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 static __rte_always_inline void
 386 vhost_flush_dequeue_packed(struct virtio_net *dev,
 387                            struct vhost_virtqueue *vq)
 388 {
 389         int shadow_count;
 390         if (!vq->shadow_used_idx)
 391                 return;
 392
 393         shadow_count = vq->last_used_idx - vq->shadow_last_used_idx;
 394         if (shadow_count <= 0)
 395                 shadow_count += vq->size;
 396
 397         if ((uint32_t)shadow_count >= (vq->size - MAX_PKT_BURST)) {
 398                 do_data_copy_dequeue(vq);
 399                 vhost_flush_dequeue_shadow_packed(dev, vq);
 400                 vhost_vring_call_packed(dev, vq);
 401         }
 402 }
 403
 404 /* avoid write operation when necessary, to lessen cache issues */
 405 #define ASSIGN_UNLESS_EQUAL(var, val) do {      \
 406         if ((var) != (val))                     \
 407                 (var) = (val);                  \
 408 } while (0)
 409
 410 static __rte_always_inline void
 411 virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
 412 {
 413         uint64_t csum_l4 = m_buf->ol_flags & PKT_TX_L4_MASK;
 414
 415         if (m_buf->ol_flags & PKT_TX_TCP_SEG)
 416                 csum_l4 |= PKT_TX_TCP_CKSUM;
 417
 418         if (csum_l4) {
 419                 net_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 420                 net_hdr->csum_start = m_buf->l2_len + m_buf->l3_len;
 421
 422                 switch (csum_l4) {
 423                 case PKT_TX_TCP_CKSUM:
 424                         net_hdr->csum_offset = (offsetof(struct rte_tcp_hdr,
 425                                                 cksum));
 426                         break;
 427                 case PKT_TX_UDP_CKSUM:
 428                         net_hdr->csum_offset = (offsetof(struct rte_udp_hdr,
 429                                                 dgram_cksum));
 430                         break;
 431                 case PKT_TX_SCTP_CKSUM:
 432                         net_hdr->csum_offset = (offsetof(struct rte_sctp_hdr,
 433                                                 cksum));
 434                         break;
 435                 }
 436         } else {
 437                 ASSIGN_UNLESS_EQUAL(net_hdr->csum_start, 0);
 438                 ASSIGN_UNLESS_EQUAL(net_hdr->csum_offset, 0);
 439                 ASSIGN_UNLESS_EQUAL(net_hdr->flags, 0);
 440         }
 441
 442         /* IP cksum verification cannot be bypassed, then calculate here */
 443         if (m_buf->ol_flags & PKT_TX_IP_CKSUM) {
 444                 struct rte_ipv4_hdr *ipv4_hdr;
 445
 446                 ipv4_hdr = rte_pktmbuf_mtod_offset(m_buf, struct rte_ipv4_hdr *,
 447                                                    m_buf->l2_len);
 448                 ipv4_hdr->hdr_checksum = 0;
 449                 ipv4_hdr->hdr_checksum = rte_ipv4_cksum(ipv4_hdr);
 450         }
 451
 452         if (m_buf->ol_flags & PKT_TX_TCP_SEG) {
 453                 if (m_buf->ol_flags & PKT_TX_IPV4)
 454                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 455                 else
 456                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 457                 net_hdr->gso_size = m_buf->tso_segsz;
 458                 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len
 459                                         + m_buf->l4_len;
 460         } else if (m_buf->ol_flags & PKT_TX_UDP_SEG) {
 461                 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
 462                 net_hdr->gso_size = m_buf->tso_segsz;
 463                 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len +
 464                         m_buf->l4_len;
 465         } else {
 466                 ASSIGN_UNLESS_EQUAL(net_hdr->gso_type, 0);
 467                 ASSIGN_UNLESS_EQUAL(net_hdr->gso_size, 0);
 468                 ASSIGN_UNLESS_EQUAL(net_hdr->hdr_len, 0);
 469         }
 470 }
 471
 472 static __rte_always_inline int
 473 map_one_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 474                 struct buf_vector *buf_vec, uint16_t *vec_idx,
 475                 uint64_t desc_iova, uint64_t desc_len, uint8_t perm)
 476 {
 477         uint16_t vec_id = *vec_idx;
 478
 479         while (desc_len) {
 480                 uint64_t desc_addr;
 481                 uint64_t desc_chunck_len = desc_len;
 482
 483                 if (unlikely(vec_id >= BUF_VECTOR_MAX))
 484                         return -1;
 485
 486                 desc_addr = vhost_iova_to_vva(dev, vq,
 487                                 desc_iova,
 488                                 &desc_chunck_len,
 489                                 perm);
 490                 if (unlikely(!desc_addr))
 491                         return -1;
 492
 493                 rte_prefetch0((void *)(uintptr_t)desc_addr);
 494
 495                 buf_vec[vec_id].buf_iova = desc_iova;
 496                 buf_vec[vec_id].buf_addr = desc_addr;
 497                 buf_vec[vec_id].buf_len  = desc_chunck_len;
 498
 499                 desc_len -= desc_chunck_len;
 500                 desc_iova += desc_chunck_len;
 501                 vec_id++;
 502         }
 503         *vec_idx = vec_id;
 504
 505         return 0;
 506 }
 507
 508 static __rte_always_inline int
 509 fill_vec_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 510                          uint32_t avail_idx, uint16_t *vec_idx,
 511                          struct buf_vector *buf_vec, uint16_t *desc_chain_head,
 512                          uint32_t *desc_chain_len, uint8_t perm)
 513 {
 514         uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
 515         uint16_t vec_id = *vec_idx;
 516         uint32_t len    = 0;
 517         uint64_t dlen;
 518         uint32_t nr_descs = vq->size;
 519         uint32_t cnt    = 0;
 520         struct vring_desc *descs = vq->desc;
 521         struct vring_desc *idesc = NULL;
 522
 523         if (unlikely(idx >= vq->size))
 524                 return -1;
 525
 526         *desc_chain_head = idx;
 527
 528         if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
 529                 dlen = vq->desc[idx].len;
 530                 nr_descs = dlen / sizeof(struct vring_desc);
 531                 if (unlikely(nr_descs > vq->size))
 532                         return -1;
 533
 534                 descs = (struct vring_desc *)(uintptr_t)
 535                         vhost_iova_to_vva(dev, vq, vq->desc[idx].addr,
 536                                                 &dlen,
 537                                                 VHOST_ACCESS_RO);
 538                 if (unlikely(!descs))
 539                         return -1;
 540
 541                 if (unlikely(dlen < vq->desc[idx].len)) {
 542                         /*
 543                          * The indirect desc table is not contiguous
 544                          * in process VA space, we have to copy it.
