lib/librte_vhost/virtio_net.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2016 Intel Corporation
   3  */
   4
   5 #include <stdint.h>
   6 #include <stdbool.h>
   7 #include <linux/virtio_net.h>
   8
   9 #include <rte_mbuf.h>
  10 #include <rte_memcpy.h>
  11 #include <rte_ether.h>
  12 #include <rte_ip.h>
  13 #include <rte_vhost.h>
  14 #include <rte_tcp.h>
  15 #include <rte_udp.h>
  16 #include <rte_sctp.h>
  17 #include <rte_arp.h>
  18 #include <rte_spinlock.h>
  19 #include <rte_malloc.h>
  20 #include <rte_vhost_async.h>
  21
  22 #include "iotlb.h"
  23 #include "vhost.h"
  24
  25 #define MAX_BATCH_LEN 256
  26
  27 #define VHOST_ASYNC_BATCH_THRESHOLD 32
  28
  29 static  __rte_always_inline bool
  30 rxvq_is_mergeable(struct virtio_net *dev)
  31 {
  32         return dev->features & (1ULL << VIRTIO_NET_F_MRG_RXBUF);
  33 }
  34
  35 static  __rte_always_inline bool
  36 virtio_net_is_inorder(struct virtio_net *dev)
  37 {
  38         return dev->features & (1ULL << VIRTIO_F_IN_ORDER);
  39 }
  40
  41 static bool
  42 is_valid_virt_queue_idx(uint32_t idx, int is_tx, uint32_t nr_vring)
  43 {
  44         return (is_tx ^ (idx & 1)) == 0 && idx < nr_vring;
  45 }
  46
  47 static inline void
  48 do_data_copy_enqueue(struct virtio_net *dev, struct vhost_virtqueue *vq)
  49 {
  50         struct batch_copy_elem *elem = vq->batch_copy_elems;
  51         uint16_t count = vq->batch_copy_nb_elems;
  52         int i;
  53
  54         for (i = 0; i < count; i++) {
  55                 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
  56                 vhost_log_cache_write_iova(dev, vq, elem[i].log_addr,
  57                                            elem[i].len);
  58                 PRINT_PACKET(dev, (uintptr_t)elem[i].dst, elem[i].len, 0);
  59         }
  60
  61         vq->batch_copy_nb_elems = 0;
  62 }
  63
  64 static inline void
  65 do_data_copy_dequeue(struct vhost_virtqueue *vq)
  66 {
  67         struct batch_copy_elem *elem = vq->batch_copy_elems;
  68         uint16_t count = vq->batch_copy_nb_elems;
  69         int i;
  70
  71         for (i = 0; i < count; i++)
  72                 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
  73
  74         vq->batch_copy_nb_elems = 0;
  75 }
  76
  77 static __rte_always_inline void
  78 do_flush_shadow_used_ring_split(struct virtio_net *dev,
  79                         struct vhost_virtqueue *vq,
  80                         uint16_t to, uint16_t from, uint16_t size)
  81 {
  82         rte_memcpy(&vq->used->ring[to],
  83                         &vq->shadow_used_split[from],
  84                         size * sizeof(struct vring_used_elem));
  85         vhost_log_cache_used_vring(dev, vq,
  86                         offsetof(struct vring_used, ring[to]),
  87                         size * sizeof(struct vring_used_elem));
  88 }
  89
  90 static __rte_always_inline void
  91 flush_shadow_used_ring_split(struct virtio_net *dev, struct vhost_virtqueue *vq)
  92 {
  93         uint16_t used_idx = vq->last_used_idx & (vq->size - 1);
  94
  95         if (used_idx + vq->shadow_used_idx <= vq->size) {
  96                 do_flush_shadow_used_ring_split(dev, vq, used_idx, 0,
  97                                           vq->shadow_used_idx);
  98         } else {
  99                 uint16_t size;
 100
 101                 /* update used ring interval [used_idx, vq->size] */
 102                 size = vq->size - used_idx;
 103                 do_flush_shadow_used_ring_split(dev, vq, used_idx, 0, size);
 104
 105                 /* update the left half used ring interval [0, left_size] */
 106                 do_flush_shadow_used_ring_split(dev, vq, 0, size,
 107                                           vq->shadow_used_idx - size);
 108         }
 109         vq->last_used_idx += vq->shadow_used_idx;
 110
 111         vhost_log_cache_sync(dev, vq);
 112
 113         __atomic_add_fetch(&vq->used->idx, vq->shadow_used_idx,
 114                            __ATOMIC_RELEASE);
 115         vq->shadow_used_idx = 0;
 116         vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
 117                 sizeof(vq->used->idx));
 118 }
 119
 120 static __rte_always_inline void
 121 async_flush_shadow_used_ring_split(struct virtio_net *dev,
 122         struct vhost_virtqueue *vq)
 123 {
 124         uint16_t used_idx = vq->last_used_idx & (vq->size - 1);
 125
 126         if (used_idx + vq->shadow_used_idx <= vq->size) {
 127                 do_flush_shadow_used_ring_split(dev, vq, used_idx, 0,
 128                                           vq->shadow_used_idx);
 129         } else {
 130                 uint16_t size;
 131
 132                 /* update used ring interval [used_idx, vq->size] */
 133                 size = vq->size - used_idx;
 134                 do_flush_shadow_used_ring_split(dev, vq, used_idx, 0, size);
 135
 136                 /* update the left half used ring interval [0, left_size] */
 137                 do_flush_shadow_used_ring_split(dev, vq, 0, size,
 138                                           vq->shadow_used_idx - size);
 139         }
 140
 141         vq->last_used_idx += vq->shadow_used_idx;
 142         vq->shadow_used_idx = 0;
 143 }
 144
 145 static __rte_always_inline void
 146 update_shadow_used_ring_split(struct vhost_virtqueue *vq,
 147                          uint16_t desc_idx, uint32_t len)
 148 {
 149         uint16_t i = vq->shadow_used_idx++;
 150
 151         vq->shadow_used_split[i].id  = desc_idx;
 152         vq->shadow_used_split[i].len = len;
 153 }
 154
 155 static __rte_always_inline void
 156 vhost_flush_enqueue_shadow_packed(struct virtio_net *dev,
 157                                   struct vhost_virtqueue *vq)
 158 {
 159         int i;
 160         uint16_t used_idx = vq->last_used_idx;
 161         uint16_t head_idx = vq->last_used_idx;
 162         uint16_t head_flags = 0;
 163
 164         /* Split loop in two to save memory barriers */
 165         for (i = 0; i < vq->shadow_used_idx; i++) {
 166                 vq->desc_packed[used_idx].id = vq->shadow_used_packed[i].id;
 167                 vq->desc_packed[used_idx].len = vq->shadow_used_packed[i].len;
 168
 169                 used_idx += vq->shadow_used_packed[i].count;
 170                 if (used_idx >= vq->size)
 171                         used_idx -= vq->size;
 172         }
 173
 174         rte_smp_wmb();
 175
 176         for (i = 0; i < vq->shadow_used_idx; i++) {
 177                 uint16_t flags;
 178
 179                 if (vq->shadow_used_packed[i].len)
 180                         flags = VRING_DESC_F_WRITE;
 181                 else
 182                         flags = 0;
 183
 184                 if (vq->used_wrap_counter) {
 185                         flags |= VRING_DESC_F_USED;
 186                         flags |= VRING_DESC_F_AVAIL;
 187                 } else {
 188                         flags &= ~VRING_DESC_F_USED;
 189                         flags &= ~VRING_DESC_F_AVAIL;
 190                 }
 191
 192                 if (i > 0) {
 193                         vq->desc_packed[vq->last_used_idx].flags = flags;
 194
 195                         vhost_log_cache_used_vring(dev, vq,
 196                                         vq->last_used_idx *
 197                                         sizeof(struct vring_packed_desc),
 198                                         sizeof(struct vring_packed_desc));
 199                 } else {
 200                         head_idx = vq->last_used_idx;
 201                         head_flags = flags;
 202                 }
 203
 204                 vq_inc_last_used_packed(vq, vq->shadow_used_packed[i].count);
 205         }
 206
 207         vq->desc_packed[head_idx].flags = head_flags;
 208
 209         vhost_log_cache_used_vring(dev, vq,
 210                                 head_idx *
 211                                 sizeof(struct vring_packed_desc),
 212                                 sizeof(struct vring_packed_desc));
 213
 214         vq->shadow_used_idx = 0;
 215         vhost_log_cache_sync(dev, vq);
 216 }
 217
 218 static __rte_always_inline void
 219 vhost_flush_dequeue_shadow_packed(struct virtio_net *dev,
 220                                   struct vhost_virtqueue *vq)
 221 {
 222         struct vring_used_elem_packed *used_elem = &vq->shadow_used_packed[0];
 223
 224         vq->desc_packed[vq->shadow_last_used_idx].id = used_elem->id;
 225         rte_smp_wmb();
 226         vq->desc_packed[vq->shadow_last_used_idx].flags = used_elem->flags;
 227
 228         vhost_log_cache_used_vring(dev, vq, vq->shadow_last_used_idx *
 229                                    sizeof(struct vring_packed_desc),
 230                                    sizeof(struct vring_packed_desc));
 231         vq->shadow_used_idx = 0;
 232         vhost_log_cache_sync(dev, vq);
 233 }
 234
 235 static __rte_always_inline void
 236 vhost_flush_enqueue_batch_packed(struct virtio_net *dev,
 237                                  struct vhost_virtqueue *vq,
 238                                  uint64_t *lens,
 239                                  uint16_t *ids)
 240 {
 241         uint16_t i;
 242         uint16_t flags;
 243
 244         if (vq->shadow_used_idx) {
 245                 do_data_copy_enqueue(dev, vq);
 246                 vhost_flush_enqueue_shadow_packed(dev, vq);
 247         }
 248
 249         flags = PACKED_DESC_ENQUEUE_USED_FLAG(vq->used_wrap_counter);
 250
 251         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
 252                 vq->desc_packed[vq->last_used_idx + i].id = ids[i];
 253                 vq->desc_packed[vq->last_used_idx + i].len = lens[i];
 254         }
 255
 256         rte_smp_wmb();
 257
 258         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
 259                 vq->desc_packed[vq->last_used_idx + i].flags = flags;
 260
 261         vhost_log_cache_used_vring(dev, vq, vq->last_used_idx *
 262                                    sizeof(struct vring_packed_desc),
 263                                    sizeof(struct vring_packed_desc) *
 264                                    PACKED_BATCH_SIZE);
 265         vhost_log_cache_sync(dev, vq);
 266
 267         vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
 268 }
 269
 270 static __rte_always_inline void
 271 vhost_shadow_dequeue_batch_packed_inorder(struct vhost_virtqueue *vq,
 272                                           uint16_t id)
 273 {
 274         vq->shadow_used_packed[0].id = id;
 275
 276         if (!vq->shadow_used_idx) {
 277                 vq->shadow_last_used_idx = vq->last_used_idx;
 278                 vq->shadow_used_packed[0].flags =
 279                         PACKED_DESC_DEQUEUE_USED_FLAG(vq->used_wrap_counter);
 280                 vq->shadow_used_packed[0].len = 0;
 281                 vq->shadow_used_packed[0].count = 1;
 282                 vq->shadow_used_idx++;
 283         }
 284
 285         vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
 286 }
 287
 288 static __rte_always_inline void
 289 vhost_shadow_dequeue_batch_packed(struct virtio_net *dev,
 290                                   struct vhost_virtqueue *vq,
 291                                   uint16_t *ids)
 292 {
 293         uint16_t flags;
 294         uint16_t i;
 295         uint16_t begin;
 296
 297         flags = PACKED_DESC_DEQUEUE_USED_FLAG(vq->used_wrap_counter);
 298
 299         if (!vq->shadow_used_idx) {
 300                 vq->shadow_last_used_idx = vq->last_used_idx;
 301                 vq->shadow_used_packed[0].id  = ids[0];
 302                 vq->shadow_used_packed[0].len = 0;
 303                 vq->shadow_used_packed[0].count = 1;
 304                 vq->shadow_used_packed[0].flags = flags;
 305                 vq->shadow_used_idx++;
 306                 begin = 1;
 307         } else
 308                 begin = 0;
 309
 310         vhost_for_each_try_unroll(i, begin, PACKED_BATCH_SIZE) {
 311                 vq->desc_packed[vq->last_used_idx + i].id = ids[i];
 312                 vq->desc_packed[vq->last_used_idx + i].len = 0;
 313         }
 314
 315         rte_smp_wmb();
 316         vhost_for_each_try_unroll(i, begin, PACKED_BATCH_SIZE)
 317                 vq->desc_packed[vq->last_used_idx + i].flags = flags;
 318
 319         vhost_log_cache_used_vring(dev, vq, vq->last_used_idx *
 320                                    sizeof(struct vring_packed_desc),
 321                                    sizeof(struct vring_packed_desc) *
 322                                    PACKED_BATCH_SIZE);
 323         vhost_log_cache_sync(dev, vq);
 324
 325         vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
 326 }
 327
 328 static __rte_always_inline void
 329 vhost_shadow_dequeue_single_packed(struct vhost_virtqueue *vq,
 330                                    uint16_t buf_id,
 331                                    uint16_t count)
 332 {
 333         uint16_t flags;
 334
 335         flags = vq->desc_packed[vq->last_used_idx].flags;
 336         if (vq->used_wrap_counter) {
 337                 flags |= VRING_DESC_F_USED;
 338                 flags |= VRING_DESC_F_AVAIL;
 339         } else {
 340                 flags &= ~VRING_DESC_F_USED;
 341                 flags &= ~VRING_DESC_F_AVAIL;
 342         }
 343
 344         if (!vq->shadow_used_idx) {
 345                 vq->shadow_last_used_idx = vq->last_used_idx;
 346
 347                 vq->shadow_used_packed[0].id  = buf_id;
 348                 vq->shadow_used_packed[0].len = 0;
 349                 vq->shadow_used_packed[0].flags = flags;
 350                 vq->shadow_used_idx++;
 351         } else {
 352                 vq->desc_packed[vq->last_used_idx].id = buf_id;
 353                 vq->desc_packed[vq->last_used_idx].len = 0;
 354                 vq->desc_packed[vq->last_used_idx].flags = flags;
 355         }
 356
 357         vq_inc_last_used_packed(vq, count);
 358 }
 359
 360 static __rte_always_inline void
 361 vhost_shadow_dequeue_single_packed_inorder(struct vhost_virtqueue *vq,
 362                                            uint16_t buf_id,
 363                                            uint16_t count)
 364 {
 365         uint16_t flags;
 366
 367         vq->shadow_used_packed[0].id = buf_id;
 368
 369         flags = vq->desc_packed[vq->last_used_idx].flags;
 370         if (vq->used_wrap_counter) {
