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
  21 #include "iotlb.h"
  22 #include "vhost.h"
  23
  24 #define MAX_PKT_BURST 32
  25
  26 #define MAX_BATCH_LEN 256
  27
  28 static  __rte_always_inline bool
  29 rxvq_is_mergeable(struct virtio_net *dev)
  30 {
  31         return dev->features & (1ULL << VIRTIO_NET_F_MRG_RXBUF);
  32 }
  33
  34 static  __rte_always_inline bool
  35 virtio_net_is_inorder(struct virtio_net *dev)
  36 {
  37         return dev->features & (1ULL << VIRTIO_F_IN_ORDER);
  38 }
  39
  40 static bool
  41 is_valid_virt_queue_idx(uint32_t idx, int is_tx, uint32_t nr_vring)
  42 {
  43         return (is_tx ^ (idx & 1)) == 0 && idx < nr_vring;
  44 }
  45
  46 static __rte_always_inline void
  47 do_flush_shadow_used_ring_split(struct virtio_net *dev,
  48                         struct vhost_virtqueue *vq,
  49                         uint16_t to, uint16_t from, uint16_t size)
  50 {
  51         rte_memcpy(&vq->used->ring[to],
  52                         &vq->shadow_used_split[from],
  53                         size * sizeof(struct vring_used_elem));
  54         vhost_log_cache_used_vring(dev, vq,
  55                         offsetof(struct vring_used, ring[to]),
  56                         size * sizeof(struct vring_used_elem));
  57 }
  58
  59 static __rte_always_inline void
  60 flush_shadow_used_ring_split(struct virtio_net *dev, struct vhost_virtqueue *vq)
  61 {
  62         uint16_t used_idx = vq->last_used_idx & (vq->size - 1);
  63
  64         if (used_idx + vq->shadow_used_idx <= vq->size) {
  65                 do_flush_shadow_used_ring_split(dev, vq, used_idx, 0,
  66                                           vq->shadow_used_idx);
  67         } else {
  68                 uint16_t size;
  69
  70                 /* update used ring interval [used_idx, vq->size] */
  71                 size = vq->size - used_idx;
  72                 do_flush_shadow_used_ring_split(dev, vq, used_idx, 0, size);
  73
  74                 /* update the left half used ring interval [0, left_size] */
  75                 do_flush_shadow_used_ring_split(dev, vq, 0, size,
  76                                           vq->shadow_used_idx - size);
  77         }
  78         vq->last_used_idx += vq->shadow_used_idx;
  79
  80         rte_smp_wmb();
  81
  82         vhost_log_cache_sync(dev, vq);
  83
  84         *(volatile uint16_t *)&vq->used->idx += vq->shadow_used_idx;
  85         vq->shadow_used_idx = 0;
  86         vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
  87                 sizeof(vq->used->idx));
  88 }
  89
  90 static __rte_always_inline void
  91 update_shadow_used_ring_split(struct vhost_virtqueue *vq,
  92                          uint16_t desc_idx, uint32_t len)
  93 {
  94         uint16_t i = vq->shadow_used_idx++;
  95
  96         vq->shadow_used_split[i].id  = desc_idx;
  97         vq->shadow_used_split[i].len = len;
  98 }
  99
 100 static __rte_always_inline void
 101 vhost_flush_enqueue_shadow_packed(struct virtio_net *dev,
 102                                   struct vhost_virtqueue *vq)
 103 {
 104         int i;
 105         uint16_t used_idx = vq->last_used_idx;
 106         uint16_t head_idx = vq->last_used_idx;
 107         uint16_t head_flags = 0;
 108
 109         /* Split loop in two to save memory barriers */
 110         for (i = 0; i < vq->shadow_used_idx; i++) {
 111                 vq->desc_packed[used_idx].id = vq->shadow_used_packed[i].id;
 112                 vq->desc_packed[used_idx].len = vq->shadow_used_packed[i].len;
 113
 114                 used_idx += vq->shadow_used_packed[i].count;
 115                 if (used_idx >= vq->size)
 116                         used_idx -= vq->size;
 117         }
 118
 119         rte_smp_wmb();
 120
 121         for (i = 0; i < vq->shadow_used_idx; i++) {
 122                 uint16_t flags;
 123
 124                 if (vq->shadow_used_packed[i].len)
 125                         flags = VRING_DESC_F_WRITE;
 126                 else
 127                         flags = 0;
 128
 129                 if (vq->used_wrap_counter) {
 130                         flags |= VRING_DESC_F_USED;
 131                         flags |= VRING_DESC_F_AVAIL;
 132                 } else {
 133                         flags &= ~VRING_DESC_F_USED;
 134                         flags &= ~VRING_DESC_F_AVAIL;
 135                 }
 136
 137                 if (i > 0) {
 138                         vq->desc_packed[vq->last_used_idx].flags = flags;
 139
 140                         vhost_log_cache_used_vring(dev, vq,
 141                                         vq->last_used_idx *
 142                                         sizeof(struct vring_packed_desc),
 143                                         sizeof(struct vring_packed_desc));
 144                 } else {
 145                         head_idx = vq->last_used_idx;
 146                         head_flags = flags;
 147                 }
 148
 149                 vq_inc_last_used_packed(vq, vq->shadow_used_packed[i].count);
 150         }
 151
 152         vq->desc_packed[head_idx].flags = head_flags;
 153
 154         vhost_log_cache_used_vring(dev, vq,
 155                                 head_idx *
 156                                 sizeof(struct vring_packed_desc),
 157                                 sizeof(struct vring_packed_desc));
 158
 159         vq->shadow_used_idx = 0;
 160         vhost_log_cache_sync(dev, vq);
 161 }
 162
 163 static __rte_always_inline void
 164 vhost_flush_dequeue_shadow_packed(struct virtio_net *dev,
 165                                   struct vhost_virtqueue *vq)
 166 {
 167         struct vring_used_elem_packed *used_elem = &vq->shadow_used_packed[0];
 168
 169         vq->desc_packed[vq->shadow_last_used_idx].id = used_elem->id;
 170         rte_smp_wmb();
 171         vq->desc_packed[vq->shadow_last_used_idx].flags = used_elem->flags;
 172
 173         vhost_log_cache_used_vring(dev, vq, vq->shadow_last_used_idx *
 174                                    sizeof(struct vring_packed_desc),
 175                                    sizeof(struct vring_packed_desc));
 176         vq->shadow_used_idx = 0;
 177         vhost_log_cache_sync(dev, vq);
 178 }
 179
 180 static __rte_always_inline void
 181 vhost_flush_enqueue_batch_packed(struct virtio_net *dev,
 182                                  struct vhost_virtqueue *vq,
 183                                  uint64_t *lens,
 184                                  uint16_t *ids)
 185 {
 186         uint16_t i;
 187         uint16_t flags;
 188
 189         flags = PACKED_DESC_ENQUEUE_USED_FLAG(vq->used_wrap_counter);
 190
 191         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
 192                 vq->desc_packed[vq->last_used_idx + i].id = ids[i];
 193                 vq->desc_packed[vq->last_used_idx + i].len = lens[i];
 194         }
 195
 196         rte_smp_wmb();
 197
 198         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
 199                 vq->desc_packed[vq->last_used_idx + i].flags = flags;
 200
 201         vhost_log_cache_used_vring(dev, vq, vq->last_used_idx *
 202                                    sizeof(struct vring_packed_desc),
 203                                    sizeof(struct vring_packed_desc) *
 204                                    PACKED_BATCH_SIZE);
 205         vhost_log_cache_sync(dev, vq);
 206
 207         vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
 208 }
 209
 210 static __rte_always_inline void
 211 vhost_shadow_dequeue_batch_packed_inorder(struct vhost_virtqueue *vq,
 212                                           uint16_t id)
 213 {
 214         vq->shadow_used_packed[0].id = id;
 215
 216         if (!vq->shadow_used_idx) {
 217                 vq->shadow_last_used_idx = vq->last_used_idx;
 218                 vq->shadow_used_packed[0].flags =
 219                         PACKED_DESC_DEQUEUE_USED_FLAG(vq->used_wrap_counter);
 220                 vq->shadow_used_packed[0].len = 0;
 221                 vq->shadow_used_packed[0].count = 1;
 222                 vq->shadow_used_idx++;
 223         }
 224
 225         vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
 226 }
 227
 228 static __rte_always_inline void
 229 vhost_shadow_dequeue_batch_packed(struct virtio_net *dev,
 230                                   struct vhost_virtqueue *vq,
 231                                   uint16_t *ids)
 232 {
 233         uint16_t flags;
 234         uint16_t i;
 235         uint16_t begin;
 236
 237         flags = PACKED_DESC_DEQUEUE_USED_FLAG(vq->used_wrap_counter);
 238
 239         if (!vq->shadow_used_idx) {
 240                 vq->shadow_last_used_idx = vq->last_used_idx;
 241                 vq->shadow_used_packed[0].id  = ids[0];
 242                 vq->shadow_used_packed[0].len = 0;
 243                 vq->shadow_used_packed[0].count = 1;
 244                 vq->shadow_used_packed[0].flags = flags;
 245                 vq->shadow_used_idx++;
 246                 begin = 1;
 247         } else
 248                 begin = 0;
 249
 250         vhost_for_each_try_unroll(i, begin, PACKED_BATCH_SIZE) {
 251                 vq->desc_packed[vq->last_used_idx + i].id = ids[i];
 252                 vq->desc_packed[vq->last_used_idx + i].len = 0;
 253         }
 254
 255         rte_smp_wmb();
 256         vhost_for_each_try_unroll(i, begin, PACKED_BATCH_SIZE)
 257                 vq->desc_packed[vq->last_used_idx + i].flags = flags;
 258
 259         vhost_log_cache_used_vring(dev, vq, vq->last_used_idx *
 260                                    sizeof(struct vring_packed_desc),
 261                                    sizeof(struct vring_packed_desc) *
 262                                    PACKED_BATCH_SIZE);
 263         vhost_log_cache_sync(dev, vq);
 264
 265         vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
 266 }
 267
 268 static __rte_always_inline void
 269 vhost_shadow_dequeue_single_packed(struct vhost_virtqueue *vq,
 270                                    uint16_t buf_id,
 271                                    uint16_t count)
 272 {
 273         uint16_t flags;
 274
 275         flags = vq->desc_packed[vq->last_used_idx].flags;
 276         if (vq->used_wrap_counter) {
 277                 flags |= VRING_DESC_F_USED;
 278                 flags |= VRING_DESC_F_AVAIL;
 279         } else {