 545                          */
 546                         idesc = vhost_alloc_copy_ind_table(dev, vq,
 547                                         vq->desc[idx].addr, vq->desc[idx].len);
 548                         if (unlikely(!idesc))
 549                                 return -1;
 550
 551                         descs = idesc;
 552                 }
 553
 554                 idx = 0;
 555         }
 556
 557         while (1) {
 558                 if (unlikely(idx >= nr_descs || cnt++ >= nr_descs)) {
 559                         free_ind_table(idesc);
 560                         return -1;
 561                 }
 562
 563                 len += descs[idx].len;
 564
 565                 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 566                                                 descs[idx].addr, descs[idx].len,
 567                                                 perm))) {
 568                         free_ind_table(idesc);
 569                         return -1;
 570                 }
 571
 572                 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
 573                         break;
 574
 575                 idx = descs[idx].next;
 576         }
 577
 578         *desc_chain_len = len;
 579         *vec_idx = vec_id;
 580
 581         if (unlikely(!!idesc))
 582                 free_ind_table(idesc);
 583
 584         return 0;
 585 }
 586
 587 /*
 588  * Returns -1 on fail, 0 on success
 589  */
 590 static inline int
 591 reserve_avail_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 592                                 uint32_t size, struct buf_vector *buf_vec,
 593                                 uint16_t *num_buffers, uint16_t avail_head,
 594                                 uint16_t *nr_vec)
 595 {
 596         uint16_t cur_idx;
 597         uint16_t vec_idx = 0;
 598         uint16_t max_tries, tries = 0;
 599
 600         uint16_t head_idx = 0;
 601         uint32_t len = 0;
 602
 603         *num_buffers = 0;
 604         cur_idx  = vq->last_avail_idx;
 605
 606         if (rxvq_is_mergeable(dev))
 607                 max_tries = vq->size - 1;
 608         else
 609                 max_tries = 1;
 610
 611         while (size > 0) {
 612                 if (unlikely(cur_idx == avail_head))
 613                         return -1;
 614                 /*
 615                  * if we tried all available ring items, and still
 616                  * can't get enough buf, it means something abnormal
 617                  * happened.
 618                  */
 619                 if (unlikely(++tries > max_tries))
 620                         return -1;
 621
 622                 if (unlikely(fill_vec_buf_split(dev, vq, cur_idx,
 623                                                 &vec_idx, buf_vec,
 624                                                 &head_idx, &len,
 625                                                 VHOST_ACCESS_RW) < 0))
 626                         return -1;
 627                 len = RTE_MIN(len, size);
 628                 update_shadow_used_ring_split(vq, head_idx, len);
 629                 size -= len;
 630
 631                 cur_idx++;
 632                 *num_buffers += 1;
 633         }
 634
 635         *nr_vec = vec_idx;
 636
 637         return 0;
 638 }
 639
 640 static __rte_always_inline int
 641 fill_vec_buf_packed_indirect(struct virtio_net *dev,
 642                         struct vhost_virtqueue *vq,
 643                         struct vring_packed_desc *desc, uint16_t *vec_idx,
 644                         struct buf_vector *buf_vec, uint32_t *len, uint8_t perm)
 645 {
 646         uint16_t i;
 647         uint32_t nr_descs;
 648         uint16_t vec_id = *vec_idx;
 649         uint64_t dlen;
 650         struct vring_packed_desc *descs, *idescs = NULL;
 651
 652         dlen = desc->len;
 653         descs = (struct vring_packed_desc *)(uintptr_t)
 654                 vhost_iova_to_vva(dev, vq, desc->addr, &dlen, VHOST_ACCESS_RO);
 655         if (unlikely(!descs))
 656                 return -1;
 657
 658         if (unlikely(dlen < desc->len)) {
 659                 /*
 660                  * The indirect desc table is not contiguous
 661                  * in process VA space, we have to copy it.
 662                  */
 663                 idescs = vhost_alloc_copy_ind_table(dev,
 664                                 vq, desc->addr, desc->len);
 665                 if (unlikely(!idescs))
 666                         return -1;
 667
 668                 descs = idescs;
 669         }
 670
 671         nr_descs =  desc->len / sizeof(struct vring_packed_desc);
 672         if (unlikely(nr_descs >= vq->size)) {
 673                 free_ind_table(idescs);
 674                 return -1;
 675         }
 676
 677         for (i = 0; i < nr_descs; i++) {
 678                 if (unlikely(vec_id >= BUF_VECTOR_MAX)) {
 679                         free_ind_table(idescs);
 680                         return -1;
 681                 }
 682
 683                 *len += descs[i].len;
 684                 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 685                                                 descs[i].addr, descs[i].len,
 686                                                 perm)))
 687                         return -1;
 688         }
 689         *vec_idx = vec_id;
 690
 691         if (unlikely(!!idescs))
 692                 free_ind_table(idescs);
 693
 694         return 0;
 695 }
 696
 697 static __rte_always_inline int
 698 fill_vec_buf_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
 699                                 uint16_t avail_idx, uint16_t *desc_count,
 700                                 struct buf_vector *buf_vec, uint16_t *vec_idx,
 701                                 uint16_t *buf_id, uint32_t *len, uint8_t perm)
 702 {
 703         bool wrap_counter = vq->avail_wrap_counter;
 704         struct vring_packed_desc *descs = vq->desc_packed;
 705         uint16_t vec_id = *vec_idx;
 706
 707         if (avail_idx < vq->last_avail_idx)
 708                 wrap_counter ^= 1;
 709
 710         /*
 711          * Perform a load-acquire barrier in desc_is_avail to
 712          * enforce the ordering between desc flags and desc
 713          * content.
 714          */
 715         if (unlikely(!desc_is_avail(&descs[avail_idx], wrap_counter)))
 716                 return -1;
 717
 718         *desc_count = 0;
 719         *len = 0;
 720
 721         while (1) {
 722                 if (unlikely(vec_id >= BUF_VECTOR_MAX))
 723                         return -1;
 724
 725                 if (unlikely(*desc_count >= vq->size))
 726                         return -1;
 727
 728                 *desc_count += 1;
 729                 *buf_id = descs[avail_idx].id;
 730
 731                 if (descs[avail_idx].flags & VRING_DESC_F_INDIRECT) {
 732                         if (unlikely(fill_vec_buf_packed_indirect(dev, vq,
 733                                                         &descs[avail_idx],
 734                                                         &vec_id, buf_vec,
 735                                                         len, perm) < 0))
 736                                 return -1;
 737                 } else {
 738                         *len += descs[avail_idx].len;
 739
 740                         if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 741                                                         descs[avail_idx].addr,
 742                                                         descs[avail_idx].len,
 743                                                         perm)))
 744                                 return -1;
 745                 }
 746
 747                 if ((descs[avail_idx].flags & VRING_DESC_F_NEXT) == 0)
 748                         break;
 749
 750                 if (++avail_idx >= vq->size) {
 751                         avail_idx -= vq->size;
 752                         wrap_counter ^= 1;
 753                 }
 754         }
 755
 756         *vec_idx = vec_id;
 757
 758         return 0;
 759 }
 760