 371                 flags |= VRING_DESC_F_USED;
 372                 flags |= VRING_DESC_F_AVAIL;
 373         } else {
 374                 flags &= ~VRING_DESC_F_USED;
 375                 flags &= ~VRING_DESC_F_AVAIL;
 376         }
 377
 378         if (!vq->shadow_used_idx) {
 379                 vq->shadow_last_used_idx = vq->last_used_idx;
 380                 vq->shadow_used_packed[0].len = 0;
 381                 vq->shadow_used_packed[0].flags = flags;
 382                 vq->shadow_used_idx++;
 383         }
 384
 385         vq_inc_last_used_packed(vq, count);
 386 }
 387
 388 static __rte_always_inline void
 389 vhost_shadow_enqueue_single_packed(struct virtio_net *dev,
 390                                    struct vhost_virtqueue *vq,
 391                                    uint32_t len[],
 392                                    uint16_t id[],
 393                                    uint16_t count[],
 394                                    uint16_t num_buffers)
 395 {
 396         uint16_t i;
 397         for (i = 0; i < num_buffers; i++) {
 398                 /* enqueue shadow flush action aligned with batch num */
 399                 if (!vq->shadow_used_idx)
 400                         vq->shadow_aligned_idx = vq->last_used_idx &
 401                                 PACKED_BATCH_MASK;
 402                 vq->shadow_used_packed[vq->shadow_used_idx].id  = id[i];
 403                 vq->shadow_used_packed[vq->shadow_used_idx].len = len[i];
 404                 vq->shadow_used_packed[vq->shadow_used_idx].count = count[i];
 405                 vq->shadow_aligned_idx += count[i];
 406                 vq->shadow_used_idx++;
 407         }
 408
 409         if (vq->shadow_aligned_idx >= PACKED_BATCH_SIZE) {
 410                 do_data_copy_enqueue(dev, vq);
 411                 vhost_flush_enqueue_shadow_packed(dev, vq);
 412         }
 413 }
 414
 415 /* avoid write operation when necessary, to lessen cache issues */
 416 #define ASSIGN_UNLESS_EQUAL(var, val) do {      \
 417         if ((var) != (val))                     \
 418                 (var) = (val);                  \
 419 } while (0)
 420
 421 static __rte_always_inline void
 422 virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
 423 {
 424         uint64_t csum_l4 = m_buf->ol_flags & PKT_TX_L4_MASK;
 425
 426         if (m_buf->ol_flags & PKT_TX_TCP_SEG)
 427                 csum_l4 |= PKT_TX_TCP_CKSUM;
 428
 429         if (csum_l4) {
 430                 net_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 431                 net_hdr->csum_start = m_buf->l2_len + m_buf->l3_len;
 432
 433                 switch (csum_l4) {
 434                 case PKT_TX_TCP_CKSUM:
 435                         net_hdr->csum_offset = (offsetof(struct rte_tcp_hdr,
 436                                                 cksum));
 437                         break;
 438                 case PKT_TX_UDP_CKSUM:
 439                         net_hdr->csum_offset = (offsetof(struct rte_udp_hdr,
 440                                                 dgram_cksum));
 441                         break;
 442                 case PKT_TX_SCTP_CKSUM:
 443                         net_hdr->csum_offset = (offsetof(struct rte_sctp_hdr,
 444                                                 cksum));
 445                         break;
 446                 }
 447         } else {
 448                 ASSIGN_UNLESS_EQUAL(net_hdr->csum_start, 0);
 449                 ASSIGN_UNLESS_EQUAL(net_hdr->csum_offset, 0);
 450                 ASSIGN_UNLESS_EQUAL(net_hdr->flags, 0);
 451         }
 452
 453         /* IP cksum verification cannot be bypassed, then calculate here */
 454         if (m_buf->ol_flags & PKT_TX_IP_CKSUM) {
 455                 struct rte_ipv4_hdr *ipv4_hdr;
 456
 457                 ipv4_hdr = rte_pktmbuf_mtod_offset(m_buf, struct rte_ipv4_hdr *,
 458                                                    m_buf->l2_len);
 459                 ipv4_hdr->hdr_checksum = 0;
 460                 ipv4_hdr->hdr_checksum = rte_ipv4_cksum(ipv4_hdr);
 461         }
 462
 463         if (m_buf->ol_flags & PKT_TX_TCP_SEG) {
 464                 if (m_buf->ol_flags & PKT_TX_IPV4)
 465                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 466                 else
 467                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 468                 net_hdr->gso_size = m_buf->tso_segsz;
 469                 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len
 470                                         + m_buf->l4_len;
 471         } else if (m_buf->ol_flags & PKT_TX_UDP_SEG) {
 472                 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
 473                 net_hdr->gso_size = m_buf->tso_segsz;
 474                 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len +
 475                         m_buf->l4_len;
 476         } else {
 477                 ASSIGN_UNLESS_EQUAL(net_hdr->gso_type, 0);
 478                 ASSIGN_UNLESS_EQUAL(net_hdr->gso_size, 0);
 479                 ASSIGN_UNLESS_EQUAL(net_hdr->hdr_len, 0);
 480         }
 481 }
 482
 483 static __rte_always_inline int
 484 map_one_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 485                 struct buf_vector *buf_vec, uint16_t *vec_idx,
 486                 uint64_t desc_iova, uint64_t desc_len, uint8_t perm)
 487 {
 488         uint16_t vec_id = *vec_idx;
 489
 490         while (desc_len) {
 491                 uint64_t desc_addr;
 492                 uint64_t desc_chunck_len = desc_len;
 493
 494                 if (unlikely(vec_id >= BUF_VECTOR_MAX))
 495                         return -1;
 496
 497                 desc_addr = vhost_iova_to_vva(dev, vq,
 498                                 desc_iova,
 499                                 &desc_chunck_len,
 500                                 perm);
 501                 if (unlikely(!desc_addr))
 502                         return -1;
 503
 504                 rte_prefetch0((void *)(uintptr_t)desc_addr);
 505
 506                 buf_vec[vec_id].buf_iova = desc_iova;
 507                 buf_vec[vec_id].buf_addr = desc_addr;
 508                 buf_vec[vec_id].buf_len  = desc_chunck_len;
 509
 510                 desc_len -= desc_chunck_len;
 511                 desc_iova += desc_chunck_len;
 512                 vec_id++;
 513         }
 514         *vec_idx = vec_id;
 515
 516         return 0;
 517 }
 518
 519 static __rte_always_inline int
 520 fill_vec_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 521                          uint32_t avail_idx, uint16_t *vec_idx,
 522                          struct buf_vector *buf_vec, uint16_t *desc_chain_head,
 523                          uint32_t *desc_chain_len, uint8_t perm)
 524 {
 525         uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
 526         uint16_t vec_id = *vec_idx;
 527         uint32_t len    = 0;
 528         uint64_t dlen;
 529         uint32_t nr_descs = vq->size;
 530         uint32_t cnt    = 0;
 531         struct vring_desc *descs = vq->desc;
 532         struct vring_desc *idesc = NULL;
 533
 534         if (unlikely(idx >= vq->size))
 535                 return -1;
 536
 537         *desc_chain_head = idx;
 538
 539         if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
 540                 dlen = vq->desc[idx].len;
 541                 nr_descs = dlen / sizeof(struct vring_desc);
 542                 if (unlikely(nr_descs > vq->size))
 543                         return -1;
 544
 545                 descs = (struct vring_desc *)(uintptr_t)
 546                         vhost_iova_to_vva(dev, vq, vq->desc[idx].addr,
 547                                                 &dlen,
 548                                                 VHOST_ACCESS_RO);
 549                 if (unlikely(!descs))
 550                         return -1;
 551
 552                 if (unlikely(dlen < vq->desc[idx].len)) {
 553                         /*
 554                          * The indirect desc table is not contiguous
 555                          * in process VA space, we have to copy it.
 556                          */
 557                         idesc = vhost_alloc_copy_ind_table(dev, vq,
 558                                         vq->desc[idx].addr, vq->desc[idx].len);
 559                         if (unlikely(!idesc))
 560                                 return -1;
 561
 562                         descs = idesc;
 563                 }
 564
 565                 idx = 0;
 566         }
 567
 568         while (1) {
 569                 if (unlikely(idx >= nr_descs || cnt++ >= nr_descs)) {
 570                         free_ind_table(idesc);
 571                         return -1;
 572                 }
 573
 574                 len += descs[idx].len;
 575
 576                 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 577                                                 descs[idx].addr, descs[idx].len,
 578                                                 perm))) {
 579                         free_ind_table(idesc);
 580                         return -1;
 581                 }
 582
 583                 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
 584                         break;
 585
 586                 idx = descs[idx].next;
 587         }
 588
 589         *desc_chain_len = len;
 590         *vec_idx = vec_id;
 591
 592         if (unlikely(!!idesc))
 593                 free_ind_table(idesc);
 594
 595         return 0;
 596 }
 597
 598 /*
 599  * Returns -1 on fail, 0 on success
 600  */
 601 static inline int
 602 reserve_avail_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 603                                 uint32_t size, struct buf_vector *buf_vec,
 604                                 uint16_t *num_buffers, uint16_t avail_head,
 605                                 uint16_t *nr_vec)
 606 {
 607         uint16_t cur_idx;
 608         uint16_t vec_idx = 0;
 609         uint16_t max_tries, tries = 0;
 610
 611         uint16_t head_idx = 0;
 612         uint32_t len = 0;
 613
 614         *num_buffers = 0;
 615         cur_idx  = vq->last_avail_idx;
 616
 617         if (rxvq_is_mergeable(dev))
 618                 max_tries = vq->size - 1;
 619         else
 620                 max_tries = 1;
 621
 622         while (size > 0) {
 623                 if (unlikely(cur_idx == avail_head))
 624                         return -1;
 625                 /*
 626                  * if we tried all available ring items, and still
 627                  * can't get enough buf, it means something abnormal
 628                  * happened.
 629                  */
 630                 if (unlikely(++tries > max_tries))
 631                         return -1;
 632
 633                 if (unlikely(fill_vec_buf_split(dev, vq, cur_idx,
 634                                                 &vec_idx, buf_vec,
 635                                                 &head_idx, &len,
 636                                                 VHOST_ACCESS_RW) < 0))
 637                         return -1;
 638                 len = RTE_MIN(len, size);
 639                 update_shadow_used_ring_split(vq, head_idx, len);
 640                 size -= len;
 641
 642                 cur_idx++;
 643                 *num_buffers += 1;
 644         }
 645
 646         *nr_vec = vec_idx;
 647
 648         return 0;
 649 }
 650
 651 static __rte_always_inline int
 652 fill_vec_buf_packed_indirect(struct virtio_net *dev,
 653                         struct vhost_virtqueue *vq,
 654                         struct vring_packed_desc *desc, uint16_t *vec_idx,
 655                         struct buf_vector *buf_vec, uint32_t *len, uint8_t perm)
 656 {
 657         uint16_t i;
 658         uint32_t nr_descs;
 659         uint16_t vec_id = *vec_idx;
 660         uint64_t dlen;
 661         struct vring_packed_desc *descs, *idescs = NULL;
 662
 663         dlen = desc->len;
 664         descs = (struct vring_packed_desc *)(uintptr_t)
 665                 vhost_iova_to_vva(dev, vq, desc->addr, &dlen, VHOST_ACCESS_RO);
 666         if (unlikely(!descs))
 667                 return -1;
 668
 669         if (unlikely(dlen < desc->len)) {
 670                 /*
 671                  * The indirect desc table is not contiguous
 672                  * in process VA space, we have to copy it.
 673                  */
 674                 idescs = vhost_alloc_copy_ind_table(dev,
 675                                 vq, desc->addr, desc->len);
 676                 if (unlikely(!idescs))
 677                         return -1;
 678
 679                 descs = idescs;
 680         }
 681
 682         nr_descs =  desc->len / sizeof(struct vring_packed_desc);
 683         if (unlikely(nr_descs >= vq->size)) {
 684                 free_ind_table(idescs);
 685                 return -1;
 686         }
 687
 688         for (i = 0; i < nr_descs; i++) {
 689                 if (unlikely(vec_id >= BUF_VECTOR_MAX)) {
 690                         free_ind_table(idescs);
 691                         return -1;
 692                 }
 693
 694                 *len += descs[i].len;
 695                 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 696                                                 descs[i].addr, descs[i].len,
 697                                                 perm)))
 698                         return -1;
 699         }
 700         *vec_idx = vec_id;
 701
 702         if (unlikely(!!idescs))
 703                 free_ind_table(idescs);
 704
 705         return 0;
 706 }
 707
 708 static __rte_always_inline int
 709 fill_vec_buf_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
 710                                 uint16_t avail_idx, uint16_t *desc_count,
 711                                 struct buf_vector *buf_vec, uint16_t *vec_idx,
 712                                 uint16_t *buf_id, uint32_t *len, uint8_t perm)
 713 {
 714         bool wrap_counter = vq->avail_wrap_counter;
 715         struct vring_packed_desc *descs = vq->desc_packed;
 716         uint16_t vec_id = *vec_idx;
 717
 718         if (avail_idx < vq->last_avail_idx)
 719                 wrap_counter ^= 1;
 720
 721         /*
 722          * Perform a load-acquire barrier in desc_is_avail to
 723          * enforce the ordering between desc flags and desc
 724          * content.