 280                 flags &= ~VRING_DESC_F_USED;
 281                 flags &= ~VRING_DESC_F_AVAIL;
 282         }
 283
 284         if (!vq->shadow_used_idx) {
 285                 vq->shadow_last_used_idx = vq->last_used_idx;
 286
 287                 vq->shadow_used_packed[0].id  = buf_id;
 288                 vq->shadow_used_packed[0].len = 0;
 289                 vq->shadow_used_packed[0].flags = flags;
 290                 vq->shadow_used_idx++;
 291         } else {
 292                 vq->desc_packed[vq->last_used_idx].id = buf_id;
 293                 vq->desc_packed[vq->last_used_idx].len = 0;
 294                 vq->desc_packed[vq->last_used_idx].flags = flags;
 295         }
 296
 297         vq_inc_last_used_packed(vq, count);
 298 }
 299
 300 static __rte_always_inline void
 301 vhost_shadow_dequeue_single_packed_inorder(struct vhost_virtqueue *vq,
 302                                            uint16_t buf_id,
 303                                            uint16_t count)
 304 {
 305         uint16_t flags;
 306
 307         vq->shadow_used_packed[0].id = buf_id;
 308
 309         flags = vq->desc_packed[vq->last_used_idx].flags;
 310         if (vq->used_wrap_counter) {
 311                 flags |= VRING_DESC_F_USED;
 312                 flags |= VRING_DESC_F_AVAIL;
 313         } else {
 314                 flags &= ~VRING_DESC_F_USED;
 315                 flags &= ~VRING_DESC_F_AVAIL;
 316         }
 317
 318         if (!vq->shadow_used_idx) {
 319                 vq->shadow_last_used_idx = vq->last_used_idx;
 320                 vq->shadow_used_packed[0].len = 0;
 321                 vq->shadow_used_packed[0].flags = flags;
 322                 vq->shadow_used_idx++;
 323         }
 324
 325         vq_inc_last_used_packed(vq, count);
 326 }
 327
 328 static inline void
 329 do_data_copy_enqueue(struct virtio_net *dev, struct vhost_virtqueue *vq)
 330 {
 331         struct batch_copy_elem *elem = vq->batch_copy_elems;
 332         uint16_t count = vq->batch_copy_nb_elems;
 333         int i;
 334
 335         for (i = 0; i < count; i++) {
 336                 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
 337                 vhost_log_cache_write_iova(dev, vq, elem[i].log_addr,
 338                                            elem[i].len);
 339                 PRINT_PACKET(dev, (uintptr_t)elem[i].dst, elem[i].len, 0);
 340         }
 341
 342         vq->batch_copy_nb_elems = 0;
 343 }
 344
 345 static inline void
 346 do_data_copy_dequeue(struct vhost_virtqueue *vq)
 347 {
 348         struct batch_copy_elem *elem = vq->batch_copy_elems;
 349         uint16_t count = vq->batch_copy_nb_elems;
 350         int i;
 351
 352         for (i = 0; i < count; i++)
 353                 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
 354
 355         vq->batch_copy_nb_elems = 0;
 356 }
 357
 358 static __rte_always_inline void
 359 vhost_shadow_enqueue_single_packed(struct virtio_net *dev,
 360                                    struct vhost_virtqueue *vq,
 361                                    uint32_t len[],
 362                                    uint16_t id[],
 363                                    uint16_t count[],
 364                                    uint16_t num_buffers)
 365 {
 366         uint16_t i;
 367         for (i = 0; i < num_buffers; i++) {
 368                 /* enqueue shadow flush action aligned with batch num */
 369                 if (!vq->shadow_used_idx)
 370                         vq->shadow_aligned_idx = vq->last_used_idx &
 371                                 PACKED_BATCH_MASK;
 372                 vq->shadow_used_packed[vq->shadow_used_idx].id  = id[i];
 373                 vq->shadow_used_packed[vq->shadow_used_idx].len = len[i];
 374                 vq->shadow_used_packed[vq->shadow_used_idx].count = count[i];
 375                 vq->shadow_aligned_idx += count[i];
 376                 vq->shadow_used_idx++;
 377         }
 378
 379         if (vq->shadow_aligned_idx >= PACKED_BATCH_SIZE) {
 380                 do_data_copy_enqueue(dev, vq);
 381                 vhost_flush_enqueue_shadow_packed(dev, vq);
 382         }
 383 }
 384
 385 /* avoid write operation when necessary, to lessen cache issues */
 386 #define ASSIGN_UNLESS_EQUAL(var, val) do {      \
 387         if ((var) != (val))                     \
 388                 (var) = (val);                  \
 389 } while (0)
 390
 391 static __rte_always_inline void
 392 virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
 393 {
 394         uint64_t csum_l4 = m_buf->ol_flags & PKT_TX_L4_MASK;
 395
 396         if (m_buf->ol_flags & PKT_TX_TCP_SEG)
 397                 csum_l4 |= PKT_TX_TCP_CKSUM;
 398
 399         if (csum_l4) {
 400                 net_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 401                 net_hdr->csum_start = m_buf->l2_len + m_buf->l3_len;
 402
 403                 switch (csum_l4) {
 404                 case PKT_TX_TCP_CKSUM:
 405                         net_hdr->csum_offset = (offsetof(struct rte_tcp_hdr,
 406                                                 cksum));
 407                         break;
 408                 case PKT_TX_UDP_CKSUM:
 409                         net_hdr->csum_offset = (offsetof(struct rte_udp_hdr,
 410                                                 dgram_cksum));
 411                         break;
 412                 case PKT_TX_SCTP_CKSUM:
 413                         net_hdr->csum_offset = (offsetof(struct rte_sctp_hdr,
 414                                                 cksum));
 415                         break;
 416                 }
 417         } else {
 418                 ASSIGN_UNLESS_EQUAL(net_hdr->csum_start, 0);
 419                 ASSIGN_UNLESS_EQUAL(net_hdr->csum_offset, 0);
 420                 ASSIGN_UNLESS_EQUAL(net_hdr->flags, 0);
 421         }
 422
 423         /* IP cksum verification cannot be bypassed, then calculate here */
 424         if (m_buf->ol_flags & PKT_TX_IP_CKSUM) {
 425                 struct rte_ipv4_hdr *ipv4_hdr;
 426
 427                 ipv4_hdr = rte_pktmbuf_mtod_offset(m_buf, struct rte_ipv4_hdr *,
 428                                                    m_buf->l2_len);
 429                 ipv4_hdr->hdr_checksum = 0;
 430                 ipv4_hdr->hdr_checksum = rte_ipv4_cksum(ipv4_hdr);
 431         }
 432
 433         if (m_buf->ol_flags & PKT_TX_TCP_SEG) {
 434                 if (m_buf->ol_flags & PKT_TX_IPV4)
 435                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 436                 else
 437                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 438                 net_hdr->gso_size = m_buf->tso_segsz;
 439                 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len
 440                                         + m_buf->l4_len;
 441         } else if (m_buf->ol_flags & PKT_TX_UDP_SEG) {
 442                 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
 443                 net_hdr->gso_size = m_buf->tso_segsz;
 444                 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len +
 445                         m_buf->l4_len;
 446         } else {
 447                 ASSIGN_UNLESS_EQUAL(net_hdr->gso_type, 0);
 448                 ASSIGN_UNLESS_EQUAL(net_hdr->gso_size, 0);
 449                 ASSIGN_UNLESS_EQUAL(net_hdr->hdr_len, 0);
 450         }
 451 }
 452
 453 static __rte_always_inline int
 454 map_one_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 455                 struct buf_vector *buf_vec, uint16_t *vec_idx,
 456                 uint64_t desc_iova, uint64_t desc_len, uint8_t perm)
 457 {
 458         uint16_t vec_id = *vec_idx;
 459
 460         while (desc_len) {
 461                 uint64_t desc_addr;
 462                 uint64_t desc_chunck_len = desc_len;
 463
 464                 if (unlikely(vec_id >= BUF_VECTOR_MAX))
 465                         return -1;
 466
 467                 desc_addr = vhost_iova_to_vva(dev, vq,
 468                                 desc_iova,
 469                                 &desc_chunck_len,
 470                                 perm);
 471                 if (unlikely(!desc_addr))
 472                         return -1;
 473
 474                 rte_prefetch0((void *)(uintptr_t)desc_addr);
 475
 476                 buf_vec[vec_id].buf_iova = desc_iova;
 477                 buf_vec[vec_id].buf_addr = desc_addr;
 478                 buf_vec[vec_id].buf_len  = desc_chunck_len;
 479
 480                 desc_len -= desc_chunck_len;
 481                 desc_iova += desc_chunck_len;
 482                 vec_id++;
 483         }
 484         *vec_idx = vec_id;
 485
 486         return 0;
 487 }
 488
 489 static __rte_always_inline int
 490 fill_vec_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 491                          uint32_t avail_idx, uint16_t *vec_idx,
 492                          struct buf_vector *buf_vec, uint16_t *desc_chain_head,
 493                          uint32_t *desc_chain_len, uint8_t perm)
 494 {
 495         uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
 496         uint16_t vec_id = *vec_idx;
 497         uint32_t len    = 0;
 498         uint64_t dlen;
 499         uint32_t nr_descs = vq->size;
 500         uint32_t cnt    = 0;
 501         struct vring_desc *descs = vq->desc;
 502         struct vring_desc *idesc = NULL;
 503
 504         if (unlikely(idx >= vq->size))
 505                 return -1;
 506
 507         *desc_chain_head = idx;
 508
 509         if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
 510                 dlen = vq->desc[idx].len;
 511                 nr_descs = dlen / sizeof(struct vring_desc);
 512                 if (unlikely(nr_descs > vq->size))
 513                         return -1;
 514
 515                 descs = (struct vring_desc *)(uintptr_t)
 516                         vhost_iova_to_vva(dev, vq, vq->desc[idx].addr,
 517                                                 &dlen,
 518                                                 VHOST_ACCESS_RO);
 519                 if (unlikely(!descs))
 520                         return -1;
 521
 522                 if (unlikely(dlen < vq->desc[idx].len)) {
 523                         /*
 524                          * The indirect desc table is not contiguous
 525                          * in process VA space, we have to copy it.