 761 static __rte_noinline void
 762 copy_vnet_hdr_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 763                 struct buf_vector *buf_vec,
 764                 struct virtio_net_hdr_mrg_rxbuf *hdr)
 765 {
 766         uint64_t len;
 767         uint64_t remain = dev->vhost_hlen;
 768         uint64_t src = (uint64_t)(uintptr_t)hdr, dst;
 769         uint64_t iova = buf_vec->buf_iova;
 770
 771         while (remain) {
 772                 len = RTE_MIN(remain,
 773                                 buf_vec->buf_len);
 774                 dst = buf_vec->buf_addr;
 775                 rte_memcpy((void *)(uintptr_t)dst,
 776                                 (void *)(uintptr_t)src,
 777                                 len);
 778
 779                 PRINT_PACKET(dev, (uintptr_t)dst,
 780                                 (uint32_t)len, 0);
 781                 vhost_log_cache_write_iova(dev, vq,
 782                                 iova, len);
 783
 784                 remain -= len;
 785                 iova += len;
 786                 src += len;
 787                 buf_vec++;
 788         }
 789 }
 790
 791 static __rte_always_inline int
 792 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 793                             struct rte_mbuf *m, struct buf_vector *buf_vec,
 794                             uint16_t nr_vec, uint16_t num_buffers)
 795 {
 796         uint32_t vec_idx = 0;
 797         uint32_t mbuf_offset, mbuf_avail;
 798         uint32_t buf_offset, buf_avail;
 799         uint64_t buf_addr, buf_iova, buf_len;
 800         uint32_t cpy_len;
 801         uint64_t hdr_addr;
 802         struct rte_mbuf *hdr_mbuf;
 803         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 804         struct virtio_net_hdr_mrg_rxbuf tmp_hdr, *hdr = NULL;
 805         int error = 0;
 806
 807         if (unlikely(m == NULL)) {
 808                 error = -1;
 809                 goto out;
 810         }
 811
 812         buf_addr = buf_vec[vec_idx].buf_addr;
 813         buf_iova = buf_vec[vec_idx].buf_iova;
 814         buf_len = buf_vec[vec_idx].buf_len;
 815
 816         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
 817                 error = -1;
 818                 goto out;
 819         }
 820
 821         hdr_mbuf = m;
 822         hdr_addr = buf_addr;
 823         if (unlikely(buf_len < dev->vhost_hlen))
 824                 hdr = &tmp_hdr;
 825         else
 826                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)hdr_addr;
 827
 828         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
 829                 dev->vid, num_buffers);
 830
 831         if (unlikely(buf_len < dev->vhost_hlen)) {
 832                 buf_offset = dev->vhost_hlen - buf_len;
 833                 vec_idx++;
 834                 buf_addr = buf_vec[vec_idx].buf_addr;
 835                 buf_iova = buf_vec[vec_idx].buf_iova;
 836                 buf_len = buf_vec[vec_idx].buf_len;
 837                 buf_avail = buf_len - buf_offset;
 838         } else {
 839                 buf_offset = dev->vhost_hlen;
 840                 buf_avail = buf_len - dev->vhost_hlen;
 841         }
 842
 843         mbuf_avail  = rte_pktmbuf_data_len(m);
 844         mbuf_offset = 0;
 845         while (mbuf_avail != 0 || m->next != NULL) {
 846                 /* done with current buf, get the next one */
 847                 if (buf_avail == 0) {
 848                         vec_idx++;
 849                         if (unlikely(vec_idx >= nr_vec)) {
 850                                 error = -1;
 851                                 goto out;
 852                         }
 853
 854                         buf_addr = buf_vec[vec_idx].buf_addr;
 855                         buf_iova = buf_vec[vec_idx].buf_iova;
 856                         buf_len = buf_vec[vec_idx].buf_len;
 857
 858                         buf_offset = 0;
 859                         buf_avail  = buf_len;
 860                 }
 861
 862                 /* done with current mbuf, get the next one */
 863                 if (mbuf_avail == 0) {
 864                         m = m->next;
 865
 866                         mbuf_offset = 0;
 867                         mbuf_avail  = rte_pktmbuf_data_len(m);
 868                 }
 869
 870                 if (hdr_addr) {
 871                         virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
 872                         if (rxvq_is_mergeable(dev))
 873                                 ASSIGN_UNLESS_EQUAL(hdr->num_buffers,
 874                                                 num_buffers);
 875
 876                         if (unlikely(hdr == &tmp_hdr)) {
 877                                 copy_vnet_hdr_to_desc(dev, vq, buf_vec, hdr);
 878                         } else {
 879                                 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
 880                                                 dev->vhost_hlen, 0);
 881                                 vhost_log_cache_write_iova(dev, vq,
 882                                                 buf_vec[0].buf_iova,
 883                                                 dev->vhost_hlen);
 884                         }
 885
 886                         hdr_addr = 0;
 887                 }
 888
 889                 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
 890
 891                 if (likely(cpy_len > MAX_BATCH_LEN ||
 892                                         vq->batch_copy_nb_elems >= vq->size)) {
 893                         rte_memcpy((void *)((uintptr_t)(buf_addr + buf_offset)),
 894                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 895                                 cpy_len);
 896                         vhost_log_cache_write_iova(dev, vq,
 897                                                    buf_iova + buf_offset,
 898                                                    cpy_len);
 899                         PRINT_PACKET(dev, (uintptr_t)(buf_addr + buf_offset),
 900                                 cpy_len, 0);
 901                 } else {
 902                         batch_copy[vq->batch_copy_nb_elems].dst =
 903                                 (void *)((uintptr_t)(buf_addr + buf_offset));
 904                         batch_copy[vq->batch_copy_nb_elems].src =
 905                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
 906                         batch_copy[vq->batch_copy_nb_elems].log_addr =
 907                                 buf_iova + buf_offset;
 908                         batch_copy[vq->batch_copy_nb_elems].len = cpy_len;
 909                         vq->batch_copy_nb_elems++;
 910                 }
 911
 912                 mbuf_avail  -= cpy_len;
 913                 mbuf_offset += cpy_len;
 914                 buf_avail  -= cpy_len;
 915                 buf_offset += cpy_len;
 916         }
 917
 918 out:
 919
 920         return error;
 921 }
 922
 923 static __rte_always_inline int
 924 vhost_enqueue_single_packed(struct virtio_net *dev,
 925                             struct vhost_virtqueue *vq,
 926                             struct rte_mbuf *pkt,
 927                             struct buf_vector *buf_vec,
 928                             uint16_t *nr_descs)
 929 {
 930         uint16_t nr_vec = 0;
 931         uint16_t avail_idx = vq->last_avail_idx;
 932         uint16_t max_tries, tries = 0;
 933         uint16_t buf_id = 0;
 934         uint32_t len = 0;
 935         uint16_t desc_count;
 936         uint32_t size = pkt->pkt_len + dev->vhost_hlen;
 937         uint16_t num_buffers = 0;
 938         uint32_t buffer_len[vq->size];
 939         uint16_t buffer_buf_id[vq->size];
 940         uint16_t buffer_desc_count[vq->size];
 941
 942         if (rxvq_is_mergeable(dev))
 943                 max_tries = vq->size - 1;
 944         else
 945                 max_tries = 1;
 946
 947         while (size > 0) {
 948                 /*
 949                  * if we tried all available ring items, and still
 950                  * can't get enough buf, it means something abnormal
 951                  * happened.