 725          */
 726         if (unlikely(!desc_is_avail(&descs[avail_idx], wrap_counter)))
 727                 return -1;
 728
 729         *desc_count = 0;
 730         *len = 0;
 731
 732         while (1) {
 733                 if (unlikely(vec_id >= BUF_VECTOR_MAX))
 734                         return -1;
 735
 736                 if (unlikely(*desc_count >= vq->size))
 737                         return -1;
 738
 739                 *desc_count += 1;
 740                 *buf_id = descs[avail_idx].id;
 741
 742                 if (descs[avail_idx].flags & VRING_DESC_F_INDIRECT) {
 743                         if (unlikely(fill_vec_buf_packed_indirect(dev, vq,
 744                                                         &descs[avail_idx],
 745                                                         &vec_id, buf_vec,
 746                                                         len, perm) < 0))
 747                                 return -1;
 748                 } else {
 749                         *len += descs[avail_idx].len;
 750
 751                         if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 752                                                         descs[avail_idx].addr,
 753                                                         descs[avail_idx].len,
 754                                                         perm)))
 755                                 return -1;
 756                 }
 757
 758                 if ((descs[avail_idx].flags & VRING_DESC_F_NEXT) == 0)
 759                         break;
 760
 761                 if (++avail_idx >= vq->size) {
 762                         avail_idx -= vq->size;
 763                         wrap_counter ^= 1;
 764                 }
 765         }
 766
 767         *vec_idx = vec_id;
 768
 769         return 0;
 770 }
 771
 772 static __rte_noinline void
 773 copy_vnet_hdr_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 774                 struct buf_vector *buf_vec,
 775                 struct virtio_net_hdr_mrg_rxbuf *hdr)
 776 {
 777         uint64_t len;
 778         uint64_t remain = dev->vhost_hlen;
 779         uint64_t src = (uint64_t)(uintptr_t)hdr, dst;
 780         uint64_t iova = buf_vec->buf_iova;
 781
 782         while (remain) {
 783                 len = RTE_MIN(remain,
 784                                 buf_vec->buf_len);
 785                 dst = buf_vec->buf_addr;
 786                 rte_memcpy((void *)(uintptr_t)dst,
 787                                 (void *)(uintptr_t)src,
 788                                 len);
 789
 790                 PRINT_PACKET(dev, (uintptr_t)dst,
 791                                 (uint32_t)len, 0);
 792                 vhost_log_cache_write_iova(dev, vq,
 793                                 iova, len);
 794
 795                 remain -= len;
 796                 iova += len;
 797                 src += len;
 798                 buf_vec++;
 799         }
 800 }
 801
 802 static __rte_always_inline int
 803 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 804                             struct rte_mbuf *m, struct buf_vector *buf_vec,
 805                             uint16_t nr_vec, uint16_t num_buffers)
 806 {
 807         uint32_t vec_idx = 0;
 808         uint32_t mbuf_offset, mbuf_avail;
 809         uint32_t buf_offset, buf_avail;
 810         uint64_t buf_addr, buf_iova, buf_len;
 811         uint32_t cpy_len;
 812         uint64_t hdr_addr;
 813         struct rte_mbuf *hdr_mbuf;
 814         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 815         struct virtio_net_hdr_mrg_rxbuf tmp_hdr, *hdr = NULL;
 816         int error = 0;
 817
 818         if (unlikely(m == NULL)) {
 819                 error = -1;
 820                 goto out;
 821         }
 822
 823         buf_addr = buf_vec[vec_idx].buf_addr;
 824         buf_iova = buf_vec[vec_idx].buf_iova;
 825         buf_len = buf_vec[vec_idx].buf_len;
 826
 827         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
 828                 error = -1;
 829                 goto out;
 830         }
 831
 832         hdr_mbuf = m;
 833         hdr_addr = buf_addr;
 834         if (unlikely(buf_len < dev->vhost_hlen))
 835                 hdr = &tmp_hdr;
 836         else
 837                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)hdr_addr;
 838
 839         VHOST_LOG_DATA(DEBUG, "(%d) RX: num merge buffers %d\n",
 840                 dev->vid, num_buffers);
 841
 842         if (unlikely(buf_len < dev->vhost_hlen)) {
 843                 buf_offset = dev->vhost_hlen - buf_len;
 844                 vec_idx++;
 845                 buf_addr = buf_vec[vec_idx].buf_addr;
 846                 buf_iova = buf_vec[vec_idx].buf_iova;
 847                 buf_len = buf_vec[vec_idx].buf_len;
 848                 buf_avail = buf_len - buf_offset;
 849         } else {
 850                 buf_offset = dev->vhost_hlen;
 851                 buf_avail = buf_len - dev->vhost_hlen;
 852         }
 853
 854         mbuf_avail  = rte_pktmbuf_data_len(m);
 855         mbuf_offset = 0;
 856         while (mbuf_avail != 0 || m->next != NULL) {
 857                 /* done with current buf, get the next one */
 858                 if (buf_avail == 0) {
 859                         vec_idx++;
 860                         if (unlikely(vec_idx >= nr_vec)) {
 861                                 error = -1;
 862                                 goto out;
 863                         }
 864
 865                         buf_addr = buf_vec[vec_idx].buf_addr;
 866                         buf_iova = buf_vec[vec_idx].buf_iova;
 867                         buf_len = buf_vec[vec_idx].buf_len;
 868
 869                         buf_offset = 0;
 870                         buf_avail  = buf_len;
 871                 }
 872
 873                 /* done with current mbuf, get the next one */
 874                 if (mbuf_avail == 0) {
 875                         m = m->next;
 876
 877                         mbuf_offset = 0;
 878                         mbuf_avail  = rte_pktmbuf_data_len(m);
 879                 }
 880
 881                 if (hdr_addr) {
 882                         virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
 883                         if (rxvq_is_mergeable(dev))
 884                                 ASSIGN_UNLESS_EQUAL(hdr->num_buffers,
 885                                                 num_buffers);
 886
 887                         if (unlikely(hdr == &tmp_hdr)) {
 888                                 copy_vnet_hdr_to_desc(dev, vq, buf_vec, hdr);
 889                         } else {
 890                                 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
 891                                                 dev->vhost_hlen, 0);
 892                                 vhost_log_cache_write_iova(dev, vq,
 893                                                 buf_vec[0].buf_iova,
 894                                                 dev->vhost_hlen);
 895                         }
 896
 897                         hdr_addr = 0;
 898                 }
 899
 900                 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
 901
 902                 if (likely(cpy_len > MAX_BATCH_LEN ||
 903                                         vq->batch_copy_nb_elems >= vq->size)) {
 904                         rte_memcpy((void *)((uintptr_t)(buf_addr + buf_offset)),
 905                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 906                                 cpy_len);
 907                         vhost_log_cache_write_iova(dev, vq,
 908                                                    buf_iova + buf_offset,
 909                                                    cpy_len);
 910                         PRINT_PACKET(dev, (uintptr_t)(buf_addr + buf_offset),
 911                                 cpy_len, 0);
 912                 } else {
 913                         batch_copy[vq->batch_copy_nb_elems].dst =
 914                                 (void *)((uintptr_t)(buf_addr + buf_offset));
 915                         batch_copy[vq->batch_copy_nb_elems].src =
 916                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
 917                         batch_copy[vq->batch_copy_nb_elems].log_addr =
 918                                 buf_iova + buf_offset;
 919                         batch_copy[vq->batch_copy_nb_elems].len = cpy_len;
 920                         vq->batch_copy_nb_elems++;
 921                 }
 922
 923                 mbuf_avail  -= cpy_len;
 924                 mbuf_offset += cpy_len;
 925                 buf_avail  -= cpy_len;
 926                 buf_offset += cpy_len;
 927         }
 928
 929 out:
 930
 931         return error;
 932 }
 933
 934 static __rte_always_inline void
 935 async_fill_vec(struct iovec *v, void *base, size_t len)
 936 {
 937         v->iov_base = base;
 938         v->iov_len = len;
 939 }
 940
 941 static __rte_always_inline void
 942 async_fill_iter(struct rte_vhost_iov_iter *it, size_t count,
 943         struct iovec *vec, unsigned long nr_seg)
 944 {
 945         it->offset = 0;
 946         it->count = count;
 947
 948         if (count) {
 949                 it->iov = vec;
 950                 it->nr_segs = nr_seg;
 951         } else {
 952                 it->iov = 0;
 953                 it->nr_segs = 0;
 954         }
 955 }
 956
 957 static __rte_always_inline void
 958 async_fill_desc(struct rte_vhost_async_desc *desc,
 959         struct rte_vhost_iov_iter *src, struct rte_vhost_iov_iter *dst)
 960 {
 961         desc->src = src;
 962         desc->dst = dst;
 963 }
 964
 965 static __rte_always_inline int
 966 async_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 967                         struct rte_mbuf *m, struct buf_vector *buf_vec,
 968                         uint16_t nr_vec, uint16_t num_buffers,
 969                         struct iovec *src_iovec, struct iovec *dst_iovec,
 970                         struct rte_vhost_iov_iter *src_it,
 971                         struct rte_vhost_iov_iter *dst_it)
 972 {
 973         uint32_t vec_idx = 0;
 974         uint32_t mbuf_offset, mbuf_avail;
 975         uint32_t buf_offset, buf_avail;
 976         uint64_t buf_addr, buf_iova, buf_len;
 977         uint32_t cpy_len, cpy_threshold;
 978         uint64_t hdr_addr;
 979         struct rte_mbuf *hdr_mbuf;
 980         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 981         struct virtio_net_hdr_mrg_rxbuf tmp_hdr, *hdr = NULL;
 982         int error = 0;
 983
 984         uint32_t tlen = 0;
 985         int tvec_idx = 0;
 986         void *hpa;
 987
 988         if (unlikely(m == NULL)) {
 989                 error = -1;
 990                 goto out;
 991         }
 992
 993         cpy_threshold = vq->async_threshold;
 994
 995         buf_addr = buf_vec[vec_idx].buf_addr;
 996         buf_iova = buf_vec[vec_idx].buf_iova;
 997         buf_len = buf_vec[vec_idx].buf_len;
 998
 999         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
1000                 error = -1;
1001                 goto out;
1002         }
1003
1004         hdr_mbuf = m;
1005         hdr_addr = buf_addr;
1006         if (unlikely(buf_len < dev->vhost_hlen))
1007                 hdr = &tmp_hdr;
1008         else
1009                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)hdr_addr;
1010
1011         VHOST_LOG_DATA(DEBUG, "(%d) RX: num merge buffers %d\n",
1012                 dev->vid, num_buffers);
1013
1014         if (unlikely(buf_len < dev->vhost_hlen)) {
1015                 buf_offset = dev->vhost_hlen - buf_len;
1016                 vec_idx++;
1017                 buf_addr = buf_vec[vec_idx].buf_addr;
1018                 buf_iova = buf_vec[vec_idx].buf_iova;
1019                 buf_len = buf_vec[vec_idx].buf_len;
1020                 buf_avail = buf_len - buf_offset;
1021         } else {
1022                 buf_offset = dev->vhost_hlen;
1023                 buf_avail = buf_len - dev->vhost_hlen;
1024         }
1025
1026         mbuf_avail  = rte_pktmbuf_data_len(m);
1027         mbuf_offset = 0;
1028
1029         while (mbuf_avail != 0 || m->next != NULL) {
1030                 /* done with current buf, get the next one */
1031                 if (buf_avail == 0) {
1032                         vec_idx++;
1033                         if (unlikely(vec_idx >= nr_vec)) {
1034                                 error = -1;
1035                                 goto out;
1036                         }
1037
1038                         buf_addr = buf_vec[vec_idx].buf_addr;
1039                         buf_iova = buf_vec[vec_idx].buf_iova;
1040                         buf_len = buf_vec[vec_idx].buf_len;
1041
1042                         buf_offset = 0;
1043                         buf_avail  = buf_len;
1044                 }
1045
1046                 /* done with current mbuf, get the next one */
1047                 if (mbuf_avail == 0) {
1048                         m = m->next;
1049
1050                         mbuf_offset = 0;
1051                         mbuf_avail  = rte_pktmbuf_data_len(m);
1052                 }
1053
1054                 if (hdr_addr) {
1055                         virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
1056                         if (rxvq_is_mergeable(dev))
1057                                 ASSIGN_UNLESS_EQUAL(hdr->num_buffers,
1058                                                 num_buffers);
1059
1060                         if (unlikely(hdr == &tmp_hdr)) {
1061                                 copy_vnet_hdr_to_desc(dev, vq, buf_vec, hdr);
1062                         } else {
1063                                 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
1064                                                 dev->vhost_hlen, 0);
1065                                 vhost_log_cache_write_iova(dev, vq,
1066                                                 buf_vec[0].buf_iova,
1067                                                 dev->vhost_hlen);
1068                         }
1069
1070                         hdr_addr = 0;
1071                 }
1072
1073                 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
1074
1075                 if (unlikely(cpy_len >= cpy_threshold)) {
1076                         hpa = (void *)(uintptr_t)gpa_to_hpa(dev,
1077                                         buf_iova + buf_offset, cpy_len);
1078
1079                         if (unlikely(!hpa)) {
1080                                 error = -1;
1081                                 goto out;
1082                         }
1083
1084                         async_fill_vec(src_iovec + tvec_idx,
1085                                 (void *)(uintptr_t)rte_pktmbuf_iova_offset(m,
1086                                                 mbuf_offset), cpy_len);
1087
1088                         async_fill_vec(dst_iovec + tvec_idx, hpa, cpy_len);
1089
1090                         tlen += cpy_len;
1091                         tvec_idx++;
1092                 } else {
1093                         if (unlikely(vq->batch_copy_nb_elems >= vq->size)) {
1094                                 rte_memcpy(
1095                                 (void *)((uintptr_t)(buf_addr + buf_offset)),
1096                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
1097                                 cpy_len);
1098
1099                                 PRINT_PACKET(dev,
1100                                         (uintptr_t)(buf_addr + buf_offset),
1101                                         cpy_len, 0);
1102                         } else {
1103                                 batch_copy[vq->batch_copy_nb_elems].dst =
1104                                 (void *)((uintptr_t)(buf_addr + buf_offset));
1105                                 batch_copy[vq->batch_copy_nb_elems].src =
1106                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
1107                                 batch_copy[vq->batch_copy_nb_elems].log_addr =
1108                                         buf_iova + buf_offset;
1109                                 batch_copy[vq->batch_copy_nb_elems].len =
1110                                         cpy_len;
1111                                 vq->batch_copy_nb_elems++;
1112                         }
1113                 }
1114
1115                 mbuf_avail  -= cpy_len;
1116                 mbuf_offset += cpy_len;
1117                 buf_avail  -= cpy_len;
1118                 buf_offset += cpy_len;
1119         }
1120
1121 out:
1122         async_fill_iter(src_it, tlen, src_iovec, tvec_idx);
1123         async_fill_iter(dst_it, tlen, dst_iovec, tvec_idx);
1124
1125         return error;
1126 }
1127
1128 static __rte_always_inline int
1129 vhost_enqueue_single_packed(struct virtio_net *dev,
1130                             struct vhost_virtqueue *vq,
1131                             struct rte_mbuf *pkt,
1132                             struct buf_vector *buf_vec,
1133                             uint16_t *nr_descs)
1134 {
1135         uint16_t nr_vec = 0;
1136         uint16_t avail_idx = vq->last_avail_idx;
1137         uint16_t max_tries, tries = 0;
1138         uint16_t buf_id = 0;
1139         uint32_t len = 0;
1140         uint16_t desc_count;
1141         uint32_t size = pkt->pkt_len + dev->vhost_hlen;
1142         uint16_t num_buffers = 0;
1143         uint32_t buffer_len[vq->size];
1144         uint16_t buffer_buf_id[vq->size];
1145         uint16_t buffer_desc_count[vq->size];
1146
1147         if (rxvq_is_mergeable(dev))
1148                 max_tries = vq->size - 1;
1149         else
1150                 max_tries = 1;
1151
1152         while (size > 0) {
1153                 /*
1154                  * if we tried all available ring items, and still
1155                  * can't get enough buf, it means something abnormal
1156                  * happened.
