d) failed to get enough desc from vring\n", dev->vid);
1787                 return -1;
1788         }
1789
1790         VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
1791                         dev->vid, vq->last_avail_idx, vq->last_avail_idx + *nr_descs);
1792
1793         return 0;
1794 }
1795
1796 static __rte_always_inline void
1797 dma_error_handler_packed(struct vhost_virtqueue *vq, uint16_t slot_idx,
1798                         uint32_t nr_err, uint32_t *pkt_idx)
1799 {
1800         uint16_t descs_err = 0;
1801         uint16_t buffers_err = 0;
1802         struct async_inflight_info *pkts_info = vq->async->pkts_info;
1803
1804         *pkt_idx -= nr_err;
1805         /* calculate the sum of buffers and descs of DMA-error packets. */
1806         while (nr_err-- > 0) {
1807                 descs_err += pkts_info[slot_idx % vq->size].descs;
1808                 buffers_err += pkts_info[slot_idx % vq->size].nr_buffers;
1809                 slot_idx--;
1810         }
1811
1812         if (vq->last_avail_idx >= descs_err) {
1813                 vq->last_avail_idx -= descs_err;
1814         } else {
1815                 vq->last_avail_idx = vq->last_avail_idx + vq->size - descs_err;
1816                 vq->avail_wrap_counter ^= 1;
1817         }
1818
1819         vq->shadow_used_idx -= buffers_err;
1820 }
1821
1822 static __rte_noinline uint32_t
1823 virtio_dev_rx_async_submit_packed(struct virtio_net *dev,
1824         struct vhost_virtqueue *vq, uint16_t queue_id,
1825         struct rte_mbuf **pkts, uint32_t count)
1826 {
1827         uint32_t pkt_idx = 0, pkt_burst_idx = 0;
1828         uint32_t remained = count;
1829         int32_t n_xfer;
1830         uint16_t num_buffers;
1831         uint16_t num_descs;
1832
1833         struct vhost_async *async = vq->async;
1834         struct rte_vhost_iov_iter *it_pool = async->it_pool;
1835         struct iovec *vec_pool = async->vec_pool;
1836         struct rte_vhost_async_desc tdes[MAX_PKT_BURST];
1837         struct iovec *src_iovec = vec_pool;
1838         struct iovec *dst_iovec = vec_pool + (VHOST_MAX_ASYNC_VEC >> 1);
1839         struct async_inflight_info *pkts_info = async->pkts_info;
1840         uint32_t n_pkts = 0, pkt_err = 0;
1841         uint16_t slot_idx = 0;
1842         uint16_t segs_await = 0;
1843         uint16_t iovec_idx = 0, it_idx = 0;
1844
1845         do {
1846                 rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
1847
1848                 num_buffers = 0;
1849                 num_descs = 0;
1850                 if (unlikely(virtio_dev_rx_async_packed(dev, vq, pkts[pkt_idx],
1851                                                 &num_descs, &num_buffers,
1852                                                 &src_iovec[iovec_idx], &dst_iovec[iovec_idx],
1853                                                 &it_pool[it_idx], &it_pool[it_idx + 1]) < 0))
1854                         break;
1855
1856                 slot_idx = (async->pkts_idx + pkt_idx) % vq->size;
1857
1858                 async_fill_desc(&tdes[pkt_burst_idx++], &it_pool[it_idx],
1859                                 &it_pool[it_idx + 1]);
1860                 pkts_info[slot_idx].descs = num_descs;
1861                 pkts_info[slot_idx].nr_buffers = num_buffers;
1862                 pkts_info[slot_idx].mbuf = pkts[pkt_idx];
1863                 iovec_idx += it_pool[it_idx].nr_segs;
1864                 segs_await += it_pool[it_idx].nr_segs;
1865                 it_idx += 2;
1866
1867                 pkt_idx++;
1868                 remained--;
1869                 vq_inc_last_avail_packed(vq, num_descs);
1870
1871                 /*
1872                  * conditions to trigger async device transfer:
1873                  * - buffered packet number reaches transfer threshold
1874                  * - unused async iov number is less than max vhost vector
1875                  */
1876                 if (unlikely(pkt_burst_idx >= VHOST_ASYNC_BATCH_THRESHOLD ||
1877                         ((VHOST_MAX_ASYNC_VEC >> 1) - segs_await < BUF_VECTOR_MAX))) {
1878                         n_xfer = async->ops.transfer_data(dev->vid,
1879                                         queue_id, tdes, 0, pkt_burst_idx);
1880                         if (likely(n_xfer >= 0)) {
1881                                 n_pkts = n_xfer;
1882                         } else {
1883                                 VHOST_LOG_DATA(ERR,
1884                                         "(%d) %s: failed to transfer data for queue id %d.\n",
1885                                         dev->vid, __func__, queue_id);
1886                                 n_pkts = 0;
1887                         }
1888
1889                         iovec_idx = 0;
1890                         it_idx = 0;
1891                         segs_await = 0;
1892
1893                         if (unlikely(n_pkts < pkt_burst_idx)) {
1894                                 /*
1895                                  * log error packets number here and do actual
1896                                  * error processing when applications poll
1897                                  * completion
1898                                  */
1899                                 pkt_err = pkt_burst_idx - n_pkts;
1900                                 pkt_burst_idx = 0;
1901                                 break;
1902                         }
1903
1904                         pkt_burst_idx = 0;
1905                 }
1906         } while (pkt_idx < count);
1907
1908         if (pkt_burst_idx) {
1909                 n_xfer = async->ops.transfer_data(dev->vid, queue_id, tdes, 0, pkt_burst_idx);
1910                 if (likely(n_xfer >= 0)) {
1911                         n_pkts = n_xfer;
1912                 } else {
1913                         VHOST_LOG_DATA(ERR, "(%d) %s: failed to transfer data for queue id %d.\n",
1914                                 dev->vid, __func__, queue_id);
1915                         n_pkts = 0;
1916                 }
1917
1918                 if (unlikely(n_pkts < pkt_burst_idx))
1919                         pkt_err = pkt_burst_idx - n_pkts;
1920         }
1921
1922         if (unlikely(pkt_err))
1923                 dma_error_handler_packed(vq, slot_idx, pkt_err, &pkt_idx);
1924
1925         if (likely(vq->shadow_used_idx)) {
1926                 /* keep used descriptors. */
1927                 store_dma_desc_info_packed(vq->shadow_used_packed, async->buffers_packed,
1928                                         vq->size, 0, async->buffer_idx_packed,
1929                                         vq->shadow_used_idx);
1930
1931                 async->buffer_idx_packed += vq->shadow_used_idx;
1932                 if (async->buffer_idx_packed >= vq->size)
1933                         async->buffer_idx_packed -= vq->size;
1934
1935                 async->pkts_idx += pkt_idx;
1936                 if (async->pkts_idx >= vq->size)