 952                  */
 953                 if (unlikely(++tries > max_tries))
 954                         return -1;
 955
 956                 if (unlikely(fill_vec_buf_packed(dev, vq,
 957                                                 avail_idx, &desc_count,
 958                                                 buf_vec, &nr_vec,
 959                                                 &buf_id, &len,
 960                                                 VHOST_ACCESS_RW) < 0))
 961                         return -1;
 962
 963                 len = RTE_MIN(len, size);
 964                 size -= len;
 965
 966                 buffer_len[num_buffers] = len;
 967                 buffer_buf_id[num_buffers] = buf_id;
 968                 buffer_desc_count[num_buffers] = desc_count;
 969                 num_buffers += 1;
 970
 971                 *nr_descs += desc_count;
 972                 avail_idx += desc_count;
 973                 if (avail_idx >= vq->size)
 974                         avail_idx -= vq->size;
 975         }
 976
 977         if (copy_mbuf_to_desc(dev, vq, pkt, buf_vec, nr_vec, num_buffers) < 0)
 978                 return -1;
 979
 980         vhost_shadow_enqueue_single_packed(dev, vq, buffer_len, buffer_buf_id,
 981                                            buffer_desc_count, num_buffers);
 982
 983         return 0;
 984 }
 985
 986 static __rte_noinline uint32_t
 987 virtio_dev_rx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 988         struct rte_mbuf **pkts, uint32_t count)
 989 {
 990         uint32_t pkt_idx = 0;
 991         uint16_t num_buffers;
 992         struct buf_vector buf_vec[BUF_VECTOR_MAX];
 993         uint16_t avail_head;
 994
 995         avail_head = *((volatile uint16_t *)&vq->avail->idx);
 996
 997         /*
 998          * The ordering between avail index and
 999          * desc reads needs to be enforced.
1000          */
1001         rte_smp_rmb();
1002
1003         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
1004
1005         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
1006                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
1007                 uint16_t nr_vec = 0;
1008
1009                 if (unlikely(reserve_avail_buf_split(dev, vq,
1010                                                 pkt_len, buf_vec, &num_buffers,
1011                                                 avail_head, &nr_vec) < 0)) {
1012                         VHOST_LOG_DEBUG(VHOST_DATA,
1013                                 "(%d) failed to get enough desc from vring\n",
1014                                 dev->vid);
1015                         vq->shadow_used_idx -= num_buffers;
1016                         break;
1017                 }
1018
1019                 VHOST_LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
1020                         dev->vid, vq->last_avail_idx,
1021                         vq->last_avail_idx + num_buffers);
1022
1023                 if (copy_mbuf_to_desc(dev, vq, pkts[pkt_idx],
1024                                                 buf_vec, nr_vec,
1025                                                 num_buffers) < 0) {
1026                         vq->shadow_used_idx -= num_buffers;
1027                         break;
1028                 }
1029
1030                 vq->last_avail_idx += num_buffers;
1031         }
1032
1033         do_data_copy_enqueue(dev, vq);
1034
1035         if (likely(vq->shadow_used_idx)) {
1036                 flush_shadow_used_ring_split(dev, vq);
1037                 vhost_vring_call_split(dev, vq);
1038         }
1039
1040         return pkt_idx;
1041 }
1042
1043 static __rte_always_inline int
1044 virtio_dev_rx_batch_packed(struct virtio_net *dev,
1045                            struct vhost_virtqueue *vq,
1046                            struct rte_mbuf **pkts)
1047 {
1048         bool wrap_counter = vq->avail_wrap_counter;
1049         struct vring_packed_desc *descs = vq->desc_packed;
1050         uint16_t avail_idx = vq->last_avail_idx;
1051         uint64_t desc_addrs[PACKED_BATCH_SIZE];
1052         struct virtio_net_hdr_mrg_rxbuf *hdrs[PACKED_BATCH_SIZE];
1053         uint32_t buf_offset = dev->vhost_hlen;
1054         uint64_t lens[PACKED_BATCH_SIZE];
1055         uint16_t ids[PACKED_BATCH_SIZE];
1056         uint16_t i;
1057
1058         if (unlikely(avail_idx & PACKED_BATCH_MASK))
1059                 return -1;
1060
1061         if (unlikely((avail_idx + PACKED_BATCH_SIZE) > vq->size))
1062                 return -1;
1063
1064         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1065                 if (unlikely(pkts[i]->next != NULL))
1066                         return -1;
1067                 if (unlikely(!desc_is_avail(&descs[avail_idx + i],
1068                                             wrap_counter)))
1069                         return -1;
1070         }
1071
1072         rte_smp_rmb();
1073
1074         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1075                 lens[i] = descs[avail_idx + i].len;
1076
1077         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1078                 if (unlikely(pkts[i]->pkt_len > (lens[i] - buf_offset)))
1079                         return -1;
1080         }
1081
1082         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1083                 desc_addrs[i] = vhost_iova_to_vva(dev, vq,
1084                                                   descs[avail_idx + i].addr,
1085                                                   &lens[i],
1086                                                   VHOST_ACCESS_RW);
1087
1088         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1089                 if (unlikely(lens[i] != descs[avail_idx + i].len))
1090                         return -1;
1091         }
1092
1093         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1094                 rte_prefetch0((void *)(uintptr_t)desc_addrs[i]);
1095                 hdrs[i] = (struct virtio_net_hdr_mrg_rxbuf *)
1096                                         (uintptr_t)desc_addrs[i];
1097                 lens[i] = pkts[i]->pkt_len + dev->vhost_hlen;
1098         }
1099
1100         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1101                 virtio_enqueue_offload(pkts[i], &hdrs[i]->hdr);
1102
1103         vq_inc_last_avail_packed(vq, PACKED_BATCH_SIZE);
1104
1105         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1106                 rte_memcpy((void *)(uintptr_t)(desc_addrs[i] + buf_offset),
1107                            rte_pktmbuf_mtod_offset(pkts[i], void *, 0),
1108                            pkts[i]->pkt_len);
1109         }
1110
1111         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1112                 ids[i] = descs[avail_idx + i].id;
1113
1114         vhost_flush_enqueue_batch_packed(dev, vq, lens, ids);
1115
1116         return 0;
1117 }
1118
1119 static __rte_always_inline int16_t
1120 virtio_dev_rx_single_packed(struct virtio_net *dev,
1121                             struct vhost_virtqueue *vq,
1122                             struct rte_mbuf *pkt)
1123 {
1124         struct buf_vector buf_vec[BUF_VECTOR_MAX];
1125         uint16_t nr_descs = 0;
1126
1127         rte_smp_rmb();
1128         if (unlikely(vhost_enqueue_single_packed(dev, vq, pkt, buf_vec,
1129                                                  &nr_descs) < 0)) {
1130                 VHOST_LOG_DEBUG(VHOST_DATA,
1131                                 "(%d) failed to get enough desc from vring\n",
1132                                 dev->vid);
1133                 return -1;
1134         }
1135
1136         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
1137                         dev->vid, vq->last_avail_idx,
1138                         vq->last_avail_idx + nr_descs);
1139
1140         vq_inc_last_avail_packed(vq, nr_descs);
1141
1142         return 0;
1143 }
1144
1145 static __rte_noinline uint32_t
1146 virtio_dev_rx_packed(struct virtio_net *dev,
1147                      struct vhost_virtqueue *vq,
1148                      struct rte_mbuf **pkts,
1149                      uint32_t count)
1150 {
1151         uint32_t pkt_idx = 0;
1152         uint32_t remained = count;
1153
1154         do {
1155                 rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
1156
1157                 if (remained >= PACKED_BATCH_SIZE) {
1158                         if (!virtio_dev_rx_batch_packed(dev, vq, pkts)) {
1159                                 pkt_idx += PACKED_BATCH_SIZE;
1160                                 remained -= PACKED_BATCH_SIZE;
1161                                 continue;
1162                         }
1163                 }
1164
1165                 if (virtio_dev_rx_single_packed(dev, vq, pkts[pkt_idx]))
1166                         break;
1167                 pkt_idx++;
1168                 remained--;
1169
1170         } while (pkt_idx < count);
1171
1172         if (vq->shadow_used_idx) {
1173                 do_data_copy_enqueue(dev, vq);
1174                 vhost_flush_enqueue_shadow_packed(dev, vq);
1175         }
1176
1177         if (pkt_idx)
1178                 vhost_vring_call_packed(dev, vq);
1179
1180         return pkt_idx;