1157                  */
1158                 if (unlikely(++tries > max_tries))
1159                         return -1;
1160
1161                 if (unlikely(fill_vec_buf_packed(dev, vq,
1162                                                 avail_idx, &desc_count,
1163                                                 buf_vec, &nr_vec,
1164                                                 &buf_id, &len,
1165                                                 VHOST_ACCESS_RW) < 0))
1166                         return -1;
1167
1168                 len = RTE_MIN(len, size);
1169                 size -= len;
1170
1171                 buffer_len[num_buffers] = len;
1172                 buffer_buf_id[num_buffers] = buf_id;
1173                 buffer_desc_count[num_buffers] = desc_count;
1174                 num_buffers += 1;
1175
1176                 *nr_descs += desc_count;
1177                 avail_idx += desc_count;
1178                 if (avail_idx >= vq->size)
1179                         avail_idx -= vq->size;
1180         }
1181
1182         if (copy_mbuf_to_desc(dev, vq, pkt, buf_vec, nr_vec, num_buffers) < 0)
1183                 return -1;
1184
1185         vhost_shadow_enqueue_single_packed(dev, vq, buffer_len, buffer_buf_id,
1186                                            buffer_desc_count, num_buffers);
1187
1188         return 0;
1189 }
1190
1191 static __rte_noinline uint32_t
1192 virtio_dev_rx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
1193         struct rte_mbuf **pkts, uint32_t count)
1194 {
1195         uint32_t pkt_idx = 0;
1196         uint16_t num_buffers;
1197         struct buf_vector buf_vec[BUF_VECTOR_MAX];
1198         uint16_t avail_head;
1199
1200         /*
1201          * The ordering between avail index and
1202          * desc reads needs to be enforced.
1203          */
1204         avail_head = __atomic_load_n(&vq->avail->idx, __ATOMIC_ACQUIRE);
1205
1206         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
1207
1208         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
1209                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
1210                 uint16_t nr_vec = 0;
1211
1212                 if (unlikely(reserve_avail_buf_split(dev, vq,
1213                                                 pkt_len, buf_vec, &num_buffers,
1214                                                 avail_head, &nr_vec) < 0)) {
1215                         VHOST_LOG_DATA(DEBUG,
1216                                 "(%d) failed to get enough desc from vring\n",
1217                                 dev->vid);
1218                         vq->shadow_used_idx -= num_buffers;
1219                         break;
1220                 }
1221
1222                 VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
1223                         dev->vid, vq->last_avail_idx,
1224                         vq->last_avail_idx + num_buffers);
1225
1226                 if (copy_mbuf_to_desc(dev, vq, pkts[pkt_idx],
1227                                                 buf_vec, nr_vec,
1228                                                 num_buffers) < 0) {
1229                         vq->shadow_used_idx -= num_buffers;
1230                         break;
1231                 }
1232
1233                 vq->last_avail_idx += num_buffers;
1234         }
1235
1236         do_data_copy_enqueue(dev, vq);
1237
1238         if (likely(vq->shadow_used_idx)) {
1239                 flush_shadow_used_ring_split(dev, vq);
1240                 vhost_vring_call_split(dev, vq);
1241         }
1242
1243         return pkt_idx;
1244 }
1245
1246 static __rte_always_inline int
1247 virtio_dev_rx_batch_packed(struct virtio_net *dev,
1248                            struct vhost_virtqueue *vq,
1249                            struct rte_mbuf **pkts)
1250 {
1251         bool wrap_counter = vq->avail_wrap_counter;
1252         struct vring_packed_desc *descs = vq->desc_packed;
1253         uint16_t avail_idx = vq->last_avail_idx;
1254         uint64_t desc_addrs[PACKED_BATCH_SIZE];
1255         struct virtio_net_hdr_mrg_rxbuf *hdrs[PACKED_BATCH_SIZE];
1256         uint32_t buf_offset = dev->vhost_hlen;
1257         uint64_t lens[PACKED_BATCH_SIZE];
1258         uint16_t ids[PACKED_BATCH_SIZE];
1259         uint16_t i;
1260
1261         if (unlikely(avail_idx & PACKED_BATCH_MASK))
1262                 return -1;
1263
1264         if (unlikely((avail_idx + PACKED_BATCH_SIZE) > vq->size))
1265                 return -1;
1266
1267         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1268                 if (unlikely(pkts[i]->next != NULL))
1269                         return -1;
1270                 if (unlikely(!desc_is_avail(&descs[avail_idx + i],
1271                                             wrap_counter)))
1272                         return -1;
1273         }
1274
1275         rte_smp_rmb();
1276
1277         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1278                 lens[i] = descs[avail_idx + i].len;
1279
1280         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1281                 if (unlikely(pkts[i]->pkt_len > (lens[i] - buf_offset)))
1282                         return -1;
1283         }
1284
1285         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1286                 desc_addrs[i] = vhost_iova_to_vva(dev, vq,
1287                                                   descs[avail_idx + i].addr,
1288                                                   &lens[i],
1289                                                   VHOST_ACCESS_RW);
1290
1291         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1292                 if (unlikely(!desc_addrs[i]))
1293                         return -1;
1294                 if (unlikely(lens[i] != descs[avail_idx + i].len))
1295                         return -1;
1296         }
1297
1298         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1299                 rte_prefetch0((void *)(uintptr_t)desc_addrs[i]);
1300                 hdrs[i] = (struct virtio_net_hdr_mrg_rxbuf *)
1301                                         (uintptr_t)desc_addrs[i];
1302                 lens[i] = pkts[i]->pkt_len + dev->vhost_hlen;
1303         }
1304
1305         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1306                 virtio_enqueue_offload(pkts[i], &hdrs[i]->hdr);
1307
1308         vq_inc_last_avail_packed(vq, PACKED_BATCH_SIZE);
1309
1310         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1311                 rte_memcpy((void *)(uintptr_t)(desc_addrs[i] + buf_offset),
1312                            rte_pktmbuf_mtod_offset(pkts[i], void *, 0),
1313                            pkts[i]->pkt_len);
1314         }
1315
1316         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1317                 vhost_log_cache_write_iova(dev, vq, descs[avail_idx + i].addr,
1318                                            lens[i]);
1319
1320         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1321                 ids[i] = descs[avail_idx + i].id;
1322
1323         vhost_flush_enqueue_batch_packed(dev, vq, lens, ids);
1324
1325         return 0;
1326 }
1327
1328 static __rte_always_inline int16_t
1329 virtio_dev_rx_single_packed(struct virtio_net *dev,
1330                             struct vhost_virtqueue *vq,
1331                             struct rte_mbuf *pkt)
1332 {
1333         struct buf_vector buf_vec[BUF_VECTOR_MAX];
1334         uint16_t nr_descs = 0;
1335
1336         rte_smp_rmb();
1337         if (unlikely(vhost_enqueue_single_packed(dev, vq, pkt, buf_vec,
1338                                                  &nr_descs) < 0)) {
1339                 VHOST_LOG_DATA(DEBUG,
1340                                 "(%d) failed to get enough desc from vring\n",
1341                                 dev->vid);
1342                 return -1;
1343         }
1344
1345         VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
1346                         dev->vid, vq->last_avail_idx,
1347                         vq->last_avail_idx + nr_descs);
1348
1349         vq_inc_last_avail_packed(vq, nr_descs);
1350
1351         return 0;
1352 }
1353
1354 static __rte_noinline uint32_t
1355 virtio_dev_rx_packed(struct virtio_net *dev,
1356                      struct vhost_virtqueue *__rte_restrict vq,
1357                      struct rte_mbuf **__rte_restrict pkts,
1358                      uint32_t count)
1359 {
1360         uint32_t pkt_idx = 0;
1361         uint32_t remained = count;
1362
1363         do {
1364                 rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
1365
1366                 if (remained >= PACKED_BATCH_SIZE) {
1367                         if (!virtio_dev_rx_batch_packed(dev, vq,
1368                                                         &pkts[pkt_idx])) {
1369                                 pkt_idx += PACKED_BATCH_SIZE;
1370                                 remained -= PACKED_BATCH_SIZE;
1371                                 continue;
1372                         }
1373                 }
1374
1375                 if (virtio_dev_rx_single_packed(dev, vq, pkts[pkt_idx]))
1376                         break;
1377                 pkt_idx++;
1378                 remained--;
1379
1380         } while (pkt_idx < count);
1381
1382         if (vq->shadow_used_idx) {
1383                 do_data_copy_enqueue(dev, vq);
1384                 vhost_flush_enqueue_shadow_packed(dev, vq);
1385         }
1386
1387         if (pkt_idx)
1388                 vhost_vring_call_packed(dev, vq);
1389
1390         return pkt_idx;
1391 }
1392
1393 static __rte_always_inline uint32_t
1394 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
1395         struct rte_mbuf **pkts, uint32_t count)
1396 {
1397         struct vhost_virtqueue *vq;
1398         uint32_t nb_tx = 0;
1399
1400         VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
1401         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
1402                 VHOST_LOG_DATA(ERR, "(%d) %s: invalid virtqueue idx %d.\n",
1403                         dev->vid, __func__, queue_id);
1404                 return 0;
1405         }
1406
1407         vq = dev->virtqueue[queue_id];
1408
1409         rte_spinlock_lock(&vq->access_lock);
1410
1411         if (unlikely(vq->enabled == 0))
1412                 goto out_access_unlock;
1413
1414         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1415                 vhost_user_iotlb_rd_lock(vq);
1416
1417         if (unlikely(vq->access_ok == 0))
1418                 if (unlikely(vring_translate(dev, vq) < 0))
1419                         goto out;
1420
1421         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
1422         if (count == 0)
1423                 goto out;
1424
1425         if (vq_is_packed(dev))
1426                 nb_tx = virtio_dev_rx_packed(dev, vq, pkts, count);
1427         else
1428                 nb_tx = virtio_dev_rx_split(dev, vq, pkts, count);
1429
1430 out:
1431         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1432                 vhost_user_iotlb_rd_unlock(vq);
1433
1434 out_access_unlock:
1435         rte_spinlock_unlock(&vq->access_lock);
1436
1437         return nb_tx;
1438 }
1439
1440 uint16_t
1441 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
1442         struct rte_mbuf **__rte_restrict pkts, uint16_t count)
1443 {
1444         struct virtio_net *dev = get_device(vid);
1445
1446         if (!dev)
1447                 return 0;
1448
1449         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
1450                 VHOST_LOG_DATA(ERR,
1451                         "(%d) %s: built-in vhost net backend is disabled.\n",
1452                         dev->vid, __func__);
1453                 return 0;
1454         }
1455
1456         return virtio_dev_rx(dev, queue_id, pkts, count);
1457 }
1458
1459 static __rte_always_inline uint16_t
1460 virtio_dev_rx_async_get_info_idx(uint16_t pkts_idx,
1461         uint16_t vq_size, uint16_t n_inflight)
1462 {
1463         return pkts_idx > n_inflight ? (pkts_idx - n_inflight) :
1464                 (vq_size - n_inflight + pkts_idx) & (vq_size - 1);
1465 }
1466
1467 static __rte_always_inline void
1468 virtio_dev_rx_async_submit_split_err(struct virtio_net *dev,
1469         struct vhost_virtqueue *vq, uint16_t queue_id,
1470         uint16_t last_idx, uint16_t shadow_idx)
1471 {
1472         uint16_t start_idx, pkts_idx, vq_size;
1473         uint64_t *async_pending_info;
1474
1475         pkts_idx = vq->async_pkts_idx;
1476         async_pending_info = vq->async_pending_info;
1477         vq_size = vq->size;
1478         start_idx = virtio_dev_rx_async_get_info_idx(pkts_idx,
1479                 vq_size, vq->async_pkts_inflight_n);
1480
1481         while (likely((start_idx & (vq_size - 1)) != pkts_idx)) {
1482                 uint64_t n_seg =
1483                         async_pending_info[(start_idx) & (vq_size - 1)] >>
1484                         ASYNC_PENDING_INFO_N_SFT;
1485
1486                 while (n_seg)
1487                         n_seg -= vq->async_ops.check_completed_copies(dev->vid,
1488                                 queue_id, 0, 1);
1489         }
1490
1491         vq->async_pkts_inflight_n = 0;
1492         vq->batch_copy_nb_elems = 0;
1493
1494         vq->shadow_used_idx = shadow_idx;
1495         vq->last_avail_idx = last_idx;
1496 }
1497
1498 static __rte_noinline uint32_t
1499 virtio_dev_rx_async_submit_split(struct virtio_net *dev,
1500         struct vhost_virtqueue *vq, uint16_t queue_id,
1501         struct rte_mbuf **pkts, uint32_t count)
1502 {
1503         uint32_t pkt_idx = 0, pkt_burst_idx = 0;
1504         uint16_t num_buffers;
1505         struct buf_vector buf_vec[BUF_VECTOR_MAX];
1506         uint16_t avail_head, last_idx, shadow_idx;
1507
1508         struct rte_vhost_iov_iter *it_pool = vq->it_pool;
1509         struct iovec *vec_pool = vq->vec_pool;
1510         struct rte_vhost_async_desc tdes[MAX_PKT_BURST];
1511         struct iovec *src_iovec = vec_pool;
1512         struct iovec *dst_iovec = vec_pool + (VHOST_MAX_ASYNC_VEC >> 1);
1513         struct rte_vhost_iov_iter *src_it = it_pool;
1514         struct rte_vhost_iov_iter *dst_it = it_pool + 1;
1515         uint16_t n_free_slot, slot_idx;
1516         int n_pkts = 0;
1517
1518         avail_head = __atomic_load_n(&vq->avail->idx, __ATOMIC_ACQUIRE);
1519         last_idx = vq->last_avail_idx;
1520         shadow_idx = vq->shadow_used_idx;
1521
1522         /*
1523          * The ordering between avail index and
1524          * desc reads needs to be enforced.