 526                          */
 527                         idesc = vhost_alloc_copy_ind_table(dev, vq,
 528                                         vq->desc[idx].addr, vq->desc[idx].len);
 529                         if (unlikely(!idesc))
 530                                 return -1;
 531
 532                         descs = idesc;
 533                 }
 534
 535                 idx = 0;
 536         }
 537
 538         while (1) {
 539                 if (unlikely(idx >= nr_descs || cnt++ >= nr_descs)) {
 540                         free_ind_table(idesc);
 541                         return -1;
 542                 }
 543
 544                 len += descs[idx].len;
 545
 546                 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 547                                                 descs[idx].addr, descs[idx].len,
 548                                                 perm))) {
 549                         free_ind_table(idesc);
 550                         return -1;
 551                 }
 552
 553                 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
 554                         break;
 555
 556                 idx = descs[idx].next;
 557         }
 558
 559         *desc_chain_len = len;
 560         *vec_idx = vec_id;
 561
 562         if (unlikely(!!idesc))
 563                 free_ind_table(idesc);
 564
 565         return 0;
 566 }
 567
 568 /*
 569  * Returns -1 on fail, 0 on success
 570  */
 571 static inline int
 572 reserve_avail_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 573                                 uint32_t size, struct buf_vector *buf_vec,
 574                                 uint16_t *num_buffers, uint16_t avail_head,
 575                                 uint16_t *nr_vec)
 576 {
 577         uint16_t cur_idx;
 578         uint16_t vec_idx = 0;
 579         uint16_t max_tries, tries = 0;
 580
 581         uint16_t head_idx = 0;
 582         uint32_t len = 0;
 583
 584         *num_buffers = 0;
 585         cur_idx  = vq->last_avail_idx;
 586
 587         if (rxvq_is_mergeable(dev))
 588                 max_tries = vq->size - 1;
 589         else
 590                 max_tries = 1;
 591
 592         while (size > 0) {
 593                 if (unlikely(cur_idx == avail_head))
 594                         return -1;
 595                 /*
 596                  * if we tried all available ring items, and still
 597                  * can't get enough buf, it means something abnormal
 598                  * happened.
 599                  */
 600                 if (unlikely(++tries > max_tries))
 601                         return -1;
 602
 603                 if (unlikely(fill_vec_buf_split(dev, vq, cur_idx,
 604                                                 &vec_idx, buf_vec,
 605                                                 &head_idx, &len,
 606                                                 VHOST_ACCESS_RW) < 0))
 607                         return -1;
 608                 len = RTE_MIN(len, size);
 609                 update_shadow_used_ring_split(vq, head_idx, len);
 610                 size -= len;
 611
 612                 cur_idx++;
 613                 *num_buffers += 1;
 614         }
 615
 616         *nr_vec = vec_idx;
 617
 618         return 0;
 619 }
 620
 621 static __rte_always_inline int
 622 fill_vec_buf_packed_indirect(struct virtio_net *dev,
 623                         struct vhost_virtqueue *vq,
 624                         struct vring_packed_desc *desc, uint16_t *vec_idx,
 625                         struct buf_vector *buf_vec, uint32_t *len, uint8_t perm)
 626 {
 627         uint16_t i;
 628         uint32_t nr_descs;
 629         uint16_t vec_id = *vec_idx;
 630         uint64_t dlen;
 631         struct vring_packed_desc *descs, *idescs = NULL;
 632
 633         dlen = desc->len;
 634         descs = (struct vring_packed_desc *)(uintptr_t)
 635                 vhost_iova_to_vva(dev, vq, desc->addr, &dlen, VHOST_ACCESS_RO);
 636         if (unlikely(!descs))
 637                 return -1;
 638
 639         if (unlikely(dlen < desc->len)) {
 640                 /*
 641                  * The indirect desc table is not contiguous
 642                  * in process VA space, we have to copy it.
 643                  */
 644                 idescs = vhost_alloc_copy_ind_table(dev,
 645                                 vq, desc->addr, desc->len);
 646                 if (unlikely(!idescs))
 647                         return -1;
 648
 649                 descs = idescs;
 650         }
 651
 652         nr_descs =  desc->len / sizeof(struct vring_packed_desc);
 653         if (unlikely(nr_descs >= vq->size)) {
 654                 free_ind_table(idescs);
 655                 return -1;
 656         }
 657
 658         for (i = 0; i < nr_descs; i++) {
 659                 if (unlikely(vec_id >= BUF_VECTOR_MAX)) {
 660                         free_ind_table(idescs);
 661                         return -1;
 662                 }
 663
 664                 *len += descs[i].len;
 665                 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 666                                                 descs[i].addr, descs[i].len,
 667                                                 perm)))
 668                         return -1;
 669         }
 670         *vec_idx = vec_id;
 671
 672         if (unlikely(!!idescs))
 673                 free_ind_table(idescs);
 674
 675         return 0;
 676 }
 677
 678 static __rte_always_inline int
 679 fill_vec_buf_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
 680                                 uint16_t avail_idx, uint16_t *desc_count,
 681                                 struct buf_vector *buf_vec, uint16_t *vec_idx,
 682                                 uint16_t *buf_id, uint32_t *len, uint8_t perm)
 683 {
 684         bool wrap_counter = vq->avail_wrap_counter;
 685         struct vring_packed_desc *descs = vq->desc_packed;
 686         uint16_t vec_id = *vec_idx;
 687
 688         if (avail_idx < vq->last_avail_idx)
 689                 wrap_counter ^= 1;
 690
 691         /*
 692          * Perform a load-acquire barrier in desc_is_avail to
 693          * enforce the ordering between desc flags and desc
 694          * content.
 695          */
 696         if (unlikely(!desc_is_avail(&descs[avail_idx], wrap_counter)))
 697                 return -1;
 698
 699         *desc_count = 0;
 700         *len = 0;
 701
 702         while (1) {
 703                 if (unlikely(vec_id >= BUF_VECTOR_MAX))
 704                         return -1;
 705
 706                 if (unlikely(*desc_count >= vq->size))
 707                         return -1;
 708
 709                 *desc_count += 1;
 710                 *buf_id = descs[avail_idx].id;
 711
 712                 if (descs[avail_idx].flags & VRING_DESC_F_INDIRECT) {
 713                         if (unlikely(fill_vec_buf_packed_indirect(dev, vq,
 714                                                         &descs[avail_idx],
 715                                                         &vec_id, buf_vec,
 716                                                         len, perm) < 0))
 717                                 return -1;
 718                 } else {
 719                         *len += descs[avail_idx].len;
 720
 721                         if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 722                                                         descs[avail_idx].addr,
 723                                                         descs[avail_idx].len,
 724                                                         perm)))
 725                                 return -1;
 726                 }
 727
 728                 if ((descs[avail_idx].flags & VRING_DESC_F_NEXT) == 0)
 729                         break;
 730
 731                 if (++avail_idx >= vq->size) {
 732                         avail_idx -= vq->size;
 733                         wrap_counter ^= 1;
 734                 }
 735         }
 736
 737         *vec_idx = vec_id;
 738
 739         return 0;
 740 }
 741
 742 static __rte_noinline void
 743 copy_vnet_hdr_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 744                 struct buf_vector *buf_vec,
 745                 struct virtio_net_hdr_mrg_rxbuf *hdr)
 746 {
 747         uint64_t len;
 748         uint64_t remain = dev->vhost_hlen;
 749         uint64_t src = (uint64_t)(uintptr_t)hdr, dst;
 750         uint64_t iova = buf_vec->buf_iova;
 751
 752         while (remain) {
 753                 len = RTE_MIN(remain,
 754                                 buf_vec->buf_len);
 755                 dst = buf_vec->buf_addr;
 756                 rte_memcpy((void *)(uintptr_t)dst,
 757                                 (void *)(uintptr_t)src,
 758                                 len);
 759
 760                 PRINT_PACKET(dev, (uintptr_t)dst,
 761                                 (uint32_t)len, 0);
 762                 vhost_log_cache_write_iova(dev, vq,
 763                                 iova, len);
 764
 765                 remain -= len;
 766                 iova += len;
 767                 src += len;
 768                 buf_vec++;
 769         }
 770 }
 771
 772 static __rte_always_inline int
 773 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 774                             struct rte_mbuf *m, struct buf_vector *buf_vec,
 775                             uint16_t nr_vec, uint16_t num_buffers)
 776 {
 777         uint32_t vec_idx = 0;
 778         uint32_t mbuf_offset, mbuf_avail;
 779         uint32_t buf_offset, buf_avail;
 780         uint64_t buf_addr, buf_iova, buf_len;
 781         uint32_t cpy_len;
 782         uint64_t hdr_addr;
 783         struct rte_mbuf *hdr_mbuf;
 784         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 785         struct virtio_net_hdr_mrg_rxbuf tmp_hdr, *hdr = NULL;
 786         int error = 0;
 787
 788         if (unlikely(m == NULL)) {
 789                 error = -1;
 790                 goto out;
 791         }
 792
 793         buf_addr = buf_vec[vec_idx].buf_addr;
 794         buf_iova = buf_vec[vec_idx].buf_iova;
 795         buf_len = buf_vec[vec_idx].buf_len;
 796
 797         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
 798                 error = -1;
 799                 goto out;
 800         }
 801
 802         hdr_mbuf = m;
 803         hdr_addr = buf_addr;
 804         if (unlikely(buf_len < dev->vhost_hlen))
 805                 hdr = &tmp_hdr;
 806         else
 807                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)hdr_addr;
 808
 809         VHOST_LOG_DATA(DEBUG, "(%d) RX: num merge buffers %d\n",
 810                 dev->vid, num_buffers);
 811
 812         if (unlikely(buf_len < dev->vhost_hlen)) {
 813                 buf_offset = dev->vhost_hlen - buf_len;
 814                 vec_idx++;
 815                 buf_addr = buf_vec[vec_idx].buf_addr;
 816                 buf_iova = buf_vec[vec_idx].buf_iova;
 817                 buf_len = buf_vec[vec_idx].buf_len;
 818                 buf_avail = buf_len - buf_offset;
 819         } else {
 820                 buf_offset = dev->vhost_hlen;
 821                 buf_avail = buf_len - dev->vhost_hlen;
 822         }
 823
 824         mbuf_avail  = rte_pktmbuf_data_len(m);
 825         mbuf_offset = 0;
 826         while (mbuf_avail != 0 || m->next != NULL) {
 827                 /* done with current buf, get the next one */
 828                 if (buf_avail == 0) {
 829                         vec_idx++;
 830                         if (unlikely(vec_idx >= nr_vec)) {
 831                                 error = -1;
 832                                 goto out;
 833                         }
 834
 835                         buf_addr = buf_vec[vec_idx].buf_addr;