1181 }
1182
1183 static __rte_always_inline uint32_t
1184 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
1185         struct rte_mbuf **pkts, uint32_t count)
1186 {
1187         struct vhost_virtqueue *vq;
1188         uint32_t nb_tx = 0;
1189
1190         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1191         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
1192                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
1193                         dev->vid, __func__, queue_id);
1194                 return 0;
1195         }
1196
1197         vq = dev->virtqueue[queue_id];
1198
1199         rte_spinlock_lock(&vq->access_lock);
1200
1201         if (unlikely(vq->enabled == 0))
1202                 goto out_access_unlock;
1203
1204         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1205                 vhost_user_iotlb_rd_lock(vq);
1206
1207         if (unlikely(vq->access_ok == 0))
1208                 if (unlikely(vring_translate(dev, vq) < 0))
1209                         goto out;
1210
1211         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
1212         if (count == 0)
1213                 goto out;
1214
1215         if (vq_is_packed(dev))
1216                 nb_tx = virtio_dev_rx_packed(dev, vq, pkts, count);
1217         else
1218                 nb_tx = virtio_dev_rx_split(dev, vq, pkts, count);
1219
1220 out:
1221         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1222                 vhost_user_iotlb_rd_unlock(vq);
1223
1224 out_access_unlock:
1225         rte_spinlock_unlock(&vq->access_lock);
1226
1227         return nb_tx;
1228 }
1229
1230 uint16_t
1231 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
1232         struct rte_mbuf **pkts, uint16_t count)
1233 {
1234         struct virtio_net *dev = get_device(vid);
1235
1236         if (!dev)
1237                 return 0;
1238
1239         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
1240                 RTE_LOG(ERR, VHOST_DATA,
1241                         "(%d) %s: built-in vhost net backend is disabled.\n",
1242                         dev->vid, __func__);
1243                 return 0;
1244         }
1245
1246         return virtio_dev_rx(dev, queue_id, pkts, count);
1247 }
1248
1249 static inline bool
1250 virtio_net_with_host_offload(struct virtio_net *dev)
1251 {
1252         if (dev->features &
1253                         ((1ULL << VIRTIO_NET_F_CSUM) |
1254                          (1ULL << VIRTIO_NET_F_HOST_ECN) |
1255                          (1ULL << VIRTIO_NET_F_HOST_TSO4) |
1256                          (1ULL << VIRTIO_NET_F_HOST_TSO6) |
1257                          (1ULL << VIRTIO_NET_F_HOST_UFO)))
1258                 return true;
1259
1260         return false;
1261 }
1262
1263 static void
1264 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
1265 {
1266         struct rte_ipv4_hdr *ipv4_hdr;
1267         struct rte_ipv6_hdr *ipv6_hdr;
1268         void *l3_hdr = NULL;
1269         struct rte_ether_hdr *eth_hdr;
1270         uint16_t ethertype;
1271
1272         eth_hdr = rte_pktmbuf_mtod(m, struct rte_ether_hdr *);
1273
1274         m->l2_len = sizeof(struct rte_ether_hdr);
1275         ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
1276
1277         if (ethertype == RTE_ETHER_TYPE_VLAN) {
1278                 struct rte_vlan_hdr *vlan_hdr =
1279                         (struct rte_vlan_hdr *)(eth_hdr + 1);
1280
1281                 m->l2_len += sizeof(struct rte_vlan_hdr);
1282                 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
1283         }
1284
1285         l3_hdr = (char *)eth_hdr + m->l2_len;
1286
1287         switch (ethertype) {
1288         case RTE_ETHER_TYPE_IPV4:
1289                 ipv4_hdr = l3_hdr;
1290                 *l4_proto = ipv4_hdr->next_proto_id;
1291                 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
1292                 *l4_hdr = (char *)l3_hdr + m->l3_len;
1293                 m->ol_flags |= PKT_TX_IPV4;
1294                 break;
1295         case RTE_ETHER_TYPE_IPV6:
1296                 ipv6_hdr = l3_hdr;
1297                 *l4_proto = ipv6_hdr->proto;
1298                 m->l3_len = sizeof(struct rte_ipv6_hdr);
1299                 *l4_hdr = (char *)l3_hdr + m->l3_len;
1300                 m->ol_flags |= PKT_TX_IPV6;
1301                 break;
1302         default:
1303                 m->l3_len = 0;
1304                 *l4_proto = 0;
1305                 *l4_hdr = NULL;
1306                 break;
1307         }
1308 }
1309
1310 static __rte_always_inline void
1311 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
1312 {
1313         uint16_t l4_proto = 0;
1314         void *l4_hdr = NULL;
1315         struct rte_tcp_hdr *tcp_hdr = NULL;
1316
1317         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
1318                 return;
1319
1320         parse_ethernet(m, &l4_proto, &l4_hdr);
1321         if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
1322                 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
1323                         switch (hdr->csum_offset) {
1324                         case (offsetof(struct rte_tcp_hdr, cksum)):
1325                                 if (l4_proto == IPPROTO_TCP)
1326                                         m->ol_flags |= PKT_TX_TCP_CKSUM;
1327                                 break;
1328                         case (offsetof(struct rte_udp_hdr, dgram_cksum)):
1329                                 if (l4_proto == IPPROTO_UDP)
1330                                         m->ol_flags |= PKT_TX_UDP_CKSUM;
1331                                 break;
1332                         case (offsetof(struct rte_sctp_hdr, cksum)):
1333                                 if (l4_proto == IPPROTO_SCTP)
1334                                         m->ol_flags |= PKT_TX_SCTP_CKSUM;
1335                                 break;
1336                         default:
1337                                 break;
1338                         }
1339                 }
1340         }
1341
1342         if (l4_hdr && hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
1343                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
1344                 case VIRTIO_NET_HDR_GSO_TCPV4:
1345                 case VIRTIO_NET_HDR_GSO_TCPV6:
1346                         tcp_hdr = l4_hdr;
1347                         m->ol_flags |= PKT_TX_TCP_SEG;
1348                         m->tso_segsz = hdr->gso_size;
1349                         m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
1350                         break;
1351                 case VIRTIO_NET_HDR_GSO_UDP:
1352                         m->ol_flags |= PKT_TX_UDP_SEG;
1353                         m->tso_segsz = hdr->gso_size;
1354                         m->l4_len = sizeof(struct rte_udp_hdr);
1355                         break;
1356                 default:
1357                         RTE_LOG(WARNING, VHOST_DATA,
1358                                 "unsupported gso type %u.\n", hdr->gso_type);
1359                         break;
1360                 }
1361         }
1362 }
1363
1364 static __rte_noinline void
1365 copy_vnet_hdr_from_desc(struct virtio_net_hdr *hdr,
1366                 struct buf_vector *buf_vec)
1367 {
1368         uint64_t len;
1369         uint64_t remain = sizeof(struct virtio_net_hdr);
1370         uint64_t src;
1371         uint64_t dst = (uint64_t)(uintptr_t)hdr;
1372
1373         while (remain) {
1374                 len = RTE_MIN(remain, buf_vec->buf_len);
1375                 src = buf_vec->buf_addr;
1376                 rte_memcpy((void *)(uintptr_t)dst,
1377                                 (void *)(uintptr_t)src, len);
1378
1379                 remain -= len;
1380                 dst += len;
1381                 buf_vec++;
1382         }
1383 }
1384
1385 static __rte_always_inline int
1386 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
1387                   struct buf_vector *buf_vec, uint16_t nr_vec,
1388                   struct rte_mbuf *m, struct rte_mempool *mbuf_pool)
1389 {
1390         uint32_t buf_avail, buf_offset;
1391         uint64_t buf_addr, buf_iova, buf_len;
1392         uint32_t mbuf_avail, mbuf_offset;
1393         uint32_t cpy_len;
1394         struct rte_mbuf *cur = m, *prev = m;
1395         struct virtio_net_hdr tmp_hdr;
1396         struct virtio_net_hdr *hdr = NULL;
1397         /* A counter to avoid desc dead loop chain */
1398         uint16_t vec_idx = 0;
1399         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
1400         int error = 0;
1401
1402         buf_addr = buf_vec[vec_idx].buf_addr;
1403         buf_iova = buf_vec[vec_idx].buf_iova;
1404         buf_len = buf_vec[vec_idx].buf_len;
1405
1406         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
1407                 error = -1;
1408                 goto out;
1409         }
1410
1411         if (virtio_net_with_host_offload(dev)) {
1412                 if (unlikely(buf_len < sizeof(struct virtio_net_hdr))) {
1413                         /*
1414                          * No luck, the virtio-net header doesn't fit
1415                          * in a contiguous virtual area.