1937                         async->pkts_idx -= vq->size;
1938
1939                 vq->shadow_used_idx = 0;
1940                 async->pkts_inflight_n += pkt_idx;
1941         }
1942
1943         return pkt_idx;
1944 }
1945
1946 static __rte_always_inline void
1947 write_back_completed_descs_split(struct vhost_virtqueue *vq, uint16_t n_descs)
1948 {
1949         struct vhost_async *async = vq->async;
1950         uint16_t nr_left = n_descs;
1951         uint16_t nr_copy;
1952         uint16_t to, from;
1953
1954         do {
1955                 from = async->last_desc_idx_split & (vq->size - 1);
1956                 nr_copy = nr_left + from <= vq->size ? nr_left : vq->size - from;
1957                 to = vq->last_used_idx & (vq->size - 1);
1958
1959                 if (to + nr_copy <= vq->size) {
1960                         rte_memcpy(&vq->used->ring[to], &async->descs_split[from],
1961                                         nr_copy * sizeof(struct vring_used_elem));
1962                 } else {
1963                         uint16_t size = vq->size - to;
1964
1965                         rte_memcpy(&vq->used->ring[to], &async->descs_split[from],
1966                                         size * sizeof(struct vring_used_elem));
1967                         rte_memcpy(&vq->used->ring[0], &async->descs_split[from + size],
1968                                         (nr_copy - size) * sizeof(struct vring_used_elem));
1969                 }
1970
1971                 async->last_desc_idx_split += nr_copy;
1972                 vq->last_used_idx += nr_copy;
1973                 nr_left -= nr_copy;
1974         } while (nr_left > 0);
1975 }
1976
1977 static __rte_always_inline void
1978 write_back_completed_descs_packed(struct vhost_virtqueue *vq,
1979                                 uint16_t n_buffers)
1980 {
1981         struct vhost_async *async = vq->async;
1982         uint16_t nr_left = n_buffers;
1983         uint16_t from, to;
1984
1985         do {
1986                 from = async->last_buffer_idx_packed;
1987                 to = (from + nr_left) % vq->size;
1988                 if (to > from) {
1989                         vhost_update_used_packed(vq, async->buffers_packed + from, to - from);
1990                         async->last_buffer_idx_packed += nr_left;
1991                         nr_left = 0;
1992                 } else {
1993                         vhost_update_used_packed(vq, async->buffers_packed + from,
1994                                 vq->size - from);
1995                         async->last_buffer_idx_packed = 0;
1996                         nr_left -= vq->size - from;
1997                 }
1998         } while (nr_left > 0);
1999 }
2000
2001 static __rte_always_inline uint16_t
2002 vhost_poll_enqueue_completed(struct virtio_net *dev, uint16_t queue_id,
2003                 struct rte_mbuf **pkts, uint16_t count)
2004 {
2005         struct vhost_virtqueue *vq;
2006         struct vhost_async *async;
2007         struct async_inflight_info *pkts_info;
2008         int32_t n_cpl;
2009         uint16_t n_pkts_cpl = 0, n_pkts_put = 0, n_descs = 0, n_buffers = 0;
2010         uint16_t start_idx, pkts_idx, vq_size;
2011         uint16_t from, i;
2012
2013         vq = dev->virtqueue[queue_id];
2014         async = vq->async;
2015         pkts_idx = async->pkts_idx % vq->size;
2016         pkts_info = async->pkts_info;
2017         vq_size = vq->size;
2018         start_idx = virtio_dev_rx_async_get_info_idx(pkts_idx,
2019                 vq_size, async->pkts_inflight_n);
2020
2021         if (count > async->last_pkts_n) {
2022                 n_cpl = async->ops.check_completed_copies(dev->vid,
2023                         queue_id, 0, count - async->last_pkts_n);
2024                 if (likely(n_cpl >= 0)) {
2025                         n_pkts_cpl = n_cpl;
2026                 } else {
2027                         VHOST_LOG_DATA(ERR,
2028                                 "(%d) %s: failed to check completed copies for queue id %d.\n",
2029                                 dev->vid, __func__, queue_id);
2030                         n_pkts_cpl = 0;
2031                 }
2032         }
2033
2034         n_pkts_cpl += async->last_pkts_n;
2035         n_pkts_put = RTE_MIN(n_pkts_cpl, count);
2036         if (unlikely(n_pkts_put == 0)) {
2037                 async->last_pkts_n = n_pkts_cpl;
2038                 return 0;
2039         }
2040
2041         if (vq_is_packed(dev)) {
2042                 for (i = 0; i < n_pkts_put; i++) {
2043                         from = (start_idx + i) % vq_size;
2044                         n_buffers += pkts_info[from].nr_buffers;
2045                         pkts[i] = pkts_info[from].mbuf;
2046                 }
2047         } else {
2048                 for (i = 0; i < n_pkts_put; i++) {
2049                         from = (start_idx + i) & (vq_size - 1);
2050                         n_descs += pkts_info[from].descs;
2051                         pkts[i] = pkts_info[from].mbuf;
2052                 }
2053         }
2054         async->last_pkts_n = n_pkts_cpl - n_pkts_put;
2055         async->pkts_inflight_n -= n_pkts_put;
2056
2057         if (likely(vq->enabled && vq->access_ok)) {
2058                 if (vq_is_packed(dev)) {
2059                         write_back_completed_descs_packed(vq, n_buffers);
2060
2061                         vhost_vring_call_packed(dev, vq);
2062                 } else {
2063                         write_back_completed_descs_split(vq, n_descs);
2064
2065                         __atomic_add_fetch(&vq->used->idx, n_descs,
2066                                         __ATOMIC_RELEASE);
2067                         vhost_vring_call_split(dev, vq);
2068                 }
2069         } else {
2070                 if (vq_is_packed(dev)) {
2071                         async->last_buffer_idx_packed += n_buffers;
2072                         if (async->last_buffer_idx_packed >= vq->size)
2073                                 async->last_buffer_idx_packed -= vq->size;
2074                 } else {
2075                         async->last_desc_idx_split += n_descs;
2076                 }
2077         }
2078
2079         return n_pkts_put;
2080 }
2081
2082 uint16_t
2083 rte_vhost_poll_enqueue_completed(int vid, uint16_t queue_id,
2084                 struct rte_mbuf **pkts, uint16_t count)
2085 {
2086         struct virtio_net *dev = get_device(vid);
2087         struct vhost_virtqueue *vq;
2088         uint16_t n_pkts_cpl = 0;
2089
2090         if (unlikely(!dev))
2091                 return 0;
2092
2093         VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
2094         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
2095                 VHOST_LOG_DATA(ERR, "(%d) %s: invalid virtqueue idx %d.\n",
2096                         dev->vid, __func__, queue_id);
2097                 return 0;
2098         }
2099
2100         vq = dev->virtqueue[queue_id];