 836                         buf_iova = buf_vec[vec_idx].buf_iova;
 837                         buf_len = buf_vec[vec_idx].buf_len;
 838
 839                         buf_offset = 0;
 840                         buf_avail  = buf_len;
 841                 }
 842
 843                 /* done with current mbuf, get the next one */
 844                 if (mbuf_avail == 0) {
 845                         m = m->next;
 846
 847                         mbuf_offset = 0;
 848                         mbuf_avail  = rte_pktmbuf_data_len(m);
 849                 }
 850
 851                 if (hdr_addr) {
 852                         virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
 853                         if (rxvq_is_mergeable(dev))
 854                                 ASSIGN_UNLESS_EQUAL(hdr->num_buffers,
 855                                                 num_buffers);
 856
 857                         if (unlikely(hdr == &tmp_hdr)) {
 858                                 copy_vnet_hdr_to_desc(dev, vq, buf_vec, hdr);
 859                         } else {
 860                                 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
 861                                                 dev->vhost_hlen, 0);
 862                                 vhost_log_cache_write_iova(dev, vq,
 863                                                 buf_vec[0].buf_iova,
 864                                                 dev->vhost_hlen);
 865                         }
 866
 867                         hdr_addr = 0;
 868                 }
 869
 870                 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
 871
 872                 if (likely(cpy_len > MAX_BATCH_LEN ||
 873                                         vq->batch_copy_nb_elems >= vq->size)) {
 874                         rte_memcpy((void *)((uintptr_t)(buf_addr + buf_offset)),
 875                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 876                                 cpy_len);
 877                         vhost_log_cache_write_iova(dev, vq,
 878                                                    buf_iova + buf_offset,
 879                                                    cpy_len);
 880                         PRINT_PACKET(dev, (uintptr_t)(buf_addr + buf_offset),
 881                                 cpy_len, 0);
 882                 } else {
 883                         batch_copy[vq->batch_copy_nb_elems].dst =
 884                                 (void *)((uintptr_t)(buf_addr + buf_offset));
 885                         batch_copy[vq->batch_copy_nb_elems].src =
 886                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
 887                         batch_copy[vq->batch_copy_nb_elems].log_addr =
 888                                 buf_iova + buf_offset;
 889                         batch_copy[vq->batch_copy_nb_elems].len = cpy_len;
 890                         vq->batch_copy_nb_elems++;
 891                 }
 892
 893                 mbuf_avail  -= cpy_len;
 894                 mbuf_offset += cpy_len;
 895                 buf_avail  -= cpy_len;
 896                 buf_offset += cpy_len;
 897         }
 898
 899 out:
 900
 901         return error;
 902 }
 903
 904 static __rte_always_inline int
 905 vhost_enqueue_single_packed(struct virtio_net *dev,
 906                             struct vhost_virtqueue *vq,
 907                             struct rte_mbuf *pkt,
 908                             struct buf_vector *buf_vec,
 909                             uint16_t *nr_descs)
 910 {
 911         uint16_t nr_vec = 0;
 912         uint16_t avail_idx = vq->last_avail_idx;
 913         uint16_t max_tries, tries = 0;
 914         uint16_t buf_id = 0;
 915         uint32_t len = 0;
 916         uint16_t desc_count;
 917         uint32_t size = pkt->pkt_len + dev->vhost_hlen;
 918         uint16_t num_buffers = 0;
 919         uint32_t buffer_len[vq->size];
 920         uint16_t buffer_buf_id[vq->size];
 921         uint16_t buffer_desc_count[vq->size];
 922
 923         if (rxvq_is_mergeable(dev))
 924                 max_tries = vq->size - 1;
 925         else
 926                 max_tries = 1;
 927
 928         while (size > 0) {
 929                 /*
 930                  * if we tried all available ring items, and still
 931                  * can't get enough buf, it means something abnormal
 932                  * happened.
 933                  */
 934                 if (unlikely(++tries > max_tries))
 935                         return -1;
 936
 937                 if (unlikely(fill_vec_buf_packed(dev, vq,
 938                                                 avail_idx, &desc_count,
 939                                                 buf_vec, &nr_vec,
 940                                                 &buf_id, &len,
 941                                                 VHOST_ACCESS_RW) < 0))
 942                         return -1;
 943
 944                 len = RTE_MIN(len, size);
 945                 size -= len;
 946
 947                 buffer_len[num_buffers] = len;
 948                 buffer_buf_id[num_buffers] = buf_id;
 949                 buffer_desc_count[num_buffers] = desc_count;
 950                 num_buffers += 1;
 951
 952                 *nr_descs += desc_count;
 953                 avail_idx += desc_count;
 954                 if (avail_idx >= vq->size)
 955                         avail_idx -= vq->size;
 956         }
 957
 958         if (copy_mbuf_to_desc(dev, vq, pkt, buf_vec, nr_vec, num_buffers) < 0)
 959                 return -1;
 960
 961         vhost_shadow_enqueue_single_packed(dev, vq, buffer_len, buffer_buf_id,
 962                                            buffer_desc_count, num_buffers);
 963
 964         return 0;
 965 }
 966
 967 static __rte_noinline uint32_t
 968 virtio_dev_rx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 969         struct rte_mbuf **pkts, uint32_t count)
 970 {
 971         uint32_t pkt_idx = 0;
 972         uint16_t num_buffers;
 973         struct buf_vector buf_vec[BUF_VECTOR_MAX];
 974         uint16_t avail_head;
 975
 976         avail_head = *((volatile uint16_t *)&vq->avail->idx);
 977
 978         /*
 979          * The ordering between avail index and
 980          * desc reads needs to be enforced.
 981          */
 982         rte_smp_rmb();
 983
 984         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
 985
 986         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
 987                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
 988                 uint16_t nr_vec = 0;
 989
 990                 if (unlikely(reserve_avail_buf_split(dev, vq,
 991                                                 pkt_len, buf_vec, &num_buffers,
 992                                                 avail_head, &nr_vec) < 0)) {
 993                         VHOST_LOG_DATA(DEBUG,
 994                                 "(%d) failed to get enough desc from vring\n",
 995                                 dev->vid);
 996                         vq->shadow_used_idx -= num_buffers;
 997                         break;
 998                 }
 999
1000                 VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
1001                         dev->vid, vq->last_avail_idx,
1002                         vq->last_avail_idx + num_buffers);
1003
1004                 if (copy_mbuf_to_desc(dev, vq, pkts[pkt_idx],
1005                                                 buf_vec, nr_vec,
1006                                                 num_buffers) < 0) {
1007                         vq->shadow_used_idx -= num_buffers;
1008                         break;
1009                 }
1010
1011                 vq->last_avail_idx += num_buffers;
1012         }
1013
1014         do_data_copy_enqueue(dev, vq);
1015
1016         if (likely(vq->shadow_used_idx)) {
1017                 flush_shadow_used_ring_split(dev, vq);
1018                 vhost_vring_call_split(dev, vq);
1019         }
1020
1021         return pkt_idx;
1022 }
1023
1024 static __rte_always_inline int
1025 virtio_dev_rx_batch_packed(struct virtio_net *dev,
1026                            struct vhost_virtqueue *vq,
1027                            struct rte_mbuf **pkts)
1028 {
1029         bool wrap_counter = vq->avail_wrap_counter;
1030         struct vring_packed_desc *descs = vq->desc_packed;
1031         uint16_t avail_idx = vq->last_avail_idx;
1032         uint64_t desc_addrs[PACKED_BATCH_SIZE];
1033         struct virtio_net_hdr_mrg_rxbuf *hdrs[PACKED_BATCH_SIZE];
1034         uint32_t buf_offset = dev->vhost_hlen;
1035         uint64_t lens[PACKED_BATCH_SIZE];
1036         uint16_t ids[PACKED_BATCH_SIZE];
1037         uint16_t i;
1038
1039         if (unlikely(avail_idx & PACKED_BATCH_MASK))
1040                 return -1;
1041
1042         if (unlikely((avail_idx + PACKED_BATCH_SIZE) > vq->size))
1043                 return -1;
1044
1045         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1046                 if (unlikely(pkts[i]->next != NULL))
1047                         return -1;
1048                 if (unlikely(!desc_is_avail(&descs[avail_idx + i],
1049                                             wrap_counter)))
1050                         return -1;
1051         }
1052
1053         rte_smp_rmb();
1054
1055         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1056                 lens[i] = descs[avail_idx + i].len;
1057
1058         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1059                 if (unlikely(pkts[i]->pkt_len > (lens[i] - buf_offset)))
1060                         return -1;
1061         }
1062
1063         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1064                 desc_addrs[i] = vhost_iova_to_vva(dev, vq,
1065                                                   descs[avail_idx + i].addr,
1066                                                   &lens[i],
1067                                                   VHOST_ACCESS_RW);
1068
1069         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1070                 if (unlikely(lens[i] != descs[avail_idx + i].len))
1071                         return -1;
1072         }
1073
1074         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1075                 rte_prefetch0((void *)(uintptr_t)desc_addrs[i]);
1076                 hdrs[i] = (struct virtio_net_hdr_mrg_rxbuf *)
1077                                         (uintptr_t)desc_addrs[i];
1078                 lens[i] = pkts[i]->pkt_len + dev->vhost_hlen;
1079         }
1080
1081         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1082                 virtio_enqueue_offload(pkts[i], &hdrs[i]->hdr);
1083
1084         vq_inc_last_avail_packed(vq, PACKED_BATCH_SIZE);
1085
1086         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1087                 rte_memcpy((void *)(uintptr_t)(desc_addrs[i] + buf_offset),
1088                            rte_pktmbuf_mtod_offset(pkts[i], void *, 0),
1089                            pkts[i]->pkt_len);
1090         }
1091
1092         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1093                 vhost_log_cache_write_iova(dev, vq, descs[avail_idx + i].addr,
1094                                            lens[i]);
1095
1096         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1097                 ids[i] = descs[avail_idx + i].id;
1098
1099         vhost_flush_enqueue_batch_packed(dev, vq, lens, ids);
1100
1101         return 0;
1102 }
1103
1104 static __rte_always_inline int16_t
1105 virtio_dev_rx_single_packed(struct virtio_net *dev,
1106                             struct vhost_virtqueue *vq,
1107                             struct rte_mbuf *pkt)
1108 {
1109         struct buf_vector buf_vec[BUF_VECTOR_MAX];
1110         uint16_t nr_descs = 0;