1416                          */
1417                         copy_vnet_hdr_from_desc(&tmp_hdr, buf_vec);
1418                         hdr = &tmp_hdr;
1419                 } else {
1420                         hdr = (struct virtio_net_hdr *)((uintptr_t)buf_addr);
1421                 }
1422         }
1423
1424         /*
1425          * A virtio driver normally uses at least 2 desc buffers
1426          * for Tx: the first for storing the header, and others
1427          * for storing the data.
1428          */
1429         if (unlikely(buf_len < dev->vhost_hlen)) {
1430                 buf_offset = dev->vhost_hlen - buf_len;
1431                 vec_idx++;
1432                 buf_addr = buf_vec[vec_idx].buf_addr;
1433                 buf_iova = buf_vec[vec_idx].buf_iova;
1434                 buf_len = buf_vec[vec_idx].buf_len;
1435                 buf_avail  = buf_len - buf_offset;
1436         } else if (buf_len == dev->vhost_hlen) {
1437                 if (unlikely(++vec_idx >= nr_vec))
1438                         goto out;
1439                 buf_addr = buf_vec[vec_idx].buf_addr;
1440                 buf_iova = buf_vec[vec_idx].buf_iova;
1441                 buf_len = buf_vec[vec_idx].buf_len;
1442
1443                 buf_offset = 0;
1444                 buf_avail = buf_len;
1445         } else {
1446                 buf_offset = dev->vhost_hlen;
1447                 buf_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
1448         }
1449
1450         PRINT_PACKET(dev,
1451                         (uintptr_t)(buf_addr + buf_offset),
1452                         (uint32_t)buf_avail, 0);
1453
1454         mbuf_offset = 0;
1455         mbuf_avail  = m->buf_len - RTE_PKTMBUF_HEADROOM;
1456         while (1) {
1457                 uint64_t hpa;
1458
1459                 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
1460
1461                 /*
1462                  * A desc buf might across two host physical pages that are
1463                  * not continuous. In such case (gpa_to_hpa returns 0), data
1464                  * will be copied even though zero copy is enabled.
1465                  */
1466                 if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
1467                                         buf_iova + buf_offset, cpy_len)))) {
1468                         cur->data_len = cpy_len;
1469                         cur->data_off = 0;
1470                         cur->buf_addr =
1471                                 (void *)(uintptr_t)(buf_addr + buf_offset);
1472                         cur->buf_iova = hpa;
1473
1474                         /*
1475                          * In zero copy mode, one mbuf can only reference data
1476                          * for one or partial of one desc buff.
1477                          */
1478                         mbuf_avail = cpy_len;
1479                 } else {
1480                         if (likely(cpy_len > MAX_BATCH_LEN ||
1481                                    vq->batch_copy_nb_elems >= vq->size ||
1482                                    (hdr && cur == m))) {
1483                                 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
1484                                                                    mbuf_offset),
1485                                            (void *)((uintptr_t)(buf_addr +
1486                                                            buf_offset)),
1487                                            cpy_len);
1488                         } else {
1489                                 batch_copy[vq->batch_copy_nb_elems].dst =
1490                                         rte_pktmbuf_mtod_offset(cur, void *,
1491                                                                 mbuf_offset);
1492                                 batch_copy[vq->batch_copy_nb_elems].src =
1493                                         (void *)((uintptr_t)(buf_addr +
1494                                                                 buf_offset));
1495                                 batch_copy[vq->batch_copy_nb_elems].len =
1496                                         cpy_len;
1497                                 vq->batch_copy_nb_elems++;
1498                         }
1499                 }
1500
1501                 mbuf_avail  -= cpy_len;
1502                 mbuf_offset += cpy_len;
1503                 buf_avail -= cpy_len;
1504                 buf_offset += cpy_len;
1505
1506                 /* This buf reaches to its end, get the next one */
1507                 if (buf_avail == 0) {
1508                         if (++vec_idx >= nr_vec)
1509                                 break;
1510
1511                         buf_addr = buf_vec[vec_idx].buf_addr;
1512                         buf_iova = buf_vec[vec_idx].buf_iova;
1513                         buf_len = buf_vec[vec_idx].buf_len;
1514
1515                         buf_offset = 0;
1516                         buf_avail  = buf_len;
1517
1518                         PRINT_PACKET(dev, (uintptr_t)buf_addr,
1519                                         (uint32_t)buf_avail, 0);
1520                 }
1521
1522                 /*
1523                  * This mbuf reaches to its end, get a new one
1524                  * to hold more data.