1525          */
1526         rte_smp_rmb();
1527
1528         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
1529
1530         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
1531                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
1532                 uint16_t nr_vec = 0;
1533
1534                 if (unlikely(reserve_avail_buf_split(dev, vq,
1535                                                 pkt_len, buf_vec, &num_buffers,
1536                                                 avail_head, &nr_vec) < 0)) {
1537                         VHOST_LOG_DATA(DEBUG,
1538                                 "(%d) failed to get enough desc from vring\n",
1539                                 dev->vid);
1540                         vq->shadow_used_idx -= num_buffers;
1541                         break;
1542                 }
1543
1544                 VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
1545                         dev->vid, vq->last_avail_idx,
1546                         vq->last_avail_idx + num_buffers);
1547
1548                 if (async_mbuf_to_desc(dev, vq, pkts[pkt_idx],
1549                                 buf_vec, nr_vec, num_buffers,
1550                                 src_iovec, dst_iovec, src_it, dst_it) < 0) {
1551                         vq->shadow_used_idx -= num_buffers;
1552                         break;
1553                 }
1554
1555                 slot_idx = (vq->async_pkts_idx + pkt_idx) & (vq->size - 1);
1556                 if (src_it->count) {
1557                         async_fill_desc(&tdes[pkt_burst_idx], src_it, dst_it);
1558                         pkt_burst_idx++;
1559                         vq->async_pending_info[slot_idx] =
1560                                 num_buffers | (src_it->nr_segs << 16);
1561                         src_iovec += src_it->nr_segs;
1562                         dst_iovec += dst_it->nr_segs;
1563                         src_it += 2;
1564                         dst_it += 2;
1565                 } else {
1566                         vq->async_pending_info[slot_idx] = num_buffers;
1567                         vq->async_pkts_inflight_n++;
1568                 }
1569
1570                 vq->last_avail_idx += num_buffers;
1571
1572                 if (pkt_burst_idx >= VHOST_ASYNC_BATCH_THRESHOLD ||
1573                                 (pkt_idx == count - 1 && pkt_burst_idx)) {
1574                         n_pkts = vq->async_ops.transfer_data(dev->vid,
1575                                         queue_id, tdes, 0, pkt_burst_idx);
1576                         src_iovec = vec_pool;
1577                         dst_iovec = vec_pool + (VHOST_MAX_ASYNC_VEC >> 1);
1578                         src_it = it_pool;
1579                         dst_it = it_pool + 1;
1580
1581                         if (unlikely(n_pkts < (int)pkt_burst_idx)) {
1582                                 vq->async_pkts_inflight_n +=
1583                                         n_pkts > 0 ? n_pkts : 0;
1584                                 virtio_dev_rx_async_submit_split_err(dev,
1585                                         vq, queue_id, last_idx, shadow_idx);
1586                                 return 0;
1587                         }
1588
1589                         pkt_burst_idx = 0;
1590                         vq->async_pkts_inflight_n += n_pkts;
1591                 }
1592         }
1593
1594         if (pkt_burst_idx) {
1595                 n_pkts = vq->async_ops.transfer_data(dev->vid,
1596                                 queue_id, tdes, 0, pkt_burst_idx);
1597                 if (unlikely(n_pkts < (int)pkt_burst_idx)) {
1598                         vq->async_pkts_inflight_n += n_pkts > 0 ? n_pkts : 0;
1599                         virtio_dev_rx_async_submit_split_err(dev, vq, queue_id,
1600                                 last_idx, shadow_idx);
1601                         return 0;
1602                 }
1603
1604                 vq->async_pkts_inflight_n += n_pkts;
1605         }
1606
1607         do_data_copy_enqueue(dev, vq);
1608
1609         n_free_slot = vq->size - vq->async_pkts_idx;
1610         if (n_free_slot > pkt_idx) {
1611                 rte_memcpy(&vq->async_pkts_pending[vq->async_pkts_idx],
1612                         pkts, pkt_idx * sizeof(uintptr_t));
1613                 vq->async_pkts_idx += pkt_idx;
1614         } else {
1615                 rte_memcpy(&vq->async_pkts_pending[vq->async_pkts_idx],
1616                         pkts, n_free_slot * sizeof(uintptr_t));
1617                 rte_memcpy(&vq->async_pkts_pending[0],
1618                         &pkts[n_free_slot],
1619                         (pkt_idx - n_free_slot) * sizeof(uintptr_t));
1620                 vq->async_pkts_idx = pkt_idx - n_free_slot;
1621         }
1622
1623         if (likely(vq->shadow_used_idx))
1624                 async_flush_shadow_used_ring_split(dev, vq);
1625
1626         return pkt_idx;
1627 }
1628
1629 uint16_t rte_vhost_poll_enqueue_completed(int vid, uint16_t queue_id,
1630                 struct rte_mbuf **pkts, uint16_t count)
1631 {
1632         struct virtio_net *dev = get_device(vid);
1633         struct vhost_virtqueue *vq;
1634         uint16_t n_pkts_cpl, n_pkts_put = 0, n_descs = 0;
1635         uint16_t start_idx, pkts_idx, vq_size;
1636         uint64_t *async_pending_info;
1637
1638         if (!dev)
1639                 return 0;
1640
1641         VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
1642         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
1643                 VHOST_LOG_DATA(ERR, "(%d) %s: invalid virtqueue idx %d.\n",
1644                         dev->vid, __func__, queue_id);
1645                 return 0;
1646         }
1647
1648         vq = dev->virtqueue[queue_id];
1649
1650         rte_spinlock_lock(&vq->access_lock);
1651
1652         pkts_idx = vq->async_pkts_idx;
1653         async_pending_info = vq->async_pending_info;
1654         vq_size = vq->size;
1655         start_idx = virtio_dev_rx_async_get_info_idx(pkts_idx,
1656                 vq_size, vq->async_pkts_inflight_n);
1657
1658         n_pkts_cpl =
1659                 vq->async_ops.check_completed_copies(vid, queue_id, 0, count);
1660
1661         rte_smp_wmb();
1662
1663         while (likely(((start_idx + n_pkts_put) & (vq_size - 1)) != pkts_idx)) {
1664                 uint64_t info = async_pending_info[
1665                         (start_idx + n_pkts_put) & (vq_size - 1)];
1666                 uint64_t n_segs;
1667                 n_pkts_put++;
1668                 n_descs += info & ASYNC_PENDING_INFO_N_MSK;
1669                 n_segs = info >> ASYNC_PENDING_INFO_N_SFT;
1670
1671                 if (n_segs) {
1672                         if (!n_pkts_cpl || n_pkts_cpl < n_segs) {
1673                                 n_pkts_put--;
1674                                 n_descs -= info & ASYNC_PENDING_INFO_N_MSK;
1675                                 if (n_pkts_cpl) {
1676                                         async_pending_info[
1677                                                 (start_idx + n_pkts_put) &
1678                                                 (vq_size - 1)] =
1679                                         ((n_segs - n_pkts_cpl) <<
1680                                          ASYNC_PENDING_INFO_N_SFT) |
1681                                         (info & ASYNC_PENDING_INFO_N_MSK);
1682                                         n_pkts_cpl = 0;
1683                                 }
1684                                 break;
1685                         }
1686                         n_pkts_cpl -= n_segs;
1687                 }
1688         }
1689
1690         if (n_pkts_put) {
1691                 vq->async_pkts_inflight_n -= n_pkts_put;
1692                 __atomic_add_fetch(&vq->used->idx, n_descs, __ATOMIC_RELEASE);
1693
1694                 vhost_vring_call_split(dev, vq);
1695         }
1696
1697         if (start_idx + n_pkts_put <= vq_size) {
1698                 rte_memcpy(pkts, &vq->async_pkts_pending[start_idx],
1699                         n_pkts_put * sizeof(uintptr_t));
1700         } else {
1701                 rte_memcpy(pkts, &vq->async_pkts_pending[start_idx],
1702                         (vq_size - start_idx) * sizeof(uintptr_t));
1703                 rte_memcpy(&pkts[vq_size - start_idx], vq->async_pkts_pending,
1704                         (n_pkts_put - vq_size + start_idx) * sizeof(uintptr_t));
1705         }
1706
1707         rte_spinlock_unlock(&vq->access_lock);
1708
1709         return n_pkts_put;
1710 }
1711
1712 static __rte_always_inline uint32_t
1713 virtio_dev_rx_async_submit(struct virtio_net *dev, uint16_t queue_id,
1714         struct rte_mbuf **pkts, uint32_t count)
1715 {
1716         struct vhost_virtqueue *vq;
1717         uint32_t nb_tx = 0;
1718         bool drawback = false;
1719
1720         VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
1721         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
1722                 VHOST_LOG_DATA(ERR, "(%d) %s: invalid virtqueue idx %d.\n",
1723                         dev->vid, __func__, queue_id);
1724                 return 0;
1725         }
1726
1727         vq = dev->virtqueue[queue_id];
1728
1729         rte_spinlock_lock(&vq->access_lock);
1730
1731         if (unlikely(vq->enabled == 0))
1732                 goto out_access_unlock;
1733
1734         if (unlikely(!vq->async_registered)) {
1735                 drawback = true;
1736                 goto out_access_unlock;
1737         }
1738
1739         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1740                 vhost_user_iotlb_rd_lock(vq);
1741
1742         if (unlikely(vq->access_ok == 0))
1743                 if (unlikely(vring_translate(dev, vq) < 0))
1744                         goto out;
1745
1746         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
1747         if (count == 0)
1748                 goto out;
1749
1750         /* TODO: packed queue not implemented */
1751         if (vq_is_packed(dev))
1752                 nb_tx = 0;
1753         else
1754                 nb_tx = virtio_dev_rx_async_submit_split(dev,
1755                                 vq, queue_id, pkts, count);
1756
1757 out:
1758         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1759                 vhost_user_iotlb_rd_unlock(vq);
1760
1761 out_access_unlock:
1762         rte_spinlock_unlock(&vq->access_lock);
1763
1764         if (drawback)
1765                 return rte_vhost_enqueue_burst(dev->vid, queue_id, pkts, count);
1766
1767         return nb_tx;
1768 }
1769
1770 uint16_t
1771 rte_vhost_submit_enqueue_burst(int vid, uint16_t queue_id,
1772                 struct rte_mbuf **pkts, uint16_t count)
1773 {
1774         struct virtio_net *dev = get_device(vid);
1775
1776         if (!dev)
1777                 return 0;
1778
1779         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
1780                 VHOST_LOG_DATA(ERR,
1781                         "(%d) %s: built-in vhost net backend is disabled.\n",
1782                         dev->vid, __func__);
1783                 return 0;
1784         }
1785
1786         return virtio_dev_rx_async_submit(dev, queue_id, pkts, count);
1787 }
1788
1789 static inline bool
1790 virtio_net_with_host_offload(struct virtio_net *dev)
1791 {
1792         if (dev->features &
1793                         ((1ULL << VIRTIO_NET_F_CSUM) |
1794                          (1ULL << VIRTIO_NET_F_HOST_ECN) |
1795                          (1ULL << VIRTIO_NET_F_HOST_TSO4) |
1796                          (1ULL << VIRTIO_NET_F_HOST_TSO6) |
1797                          (1ULL << VIRTIO_NET_F_HOST_UFO)))
1798                 return true;
1799
1800         return false;
1801 }
1802
1803 static void
1804 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
1805 {
1806         struct rte_ipv4_hdr *ipv4_hdr;
1807         struct rte_ipv6_hdr *ipv6_hdr;
1808         void *l3_hdr = NULL;
1809         struct rte_ether_hdr *eth_hdr;
1810         uint16_t ethertype;
1811
1812         eth_hdr = rte_pktmbuf_mtod(m, struct rte_ether_hdr *);
1813
1814         m->l2_len = sizeof(struct rte_ether_hdr);
1815         ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
1816
1817         if (ethertype == RTE_ETHER_TYPE_VLAN) {
1818                 struct rte_vlan_hdr *vlan_hdr =
1819                         (struct rte_vlan_hdr *)(eth_hdr + 1);
1820
1821                 m->l2_len += sizeof(struct rte_vlan_hdr);
1822                 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
1823         }
1824
1825         l3_hdr = (char *)eth_hdr + m->l2_len;
1826
1827         switch (ethertype) {
1828         case RTE_ETHER_TYPE_IPV4:
1829                 ipv4_hdr = l3_hdr;
1830                 *l4_proto = ipv4_hdr->next_proto_id;
1831                 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
1832                 *l4_hdr = (char *)l3_hdr + m->l3_len;
1833                 m->ol_flags |= PKT_TX_IPV4;
1834                 break;
1835         case RTE_ETHER_TYPE_IPV6:
1836                 ipv6_hdr = l3_hdr;