2101
2102         if (unlikely(!vq->async)) {
2103                 VHOST_LOG_DATA(ERR, "(%d) %s: async not registered for queue id %d.\n",
2104                         dev->vid, __func__, queue_id);
2105                 return 0;
2106         }
2107
2108         rte_spinlock_lock(&vq->access_lock);
2109
2110         n_pkts_cpl = vhost_poll_enqueue_completed(dev, queue_id, pkts, count);
2111
2112         rte_spinlock_unlock(&vq->access_lock);
2113
2114         return n_pkts_cpl;
2115 }
2116
2117 uint16_t
2118 rte_vhost_clear_queue_thread_unsafe(int vid, uint16_t queue_id,
2119                 struct rte_mbuf **pkts, uint16_t count)
2120 {
2121         struct virtio_net *dev = get_device(vid);
2122         struct vhost_virtqueue *vq;
2123         uint16_t n_pkts_cpl = 0;
2124
2125         if (!dev)
2126                 return 0;
2127
2128         VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
2129         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
2130                 VHOST_LOG_DATA(ERR, "(%d) %s: invalid virtqueue idx %d.\n",
2131                         dev->vid, __func__, queue_id);
2132                 return 0;
2133         }
2134
2135         vq = dev->virtqueue[queue_id];
2136
2137         if (unlikely(!vq->async)) {
2138                 VHOST_LOG_DATA(ERR, "(%d) %s: async not registered for queue id %d.\n",
2139                         dev->vid, __func__, queue_id);
2140                 return 0;
2141         }
2142
2143         n_pkts_cpl = vhost_poll_enqueue_completed(dev, queue_id, pkts, count);
2144
2145         return n_pkts_cpl;
2146 }
2147
2148 static __rte_always_inline uint32_t
2149 virtio_dev_rx_async_submit(struct virtio_net *dev, uint16_t queue_id,
2150         struct rte_mbuf **pkts, uint32_t count)
2151 {
2152         struct vhost_virtqueue *vq;
2153         uint32_t nb_tx = 0;
2154
2155         VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
2156         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
2157                 VHOST_LOG_DATA(ERR, "(%d) %s: invalid virtqueue idx %d.\n",
2158                         dev->vid, __func__, queue_id);
2159                 return 0;
2160         }
2161
2162         vq = dev->virtqueue[queue_id];
2163
2164         rte_spinlock_lock(&vq->access_lock);
2165
2166         if (unlikely(!vq->enabled || !vq->async))
2167                 goto out_access_unlock;
2168
2169         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
2170                 vhost_user_iotlb_rd_lock(vq);
2171
2172         if (unlikely(!vq->access_ok))
2173                 if (unlikely(vring_translate(dev, vq) < 0))
2174                         goto out;
2175
2176         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
2177         if (count == 0)
2178                 goto out;
2179
2180         if (vq_is_packed(dev))
2181                 nb_tx = virtio_dev_rx_async_submit_packed(dev, vq, queue_id,
2182                                 pkts, count);
2183         else
2184                 nb_tx = virtio_dev_rx_async_submit_split(dev, vq, queue_id,
2185                                 pkts, count);
2186
2187 out:
2188         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
2189                 vhost_user_iotlb_rd_unlock(vq);
2190
2191 out_access_unlock:
2192         rte_spinlock_unlock(&vq->access_lock);
2193
2194         return nb_tx;
2195 }
2196
2197 uint16_t
2198 rte_vhost_submit_enqueue_burst(int vid, uint16_t queue_id,
2199                 struct rte_mbuf **pkts, uint16_t count)
2200 {
2201         struct virtio_net *dev = get_device(vid);
2202
2203         if (!dev)
2204                 return 0;
2205
2206         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
2207                 VHOST_LOG_DATA(ERR,
2208                         "(%d) %s: built-in vhost net backend is disabled.\n",
2209                         dev->vid, __func__);
2210                 return 0;
2211         }
2212
2213         return virtio_dev_rx_async_submit(dev, queue_id, pkts, count);
2214 }
2215
2216 static inline bool
2217 virtio_net_with_host_offload(struct virtio_net *dev)
2218 {
2219         if (dev->features &
2220                         ((1ULL << VIRTIO_NET_F_CSUM) |
2221                          (1ULL << VIRTIO_NET_F_HOST_ECN) |
2222                          (1ULL << VIRTIO_NET_F_HOST_TSO4) |
2223                          (1ULL << VIRTIO_NET_F_HOST_TSO6) |
2224                          (1ULL << VIRTIO_NET_F_HOST_UFO)))
2225                 return true;
2226
2227         return false;
2228 }
2229
2230 static int
2231 parse_headers(struct rte_mbuf *m, uint8_t *l4_proto)
2232 {
2233         struct rte_ipv4_hdr *ipv4_hdr;
2234         struct rte_ipv6_hdr *ipv6_hdr;
2235         struct rte_ether_hdr *eth_hdr;
2236         uint16_t ethertype;
2237         uint16_t data_len = rte_pktmbuf_data_len(m);
2238
2239         if (data_len < sizeof(struct rte_ether_hdr))
2240                 return -EINVAL;
2241
2242         eth_hdr = rte_pktmbuf_mtod(m, struct rte_ether_hdr *);
2243
2244         m->l2_len = sizeof(struct rte_ether_hdr);
2245         ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
2246
2247         if (ethertype == RTE_ETHER_TYPE_VLAN) {
2248                 if (data_len < sizeof(struct rte_ether_hdr) +
2249                                 sizeof(struct rte_vlan_hdr))
2250                         goto error;
2251
2252                 struct rte_vlan_hdr *vlan_hdr =
2253                         (struct rte_vlan_hdr *)(eth_hdr + 1);
2254
2255                 m->l2_len += sizeof(struct rte_vlan_hdr);
2256                 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
2257         }
2258
2259         switch (ethertype) {
2260         case RTE_ETHER_TYPE_IPV4:
2261                 if (data_len < m->l2_len + sizeof(struct rte_ipv4_hdr))
2262                         goto error;
2263                 ipv4_hdr = rte_pktmbuf_mtod_offset(m, struct rte_ipv4_hdr *,
2264                                 m->l2_len);
2265                 m->l3_len = rte_ipv4_hdr_len(ipv4_hdr);
2266                 if (data_len < m->l2_len + m->l3_len)
2267                         goto error;
2268                 m->ol_flags |= RTE_MBUF_F_TX_IPV4;
2269                 *l4_proto = ipv4_hdr->next_proto_id;
2270                 break;
2271         case RTE_ETHER_TYPE_IPV6:
2272                 if (data_len < m->l2_len + sizeof(struct rte_ipv6_hdr))
2273                         goto error;
2274                 ipv6_hdr = rte_pktmbuf_mtod_offset(m, struct rte_ipv6_hdr *,
2275                                 m->l2_len);
2276                 m->l3_len = sizeof(struct rte_ipv6_hdr);
2277                 m->ol_flags |= RTE_MBUF_F_TX_IPV6;
2278                 *l4_proto = ipv6_hdr->proto;
2279                 break;
2280         default:
2281                 /* a valid L3 header is needed for further L4 parsing */
2282                 goto error;
2283         }
2284
2285         /* both CSUM and GSO need a valid L4 header */