1111
1112         rte_smp_rmb();
1113         if (unlikely(vhost_enqueue_single_packed(dev, vq, pkt, buf_vec,
1114                                                  &nr_descs) < 0)) {
1115                 VHOST_LOG_DATA(DEBUG,
1116                                 "(%d) failed to get enough desc from vring\n",
1117                                 dev->vid);
1118                 return -1;
1119         }
1120
1121         VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
1122                         dev->vid, vq->last_avail_idx,
1123                         vq->last_avail_idx + nr_descs);
1124
1125         vq_inc_last_avail_packed(vq, nr_descs);
1126
1127         return 0;
1128 }
1129
1130 static __rte_noinline uint32_t
1131 virtio_dev_rx_packed(struct virtio_net *dev,
1132                      struct vhost_virtqueue *vq,
1133                      struct rte_mbuf **pkts,
1134                      uint32_t count)
1135 {
1136         uint32_t pkt_idx = 0;
1137         uint32_t remained = count;
1138
1139         do {
1140                 rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
1141
1142                 if (remained >= PACKED_BATCH_SIZE) {
1143                         if (!virtio_dev_rx_batch_packed(dev, vq,
1144                                                         &pkts[pkt_idx])) {
1145                                 pkt_idx += PACKED_BATCH_SIZE;
1146                                 remained -= PACKED_BATCH_SIZE;
1147                                 continue;
1148                         }
1149                 }
1150
1151                 if (virtio_dev_rx_single_packed(dev, vq, pkts[pkt_idx]))
1152                         break;
1153                 pkt_idx++;
1154                 remained--;
1155
1156         } while (pkt_idx < count);
1157
1158         if (vq->shadow_used_idx) {
1159                 do_data_copy_enqueue(dev, vq);
1160                 vhost_flush_enqueue_shadow_packed(dev, vq);
1161         }
1162
1163         if (pkt_idx)
1164                 vhost_vring_call_packed(dev, vq);
1165
1166         return pkt_idx;
1167 }
1168
1169 static __rte_always_inline uint32_t
1170 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
1171         struct rte_mbuf **pkts, uint32_t count)
1172 {
1173         struct vhost_virtqueue *vq;
1174         uint32_t nb_tx = 0;
1175
1176         VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
1177         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
1178                 VHOST_LOG_DATA(ERR, "(%d) %s: invalid virtqueue idx %d.\n",
1179                         dev->vid, __func__, queue_id);
1180                 return 0;
1181         }
1182
1183         vq = dev->virtqueue[queue_id];
1184
1185         rte_spinlock_lock(&vq->access_lock);
1186
1187         if (unlikely(vq->enabled == 0))
1188                 goto out_access_unlock;
1189
1190         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1191                 vhost_user_iotlb_rd_lock(vq);
1192
1193         if (unlikely(vq->access_ok == 0))
1194                 if (unlikely(vring_translate(dev, vq) < 0))
1195                         goto out;
1196
1197         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
1198         if (count == 0)
1199                 goto out;
1200
1201         if (vq_is_packed(dev))
1202                 nb_tx = virtio_dev_rx_packed(dev, vq, pkts, count);
1203         else
1204                 nb_tx = virtio_dev_rx_split(dev, vq, pkts, count);
1205
1206 out:
1207         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1208                 vhost_user_iotlb_rd_unlock(vq);
1209
1210 out_access_unlock:
1211         rte_spinlock_unlock(&vq->access_lock);
1212
1213         return nb_tx;
1214 }
1215
1216 uint16_t
1217 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
1218         struct rte_mbuf **pkts, uint16_t count)
1219 {
1220         struct virtio_net *dev = get_device(vid);
1221
1222         if (!dev)
1223                 return 0;
1224
1225         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
1226                 VHOST_LOG_DATA(ERR,
1227                         "(%d) %s: built-in vhost net backend is disabled.\n",
1228                         dev->vid, __func__);
1229                 return 0;
1230         }
1231
1232         return virtio_dev_rx(dev, queue_id, pkts, count);
1233 }
1234
1235 static inline bool
1236 virtio_net_with_host_offload(struct virtio_net *dev)
1237 {
1238         if (dev->features &
1239                         ((1ULL << VIRTIO_NET_F_CSUM) |
1240                          (1ULL << VIRTIO_NET_F_HOST_ECN) |
1241                          (1ULL << VIRTIO_NET_F_HOST_TSO4) |
1242                          (1ULL << VIRTIO_NET_F_HOST_TSO6) |
1243                          (1ULL << VIRTIO_NET_F_HOST_UFO)))
1244                 return true;
1245
1246         return false;
1247 }
1248
1249 static void
1250 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
1251 {
1252         struct rte_ipv4_hdr *ipv4_hdr;
1253         struct rte_ipv6_hdr *ipv6_hdr;
1254         void *l3_hdr = NULL;
1255         struct rte_ether_hdr *eth_hdr;
1256         uint16_t ethertype;
1257
1258         eth_hdr = rte_pktmbuf_mtod(m, struct rte_ether_hdr *);
1259
1260         m->l2_len = sizeof(struct rte_ether_hdr);
1261         ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
1262
1263         if (ethertype == RTE_ETHER_TYPE_VLAN) {
1264                 struct rte_vlan_hdr *vlan_hdr =
1265                         (struct rte_vlan_hdr *)(eth_hdr + 1);
1266
1267                 m->l2_len += sizeof(struct rte_vlan_hdr);
1268                 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
1269         }
1270
1271         l3_hdr = (char *)eth_hdr + m->l2_len;
1272
1273         switch (ethertype) {
1274         case RTE_ETHER_TYPE_IPV4:
1275                 ipv4_hdr = l3_hdr;
1276                 *l4_proto = ipv4_hdr->next_proto_id;
1277                 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
1278                 *l4_hdr = (char *)l3_hdr + m->l3_len;
1279                 m->ol_flags |= PKT_TX_IPV4;
1280                 break;
1281         case RTE_ETHER_TYPE_IPV6:
1282                 ipv6_hdr = l3_hdr;
1283                 *l4_proto = ipv6_hdr->proto;
1284                 m->l3_len = sizeof(struct rte_ipv6_hdr);
1285                 *l4_hdr = (char *)l3_hdr + m->l3_len;
1286                 m->ol_flags |= PKT_TX_IPV6;
1287                 break;
1288         default:
1289                 m->l3_len = 0;
1290                 *l4_proto = 0;
1291                 *l4_hdr = NULL;
1292                 break;
1293         }
1294 }
1295
1296 static __rte_always_inline void
1297 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
1298 {
1299         uint16_t l4_proto = 0;
1300         void *l4_hdr = NULL;
1301         struct rte_tcp_hdr *tcp_hdr = NULL;
1302
1303         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
1304                 return;
1305
1306         parse_ethernet(m, &l4_proto, &l4_hdr);
1307         if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
1308                 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
1309                         switch (hdr->csum_offset) {
1310                         case (offsetof(struct rte_tcp_hdr, cksum)):
1311                                 if (l4_proto == IPPROTO_TCP)
1312                                         m->ol_flags |= PKT_TX_TCP_CKSUM;
1313                                 break;
1314                         case (offsetof(struct rte_udp_hdr, dgram_cksum)):
1315                                 if (l4_proto == IPPROTO_UDP)
1316                                         m->ol_flags |= PKT_TX_UDP_CKSUM;
1317                                 break;
1318                         case (offsetof(struct rte_sctp_hdr, cksum)):
1319                                 if (l4_proto == IPPROTO_SCTP)
1320                                         m->ol_flags |= PKT_TX_SCTP_CKSUM;
1321                                 break;
1322                         default:
1323                                 break;
1324                         }
1325                 }
1326         }
1327
1328         if (l4_hdr && hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
1329                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
1330                 case VIRTIO_NET_HDR_GSO_TCPV4:
1331                 case VIRTIO_NET_HDR_GSO_TCPV6:
1332                         tcp_hdr = l4_hdr;
1333                         m->ol_flags |= PKT_TX_TCP_SEG;
1334                         m->tso_segsz = hdr->gso_size;
1335                         m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
1336                         break;
1337                 case VIRTIO_NET_HDR_GSO_UDP:
1338                         m->ol_flags |= PKT_TX_UDP_SEG;
1339                         m->tso_segsz = hdr->gso_size;
1340                         m->l4_len = sizeof(struct rte_udp_hdr);
1341                         break;
1342                 default:
1343                         VHOST_LOG_DATA(WARNING,
1344                                 "unsupported gso type %u.\n", hdr->gso_type);
1345                         break;
1346                 }
1347         }
1348 }
1349
1350 static __rte_noinline void
1351 copy_vnet_hdr_from_desc(struct virtio_net_hdr *hdr,
1352                 struct buf_vector *buf_vec)
1353 {
1354         uint64_t len;
1355         uint64_t remain = sizeof(struct virtio_net_hdr);
1356         uint64_t src;
1357         uint64_t dst = (uint64_t)(uintptr_t)hdr;
1358
1359         while (remain) {
1360                 len = RTE_MIN(remain, buf_vec->buf_len);
1361                 src = buf_vec->buf_addr;
1362                 rte_memcpy((void *)(uintptr_t)dst,
1363                                 (void *)(uintptr_t)src, len);
1364
1365                 remain -= len;
1366                 dst += len;
1367                 buf_vec++;
1368         }
1369 }
1370
1371 static __rte_always_inline int
1372 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
1373                   struct buf_vector *buf_vec, uint16_t nr_vec,
1374                   struct rte_mbuf *m, struct rte_mempool *mbuf_pool)
1375 {
1376         uint32_t buf_avail, buf_offset;
1377         uint64_t buf_addr, buf_iova, buf_len;
1378         uint32_t mbuf_avail, mbuf_offset;
1379         uint32_t cpy_len;
1380         struct rte_mbuf *cur = m, *prev = m;
1381         struct virtio_net_hdr tmp_hdr;
1382         struct virtio_net_hdr *hdr = NULL;
1383         /* A counter to avoid desc dead loop chain */
1384         uint16_t vec_idx = 0;
1385         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
1386         int error = 0;
1387
1388         buf_addr = buf_vec[vec_idx].buf_addr;
1389         buf_iova = buf_vec[vec_idx].buf_iova;
1390         buf_len = buf_vec[vec_idx].buf_len;
1391
1392         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
1393                 error = -1;
1394                 goto out;
1395         }
1396
1397         if (virtio_net_with_host_offload(dev)) {
1398                 if (unlikely(buf_len < sizeof(struct virtio_net_hdr))) {
1399                         /*
1400                          * No luck, the virtio-net header doesn't fit
1401                          * in a contiguous virtual area.
1402                          */
1403                         copy_vnet_hdr_from_desc(&tmp_hdr, buf_vec);
1404                         hdr = &tmp_hdr;
1405                 } else {
1406                         hdr = (struct virtio_net_hdr *)((uintptr_t)buf_addr);
1407                 }
1408         }
1409
1410         /*
1411          * A virtio driver normally uses at least 2 desc buffers
1412          * for Tx: the first for storing the header, and others
1413          * for storing the data.