1837                 *l4_proto = ipv6_hdr->proto;
1838                 m->l3_len = sizeof(struct rte_ipv6_hdr);
1839                 *l4_hdr = (char *)l3_hdr + m->l3_len;
1840                 m->ol_flags |= PKT_TX_IPV6;
1841                 break;
1842         default:
1843                 m->l3_len = 0;
1844                 *l4_proto = 0;
1845                 *l4_hdr = NULL;
1846                 break;
1847         }
1848 }
1849
1850 static __rte_always_inline void
1851 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
1852 {
1853         uint16_t l4_proto = 0;
1854         void *l4_hdr = NULL;
1855         struct rte_tcp_hdr *tcp_hdr = NULL;
1856
1857         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
1858                 return;
1859
1860         parse_ethernet(m, &l4_proto, &l4_hdr);
1861         if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
1862                 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
1863                         switch (hdr->csum_offset) {
1864                         case (offsetof(struct rte_tcp_hdr, cksum)):
1865                                 if (l4_proto == IPPROTO_TCP)
1866                                         m->ol_flags |= PKT_TX_TCP_CKSUM;
1867                                 break;
1868                         case (offsetof(struct rte_udp_hdr, dgram_cksum)):
1869                                 if (l4_proto == IPPROTO_UDP)
1870                                         m->ol_flags |= PKT_TX_UDP_CKSUM;
1871                                 break;
1872                         case (offsetof(struct rte_sctp_hdr, cksum)):
1873                                 if (l4_proto == IPPROTO_SCTP)
1874                                         m->ol_flags |= PKT_TX_SCTP_CKSUM;
1875                                 break;
1876                         default:
1877                                 break;
1878                         }
1879                 }
1880         }
1881
1882         if (l4_hdr && hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
1883                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
1884                 case VIRTIO_NET_HDR_GSO_TCPV4:
1885                 case VIRTIO_NET_HDR_GSO_TCPV6:
1886                         tcp_hdr = l4_hdr;
1887                         m->ol_flags |= PKT_TX_TCP_SEG;
1888                         m->tso_segsz = hdr->gso_size;
1889                         m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
1890                         break;
1891                 case VIRTIO_NET_HDR_GSO_UDP:
1892                         m->ol_flags |= PKT_TX_UDP_SEG;
1893                         m->tso_segsz = hdr->gso_size;
1894                         m->l4_len = sizeof(struct rte_udp_hdr);
1895                         break;
1896                 default:
1897                         VHOST_LOG_DATA(WARNING,
1898                                 "unsupported gso type %u.\n", hdr->gso_type);
1899                         break;
1900                 }
1901         }
1902 }
1903
1904 static __rte_noinline void
1905 copy_vnet_hdr_from_desc(struct virtio_net_hdr *hdr,
1906                 struct buf_vector *buf_vec)
1907 {
1908         uint64_t len;
1909         uint64_t remain = sizeof(struct virtio_net_hdr);
1910         uint64_t src;
1911         uint64_t dst = (uint64_t)(uintptr_t)hdr;
1912
1913         while (remain) {
1914                 len = RTE_MIN(remain, buf_vec->buf_len);
1915                 src = buf_vec->buf_addr;
1916                 rte_memcpy((void *)(uintptr_t)dst,
1917                                 (void *)(uintptr_t)src, len);
1918
1919                 remain -= len;
1920                 dst += len;
1921                 buf_vec++;
1922         }
1923 }
1924
1925 static __rte_always_inline int
1926 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
1927                   struct buf_vector *buf_vec, uint16_t nr_vec,
1928                   struct rte_mbuf *m, struct rte_mempool *mbuf_pool)
1929 {
1930         uint32_t buf_avail, buf_offset;
1931         uint64_t buf_addr, buf_iova, buf_len;
1932         uint32_t mbuf_avail, mbuf_offset;
1933         uint32_t cpy_len;
1934         struct rte_mbuf *cur = m, *prev = m;
1935         struct virtio_net_hdr tmp_hdr;
1936         struct virtio_net_hdr *hdr = NULL;
1937         /* A counter to avoid desc dead loop chain */
1938         uint16_t vec_idx = 0;
1939         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
1940         int error = 0;
1941
1942         buf_addr = buf_vec[vec_idx].buf_addr;
1943         buf_iova = buf_vec[vec_idx].buf_iova;
1944         buf_len = buf_vec[vec_idx].buf_len;
1945
1946         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
1947                 error = -1;
1948                 goto out;
1949         }
1950
1951         if (virtio_net_with_host_offload(dev)) {
1952                 if (unlikely(buf_len < sizeof(struct virtio_net_hdr))) {
1953                         /*
1954                          * No luck, the virtio-net header doesn't fit
1955                          * in a contiguous virtual area.
1956                          */
1957                         copy_vnet_hdr_from_desc(&tmp_hdr, buf_vec);
1958                         hdr = &tmp_hdr;
1959                 } else {
1960                         hdr = (struct virtio_net_hdr *)((uintptr_t)buf_addr);
1961                 }
1962         }
1963
1964         /*
1965          * A virtio driver normally uses at least 2 desc buffers
1966          * for Tx: the first for storing the header, and others
1967          * for storing the data.
1968          */
1969         if (unlikely(buf_len < dev->vhost_hlen)) {
1970                 buf_offset = dev->vhost_hlen - buf_len;
1971                 vec_idx++;
1972                 buf_addr = buf_vec[vec_idx].buf_addr;
1973                 buf_iova = buf_vec[vec_idx].buf_iova;
1974                 buf_len = buf_vec[vec_idx].buf_len;
1975                 buf_avail  = buf_len - buf_offset;
1976         } else if (buf_len == dev->vhost_hlen) {
1977                 if (unlikely(++vec_idx >= nr_vec))
1978                         goto out;
1979                 buf_addr = buf_vec[vec_idx].buf_addr;
1980                 buf_iova = buf_vec[vec_idx].buf_iova;
1981                 buf_len = buf_vec[vec_idx].buf_len;
1982
1983                 buf_offset = 0;
1984                 buf_avail = buf_len;
1985         } else {
1986                 buf_offset = dev->vhost_hlen;
1987                 buf_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
1988         }
1989
1990         PRINT_PACKET(dev,
1991                         (uintptr_t)(buf_addr + buf_offset),
1992                         (uint32_t)buf_avail, 0);
1993
1994         mbuf_offset = 0;
1995         mbuf_avail  = m->buf_len - RTE_PKTMBUF_HEADROOM;
1996         while (1) {
1997                 uint64_t hpa;
1998
1999                 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
2000
2001                 /*
2002                  * A desc buf might across two host physical pages that are
2003                  * not continuous. In such case (gpa_to_hpa returns 0), data
2004                  * will be copied even though zero copy is enabled.
2005                  */
2006                 if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
2007                                         buf_iova + buf_offset, cpy_len)))) {
2008                         cur->data_len = cpy_len;
2009                         cur->data_off = 0;
2010                         cur->buf_addr =
2011                                 (void *)(uintptr_t)(buf_addr + buf_offset);
2012                         cur->buf_iova = hpa;
2013
2014                         /*
2015                          * In zero copy mode, one mbuf can only reference data
2016                          * for one or partial of one desc buff.
2017                          */
2018                         mbuf_avail = cpy_len;
2019                 } else {
2020                         if (likely(cpy_len > MAX_BATCH_LEN ||
2021                                    vq->batch_copy_nb_elems >= vq->size ||
2022                                    (hdr && cur == m))) {
2023                                 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
2024                                                                    mbuf_offset),
2025                                            (void *)((uintptr_t)(buf_addr +
2026                                                            buf_offset)),
2027                                            cpy_len);
2028                         } else {
2029                                 batch_copy[vq->batch_copy_nb_elems].dst =
2030                                         rte_pktmbuf_mtod_offset(cur, void *,
2031                                                                 mbuf_offset);
2032                                 batch_copy[vq->batch_copy_nb_elems].src =
2033                                         (void *)((uintptr_t)(buf_addr +
2034                                                                 buf_offset));
2035                                 batch_copy[vq->batch_copy_nb_elems].len =
2036                                         cpy_len;
2037                                 vq->batch_copy_nb_elems++;
2038                         }
2039                 }
2040
2041                 mbuf_avail  -= cpy_len;
2042                 mbuf_offset += cpy_len;
2043                 buf_avail -= cpy_len;
2044                 buf_offset += cpy_len;
2045
2046                 /* This buf reaches to its end, get the next one */
2047                 if (buf_avail == 0) {
2048                         if (++vec_idx >= nr_vec)
2049                                 break;
2050
2051                         buf_addr = buf_vec[vec_idx].buf_addr;
2052                         buf_iova = buf_vec[vec_idx].buf_iova;
2053                         buf_len = buf_vec[vec_idx].buf_len;
2054
2055                         buf_offset = 0;
2056                         buf_avail  = buf_len;
2057
2058                         PRINT_PACKET(dev, (uintptr_t)buf_addr,
2059                                         (uint32_t)buf_avail, 0);
2060                 }
2061
2062                 /*
2063                  * This mbuf reaches to its end, get a new one
2064                  * to hold more data.
2065                  */
2066                 if (mbuf_avail == 0) {
2067                         cur = rte_pktmbuf_alloc(mbuf_pool);
2068                         if (unlikely(cur == NULL)) {
2069                                 VHOST_LOG_DATA(ERR, "Failed to "
2070                                         "allocate memory for mbuf.\n");
2071                                 error = -1;
2072                                 goto out;
2073                         }
2074                         if (unlikely(dev->dequeue_zero_copy))
2075                                 rte_mbuf_refcnt_update(cur, 1);
2076
2077                         prev->next = cur;
2078                         prev->data_len = mbuf_offset;
2079                         m->nb_segs += 1;
2080                         m->pkt_len += mbuf_offset;
2081                         prev = cur;
2082
2083                         mbuf_offset = 0;
2084                         mbuf_avail  = cur->buf_len - RTE_PKTMBUF_HEADROOM;
2085                 }
2086         }
2087
2088         prev->data_len = mbuf_offset;
2089         m->pkt_len    += mbuf_offset;
2090
2091         if (hdr)
2092                 vhost_dequeue_offload(hdr, m);
2093
2094 out:
2095
2096         return error;
2097 }
2098
2099 static __rte_always_inline struct zcopy_mbuf *
2100 get_zmbuf(struct vhost_virtqueue *vq)
2101 {
2102         uint16_t i;
2103         uint16_t last;
2104         int tries = 0;
2105
2106         /* search [last_zmbuf_idx, zmbuf_size) */
2107         i = vq->last_zmbuf_idx;
2108         last = vq->zmbuf_size;
2109
2110 again:
2111         for (; i < last; i++) {
2112                 if (vq->zmbufs[i].in_use == 0) {
2113                         vq->last_zmbuf_idx = i + 1;
2114                         vq->zmbufs[i].in_use = 1;
2115                         return &vq->zmbufs[i];
2116                 }
2117         }
2118
2119         tries++;
2120         if (tries == 1) {
2121                 /* search [0, last_zmbuf_idx) */
2122                 i = 0;
2123                 last = vq->last_zmbuf_idx;
2124                 goto again;
2125         }
2126
2127         return NULL;
2128 }
2129
2130 static void
2131 virtio_dev_extbuf_free(void *addr __rte_unused, void *opaque)
2132 {
2133         rte_free(opaque);
2134 }
2135
2136 static int
2137 virtio_dev_extbuf_alloc(struct rte_mbuf *pkt, uint32_t size)
2138 {
2139         struct rte_mbuf_ext_shared_info *shinfo = NULL;
2140         uint32_t total_len = RTE_PKTMBUF_HEADROOM + size;
2141         uint16_t buf_len;
2142         rte_iova_t iova;
2143         void *buf;
2144
2145         /* Try to use pkt buffer to store shinfo to reduce the amount of memory
2146          * required, otherwise store shinfo in the new buffer.