2286         switch (*l4_proto) {
2287         case IPPROTO_TCP:
2288                 if (data_len < m->l2_len + m->l3_len +
2289                                 sizeof(struct rte_tcp_hdr))
2290                         goto error;
2291                 break;
2292         case IPPROTO_UDP:
2293                 if (data_len < m->l2_len + m->l3_len +
2294                                 sizeof(struct rte_udp_hdr))
2295                         goto error;
2296                 break;
2297         case IPPROTO_SCTP:
2298                 if (data_len < m->l2_len + m->l3_len +
2299                                 sizeof(struct rte_sctp_hdr))
2300                         goto error;
2301                 break;
2302         default:
2303                 goto error;
2304         }
2305
2306         return 0;
2307
2308 error:
2309         m->l2_len = 0;
2310         m->l3_len = 0;
2311         m->ol_flags = 0;
2312         return -EINVAL;
2313 }
2314
2315 static __rte_always_inline void
2316 vhost_dequeue_offload_legacy(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
2317 {
2318         uint8_t l4_proto = 0;
2319         struct rte_tcp_hdr *tcp_hdr = NULL;
2320         uint16_t tcp_len;
2321         uint16_t data_len = rte_pktmbuf_data_len(m);
2322
2323         if (parse_headers(m, &l4_proto) < 0)
2324                 return;
2325
2326         if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
2327                 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
2328                         switch (hdr->csum_offset) {
2329                         case (offsetof(struct rte_tcp_hdr, cksum)):
2330                                 if (l4_proto != IPPROTO_TCP)
2331                                         goto error;
2332                                 m->ol_flags |= RTE_MBUF_F_TX_TCP_CKSUM;
2333                                 break;
2334                         case (offsetof(struct rte_udp_hdr, dgram_cksum)):
2335                                 if (l4_proto != IPPROTO_UDP)
2336                                         goto error;
2337                                 m->ol_flags |= RTE_MBUF_F_TX_UDP_CKSUM;
2338                                 break;
2339                         case (offsetof(struct rte_sctp_hdr, cksum)):
2340                                 if (l4_proto != IPPROTO_SCTP)
2341                                         goto error;
2342                                 m->ol_flags |= RTE_MBUF_F_TX_SCTP_CKSUM;
2343                                 break;
2344                         default:
2345                                 goto error;
2346                         }
2347                 } else {
2348                         goto error;
2349                 }
2350         }
2351
2352         if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
2353                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
2354                 case VIRTIO_NET_HDR_GSO_TCPV4:
2355                 case VIRTIO_NET_HDR_GSO_TCPV6:
2356                         if (l4_proto != IPPROTO_TCP)
2357                                 goto error;
2358                         tcp_hdr = rte_pktmbuf_mtod_offset(m,
2359                                         struct rte_tcp_hdr *,
2360                                         m->l2_len + m->l3_len);
2361                         tcp_len = (tcp_hdr->data_off & 0xf0) >> 2;
2362                         if (data_len < m->l2_len + m->l3_len + tcp_len)
2363                                 goto error;
2364                         m->ol_flags |= RTE_MBUF_F_TX_TCP_SEG;
2365                         m->tso_segsz = hdr->gso_size;
2366                         m->l4_len = tcp_len;
2367                         break;
2368                 case VIRTIO_NET_HDR_GSO_UDP:
2369                         if (l4_proto != IPPROTO_UDP)
2370                                 goto error;
2371                         m->ol_flags |= RTE_MBUF_F_TX_UDP_SEG;
2372                         m->tso_segsz = hdr->gso_size;
2373                         m->l4_len = sizeof(struct rte_udp_hdr);
2374                         break;
2375                 default:
2376                         VHOST_LOG_DATA(WARNING,
2377                                 "unsupported gso type %u.\n", hdr->gso_type);
2378                         goto error;
2379                 }
2380         }
2381         return;
2382
2383 error:
2384         m->l2_len = 0;
2385         m->l3_len = 0;
2386         m->ol_flags = 0;
2387 }
2388
2389 static __rte_always_inline void
2390 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m,
2391         bool legacy_ol_flags)
2392 {
2393         struct rte_net_hdr_lens hdr_lens;
2394         int l4_supported = 0;
2395         uint32_t ptype;
2396
2397         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
2398                 return;
2399
2400         if (legacy_ol_flags) {
2401                 vhost_dequeue_offload_legacy(hdr, m);
2402                 return;
2403         }
2404
2405         m->ol_flags |= RTE_MBUF_F_RX_IP_CKSUM_UNKNOWN;
2406
2407         ptype = rte_net_get_ptype(m, &hdr_lens, RTE_PTYPE_ALL_MASK);
2408         m->packet_type = ptype;
2409         if ((ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_TCP ||
2410             (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_UDP ||
2411             (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_SCTP)
2412                 l4_supported = 1;
2413
2414         /* According to Virtio 1.1 spec, the device only needs to look at
2415          * VIRTIO_NET_HDR_F_NEEDS_CSUM in the packet transmission path.
2416          * This differs from the processing incoming packets path where the
2417          * driver could rely on VIRTIO_NET_HDR_F_DATA_VALID flag set by the
2418          * device.
2419          *
2420          * 5.1.6.2.1 Driver Requirements: Packet Transmission
2421          * The driver MUST NOT set the VIRTIO_NET_HDR_F_DATA_VALID and
2422          * VIRTIO_NET_HDR_F_RSC_INFO bits in flags.
2423          *
2424          * 5.1.6.2.2 Device Requirements: Packet Transmission
2425          * The device MUST ignore flag bits that it does not recognize.
2426          */
2427         if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
2428                 uint32_t hdrlen;
2429
2430                 hdrlen = hdr_lens.l2_len + hdr_lens.l3_len + hdr_lens.l4_len;
2431                 if (hdr->csum_start <= hdrlen && l4_supported != 0) {
2432                         m->ol_flags |= RTE_MBUF_F_RX_L4_CKSUM_NONE;
2433                 } else {
2434                         /* Unknown proto or tunnel, do sw cksum. We can assume
2435                          * the cksum field is in the first segment since the
2436                          * buffers we provided to the host are large enough.
2437                          * In case of SCTP, this will be wrong since it's a CRC
2438                          * but there's nothing we can do.