1414          */
1415         if (unlikely(buf_len < dev->vhost_hlen)) {
1416                 buf_offset = dev->vhost_hlen - buf_len;
1417                 vec_idx++;
1418                 buf_addr = buf_vec[vec_idx].buf_addr;
1419                 buf_iova = buf_vec[vec_idx].buf_iova;
1420                 buf_len = buf_vec[vec_idx].buf_len;
1421                 buf_avail  = buf_len - buf_offset;
1422         } else if (buf_len == dev->vhost_hlen) {
1423                 if (unlikely(++vec_idx >= nr_vec))
1424                         goto out;
1425                 buf_addr = buf_vec[vec_idx].buf_addr;
1426                 buf_iova = buf_vec[vec_idx].buf_iova;
1427                 buf_len = buf_vec[vec_idx].buf_len;
1428
1429                 buf_offset = 0;
1430                 buf_avail = buf_len;
1431         } else {
1432                 buf_offset = dev->vhost_hlen;
1433                 buf_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
1434         }
1435
1436         PRINT_PACKET(dev,
1437                         (uintptr_t)(buf_addr + buf_offset),
1438                         (uint32_t)buf_avail, 0);
1439
1440         mbuf_offset = 0;
1441         mbuf_avail  = m->buf_len - RTE_PKTMBUF_HEADROOM;
1442         while (1) {
1443                 uint64_t hpa;
1444
1445                 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
1446
1447                 /*
1448                  * A desc buf might across two host physical pages that are
1449                  * not continuous. In such case (gpa_to_hpa returns 0), data
1450                  * will be copied even though zero copy is enabled.
1451                  */
1452                 if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
1453                                         buf_iova + buf_offset, cpy_len)))) {
1454                         cur->data_len = cpy_len;
1455                         cur->data_off = 0;
1456                         cur->buf_addr =
1457                                 (void *)(uintptr_t)(buf_addr + buf_offset);
1458                         cur->buf_iova = hpa;
1459
1460                         /*
1461                          * In zero copy mode, one mbuf can only reference data
1462                          * for one or partial of one desc buff.
1463                          */
1464                         mbuf_avail = cpy_len;
1465                 } else {
1466                         if (likely(cpy_len > MAX_BATCH_LEN ||
1467                                    vq->batch_copy_nb_elems >= vq->size ||
1468                                    (hdr && cur == m))) {
1469                                 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
1470                                                                    mbuf_offset),
1471                                            (void *)((uintptr_t)(buf_addr +
1472                                                            buf_offset)),
1473                                            cpy_len);
1474                         } else {
1475                                 batch_copy[vq->batch_copy_nb_elems].dst =
1476                                         rte_pktmbuf_mtod_offset(cur, void *,
1477                                                                 mbuf_offset);
1478                                 batch_copy[vq->batch_copy_nb_elems].src =
1479                                         (void *)((uintptr_t)(buf_addr +
1480                                                                 buf_offset));
1481                                 batch_copy[vq->batch_copy_nb_elems].len =
1482                                         cpy_len;
1483                                 vq->batch_copy_nb_elems++;
1484                         }
1485                 }
1486
1487                 mbuf_avail  -= cpy_len;
1488                 mbuf_offset += cpy_len;
1489                 buf_avail -= cpy_len;
1490                 buf_offset += cpy_len;
1491
1492                 /* This buf reaches to its end, get the next one */
1493                 if (buf_avail == 0) {
1494                         if (++vec_idx >= nr_vec)
1495                                 break;
1496
1497                         buf_addr = buf_vec[vec_idx].buf_addr;
1498                         buf_iova = buf_vec[vec_idx].buf_iova;
1499                         buf_len = buf_vec[vec_idx].buf_len;
1500
1501                         buf_offset = 0;
1502                         buf_avail  = buf_len;
1503
1504                         PRINT_PACKET(dev, (uintptr_t)buf_addr,
1505                                         (uint32_t)buf_avail, 0);
1506                 }
1507
1508                 /*
1509                  * This mbuf reaches to its end, get a new one
1510                  * to hold more data.
1511                  */
1512                 if (mbuf_avail == 0) {
1513                         cur = rte_pktmbuf_alloc(mbuf_pool);
1514                         if (unlikely(cur == NULL)) {
1515                                 VHOST_LOG_DATA(ERR, "Failed to "
1516                                         "allocate memory for mbuf.\n");
1517                                 error = -1;
1518                                 goto out;
1519                         }
1520                         if (unlikely(dev->dequeue_zero_copy))
1521                                 rte_mbuf_refcnt_update(cur, 1);
1522
1523                         prev->next = cur;
1524                         prev->data_len = mbuf_offset;
1525                         m->nb_segs += 1;
1526                         m->pkt_len += mbuf_offset;
1527                         prev = cur;
1528
1529                         mbuf_offset = 0;
1530                         mbuf_avail  = cur->buf_len - RTE_PKTMBUF_HEADROOM;
1531                 }
1532         }
1533
1534         prev->data_len = mbuf_offset;
1535         m->pkt_len    += mbuf_offset;
1536
1537         if (hdr)
1538                 vhost_dequeue_offload(hdr, m);
1539
1540 out:
1541
1542         return error;
1543 }
1544
1545 static __rte_always_inline struct zcopy_mbuf *
1546 get_zmbuf(struct vhost_virtqueue *vq)
1547 {
1548         uint16_t i;
1549         uint16_t last;
1550         int tries = 0;
1551
1552         /* search [last_zmbuf_idx, zmbuf_size) */
1553         i = vq->last_zmbuf_idx;
1554         last = vq->zmbuf_size;
1555
1556 again:
1557         for (; i < last; i++) {
1558                 if (vq->zmbufs[i].in_use == 0) {
1559                         vq->last_zmbuf_idx = i + 1;
1560                         vq->zmbufs[i].in_use = 1;
1561                         return &vq->zmbufs[i];
1562                 }
1563         }
1564
1565         tries++;
1566         if (tries == 1) {
1567                 /* search [0, last_zmbuf_idx) */
1568                 i = 0;
1569                 last = vq->last_zmbuf_idx;
1570                 goto again;
1571         }
1572
1573         return NULL;
1574 }
1575
1576 static void
1577 virtio_dev_extbuf_free(void *addr __rte_unused, void *opaque)
1578 {
1579         rte_free(opaque);
1580 }
1581
1582 static int
1583 virtio_dev_extbuf_alloc(struct rte_mbuf *pkt, uint32_t size)
1584 {
1585         struct rte_mbuf_ext_shared_info *shinfo = NULL;
1586         uint32_t total_len = RTE_PKTMBUF_HEADROOM + size;
1587         uint16_t buf_len;
1588         rte_iova_t iova;
1589         void *buf;
1590
1591         /* Try to use pkt buffer to store shinfo to reduce the amount of memory
1592          * required, otherwise store shinfo in the new buffer.
1593          */
1594         if (rte_pktmbuf_tailroom(pkt) >= sizeof(*shinfo))
1595                 shinfo = rte_pktmbuf_mtod(pkt,
1596                                           struct rte_mbuf_ext_shared_info *);
1597         else {
1598                 total_len += sizeof(*shinfo) + sizeof(uintptr_t);
1599                 total_len = RTE_ALIGN_CEIL(total_len, sizeof(uintptr_t));
1600         }
1601
1602         if (unlikely(total_len > UINT16_MAX))
1603                 return -ENOSPC;
1604
1605         buf_len = total_len;
1606         buf = rte_malloc(NULL, buf_len, RTE_CACHE_LINE_SIZE);
1607         if (unlikely(buf == NULL))
1608                 return -ENOMEM;
1609
1610         /* Initialize shinfo */
1611         if (shinfo) {
1612                 shinfo->free_cb = virtio_dev_extbuf_free;
1613                 shinfo->fcb_opaque = buf;
1614                 rte_mbuf_ext_refcnt_set(shinfo, 1);
1615         } else {
1616                 shinfo = rte_pktmbuf_ext_shinfo_init_helper(buf, &buf_len,
1617                                               virtio_dev_extbuf_free, buf);
1618                 if (unlikely(shinfo == NULL)) {
1619                         rte_free(buf);
1620                         VHOST_LOG_DATA(ERR, "Failed to init shinfo\n");
1621                         return -1;
1622                 }
1623         }
1624
1625         iova = rte_malloc_virt2iova(buf);
1626         rte_pktmbuf_attach_extbuf(pkt, buf, iova, buf_len, shinfo);
1627         rte_pktmbuf_reset_headroom(pkt);
1628
1629         return 0;
1630 }
1631
1632 /*
1633  * Allocate a host supported pktmbuf.
1634  */
1635 static __rte_always_inline struct rte_mbuf *
1636 virtio_dev_pktmbuf_alloc(struct virtio_net *dev, struct rte_mempool *mp,
1637                          uint32_t data_len)
1638 {
1639         struct rte_mbuf *pkt = rte_pktmbuf_alloc(mp);
1640
1641         if (unlikely(pkt == NULL)) {
1642                 VHOST_LOG_DATA(ERR,
1643                         "Failed to allocate memory for mbuf.\n");
1644                 return NULL;
1645         }
1646
1647         if (rte_pktmbuf_tailroom(pkt) >= data_len)
1648                 return pkt;
1649
1650         /* attach an external buffer if supported */
1651         if (dev->extbuf && !virtio_dev_extbuf_alloc(pkt, data_len))
1652                 return pkt;
1653
1654         /* check if chained buffers are allowed */
1655         if (!dev->linearbuf)
1656                 return pkt;
1657
1658         /* Data doesn't fit into the buffer and the host supports
1659          * only linear buffers
1660          */
1661         rte_pktmbuf_free(pkt);
1662
1663         return NULL;
1664 }
1665
1666 static __rte_noinline uint16_t
1667 virtio_dev_tx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
1668         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1669 {
1670         uint16_t i;
1671         uint16_t free_entries;
1672
1673         if (unlikely(dev->dequeue_zero_copy)) {
1674                 struct zcopy_mbuf *zmbuf, *next;
1675
1676                 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1677                      zmbuf != NULL; zmbuf = next) {
1678                         next = TAILQ_NEXT(zmbuf, next);
1679
1680                         if (mbuf_is_consumed(zmbuf->mbuf)) {
1681                                 update_shadow_used_ring_split(vq,
1682                                                 zmbuf->desc_idx, 0);
1683                                 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1684                                 restore_mbuf(zmbuf->mbuf);
1685                                 rte_pktmbuf_free(zmbuf->mbuf);
1686                                 put_zmbuf(zmbuf);
1687                                 vq->nr_zmbuf -= 1;
1688                         }
1689                 }
1690
1691                 if (likely(vq->shadow_used_idx)) {
1692                         flush_shadow_used_ring_split(dev, vq);
1693                         vhost_vring_call_split(dev, vq);
1694                 }
1695         }
1696
1697         free_entries = *((volatile uint16_t *)&vq->avail->idx) -
1698                         vq->last_avail_idx;
1699         if (free_entries == 0)
1700                 return 0;
1701
1702         /*
1703          * The ordering between avail index and
1704          * desc reads needs to be enforced.