1525                  */
1526                 if (mbuf_avail == 0) {
1527                         cur = rte_pktmbuf_alloc(mbuf_pool);
1528                         if (unlikely(cur == NULL)) {
1529                                 RTE_LOG(ERR, VHOST_DATA, "Failed to "
1530                                         "allocate memory for mbuf.\n");
1531                                 error = -1;
1532                                 goto out;
1533                         }
1534                         if (unlikely(dev->dequeue_zero_copy))
1535                                 rte_mbuf_refcnt_update(cur, 1);
1536
1537                         prev->next = cur;
1538                         prev->data_len = mbuf_offset;
1539                         m->nb_segs += 1;
1540                         m->pkt_len += mbuf_offset;
1541                         prev = cur;
1542
1543                         mbuf_offset = 0;
1544                         mbuf_avail  = cur->buf_len - RTE_PKTMBUF_HEADROOM;
1545                 }
1546         }
1547
1548         prev->data_len = mbuf_offset;
1549         m->pkt_len    += mbuf_offset;
1550
1551         if (hdr)
1552                 vhost_dequeue_offload(hdr, m);
1553
1554 out:
1555
1556         return error;
1557 }
1558
1559 static __rte_always_inline struct zcopy_mbuf *
1560 get_zmbuf(struct vhost_virtqueue *vq)
1561 {
1562         uint16_t i;
1563         uint16_t last;
1564         int tries = 0;
1565
1566         /* search [last_zmbuf_idx, zmbuf_size) */
1567         i = vq->last_zmbuf_idx;
1568         last = vq->zmbuf_size;
1569
1570 again:
1571         for (; i < last; i++) {
1572                 if (vq->zmbufs[i].in_use == 0) {
1573                         vq->last_zmbuf_idx = i + 1;
1574                         vq->zmbufs[i].in_use = 1;
1575                         return &vq->zmbufs[i];
1576                 }
1577         }
1578
1579         tries++;
1580         if (tries == 1) {
1581                 /* search [0, last_zmbuf_idx) */
1582                 i = 0;
1583                 last = vq->last_zmbuf_idx;
1584                 goto again;
1585         }
1586
1587         return NULL;
1588 }
1589
1590 static void
1591 virtio_dev_extbuf_free(void *addr __rte_unused, void *opaque)
1592 {
1593         rte_free(opaque);
1594 }
1595
1596 static int
1597 virtio_dev_extbuf_alloc(struct rte_mbuf *pkt, uint32_t size)
1598 {
1599         struct rte_mbuf_ext_shared_info *shinfo = NULL;
1600         uint32_t total_len = RTE_PKTMBUF_HEADROOM + size;
1601         uint16_t buf_len;
1602         rte_iova_t iova;
1603         void *buf;
1604
1605         /* Try to use pkt buffer to store shinfo to reduce the amount of memory
1606          * required, otherwise store shinfo in the new buffer.
1607          */
1608         if (rte_pktmbuf_tailroom(pkt) >= sizeof(*shinfo))
1609                 shinfo = rte_pktmbuf_mtod(pkt,
1610                                           struct rte_mbuf_ext_shared_info *);
1611         else {
1612                 total_len += sizeof(*shinfo) + sizeof(uintptr_t);
1613                 total_len = RTE_ALIGN_CEIL(total_len, sizeof(uintptr_t));
1614         }
1615
1616         if (unlikely(total_len > UINT16_MAX))
1617                 return -ENOSPC;
1618
1619         buf_len = total_len;
1620         buf = rte_malloc(NULL, buf_len, RTE_CACHE_LINE_SIZE);
1621         if (unlikely(buf == NULL))
1622                 return -ENOMEM;
1623
1624         /* Initialize shinfo */
1625         if (shinfo) {
1626                 shinfo->free_cb = virtio_dev_extbuf_free;
1627                 shinfo->fcb_opaque = buf;
1628                 rte_mbuf_ext_refcnt_set(shinfo, 1);
1629         } else {
1630                 shinfo = rte_pktmbuf_ext_shinfo_init_helper(buf, &buf_len,
1631                                               virtio_dev_extbuf_free, buf);
1632                 if (unlikely(shinfo == NULL)) {
1633                         rte_free(buf);
1634                         RTE_LOG(ERR, VHOST_DATA, "Failed to init shinfo\n");
1635                         return -1;
1636                 }
1637         }
1638
1639         iova = rte_malloc_virt2iova(buf);
1640         rte_pktmbuf_attach_extbuf(pkt, buf, iova, buf_len, shinfo);
1641         rte_pktmbuf_reset_headroom(pkt);
1642
1643         return 0;
1644 }
1645
1646 /*
1647  * Allocate a host supported pktmbuf.
1648  */
1649 static __rte_always_inline struct rte_mbuf *
1650 virtio_dev_pktmbuf_alloc(struct virtio_net *dev, struct rte_mempool *mp,
1651                          uint32_t data_len)
1652 {
1653         struct rte_mbuf *pkt = rte_pktmbuf_alloc(mp);
1654
1655         if (unlikely(pkt == NULL)) {
1656                 RTE_LOG(ERR, VHOST_DATA,
1657                         "Failed to allocate memory for mbuf.\n");
1658                 return NULL;
1659         }
1660
1661         if (rte_pktmbuf_tailroom(pkt) >= data_len)
1662                 return pkt;
1663
1664         /* attach an external buffer if supported */
1665         if (dev->extbuf && !virtio_dev_extbuf_alloc(pkt, data_len))
1666                 return pkt;
1667
1668         /* check if chained buffers are allowed */
1669         if (!dev->linearbuf)
1670                 return pkt;
1671
1672         /* Data doesn't fit into the buffer and the host supports
1673          * only linear buffers
1674          */
1675         rte_pktmbuf_free(pkt);
1676
1677         return NULL;
1678 }
1679
1680 static __rte_noinline uint16_t
1681 virtio_dev_tx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
1682         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1683 {
1684         uint16_t i;
1685         uint16_t free_entries;
1686
1687         if (unlikely(dev->dequeue_zero_copy)) {
1688                 struct zcopy_mbuf *zmbuf, *next;
1689
1690                 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1691                      zmbuf != NULL; zmbuf = next) {
1692                         next = TAILQ_NEXT(zmbuf, next);
1693
1694                         if (mbuf_is_consumed(zmbuf->mbuf)) {
1695                                 update_shadow_used_ring_split(vq,
1696                                                 zmbuf->desc_idx, 0);
1697                                 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1698                                 restore_mbuf(zmbuf->mbuf);
1699                                 rte_pktmbuf_free(zmbuf->mbuf);
1700                                 put_zmbuf(zmbuf);
1701                                 vq->nr_zmbuf -= 1;
1702                         }
1703                 }
1704
1705                 if (likely(vq->shadow_used_idx)) {
1706                         flush_shadow_used_ring_split(dev, vq);
1707                         vhost_vring_call_split(dev, vq);
1708                 }
1709         }
1710
1711         free_entries = *((volatile uint16_t *)&vq->avail->idx) -
1712                         vq->last_avail_idx;
1713         if (free_entries == 0)
1714                 return 0;
1715
1716         /*
1717          * The ordering between avail index and
1718          * desc reads needs to be enforced.
1719          */
1720         rte_smp_rmb();
1721
1722         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
1723
1724         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1725
1726         count = RTE_MIN(count, MAX_PKT_BURST);
1727         count = RTE_MIN(count, free_entries);
1728         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1729                         dev->vid, count);
1730
1731         for (i = 0; i < count; i++) {
1732                 struct buf_vector buf_vec[BUF_VECTOR_MAX];
1733                 uint16_t head_idx;
1734                 uint32_t buf_len;
1735                 uint16_t nr_vec = 0;
1736                 int err;
1737
1738                 if (unlikely(fill_vec_buf_split(dev, vq,
1739                                                 vq->last_avail_idx + i,
1740                                                 &nr_vec, buf_vec,
1741                                                 &head_idx, &buf_len,
1742                                                 VHOST_ACCESS_RO) < 0))
1743                         break;
1744
1745                 if (likely(dev->dequeue_zero_copy == 0))
1746                         update_shadow_used_ring_split(vq, head_idx, 0);
1747
1748                 pkts[i] = virtio_dev_pktmbuf_alloc(dev, mbuf_pool, buf_len);
1749                 if (unlikely(pkts[i] == NULL))
1750                         break;
1751
1752                 err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts[i],
1753                                 mbuf_pool);
1754                 if (unlikely(err)) {
1755                         rte_pktmbuf_free(pkts[i]);
1756                         break;
1757                 }
1758
1759                 if (unlikely(dev->dequeue_zero_copy)) {
1760                         struct zcopy_mbuf *zmbuf;
1761
1762                         zmbuf = get_zmbuf(vq);
1763                         if (!zmbuf) {
1764                                 rte_pktmbuf_free(pkts[i]);
1765                                 break;
1766                         }
1767                         zmbuf->mbuf = pkts[i];
1768                         zmbuf->desc_idx = head_idx;
1769
1770                         /*
1771                          * Pin lock the mbuf; we will check later to see
1772                          * whether the mbuf is freed (when we are the last
1773                          * user) or not. If that's the case, we then could
1774                          * update the used ring safely.