2439                          */
2440                         uint16_t csum = 0, off;
2441
2442                         if (rte_raw_cksum_mbuf(m, hdr->csum_start,
2443                                         rte_pktmbuf_pkt_len(m) - hdr->csum_start, &csum) < 0)
2444                                 return;
2445                         if (likely(csum != 0xffff))
2446                                 csum = ~csum;
2447                         off = hdr->csum_offset + hdr->csum_start;
2448                         if (rte_pktmbuf_data_len(m) >= off + 1)
2449                                 *rte_pktmbuf_mtod_offset(m, uint16_t *, off) = csum;
2450                 }
2451         }
2452
2453         if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
2454                 if (hdr->gso_size == 0)
2455                         return;
2456
2457                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
2458                 case VIRTIO_NET_HDR_GSO_TCPV4:
2459                 case VIRTIO_NET_HDR_GSO_TCPV6:
2460                         if ((ptype & RTE_PTYPE_L4_MASK) != RTE_PTYPE_L4_TCP)
2461                                 break;
2462                         m->ol_flags |= RTE_MBUF_F_RX_LRO | RTE_MBUF_F_RX_L4_CKSUM_NONE;
2463                         m->tso_segsz = hdr->gso_size;
2464                         break;
2465                 case VIRTIO_NET_HDR_GSO_UDP:
2466                         if ((ptype & RTE_PTYPE_L4_MASK) != RTE_PTYPE_L4_UDP)
2467                                 break;
2468                         m->ol_flags |= RTE_MBUF_F_RX_LRO | RTE_MBUF_F_RX_L4_CKSUM_NONE;
2469                         m->tso_segsz = hdr->gso_size;
2470                         break;
2471                 default:
2472                         break;
2473                 }
2474         }
2475 }
2476
2477 static __rte_noinline void
2478 copy_vnet_hdr_from_desc(struct virtio_net_hdr *hdr,
2479                 struct buf_vector *buf_vec)
2480 {
2481         uint64_t len;
2482         uint64_t remain = sizeof(struct virtio_net_hdr);
2483         uint64_t src;
2484         uint64_t dst = (uint64_t)(uintptr_t)hdr;
2485
2486         while (remain) {
2487                 len = RTE_MIN(remain, buf_vec->buf_len);
2488                 src = buf_vec->buf_addr;
2489                 rte_memcpy((void *)(uintptr_t)dst,
2490                                 (void *)(uintptr_t)src, len);
2491
2492                 remain -= len;
2493                 dst += len;
2494                 buf_vec++;
2495         }
2496 }
2497
2498 static __rte_always_inline int
2499 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
2500                   struct buf_vector *buf_vec, uint16_t nr_vec,
2501                   struct rte_mbuf *m, struct rte_mempool *mbuf_pool,
2502                   bool legacy_ol_flags)
2503 {
2504         uint32_t buf_avail, buf_offset;
2505         uint64_t buf_addr, buf_len;
2506         uint32_t mbuf_avail, mbuf_offset;
2507         uint32_t cpy_len;
2508         struct rte_mbuf *cur = m, *prev = m;
2509         struct virtio_net_hdr tmp_hdr;
2510         struct virtio_net_hdr *hdr = NULL;
2511         /* A counter to avoid desc dead loop chain */
2512         uint16_t vec_idx = 0;
2513         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
2514         int error = 0;
2515
2516         buf_addr = buf_vec[vec_idx].buf_addr;
2517         buf_len = buf_vec[vec_idx].buf_len;
2518
2519         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
2520                 error = -1;
2521                 goto out;
2522         }
2523
2524         if (virtio_net_with_host_offload(dev)) {
2525                 if (unlikely(buf_len < sizeof(struct virtio_net_hdr))) {
2526                         /*
2527                          * No luck, the virtio-net header doesn't fit
2528                          * in a contiguous virtual area.
2529                          */
2530                         copy_vnet_hdr_from_desc(&tmp_hdr, buf_vec);
2531                         hdr = &tmp_hdr;
2532                 } else {
2533                         hdr = (struct virtio_net_hdr *)((uintptr_t)buf_addr);
2534                 }
2535         }
2536
2537         /*
2538          * A virtio driver normally uses at least 2 desc buffers
2539          * for Tx: the first for storing the header, and others
2540          * for storing the data.
2541          */
2542         if (unlikely(buf_len < dev->vhost_hlen)) {
2543                 buf_offset = dev->vhost_hlen - buf_len;
2544                 vec_idx++;
2545                 buf_addr = buf_vec[vec_idx].buf_addr;
2546                 buf_len = buf_vec[vec_idx].buf_len;
2547                 buf_avail  = buf_len - buf_offset;
2548         } else if (buf_len == dev->vhost_hlen) {
2549                 if (unlikely(++vec_idx >= nr_vec))
2550                         goto out;
2551                 buf_addr = buf_vec[vec_idx].buf_addr;
2552                 buf_len = buf_vec[vec_idx].buf_len;
2553
2554                 buf_offset = 0;
2555                 buf_avail = buf_len;
2556         } else {
2557                 buf_offset = dev->vhost_hlen;
2558                 buf_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
2559         }
2560
2561         PRINT_PACKET(dev,
2562                         (uintptr_t)(buf_addr + buf_offset),
2563                         (uint32_t)buf_avail, 0);
2564
2565         mbuf_offset = 0;
2566         mbuf_avail  = m->buf_len - RTE_PKTMBUF_HEADROOM;
2567         while (1) {
2568                 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
2569
2570                 if (likely(cpy_len > MAX_BATCH_LEN ||
2571                                         vq->batch_copy_nb_elems >= vq->size ||
2572                                         (hdr && cur == m))) {
2573                         rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
2574                                                 mbuf_offset),
2575                                         (void *)((uintptr_t)(buf_addr +
2576                                                         buf_offset)), cpy_len);
2577                 } else {
2578                         batch_copy[vq->batch_copy_nb_elems].dst =
2579                                 rte_pktmbuf_mtod_offset(cur, void *,
2580                                                 mbuf_offset);
2581                         batch_copy[vq->batch_copy_nb_elems].src =
2582                                 (void *)((uintptr_t)(buf_addr + buf_offset));
2583                         batch_copy[vq->batch_copy_nb_elems].len = cpy_len;
2584                         vq->batch_copy_nb_elems++;
2585                 }
2586
2587                 mbuf_avail  -= cpy_len;
2588                 mbuf_offset += cpy_len;
2589                 buf_avail -= cpy_len;
2590                 buf_offset += cpy_len;
2591
2592                 /* This buf reaches to its end, get the next one */
2593                 if (buf_avail == 0) {
2594                         if (++vec_idx >= nr_vec)
2595                                 break;
2596
2597                         buf_addr = buf_vec[vec_idx].buf_addr;
2598                         buf_len = buf_vec[vec_idx].buf_len;
2599
2600                         buf_offset = 0;
2601                         buf_avail  = buf_len;
2602
2603                         PRINT_PACKET(dev, (uintptr_t)buf_addr,
2604                                         (uint32_t)buf_avail, 0);
2605                 }
2606
2607                 /*
2608                  * This mbuf reaches to its end, get a new one
2609                  * to hold more data.