1705          */
1706         rte_smp_rmb();
1707
1708         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
1709
1710         VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
1711
1712         count = RTE_MIN(count, MAX_PKT_BURST);
1713         count = RTE_MIN(count, free_entries);
1714         VHOST_LOG_DATA(DEBUG, "(%d) about to dequeue %u buffers\n",
1715                         dev->vid, count);
1716
1717         for (i = 0; i < count; i++) {
1718                 struct buf_vector buf_vec[BUF_VECTOR_MAX];
1719                 uint16_t head_idx;
1720                 uint32_t buf_len;
1721                 uint16_t nr_vec = 0;
1722                 int err;
1723
1724                 if (unlikely(fill_vec_buf_split(dev, vq,
1725                                                 vq->last_avail_idx + i,
1726                                                 &nr_vec, buf_vec,
1727                                                 &head_idx, &buf_len,
1728                                                 VHOST_ACCESS_RO) < 0))
1729                         break;
1730
1731                 if (likely(dev->dequeue_zero_copy == 0))
1732                         update_shadow_used_ring_split(vq, head_idx, 0);
1733
1734                 pkts[i] = virtio_dev_pktmbuf_alloc(dev, mbuf_pool, buf_len);
1735                 if (unlikely(pkts[i] == NULL))
1736                         break;
1737
1738                 err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts[i],
1739                                 mbuf_pool);
1740                 if (unlikely(err)) {
1741                         rte_pktmbuf_free(pkts[i]);
1742                         break;
1743                 }
1744
1745                 if (unlikely(dev->dequeue_zero_copy)) {
1746                         struct zcopy_mbuf *zmbuf;
1747
1748                         zmbuf = get_zmbuf(vq);
1749                         if (!zmbuf) {
1750                                 rte_pktmbuf_free(pkts[i]);
1751                                 break;
1752                         }
1753                         zmbuf->mbuf = pkts[i];
1754                         zmbuf->desc_idx = head_idx;
1755
1756                         /*
1757                          * Pin lock the mbuf; we will check later to see
1758                          * whether the mbuf is freed (when we are the last
1759                          * user) or not. If that's the case, we then could
1760                          * update the used ring safely.
1761                          */
1762                         rte_mbuf_refcnt_update(pkts[i], 1);
1763
1764                         vq->nr_zmbuf += 1;
1765                         TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1766                 }
1767         }
1768         vq->last_avail_idx += i;
1769
1770         if (likely(dev->dequeue_zero_copy == 0)) {
1771                 do_data_copy_dequeue(vq);
1772                 if (unlikely(i < count))
1773                         vq->shadow_used_idx = i;
1774                 if (likely(vq->shadow_used_idx)) {
1775                         flush_shadow_used_ring_split(dev, vq);
1776                         vhost_vring_call_split(dev, vq);
1777                 }
1778         }
1779
1780         return i;
1781 }
1782
1783 static __rte_always_inline int
1784 vhost_reserve_avail_batch_packed(struct virtio_net *dev,
1785                                  struct vhost_virtqueue *vq,
1786                                  struct rte_mempool *mbuf_pool,
1787                                  struct rte_mbuf **pkts,
1788                                  uint16_t avail_idx,
1789                                  uintptr_t *desc_addrs,
1790                                  uint16_t *ids)
1791 {
1792         bool wrap = vq->avail_wrap_counter;
1793         struct vring_packed_desc *descs = vq->desc_packed;
1794         struct virtio_net_hdr *hdr;
1795         uint64_t lens[PACKED_BATCH_SIZE];
1796         uint64_t buf_lens[PACKED_BATCH_SIZE];
1797         uint32_t buf_offset = dev->vhost_hlen;
1798         uint16_t flags, i;
1799
1800         if (unlikely(avail_idx & PACKED_BATCH_MASK))
1801                 return -1;
1802         if (unlikely((avail_idx + PACKED_BATCH_SIZE) > vq->size))
1803                 return -1;
1804
1805         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1806                 flags = descs[avail_idx + i].flags;
1807                 if (unlikely((wrap != !!(flags & VRING_DESC_F_AVAIL)) ||
1808                              (wrap == !!(flags & VRING_DESC_F_USED))  ||
1809                              (flags & PACKED_DESC_SINGLE_DEQUEUE_FLAG)))
1810                         return -1;
1811         }
1812
1813         rte_smp_rmb();
1814
1815         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1816                 lens[i] = descs[avail_idx + i].len;
1817
1818         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1819                 desc_addrs[i] = vhost_iova_to_vva(dev, vq,
1820                                                   descs[avail_idx + i].addr,
1821                                                   &lens[i], VHOST_ACCESS_RW);
1822         }
1823
1824         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1825                 if (unlikely((lens[i] != descs[avail_idx + i].len)))
1826                         return -1;
1827         }
1828
1829         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1830                 pkts[i] = virtio_dev_pktmbuf_alloc(dev, mbuf_pool, lens[i]);
1831                 if (!pkts[i])
1832                         goto free_buf;
1833         }
1834
1835         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1836                 buf_lens[i] = pkts[i]->buf_len - pkts[i]->data_off;
1837
1838         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1839                 if (unlikely(buf_lens[i] < (lens[i] - buf_offset)))
1840                         goto free_buf;
1841         }
1842
1843         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1844                 pkts[i]->pkt_len = descs[avail_idx + i].len - buf_offset;
1845                 pkts[i]->data_len = pkts[i]->pkt_len;
1846                 ids[i] = descs[avail_idx + i].id;
1847         }
1848
1849         if (virtio_net_with_host_offload(dev)) {
1850                 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1851                         hdr = (struct virtio_net_hdr *)(desc_addrs[i]);
1852                         vhost_dequeue_offload(hdr, pkts[i]);
1853                 }
1854         }
1855
1856         return 0;
1857
1858 free_buf:
1859         for (i = 0; i < PACKED_BATCH_SIZE; i++)
1860                 rte_pktmbuf_free(pkts[i]);
1861
1862         return -1;
1863 }
1864
1865 static __rte_always_inline int
1866 virtio_dev_tx_batch_packed(struct virtio_net *dev,
1867                            struct vhost_virtqueue *vq,
1868                            struct rte_mempool *mbuf_pool,
1869                            struct rte_mbuf **pkts)
1870 {
1871         uint16_t avail_idx = vq->last_avail_idx;
1872         uint32_t buf_offset = dev->vhost_hlen;
1873         uintptr_t desc_addrs[PACKED_BATCH_SIZE];
1874         uint16_t ids[PACKED_BATCH_SIZE];
1875         uint16_t i;
1876
1877         if (vhost_reserve_avail_batch_packed(dev, vq, mbuf_pool, pkts,
1878                                              avail_idx, desc_addrs, ids))
1879                 return -1;
1880
1881         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1882                 rte_prefetch0((void *)(uintptr_t)desc_addrs[i]);
1883
1884         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1885                 rte_memcpy(rte_pktmbuf_mtod_offset(pkts[i], void *, 0),
1886                            (void *)(uintptr_t)(desc_addrs[i] + buf_offset),
1887                            pkts[i]->pkt_len);
1888
1889         if (virtio_net_is_inorder(dev))
1890                 vhost_shadow_dequeue_batch_packed_inorder(vq,
1891                         ids[PACKED_BATCH_SIZE - 1]);
1892         else
1893                 vhost_shadow_dequeue_batch_packed(dev, vq, ids);
1894
1895         vq_inc_last_avail_packed(vq, PACKED_BATCH_SIZE);
1896
1897         return 0;
1898 }
1899
1900 static __rte_always_inline int
1901 vhost_dequeue_single_packed(struct virtio_net *dev,
1902                             struct vhost_virtqueue *vq,
1903                             struct rte_mempool *mbuf_pool,
1904                             struct rte_mbuf **pkts,
1905                             uint16_t *buf_id,
1906                             uint16_t *desc_count)
1907 {
1908         struct buf_vector buf_vec[BUF_VECTOR_MAX];
1909         uint32_t buf_len;
1910         uint16_t nr_vec = 0;
1911         int err;
1912
1913         if (unlikely(fill_vec_buf_packed(dev, vq,
1914                                          vq->last_avail_idx, desc_count,
1915                                          buf_vec, &nr_vec,
1916                                          buf_id, &buf_len,
1917                                          VHOST_ACCESS_RO) < 0))
1918                 return -1;
1919
1920         *pkts = virtio_dev_pktmbuf_alloc(dev, mbuf_pool, buf_len);
1921         if (unlikely(*pkts == NULL)) {
1922                 VHOST_LOG_DATA(ERR,
1923                         "Failed to allocate memory for mbuf.\n");
1924                 return -1;
1925         }
1926
1927         err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, *pkts,
1928                                 mbuf_pool);
1929         if (unlikely(err)) {
1930                 rte_pktmbuf_free(*pkts);
1931                 return -1;
1932         }
1933
1934         return 0;
1935 }
1936
1937 static __rte_always_inline int
1938 virtio_dev_tx_single_packed(struct virtio_net *dev,
1939                             struct vhost_virtqueue *vq,
1940                             struct rte_mempool *mbuf_pool,
1941                             struct rte_mbuf **pkts)
1942 {
1943
1944         uint16_t buf_id, desc_count;
1945
1946         if (vhost_dequeue_single_packed(dev, vq, mbuf_pool, pkts, &buf_id,
1947                                         &desc_count))
1948                 return -1;
1949
1950         if (virtio_net_is_inorder(dev))
1951                 vhost_shadow_dequeue_single_packed_inorder(vq, buf_id,
1952                                                            desc_count);
1953         else
1954                 vhost_shadow_dequeue_single_packed(vq, buf_id, desc_count);
1955
1956         vq_inc_last_avail_packed(vq, desc_count);
1957
1958         return 0;
1959 }
1960
1961 static __rte_always_inline int
1962 virtio_dev_tx_batch_packed_zmbuf(struct virtio_net *dev,
1963                                  struct vhost_virtqueue *vq,
1964                                  struct rte_mempool *mbuf_pool,
1965                                  struct rte_mbuf **pkts)
1966 {
1967         struct zcopy_mbuf *zmbufs[PACKED_BATCH_SIZE];
1968         uintptr_t desc_addrs[PACKED_BATCH_SIZE];
1969         uint16_t ids[PACKED_BATCH_SIZE];
1970         uint16_t i;
1971