2147          */
2148         if (rte_pktmbuf_tailroom(pkt) >= sizeof(*shinfo))
2149                 shinfo = rte_pktmbuf_mtod(pkt,
2150                                           struct rte_mbuf_ext_shared_info *);
2151         else {
2152                 total_len += sizeof(*shinfo) + sizeof(uintptr_t);
2153                 total_len = RTE_ALIGN_CEIL(total_len, sizeof(uintptr_t));
2154         }
2155
2156         if (unlikely(total_len > UINT16_MAX))
2157                 return -ENOSPC;
2158
2159         buf_len = total_len;
2160         buf = rte_malloc(NULL, buf_len, RTE_CACHE_LINE_SIZE);
2161         if (unlikely(buf == NULL))
2162                 return -ENOMEM;
2163
2164         /* Initialize shinfo */
2165         if (shinfo) {
2166                 shinfo->free_cb = virtio_dev_extbuf_free;
2167                 shinfo->fcb_opaque = buf;
2168                 rte_mbuf_ext_refcnt_set(shinfo, 1);
2169         } else {
2170                 shinfo = rte_pktmbuf_ext_shinfo_init_helper(buf, &buf_len,
2171                                               virtio_dev_extbuf_free, buf);
2172                 if (unlikely(shinfo == NULL)) {
2173                         rte_free(buf);
2174                         VHOST_LOG_DATA(ERR, "Failed to init shinfo\n");
2175                         return -1;
2176                 }
2177         }
2178
2179         iova = rte_malloc_virt2iova(buf);
2180         rte_pktmbuf_attach_extbuf(pkt, buf, iova, buf_len, shinfo);
2181         rte_pktmbuf_reset_headroom(pkt);
2182
2183         return 0;
2184 }
2185
2186 /*
2187  * Allocate a host supported pktmbuf.
2188  */
2189 static __rte_always_inline struct rte_mbuf *
2190 virtio_dev_pktmbuf_alloc(struct virtio_net *dev, struct rte_mempool *mp,
2191                          uint32_t data_len)
2192 {
2193         struct rte_mbuf *pkt = rte_pktmbuf_alloc(mp);
2194
2195         if (unlikely(pkt == NULL)) {
2196                 VHOST_LOG_DATA(ERR,
2197                         "Failed to allocate memory for mbuf.\n");
2198                 return NULL;
2199         }
2200
2201         if (rte_pktmbuf_tailroom(pkt) >= data_len)
2202                 return pkt;
2203
2204         /* attach an external buffer if supported */
2205         if (dev->extbuf && !virtio_dev_extbuf_alloc(pkt, data_len))
2206                 return pkt;
2207
2208         /* check if chained buffers are allowed */
2209         if (!dev->linearbuf)
2210                 return pkt;
2211
2212         /* Data doesn't fit into the buffer and the host supports
2213          * only linear buffers
2214          */
2215         rte_pktmbuf_free(pkt);
2216
2217         return NULL;
2218 }
2219
2220 static __rte_noinline uint16_t
2221 virtio_dev_tx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
2222         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
2223 {
2224         uint16_t i;
2225         uint16_t free_entries;
2226         uint16_t dropped = 0;
2227         static bool allocerr_warned;
2228
2229         if (unlikely(dev->dequeue_zero_copy)) {
2230                 struct zcopy_mbuf *zmbuf, *next;
2231
2232                 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
2233                      zmbuf != NULL; zmbuf = next) {
2234                         next = TAILQ_NEXT(zmbuf, next);
2235
2236                         if (mbuf_is_consumed(zmbuf->mbuf)) {
2237                                 update_shadow_used_ring_split(vq,
2238                                                 zmbuf->desc_idx, 0);
2239                                 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
2240                                 restore_mbuf(zmbuf->mbuf);
2241                                 rte_pktmbuf_free(zmbuf->mbuf);
2242                                 put_zmbuf(zmbuf);
2243                                 vq->nr_zmbuf -= 1;
2244                         }
2245                 }
2246
2247                 if (likely(vq->shadow_used_idx)) {
2248                         flush_shadow_used_ring_split(dev, vq);
2249                         vhost_vring_call_split(dev, vq);
2250                 }
2251         }
2252
2253         /*
2254          * The ordering between avail index and
2255          * desc reads needs to be enforced.
2256          */
2257         free_entries = __atomic_load_n(&vq->avail->idx, __ATOMIC_ACQUIRE) -
2258                         vq->last_avail_idx;
2259         if (free_entries == 0)
2260                 return 0;
2261
2262         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
2263
2264         VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
2265
2266         count = RTE_MIN(count, MAX_PKT_BURST);
2267         count = RTE_MIN(count, free_entries);
2268         VHOST_LOG_DATA(DEBUG, "(%d) about to dequeue %u buffers\n",
2269                         dev->vid, count);
2270
2271         for (i = 0; i < count; i++) {
2272                 struct buf_vector buf_vec[BUF_VECTOR_MAX];
2273                 uint16_t head_idx;
2274                 uint32_t buf_len;
2275                 uint16_t nr_vec = 0;
2276                 int err;
2277
2278                 if (unlikely(fill_vec_buf_split(dev, vq,
2279                                                 vq->last_avail_idx + i,
2280                                                 &nr_vec, buf_vec,
2281                                                 &head_idx, &buf_len,
2282                                                 VHOST_ACCESS_RO) < 0))
2283                         break;
2284
2285                 if (likely(dev->dequeue_zero_copy == 0))
2286                         update_shadow_used_ring_split(vq, head_idx, 0);
2287
2288                 pkts[i] = virtio_dev_pktmbuf_alloc(dev, mbuf_pool, buf_len);
2289                 if (unlikely(pkts[i] == NULL)) {
2290                         /*
2291                          * mbuf allocation fails for jumbo packets when external
2292                          * buffer allocation is not allowed and linear buffer
2293                          * is required. Drop this packet.
2294                          */
2295                         if (!allocerr_warned) {
2296                                 VHOST_LOG_DATA(ERR,
2297                                         "Failed mbuf alloc of size %d from %s on %s.\n",
2298                                         buf_len, mbuf_pool->name, dev->ifname);
2299                                 allocerr_warned = true;
2300                         }
2301                         dropped += 1;
2302                         i++;
2303                         break;
2304                 }
2305
2306                 err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts[i],
2307                                 mbuf_pool);
2308                 if (unlikely(err)) {
2309                         rte_pktmbuf_free(pkts[i]);
2310                         if (!allocerr_warned) {
2311                                 VHOST_LOG_DATA(ERR,
2312                                         "Failed to copy desc to mbuf on %s.\n",
2313                                         dev->ifname);
2314                                 allocerr_warned = true;
2315                         }
2316                         dropped += 1;
2317                         i++;
2318                         break;
2319                 }
2320
2321                 if (unlikely(dev->dequeue_zero_copy)) {
2322                         struct zcopy_mbuf *zmbuf;
2323
2324                         zmbuf = get_zmbuf(vq);
2325                         if (!zmbuf) {
2326                                 rte_pktmbuf_free(pkts[i]);
2327                                 dropped += 1;
2328                                 i++;
2329                                 break;
2330                         }
2331                         zmbuf->mbuf = pkts[i];
2332                         zmbuf->desc_idx = head_idx;
2333
2334                         /*
2335                          * Pin lock the mbuf; we will check later to see
2336                          * whether the mbuf is freed (when we are the last
2337                          * user) or not. If that's the case, we then could
2338                          * update the used ring safely.
2339                          */
2340                         rte_mbuf_refcnt_update(pkts[i], 1);
2341
2342                         vq->nr_zmbuf += 1;
2343                         TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
2344                 }
2345         }
2346         vq->last_avail_idx += i;
2347
2348         if (likely(dev->dequeue_zero_copy == 0)) {
2349                 do_data_copy_dequeue(vq);
2350                 if (unlikely(i < count))
2351                         vq->shadow_used_idx = i;
2352                 if (likely(vq->shadow_used_idx)) {
2353                         flush_shadow_used_ring_split(dev, vq);
2354                         vhost_vring_call_split(dev, vq);
2355                 }
2356         }
2357
2358         return (i - dropped);
2359 }
2360
2361 static __rte_always_inline int
2362 vhost_reserve_avail_batch_packed(struct virtio_net *dev,
2363                                  struct vhost_virtqueue *vq,
2364                                  struct rte_mempool *mbuf_pool,
2365                                  struct rte_mbuf **pkts,
2366                                  uint16_t avail_idx,
2367                                  uintptr_t *desc_addrs,
2368                                  uint16_t *ids)
2369 {
2370         bool wrap = vq->avail_wrap_counter;
2371         struct vring_packed_desc *descs = vq->desc_packed;
2372         struct virtio_net_hdr *hdr;
2373         uint64_t lens[PACKED_BATCH_SIZE];
2374         uint64_t buf_lens[PACKED_BATCH_SIZE];
2375         uint32_t buf_offset = dev->vhost_hlen;
2376         uint16_t flags, i;
2377
2378         if (unlikely(avail_idx & PACKED_BATCH_MASK))
2379                 return -1;
2380         if (unlikely((avail_idx + PACKED_BATCH_SIZE) > vq->size))
2381                 return -1;
2382
2383         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2384                 flags = descs[avail_idx + i].flags;
2385                 if (unlikely((wrap != !!(flags & VRING_DESC_F_AVAIL)) ||
2386                              (wrap == !!(flags & VRING_DESC_F_USED))  ||
2387                              (flags & PACKED_DESC_SINGLE_DEQUEUE_FLAG)))
2388                         return -1;
2389         }
2390
2391         rte_smp_rmb();
2392
2393         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2394                 lens[i] = descs[avail_idx + i].len;
2395
2396         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2397                 desc_addrs[i] = vhost_iova_to_vva(dev, vq,
2398                                                   descs[avail_idx + i].addr,
2399                                                   &lens[i], VHOST_ACCESS_RW);
2400         }
2401
2402         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2403                 if (unlikely(!desc_addrs[i]))
2404                         return -1;
2405                 if (unlikely((lens[i] != descs[avail_idx + i].len)))
2406                         return -1;
2407         }
2408
2409         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2410                 pkts[i] = virtio_dev_pktmbuf_alloc(dev, mbuf_pool, lens[i]);
2411                 if (!pkts[i])
2412                         goto free_buf;
2413         }
2414
2415         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2416                 buf_lens[i] = pkts[i]->buf_len - pkts[i]->data_off;
2417
2418         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2419                 if (unlikely(buf_lens[i] < (lens[i] - buf_offset)))
2420                         goto free_buf;
2421         }
2422
2423         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2424                 pkts[i]->pkt_len = descs[avail_idx + i].len - buf_offset;
2425                 pkts[i]->data_len = pkts[i]->pkt_len;
2426                 ids[i] = descs[avail_idx + i].id;
2427         }
2428
2429         if (virtio_net_with_host_offload(dev)) {
2430                 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2431                         hdr = (struct virtio_net_hdr *)(desc_addrs[i]);
2432                         vhost_dequeue_offload(hdr, pkts[i]);
2433                 }
2434         }
2435
2436         return 0;
2437
2438 free_buf:
2439         for (i = 0; i < PACKED_BATCH_SIZE; i++)
2440                 rte_pktmbuf_free(pkts[i]);
2441
2442         return -1;
2443 }
2444
2445 static __rte_always_inline int
2446 virtio_dev_tx_batch_packed(struct virtio_net *dev,
2447                            struct vhost_virtqueue *vq,
2448                            struct rte_mempool *mbuf_pool,
2449                            struct rte_mbuf **pkts)
2450 {
2451         uint16_t avail_idx = vq->last_avail_idx;
2452         uint32_t buf_offset = dev->vhost_hlen;
2453         uintptr_t desc_addrs[PACKED_BATCH_SIZE];
2454         uint16_t ids[PACKED_BATCH_SIZE];
2455         uint16_t i;
2456
2457         if (vhost_reserve_avail_batch_packed(dev, vq, mbuf_pool, pkts,
2458                                              avail_idx, desc_addrs, ids))
2459                 return -1;
2460
2461         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2462                 rte_prefetch0((void *)(uintptr_t)desc_addrs[i]);
2463
2464         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2465                 rte_memcpy(rte_pktmbuf_mtod_offset(pkts[i], void *, 0),
2466                            (void *)(uintptr_t)(desc_addrs[i] + buf_offset),
2467                            pkts[i]->pkt_len);
2468
2469         if (virtio_net_is_inorder(dev))
2470                 vhost_shadow_dequeue_batch_packed_inorder(vq,
2471                         ids[PACKED_BATCH_SIZE - 1]);
2472         else
2473                 vhost_shadow_dequeue_batch_packed(dev, vq, ids);
2474
2475         vq_inc_last_avail_packed(vq, PACKED_BATCH_SIZE);
2476
2477         return 0;
2478 }
2479
2480 static __rte_always_inline int
2481 vhost_dequeue_single_packed(struct virtio_net *dev,
2482                             struct vhost_virtqueue *vq,
2483                             struct rte_mempool *mbuf_pool,
2484                             struct rte_mbuf **pkts,
2485                             uint16_t *buf_id,
2486                             uint16_t *desc_count)
2487 {
2488         struct buf_vector buf_vec[BUF_VECTOR_MAX];