2610                  */
2611                 if (mbuf_avail == 0) {
2612                         cur = rte_pktmbuf_alloc(mbuf_pool);
2613                         if (unlikely(cur == NULL)) {
2614                                 VHOST_LOG_DATA(ERR, "Failed to "
2615                                         "allocate memory for mbuf.\n");
2616                                 error = -1;
2617                                 goto out;
2618                         }
2619
2620                         prev->next = cur;
2621                         prev->data_len = mbuf_offset;
2622                         m->nb_segs += 1;
2623                         m->pkt_len += mbuf_offset;
2624                         prev = cur;
2625
2626                         mbuf_offset = 0;
2627                         mbuf_avail  = cur->buf_len - RTE_PKTMBUF_HEADROOM;
2628                 }
2629         }
2630
2631         prev->data_len = mbuf_offset;
2632         m->pkt_len    += mbuf_offset;
2633
2634         if (hdr)
2635                 vhost_dequeue_offload(hdr, m, legacy_ol_flags);
2636
2637 out:
2638
2639         return error;
2640 }
2641
2642 static void
2643 virtio_dev_extbuf_free(void *addr __rte_unused, void *opaque)
2644 {
2645         rte_free(opaque);
2646 }
2647
2648 static int
2649 virtio_dev_extbuf_alloc(struct rte_mbuf *pkt, uint32_t size)
2650 {
2651         struct rte_mbuf_ext_shared_info *shinfo = NULL;
2652         uint32_t total_len = RTE_PKTMBUF_HEADROOM + size;
2653         uint16_t buf_len;
2654         rte_iova_t iova;
2655         void *buf;
2656
2657         total_len += sizeof(*shinfo) + sizeof(uintptr_t);
2658         total_len = RTE_ALIGN_CEIL(total_len, sizeof(uintptr_t));
2659
2660         if (unlikely(total_len > UINT16_MAX))
2661                 return -ENOSPC;
2662
2663         buf_len = total_len;
2664         buf = rte_malloc(NULL, buf_len, RTE_CACHE_LINE_SIZE);
2665         if (unlikely(buf == NULL))
2666                 return -ENOMEM;
2667
2668         /* Initialize shinfo */
2669         shinfo = rte_pktmbuf_ext_shinfo_init_helper(buf, &buf_len,
2670                                                 virtio_dev_extbuf_free, buf);
2671         if (unlikely(shinfo == NULL)) {
2672                 rte_free(buf);
2673                 VHOST_LOG_DATA(ERR, "Failed to init shinfo\n");
2674                 return -1;
2675         }
2676
2677         iova = rte_malloc_virt2iova(buf);
2678         rte_pktmbuf_attach_extbuf(pkt, buf, iova, buf_len, shinfo);
2679         rte_pktmbuf_reset_headroom(pkt);
2680
2681         return 0;
2682 }
2683
2684 /*
2685  * Prepare a host supported pktmbuf.
2686  */
2687 static __rte_always_inline int
2688 virtio_dev_pktmbuf_prep(struct virtio_net *dev, struct rte_mbuf *pkt,
2689                          uint32_t data_len)
2690 {
2691         if (rte_pktmbuf_tailroom(pkt) >= data_len)
2692                 return 0;
2693
2694         /* attach an external buffer if supported */
2695         if (dev->extbuf && !virtio_dev_extbuf_alloc(pkt, data_len))
2696                 return 0;
2697
2698         /* check if chained buffers are allowed */
2699         if (!dev->linearbuf)
2700                 return 0;
2701
2702         return -1;
2703 }
2704
2705 __rte_always_inline
2706 static uint16_t
2707 virtio_dev_tx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
2708         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count,
2709         bool legacy_ol_flags)
2710 {
2711         uint16_t i;
2712         uint16_t free_entries;
2713         uint16_t dropped = 0;
2714         static bool allocerr_warned;
2715
2716         /*
2717          * The ordering between avail index and
2718          * desc reads needs to be enforced.
2719          */
2720         free_entries = __atomic_load_n(&vq->avail->idx, __ATOMIC_ACQUIRE) -
2721                         vq->last_avail_idx;
2722         if (free_entries == 0)
2723                 return 0;
2724
2725         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
2726
2727         VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
2728
2729         count = RTE_MIN(count, MAX_PKT_BURST);
2730         count = RTE_MIN(count, free_entries);
2731         VHOST_LOG_DATA(DEBUG, "(%d) about to dequeue %u buffers\n",
2732                         dev->vid, count);
2733
2734         if (rte_pktmbuf_alloc_bulk(mbuf_pool, pkts, count))
2735                 return 0;
2736
2737         for (i = 0; i < count; i++) {
2738                 struct buf_vector buf_vec[BUF_VECTOR_MAX];
2739                 uint16_t head_idx;
2740                 uint32_t buf_len;
2741                 uint16_t nr_vec = 0;
2742                 int err;
2743
2744                 if (unlikely(fill_vec_buf_split(dev, vq,
2745                                                 vq->last_avail_idx + i,
2746                                                 &nr_vec, buf_vec,
2747                                                 &head_idx, &buf_len,
2748                                                 VHOST_ACCESS_RO) < 0))
2749                         break;
2750
2751                 update_shadow_used_ring_split(vq, head_idx, 0);
2752
2753                 err = virtio_dev_pktmbuf_prep(dev, pkts[i], buf_len);
2754                 if (unlikely(err)) {
2755                         /*
2756                          * mbuf allocation fails for jumbo packets when external
2757                          * buffer allocation is not allowed and linear buffer
2758                          * is required. Drop this packet.
2759                          */
2760                         if (!allocerr_warned) {
2761                                 VHOST_LOG_DATA(ERR,
2762                                         "Failed mbuf alloc of size %d from %s on %s.\n",
2763                                         buf_len, mbuf_pool->name, dev->ifname);
2764                                 allocerr_warned = true;
2765                         }
2766                         dropped += 1;
2767                         i++;
2768                         break;
2769                 }
2770
2771                 err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts[i],
2772                                 mbuf_pool, legacy_ol_flags);
2773                 if (unlikely(err)) {
2774                         if (!allocerr_warned) {
2775                                 VHOST_LOG_DATA(ERR,
2776                                         "Failed to copy desc to mbuf on %s.\n",
2777                                         dev->ifname);
2778                                 allocerr_warned = true;
2779                         }
2780                         dropped += 1;
2781                         i++;
2782                         break;
2783                 }
2784         }
2785
2786         if (dropped)
2787                 rte_pktmbuf_free_bulk(&pkts[i - 1], count - i + 1);
2788
2789         vq->last_avail_idx += i;
2790
2791         do_data_copy_dequeue(vq);
2792         if (unlikely(i < count))
2793                 vq->shadow_used_idx = i;
2794         if (likely(vq->shadow_used_idx)) {
2795                 flush_shadow_used_ring_split(dev, vq);
2796                 vhost_vring_call_split(dev, vq);
2797         }
2798
2799         return (i - dropped);
2800 }
2801
2802 __rte_noinline
2803 static uint16_t
2804 virtio_dev_tx_split_legacy(struct virtio_net *dev,
2805         struct vhost_virtqueue *vq, struct rte_mempool *mbuf_pool,
2806         struct rte_mbuf **pkts, uint16_t count)
2807 {
2808         return virtio_dev_tx_split(dev, vq, mbuf_pool, pkts, count, true);