1972         uint16_t avail_idx = vq->last_avail_idx;
1973
1974         if (vhost_reserve_avail_batch_packed(dev, vq, mbuf_pool, pkts,
1975                                              avail_idx, desc_addrs, ids))
1976                 return -1;
1977
1978         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1979                 zmbufs[i] = get_zmbuf(vq);
1980
1981         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1982                 if (!zmbufs[i])
1983                         goto free_pkt;
1984         }
1985
1986         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1987                 zmbufs[i]->mbuf = pkts[i];
1988                 zmbufs[i]->desc_idx = ids[i];
1989                 zmbufs[i]->desc_count = 1;
1990         }
1991
1992         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1993                 rte_mbuf_refcnt_update(pkts[i], 1);
1994
1995         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1996                 TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbufs[i], next);
1997
1998         vq->nr_zmbuf += PACKED_BATCH_SIZE;
1999         vq_inc_last_avail_packed(vq, PACKED_BATCH_SIZE);
2000
2001         return 0;
2002
2003 free_pkt:
2004         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2005                 rte_pktmbuf_free(pkts[i]);
2006
2007         return -1;
2008 }
2009
2010 static __rte_always_inline int
2011 virtio_dev_tx_single_packed_zmbuf(struct virtio_net *dev,
2012                                   struct vhost_virtqueue *vq,
2013                                   struct rte_mempool *mbuf_pool,
2014                                   struct rte_mbuf **pkts)
2015 {
2016         uint16_t buf_id, desc_count;
2017         struct zcopy_mbuf *zmbuf;
2018
2019         if (vhost_dequeue_single_packed(dev, vq, mbuf_pool, pkts, &buf_id,
2020                                         &desc_count))
2021                 return -1;
2022
2023         zmbuf = get_zmbuf(vq);
2024         if (!zmbuf) {
2025                 rte_pktmbuf_free(*pkts);
2026                 return -1;
2027         }
2028         zmbuf->mbuf = *pkts;
2029         zmbuf->desc_idx = buf_id;
2030         zmbuf->desc_count = desc_count;
2031
2032         rte_mbuf_refcnt_update(*pkts, 1);
2033
2034         vq->nr_zmbuf += 1;
2035         TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
2036
2037         vq_inc_last_avail_packed(vq, desc_count);
2038         return 0;
2039 }
2040
2041 static __rte_always_inline void
2042 free_zmbuf(struct vhost_virtqueue *vq)
2043 {
2044         struct zcopy_mbuf *next = NULL;
2045         struct zcopy_mbuf *zmbuf;
2046
2047         for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
2048              zmbuf != NULL; zmbuf = next) {
2049                 next = TAILQ_NEXT(zmbuf, next);
2050
2051                 uint16_t last_used_idx = vq->last_used_idx;
2052
2053                 if (mbuf_is_consumed(zmbuf->mbuf)) {
2054                         uint16_t flags;
2055                         flags = vq->desc_packed[last_used_idx].flags;
2056                         if (vq->used_wrap_counter) {
2057                                 flags |= VRING_DESC_F_USED;
2058                                 flags |= VRING_DESC_F_AVAIL;
2059                         } else {
2060                                 flags &= ~VRING_DESC_F_USED;
2061                                 flags &= ~VRING_DESC_F_AVAIL;
2062                         }
2063
2064                         vq->desc_packed[last_used_idx].id = zmbuf->desc_idx;
2065                         vq->desc_packed[last_used_idx].len = 0;
2066
2067                         rte_smp_wmb();
2068                         vq->desc_packed[last_used_idx].flags = flags;
2069
2070                         vq_inc_last_used_packed(vq, zmbuf->desc_count);
2071
2072                         TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
2073                         restore_mbuf(zmbuf->mbuf);
2074                         rte_pktmbuf_free(zmbuf->mbuf);
2075                         put_zmbuf(zmbuf);
2076                         vq->nr_zmbuf -= 1;
2077                 }
2078         }
2079 }
2080
2081 static __rte_noinline uint16_t
2082 virtio_dev_tx_packed_zmbuf(struct virtio_net *dev,
2083                            struct vhost_virtqueue *vq,
2084                            struct rte_mempool *mbuf_pool,
2085                            struct rte_mbuf **pkts,
2086                            uint32_t count)
2087 {
2088         uint32_t pkt_idx = 0;
2089         uint32_t remained = count;
2090
2091         free_zmbuf(vq);
2092
2093         do {
2094                 if (remained >= PACKED_BATCH_SIZE) {
2095                         if (!virtio_dev_tx_batch_packed_zmbuf(dev, vq,
2096                                 mbuf_pool, &pkts[pkt_idx])) {
2097                                 pkt_idx += PACKED_BATCH_SIZE;
2098                                 remained -= PACKED_BATCH_SIZE;
2099                                 continue;
2100                         }
2101                 }
2102
2103                 if (virtio_dev_tx_single_packed_zmbuf(dev, vq, mbuf_pool,
2104                                                       &pkts[pkt_idx]))
2105                         break;
2106                 pkt_idx++;
2107                 remained--;
2108
2109         } while (remained);
2110
2111         if (pkt_idx)
2112                 vhost_vring_call_packed(dev, vq);
2113
2114         return pkt_idx;
2115 }
2116
2117 static __rte_noinline uint16_t
2118 virtio_dev_tx_packed(struct virtio_net *dev,
2119                      struct vhost_virtqueue *vq,
2120                      struct rte_mempool *mbuf_pool,
2121                      struct rte_mbuf **pkts,
2122                      uint32_t count)
2123 {
2124         uint32_t pkt_idx = 0;
2125         uint32_t remained = count;
2126
2127         do {
2128                 rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
2129
2130                 if (remained >= PACKED_BATCH_SIZE) {
2131                         if (!virtio_dev_tx_batch_packed(dev, vq, mbuf_pool,
2132                                                         &pkts[pkt_idx])) {
2133                                 pkt_idx += PACKED_BATCH_SIZE;
2134                                 remained -= PACKED_BATCH_SIZE;
2135                                 continue;
2136                         }
2137                 }
2138
2139                 if (virtio_dev_tx_single_packed(dev, vq, mbuf_pool,
2140                                                 &pkts[pkt_idx]))
2141                         break;
2142                 pkt_idx++;
2143                 remained--;
2144
2145         } while (remained);
2146
2147         if (vq->shadow_used_idx) {
2148                 do_data_copy_dequeue(vq);
2149
2150                 vhost_flush_dequeue_shadow_packed(dev, vq);
2151                 vhost_vring_call_packed(dev, vq);
2152         }
2153
2154         return pkt_idx;
2155 }
2156
2157 uint16_t
2158 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
2159         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
2160 {
2161         struct virtio_net *dev;
2162         struct rte_mbuf *rarp_mbuf = NULL;
2163         struct vhost_virtqueue *vq;
2164
2165         dev = get_device(vid);
2166         if (!dev)
2167                 return 0;
2168
2169         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
2170                 VHOST_LOG_DATA(ERR,
2171                         "(%d) %s: built-in vhost net backend is disabled.\n",
2172                         dev->vid, __func__);
2173                 return 0;
2174         }
2175
2176         if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
2177                 VHOST_LOG_DATA(ERR,
2178                         "(%d) %s: invalid virtqueue idx %d.\n",
2179                         dev->vid, __func__, queue_id);
2180                 return 0;
2181         }
2182
2183         vq = dev->virtqueue[queue_id];
2184
2185         if (unlikely(rte_spinlock_trylock(&vq->access_lock) == 0))
2186                 return 0;
2187
2188         if (unlikely(vq->enabled == 0)) {
2189                 count = 0;
2190                 goto out_access_unlock;
2191         }
2192
2193         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
2194                 vhost_user_iotlb_rd_lock(vq);
2195
2196         if (unlikely(vq->access_ok == 0))
2197                 if (unlikely(vring_translate(dev, vq) < 0)) {
2198                         count = 0;
2199                         goto out;
2200                 }
2201
2202         /*
2203          * Construct a RARP broadcast packet, and inject it to the "pkts"
2204          * array, to looks like that guest actually send such packet.
2205          *
2206          * Check user_send_rarp() for more information.
2207          *
2208          * broadcast_rarp shares a cacheline in the virtio_net structure
2209          * with some fields that are accessed during enqueue and
2210          * rte_atomic16_cmpset() causes a write if using cmpxchg. This could
2211          * result in false sharing between enqueue and dequeue.
2212          *
2213          * Prevent unnecessary false sharing by reading broadcast_rarp first
2214          * and only performing cmpset if the read indicates it is likely to
2215          * be set.
2216          */
2217         if (unlikely(rte_atomic16_read(&dev->broadcast_rarp) &&
2218                         rte_atomic16_cmpset((volatile uint16_t *)
2219                                 &dev->broadcast_rarp.cnt, 1, 0))) {
2220
2221                 rarp_mbuf = rte_net_make_rarp_packet(mbuf_pool, &dev->mac);
2222                 if (rarp_mbuf == NULL) {
2223                         VHOST_LOG_DATA(ERR, "Failed to make RARP packet.\n");
2224                         count = 0;
2225                         goto out;
2226                 }
2227                 count -= 1;
2228         }
2229
2230         if (vq_is_packed(dev)) {
2231                 if (unlikely(dev->dequeue_zero_copy))
2232                         count = virtio_dev_tx_packed_zmbuf(dev, vq, mbuf_pool,
2233                                                            pkts, count);
2234                 else
2235                         count = virtio_dev_tx_packed(dev, vq, mbuf_pool, pkts,
2236                                                      count);
2237         } else
2238                 count = virtio_dev_tx_split(dev, vq, mbuf_pool, pkts, count);
2239
2240 out:
2241         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
2242                 vhost_user_iotlb_rd_unlock(vq);
2243
2244 out_access_unlock:
2245         rte_spinlock_unlock(&vq->access_lock);
2246
2247         if (unlikely(rarp_mbuf != NULL)) {
2248                 /*
2249                  * Inject it to the head of "pkts" array, so that switch's mac
2250                  * learning table will get updated first.
2251                  */
2252                 memmove(&pkts[1], pkts, count * sizeof(struct rte_mbuf *));
2253                 pkts[0] = rarp_mbuf;
2254                 count += 1;
2255         }
2256
2257         return count;
2258 }