2489         uint32_t buf_len;
2490         uint16_t nr_vec = 0;
2491         int err;
2492         static bool allocerr_warned;
2493
2494         if (unlikely(fill_vec_buf_packed(dev, vq,
2495                                          vq->last_avail_idx, desc_count,
2496                                          buf_vec, &nr_vec,
2497                                          buf_id, &buf_len,
2498                                          VHOST_ACCESS_RO) < 0))
2499                 return -1;
2500
2501         *pkts = virtio_dev_pktmbuf_alloc(dev, mbuf_pool, buf_len);
2502         if (unlikely(*pkts == NULL)) {
2503                 if (!allocerr_warned) {
2504                         VHOST_LOG_DATA(ERR,
2505                                 "Failed mbuf alloc of size %d from %s on %s.\n",
2506                                 buf_len, mbuf_pool->name, dev->ifname);
2507                         allocerr_warned = true;
2508                 }
2509                 return -1;
2510         }
2511
2512         err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, *pkts,
2513                                 mbuf_pool);
2514         if (unlikely(err)) {
2515                 if (!allocerr_warned) {
2516                         VHOST_LOG_DATA(ERR,
2517                                 "Failed to copy desc to mbuf on %s.\n",
2518                                 dev->ifname);
2519                         allocerr_warned = true;
2520                 }
2521                 rte_pktmbuf_free(*pkts);
2522                 return -1;
2523         }
2524
2525         return 0;
2526 }
2527
2528 static __rte_always_inline int
2529 virtio_dev_tx_single_packed(struct virtio_net *dev,
2530                             struct vhost_virtqueue *vq,
2531                             struct rte_mempool *mbuf_pool,
2532                             struct rte_mbuf **pkts)
2533 {
2534
2535         uint16_t buf_id, desc_count = 0;
2536         int ret;
2537
2538         ret = vhost_dequeue_single_packed(dev, vq, mbuf_pool, pkts, &buf_id,
2539                                         &desc_count);
2540
2541         if (likely(desc_count > 0)) {
2542                 if (virtio_net_is_inorder(dev))
2543                         vhost_shadow_dequeue_single_packed_inorder(vq, buf_id,
2544                                                                    desc_count);
2545                 else
2546                         vhost_shadow_dequeue_single_packed(vq, buf_id,
2547                                         desc_count);
2548
2549                 vq_inc_last_avail_packed(vq, desc_count);
2550         }
2551
2552         return ret;
2553 }
2554
2555 static __rte_always_inline int
2556 virtio_dev_tx_batch_packed_zmbuf(struct virtio_net *dev,
2557                                  struct vhost_virtqueue *vq,
2558                                  struct rte_mempool *mbuf_pool,
2559                                  struct rte_mbuf **pkts)
2560 {
2561         struct zcopy_mbuf *zmbufs[PACKED_BATCH_SIZE];
2562         uintptr_t desc_addrs[PACKED_BATCH_SIZE];
2563         uint16_t ids[PACKED_BATCH_SIZE];
2564         uint16_t i;
2565
2566         uint16_t avail_idx = vq->last_avail_idx;
2567
2568         if (vhost_reserve_avail_batch_packed(dev, vq, mbuf_pool, pkts,
2569                                              avail_idx, desc_addrs, ids))
2570                 return -1;
2571
2572         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2573                 zmbufs[i] = get_zmbuf(vq);
2574
2575         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2576                 if (!zmbufs[i])
2577                         goto free_pkt;
2578         }
2579
2580         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2581                 zmbufs[i]->mbuf = pkts[i];
2582                 zmbufs[i]->desc_idx = ids[i];
2583                 zmbufs[i]->desc_count = 1;
2584         }
2585
2586         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2587                 rte_mbuf_refcnt_update(pkts[i], 1);
2588
2589         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2590                 TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbufs[i], next);
2591
2592         vq->nr_zmbuf += PACKED_BATCH_SIZE;
2593         vq_inc_last_avail_packed(vq, PACKED_BATCH_SIZE);
2594
2595         return 0;
2596
2597 free_pkt:
2598         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2599                 rte_pktmbuf_free(pkts[i]);
2600
2601         return -1;
2602 }
2603
2604 static __rte_always_inline int
2605 virtio_dev_tx_single_packed_zmbuf(struct virtio_net *dev,
2606                                   struct vhost_virtqueue *vq,
2607                                   struct rte_mempool *mbuf_pool,
2608                                   struct rte_mbuf **pkts)
2609 {
2610         uint16_t buf_id, desc_count;
2611         struct zcopy_mbuf *zmbuf;
2612
2613         if (vhost_dequeue_single_packed(dev, vq, mbuf_pool, pkts, &buf_id,
2614                                         &desc_count))
2615                 return -1;
2616
2617         zmbuf = get_zmbuf(vq);
2618         if (!zmbuf) {
2619                 rte_pktmbuf_free(*pkts);
2620                 return -1;
2621         }
2622         zmbuf->mbuf = *pkts;
2623         zmbuf->desc_idx = buf_id;
2624         zmbuf->desc_count = desc_count;
2625
2626         rte_mbuf_refcnt_update(*pkts, 1);
2627
2628         vq->nr_zmbuf += 1;
2629         TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
2630
2631         vq_inc_last_avail_packed(vq, desc_count);
2632         return 0;
2633 }
2634
2635 static __rte_always_inline void
2636 free_zmbuf(struct vhost_virtqueue *vq)
2637 {
2638         struct zcopy_mbuf *next = NULL;
2639         struct zcopy_mbuf *zmbuf;
2640
2641         for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
2642              zmbuf != NULL; zmbuf = next) {
2643                 next = TAILQ_NEXT(zmbuf, next);
2644
2645                 uint16_t last_used_idx = vq->last_used_idx;
2646
2647                 if (mbuf_is_consumed(zmbuf->mbuf)) {
2648                         uint16_t flags;
2649                         flags = vq->desc_packed[last_used_idx].flags;
2650                         if (vq->used_wrap_counter) {
2651                                 flags |= VRING_DESC_F_USED;
2652                                 flags |= VRING_DESC_F_AVAIL;
2653                         } else {
2654                                 flags &= ~VRING_DESC_F_USED;
2655                                 flags &= ~VRING_DESC_F_AVAIL;
2656                         }
2657
2658                         vq->desc_packed[last_used_idx].id = zmbuf->desc_idx;
2659                         vq->desc_packed[last_used_idx].len = 0;
2660
2661                         rte_smp_wmb();
2662                         vq->desc_packed[last_used_idx].flags = flags;
2663
2664                         vq_inc_last_used_packed(vq, zmbuf->desc_count);
2665
2666                         TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
2667                         restore_mbuf(zmbuf->mbuf);
2668                         rte_pktmbuf_free(zmbuf->mbuf);
2669                         put_zmbuf(zmbuf);
2670                         vq->nr_zmbuf -= 1;
2671                 }
2672         }
2673 }
2674
2675 static __rte_noinline uint16_t
2676 virtio_dev_tx_packed_zmbuf(struct virtio_net *dev,
2677                            struct vhost_virtqueue *__rte_restrict vq,
2678                            struct rte_mempool *mbuf_pool,
2679                            struct rte_mbuf **__rte_restrict pkts,
2680                            uint32_t count)
2681 {
2682         uint32_t pkt_idx = 0;
2683         uint32_t remained = count;
2684
2685         free_zmbuf(vq);
2686
2687         do {
2688                 if (remained >= PACKED_BATCH_SIZE) {
2689                         if (!virtio_dev_tx_batch_packed_zmbuf(dev, vq,
2690                                 mbuf_pool, &pkts[pkt_idx])) {
2691                                 pkt_idx += PACKED_BATCH_SIZE;
2692                                 remained -= PACKED_BATCH_SIZE;
2693                                 continue;
2694                         }
2695                 }
2696
2697                 if (virtio_dev_tx_single_packed_zmbuf(dev, vq, mbuf_pool,
2698                                                       &pkts[pkt_idx]))
2699                         break;
2700                 pkt_idx++;
2701                 remained--;
2702
2703         } while (remained);
2704
2705         if (pkt_idx)
2706                 vhost_vring_call_packed(dev, vq);
2707
2708         return pkt_idx;
2709 }
2710
2711 static __rte_noinline uint16_t
2712 virtio_dev_tx_packed(struct virtio_net *dev,
2713                      struct vhost_virtqueue *__rte_restrict vq,
2714                      struct rte_mempool *mbuf_pool,
2715                      struct rte_mbuf **__rte_restrict pkts,
2716                      uint32_t count)
2717 {
2718         uint32_t pkt_idx = 0;
2719         uint32_t remained = count;
2720
2721         do {
2722                 rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
2723
2724                 if (remained >= PACKED_BATCH_SIZE) {
2725                         if (!virtio_dev_tx_batch_packed(dev, vq, mbuf_pool,
2726                                                         &pkts[pkt_idx])) {
2727                                 pkt_idx += PACKED_BATCH_SIZE;
2728                                 remained -= PACKED_BATCH_SIZE;
2729                                 continue;
2730                         }
2731                 }
2732
2733                 if (virtio_dev_tx_single_packed(dev, vq, mbuf_pool,
2734                                                 &pkts[pkt_idx]))
2735                         break;
2736                 pkt_idx++;
2737                 remained--;
2738
2739         } while (remained);
2740
2741         if (vq->shadow_used_idx) {
2742                 do_data_copy_dequeue(vq);
2743
2744                 vhost_flush_dequeue_shadow_packed(dev, vq);
2745                 vhost_vring_call_packed(dev, vq);
2746         }
2747
2748         return pkt_idx;
2749 }
2750
2751 uint16_t
2752 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
2753         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
2754 {
2755         struct virtio_net *dev;
2756         struct rte_mbuf *rarp_mbuf = NULL;
2757         struct vhost_virtqueue *vq;
2758         int16_t success = 1;
2759
2760         dev = get_device(vid);
2761         if (!dev)
2762                 return 0;
2763
2764         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
2765                 VHOST_LOG_DATA(ERR,
2766                         "(%d) %s: built-in vhost net backend is disabled.\n",
2767                         dev->vid, __func__);
2768                 return 0;
2769         }
2770
2771         if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
2772                 VHOST_LOG_DATA(ERR,
2773                         "(%d) %s: invalid virtqueue idx %d.\n",
2774                         dev->vid, __func__, queue_id);
2775                 return 0;
2776         }
2777
2778         vq = dev->virtqueue[queue_id];
2779
2780         if (unlikely(rte_spinlock_trylock(&vq->access_lock) == 0))
2781                 return 0;
2782
2783         if (unlikely(vq->enabled == 0)) {
2784                 count = 0;
2785                 goto out_access_unlock;
2786         }
2787
2788         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
2789                 vhost_user_iotlb_rd_lock(vq);
2790
2791         if (unlikely(vq->access_ok == 0))
2792                 if (unlikely(vring_translate(dev, vq) < 0)) {
2793                         count = 0;
2794                         goto out;
2795                 }
2796
2797         /*
2798          * Construct a RARP broadcast packet, and inject it to the "pkts"
2799          * array, to looks like that guest actually send such packet.
2800          *
2801          * Check user_send_rarp() for more information.
2802          *
2803          * broadcast_rarp shares a cacheline in the virtio_net structure
2804          * with some fields that are accessed during enqueue and
2805          * __atomic_compare_exchange_n causes a write if performed compare
2806          * and exchange. This could result in false sharing between enqueue
2807          * and dequeue.
2808          *
2809          * Prevent unnecessary false sharing by reading broadcast_rarp first
2810          * and only performing compare and exchange if the read indicates it
2811          * is likely to be set.
2812          */
2813         if (unlikely(__atomic_load_n(&dev->broadcast_rarp, __ATOMIC_ACQUIRE) &&
2814                         __atomic_compare_exchange_n(&dev->broadcast_rarp,
2815                         &success, 0, 0, __ATOMIC_RELEASE, __ATOMIC_RELAXED))) {
2816
2817                 rarp_mbuf = rte_net_make_rarp_packet(mbuf_pool, &dev->mac);
2818                 if (rarp_mbuf == NULL) {
2819                         VHOST_LOG_DATA(ERR, "Failed to make RARP packet.\n");
2820                         count = 0;
2821                         goto out;
2822                 }
2823                 count -= 1;
2824         }
2825
2826         if (vq_is_packed(dev)) {
2827                 if (unlikely(dev->dequeue_zero_copy))
2828                         count = virtio_dev_tx_packed_zmbuf(dev, vq, mbuf_pool,
2829                                                            pkts, count);
2830                 else
2831                         count = virtio_dev_tx_packed(dev, vq, mbuf_pool, pkts,
2832                                                      count);
2833         } else
2834                 count = virtio_dev_tx_split(dev, vq, mbuf_pool, pkts, count);
2835
2836 out:
2837         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
2838                 vhost_user_iotlb_rd_unlock(vq);
2839
2840 out_access_unlock:
2841         rte_spinlock_unlock(&vq->access_lock);
2842
2843         if (unlikely(rarp_mbuf != NULL)) {
2844                 /*
2845                  * Inject it to the head of "pkts" array, so that switch's mac
2846                  * learning table will get updated first.
2847                  */
2848                 memmove(&pkts[1], pkts, count * sizeof(struct rte_mbuf *));
2849                 pkts[0] = rarp_mbuf;
2850                 count += 1;
2851         }
2852
2853         return count;
2854 }