2809 }
2810
2811 __rte_noinline
2812 static uint16_t
2813 virtio_dev_tx_split_compliant(struct virtio_net *dev,
2814         struct vhost_virtqueue *vq, struct rte_mempool *mbuf_pool,
2815         struct rte_mbuf **pkts, uint16_t count)
2816 {
2817         return virtio_dev_tx_split(dev, vq, mbuf_pool, pkts, count, false);
2818 }
2819
2820 static __rte_always_inline int
2821 vhost_reserve_avail_batch_packed(struct virtio_net *dev,
2822                                  struct vhost_virtqueue *vq,
2823                                  struct rte_mbuf **pkts,
2824                                  uint16_t avail_idx,
2825                                  uintptr_t *desc_addrs,
2826                                  uint16_t *ids)
2827 {
2828         bool wrap = vq->avail_wrap_counter;
2829         struct vring_packed_desc *descs = vq->desc_packed;
2830         uint64_t lens[PACKED_BATCH_SIZE];
2831         uint64_t buf_lens[PACKED_BATCH_SIZE];
2832         uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2833         uint16_t flags, i;
2834
2835         if (unlikely(avail_idx & PACKED_BATCH_MASK))
2836                 return -1;
2837         if (unlikely((avail_idx + PACKED_BATCH_SIZE) > vq->size))
2838                 return -1;
2839
2840         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2841                 flags = descs[avail_idx + i].flags;
2842                 if (unlikely((wrap != !!(flags & VRING_DESC_F_AVAIL)) ||
2843                              (wrap == !!(flags & VRING_DESC_F_USED))  ||
2844                              (flags & PACKED_DESC_SINGLE_DEQUEUE_FLAG)))
2845                         return -1;
2846         }
2847
2848         rte_atomic_thread_fence(__ATOMIC_ACQUIRE);
2849
2850         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2851                 lens[i] = descs[avail_idx + i].len;
2852
2853         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2854                 desc_addrs[i] = vhost_iova_to_vva(dev, vq,
2855                                                   descs[avail_idx + i].addr,
2856                                                   &lens[i], VHOST_ACCESS_RW);
2857         }
2858
2859         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2860                 if (unlikely(!desc_addrs[i]))
2861                         return -1;
2862                 if (unlikely((lens[i] != descs[avail_idx + i].len)))
2863                         return -1;
2864         }
2865
2866         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2867                 if (virtio_dev_pktmbuf_prep(dev, pkts[i], lens[i]))
2868                         goto err;
2869         }
2870
2871         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2872                 buf_lens[i] = pkts[i]->buf_len - pkts[i]->data_off;
2873
2874         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2875                 if (unlikely(buf_lens[i] < (lens[i] - buf_offset)))
2876                         goto err;
2877         }
2878
2879         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2880                 pkts[i]->pkt_len = lens[i] - buf_offset;
2881                 pkts[i]->data_len = pkts[i]->pkt_len;
2882                 ids[i] = descs[avail_idx + i].id;
2883         }
2884
2885         return 0;
2886
2887 err:
2888         return -1;
2889 }
2890
2891 static __rte_always_inline int
2892 virtio_dev_tx_batch_packed(struct virtio_net *dev,
2893                            struct vhost_virtqueue *vq,
2894                            struct rte_mbuf **pkts,
2895                            bool legacy_ol_flags)
2896 {
2897         uint16_t avail_idx = vq->last_avail_idx;
2898         uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2899         struct virtio_net_hdr *hdr;
2900         uintptr_t desc_addrs[PACKED_BATCH_SIZE];
2901         uint16_t ids[PACKED_BATCH_SIZE];
2902         uint16_t i;
2903
2904         if (vhost_reserve_avail_batch_packed(dev, vq, pkts, avail_idx,
2905                                              desc_addrs, ids))
2906                 return -1;
2907
2908         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2909                 rte_prefetch0((void *)(uintptr_t)desc_addrs[i]);
2910
2911         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2912                 rte_memcpy(rte_pktmbuf_mtod_offset(pkts[i], void *, 0),
2913                            (void *)(uintptr_t)(desc_addrs[i] + buf_offset),
2914                            pkts[i]->pkt_len);
2915
2916         if (virtio_net_with_host_offload(dev)) {
2917                 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2918                         hdr = (struct virtio_net_hdr *)(desc_addrs[i]);
2919                         vhost_dequeue_offload(hdr, pkts[i], legacy_ol_flags);
2920                 }
2921         }
2922
2923         if (virtio_net_is_inorder(dev))
2924                 vhost_shadow_dequeue_batch_packed_inorder(vq,
2925                         ids[PACKED_BATCH_SIZE - 1]);
2926         else
2927                 vhost_shadow_dequeue_batch_packed(dev, vq, ids);
2928
2929         vq_inc_last_avail_packed(vq, PACKED_BATCH_SIZE);
2930
2931         return 0;
2932 }
2933
2934 static __rte_always_inline int
2935 vhost_dequeue_single_packed(struct virtio_net *dev,
2936                             struct vhost_virtqueue *vq,
2937                             struct rte_mempool *mbuf_pool,
2938                             struct rte_mbuf *pkts,
2939                             uint16_t *buf_id,
2940                             uint16_t *desc_count,
2941                             bool legacy_ol_flags)
2942 {
2943         struct buf_vector buf_vec[BUF_VECTOR_MAX];
2944         uint32_t buf_len;
2945         uint16_t nr_vec = 0;
2946         int err;
2947         static bool allocerr_warned;
2948
2949         if (unlikely(fill_vec_buf_packed(dev, vq,
2950                                          vq->last_avail_idx, desc_count,
2951                                          buf_vec, &nr_vec,
2952                                          buf_id, &buf_len,
2953                                          VHOST_ACCESS_RO) < 0))
2954                 return -1;
2955
2956         if (unlikely(virtio_dev_pktmbuf_prep(dev, pkts, buf_len))) {
2957                 if (!allocerr_warned) {
2958                         VHOST_LOG_DATA(ERR,
2959                                 "Failed mbuf alloc of size %d from %s on %s.\n",
2960                                 buf_len, mbuf_pool->name, dev->ifname);
2961                         allocerr_warned = true;
2962                 }
2963                 return -1;
2964         }
2965
2966         err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts,
2967                                 mbuf_pool, legacy_ol_flags);
2968         if (unlikely(err)) {
2969                 if (!allocerr_warned) {
2970                         VHOST_LOG_DATA(ERR,
2971                                 "Failed to copy desc to mbuf on %s.\n",
2972                                 dev->ifname);
2973                         allocerr_warned = true;
2974                 }
2975                 return -1;
2976         }
2977
2978         return 0;
2979 }
2980
2981 static __rte_always_inline int
2982 virtio_dev_tx_single_packed(struct virtio_net *dev,
2983                             struct vhost_virtqueue *vq,
2984                             struct rte_mempool *mbuf_pool,
2985                             struct rte_mbuf *pkts,
2986                             bool legacy_ol_flags)
2987 {
2988
2989         uint16_t buf_id, desc_count = 0;
2990         int ret;
2991
2992         ret = vhost_dequeue_single_packed(dev, vq, mbuf_pool, pkts, &buf_id,
2993                                         &desc_count, legacy_ol_flags);
2994
2995         if (likely(desc_count > 0)) {
2996                 if (virtio_net_is_inorder(dev))
2997                         vhost_shadow_dequeue_single_packed_inorder(vq, buf_id,
2998                                                                    desc_count);
2999                 else
3000                         vhost_shadow_dequeue_single_packed(vq, buf_id,
3001                                         desc_count);
3002
3003                 vq_inc_last_avail_packed(vq, desc_count);
3004         }
3005
3006         return ret;
3007 }
3008
3009 __rte_always_inline
3010 static uint16_t
3011 virtio_dev_tx_packed(struct virtio_net *dev,
3012                      struct vhost_virtqueue *__rte_restrict vq,
3013                      struct rte_mempool *mbuf_pool,
3014                      struct rte_mbuf **__rte_restrict pkts,
3015                      uint32_t count,
3016                      bool legacy_ol_flags)
3017 {
3018         uint32_t pkt_idx = 0;
3019
3020         if (rte_pktmbuf_alloc_bulk(mbuf_pool, pkts, count))
3021                 return 0;
3022
3023         do {
3024                 rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
3025
3026                 if (count - pkt_idx >= PACKED_BATCH_SIZE) {
3027                         if (!virtio_dev_tx_batch_packed(dev, vq,
3028                                                         &pkts[pkt_idx],
3029                                                         legacy_ol_flags)) {
3030                                 pkt_idx += PACKED_BATCH_SIZE;
3031                                 continue;
3032                         }
3033                 }
3034
3035                 if (virtio_dev_tx_single_packed(dev, vq, mbuf_pool,
3036                                                 pkts[pkt_idx],
3037                                                 legacy_ol_flags))
3038                         break;
3039                 pkt_idx++;
3040         } while (pkt_idx < count);
3041
3042         if (pkt_idx != count)
3043                 rte_pktmbuf_free_bulk(&pkts[pkt_idx], count - pkt_idx);
3044
3045         if (vq->shadow_used_idx) {
3046                 do_data_copy_dequeue(vq);
3047
3048                 vhost_flush_dequeue_shadow_packed(dev, vq);
3049                 vhost_vring_call_packed(dev, vq);
3050         }
3051
3052         return pkt_idx;
3053 }
3054
3055 __rte_noinline
3056 static uint16_t
3057 virtio_dev_tx_packed_legacy(struct virtio_net *dev,
3058         struct vhost_virtqueue *__rte_restrict vq, struct rte_mempool *mbuf_pool,
3059         struct rte_mbuf **__rte_restrict pkts, uint32_t count)
3060 {
3061         return virtio_dev_tx_packed(dev, vq, mbuf_pool, pkts, count, true);
3062 }
3063
3064 __rte_noinline
3065 static uint16_t
3066 virtio_dev_tx_packed_compliant(struct virtio_net *dev,
3067         struct vhost_virtqueue *__rte_restrict vq, struct rte_mempool *mbuf_pool,
3068         struct rte_mbuf **__rte_restrict pkts, uint32_t count)
3069 {
3070         return virtio_dev_tx_packed(dev, vq, mbuf_pool, pkts, count, false);
3071 }
3072
3073 uint16_t
3074 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
3075         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
3076 {
3077         struct virtio_net *dev;
3078         struct rte_mbuf *rarp_mbuf = NULL;
3079         struct vhost_virtqueue *vq;
3080         int16_t success = 1;
3081
3082         dev = get_device(vid);
3083         if (!dev)
3084                 return 0;
3085
3086         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
3087                 VHOST_LOG_DATA(ERR,
3088                         "(%d) %s: built-in vhost net backend is disabled.\n",
3089                         dev->vid, __func__);
3090                 return 0;
3091         }
3092
3093         if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
3094                 VHOST_LOG_DATA(ERR,
3095                         "(%d) %s: invalid virtqueue idx %d.\n",
3096                         dev->vid, __func__, queue_id);
3097                 return 0;
3098         }
3099
3100         vq = dev->virtqueue[queue_id];
3101
3102         if (unlikely(rte_spinlock_trylock(&vq->access_lock) == 0))
3103                 return 0;
3104
3105         if (unlikely(!vq->enabled)) {
3106                 count = 0;
3107                 goto out_access_unlock;
3108         }
3109
3110         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
3111                 vhost_user_iotlb_rd_lock(vq);
3112
3113         if (unlikely(!vq->access_ok))
3114                 if (unlikely(vring_translate(dev, vq) < 0)) {
3115                         count = 0;
3116                         goto out;
3117                 }
3118
3119         /*
3120          * Construct a RARP broadcast packet, and inject it to the "pkts"
3121          * array, to looks like that guest actually send such packet.
3122          *
3123          * Check user_send_rarp() for more information.
3124          *
3125          * broadcast_rarp shares a cacheline in the virtio_net structure
3126          * with some fields that are accessed during enqueue and
3127          * __atomic_compare_exchange_n causes a write if performed compare
3128          * and exchange. This could result in false sharing between enqueue
3129          * and dequeue.
3130          *
3131          * Prevent unnecessary false sharing by reading broadcast_rarp first
3132          * and only performing compare and exchange if the read indicates it
3133          * is likely to be set.
3134          */
3135         if (unlikely(__atomic_load_n(&dev->broadcast_rarp, __ATOMIC_ACQUIRE) &&
3136                         __atomic_compare_exchange_n(&dev->broadcast_rarp,
3137                         &success, 0, 0, __ATOMIC_RELEASE, __ATOMIC_RELAXED))) {
3138
3139                 rarp_mbuf = rte_net_make_rarp_packet(mbuf_pool, &dev->mac);
3140                 if (rarp_mbuf == NULL) {
3141                         VHOST_LOG_DATA(ERR, "Failed to make RARP packet.\n");
3142                         count = 0;
3143                         goto out;
3144                 }
3145                 /*
3146                  * Inject it to the head of "pkts" array, so that switch's mac
3147                  * learning table will get updated first.
3148                  */
3149                 pkts[0] = rarp_mbuf;
3150                 pkts++;
3151                 count -= 1;
3152         }
3153
3154         if (vq_is_packed(dev)) {
3155                 if (dev->flags & VIRTIO_DEV_LEGACY_OL_FLAGS)
3156                         count = virtio_dev_tx_packed_legacy(dev, vq, mbuf_pool, pkts, count);
3157                 else
3158                         count = virtio_dev_tx_packed_compliant(dev, vq, mbuf_pool, pkts, count);
3159         } else {
3160                 if (dev->flags & VIRTIO_DEV_LEGACY_OL_FLAGS)
3161                         count = virtio_dev_tx_split_legacy(dev, vq, mbuf_pool, pkts, count);
3162                 else
3163                         count = virtio_dev_tx_split_compliant(dev, vq, mbuf_pool, pkts, count);
3164         }
3165
3166 out:
3167         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
3168                 vhost_user_iotlb_rd_unlock(vq);
3169
3170 out_access_unlock:
3171         rte_spinlock_unlock(&vq->access_lock);
3172
3173         if (unlikely(rarp_mbuf != NULL))
3174                 count += 1;
3175
3176         return count;
3177 }