1775                          */
1776                         rte_mbuf_refcnt_update(pkts[i], 1);
1777
1778                         vq->nr_zmbuf += 1;
1779                         TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1780                 }
1781         }
1782         vq->last_avail_idx += i;
1783
1784         if (likely(dev->dequeue_zero_copy == 0)) {
1785                 do_data_copy_dequeue(vq);
1786                 if (unlikely(i < count))
1787                         vq->shadow_used_idx = i;
1788                 if (likely(vq->shadow_used_idx)) {
1789                         flush_shadow_used_ring_split(dev, vq);
1790                         vhost_vring_call_split(dev, vq);
1791                 }
1792         }
1793
1794         return i;
1795 }
1796
1797 static __rte_always_inline int
1798 vhost_reserve_avail_batch_packed(struct virtio_net *dev,
1799                                  struct vhost_virtqueue *vq,
1800                                  struct rte_mempool *mbuf_pool,
1801                                  struct rte_mbuf **pkts,
1802                                  uint16_t avail_idx,
1803                                  uintptr_t *desc_addrs,
1804                                  uint16_t *ids)
1805 {
1806         bool wrap = vq->avail_wrap_counter;
1807         struct vring_packed_desc *descs = vq->desc_packed;
1808         struct virtio_net_hdr *hdr;
1809         uint64_t lens[PACKED_BATCH_SIZE];
1810         uint64_t buf_lens[PACKED_BATCH_SIZE];
1811         uint32_t buf_offset = dev->vhost_hlen;
1812         uint16_t flags, i;
1813
1814         if (unlikely(avail_idx & PACKED_BATCH_MASK))
1815                 return -1;
1816         if (unlikely((avail_idx + PACKED_BATCH_SIZE) > vq->size))
1817                 return -1;
1818
1819         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1820                 flags = descs[avail_idx + i].flags;
1821                 if (unlikely((wrap != !!(flags & VRING_DESC_F_AVAIL)) ||
1822                              (wrap == !!(flags & VRING_DESC_F_USED))  ||
1823                              (flags & PACKED_DESC_SINGLE_DEQUEUE_FLAG)))
1824                         return -1;
1825         }
1826
1827         rte_smp_rmb();
1828
1829         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1830                 lens[i] = descs[avail_idx + i].len;
1831
1832         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1833                 desc_addrs[i] = vhost_iova_to_vva(dev, vq,
1834                                                   descs[avail_idx + i].addr,
1835                                                   &lens[i], VHOST_ACCESS_RW);
1836         }
1837
1838         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1839                 if (unlikely((lens[i] != descs[avail_idx + i].len)))
1840                         return -1;
1841         }
1842
1843         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1844                 pkts[i] = virtio_dev_pktmbuf_alloc(dev, mbuf_pool, lens[i]);
1845                 if (!pkts[i])
1846                         goto free_buf;
1847         }
1848
1849         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1850                 buf_lens[i] = pkts[i]->buf_len - pkts[i]->data_off;
1851
1852         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1853                 if (unlikely(buf_lens[i] < (lens[i] - buf_offset)))
1854                         goto free_buf;
1855         }
1856
1857         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1858                 pkts[i]->pkt_len = descs[avail_idx + i].len - buf_offset;
1859                 pkts[i]->data_len = pkts[i]->pkt_len;
1860                 ids[i] = descs[avail_idx + i].id;
1861         }
1862
1863         if (virtio_net_with_host_offload(dev)) {
1864                 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1865                         hdr = (struct virtio_net_hdr *)(desc_addrs[i]);
1866                         vhost_dequeue_offload(hdr, pkts[i]);
1867                 }
1868         }
1869
1870         return 0;
1871
1872 free_buf:
1873         for (i = 0; i < PACKED_BATCH_SIZE; i++)
1874                 rte_pktmbuf_free(pkts[i]);
1875
1876         return -1;
1877 }
1878
1879 static __rte_always_inline int
1880 virtio_dev_tx_batch_packed(struct virtio_net *dev,
1881                            struct vhost_virtqueue *vq,
1882                            struct rte_mempool *mbuf_pool,
1883                            struct rte_mbuf **pkts)
1884 {
1885         uint16_t avail_idx = vq->last_avail_idx;
1886         uint32_t buf_offset = dev->vhost_hlen;
1887         uintptr_t desc_addrs[PACKED_BATCH_SIZE];
1888         uint16_t ids[PACKED_BATCH_SIZE];
1889         uint16_t i;
1890
1891         if (vhost_reserve_avail_batch_packed(dev, vq, mbuf_pool, pkts,
1892                                              avail_idx, desc_addrs, ids))
1893                 return -1;
1894
1895         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1896                 rte_prefetch0((void *)(uintptr_t)desc_addrs[i]);
1897
1898         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1899                 rte_memcpy(rte_pktmbuf_mtod_offset(pkts[i], void *, 0),
1900                            (void *)(uintptr_t)(desc_addrs[i] + buf_offset),
1901                            pkts[i]->pkt_len);
1902
1903         if (virtio_net_is_inorder(dev))
1904                 vhost_shadow_dequeue_batch_packed_inorder(vq,
1905                         ids[PACKED_BATCH_SIZE - 1]);
1906         else
1907                 vhost_shadow_dequeue_batch_packed(dev, vq, ids);
1908
1909         vq_inc_last_avail_packed(vq, PACKED_BATCH_SIZE);
1910
1911         return 0;
1912 }
1913
1914 static __rte_always_inline int
1915 vhost_dequeue_single_packed(struct virtio_net *dev,
1916                             struct vhost_virtqueue *vq,
1917                             struct rte_mempool *mbuf_pool,
1918                             struct rte_mbuf **pkts,
1919                             uint16_t *buf_id,
1920                             uint16_t *desc_count)
1921 {
1922         struct buf_vector buf_vec[BUF_VECTOR_MAX];
1923         uint32_t buf_len;
1924         uint16_t nr_vec = 0;
1925         int err;
1926
1927         if (unlikely(fill_vec_buf_packed(dev, vq,
1928                                          vq->last_avail_idx, desc_count,
1929                                          buf_vec, &nr_vec,
1930                                          buf_id, &buf_len,
1931                                          VHOST_ACCESS_RO) < 0))
1932                 return -1;
1933
1934         *pkts = virtio_dev_pktmbuf_alloc(dev, mbuf_pool, buf_len);
1935         if (unlikely(*pkts == NULL)) {
1936                 RTE_LOG(ERR, VHOST_DATA,
1937                         "Failed to allocate memory for mbuf.\n");
1938                 return -1;
1939         }
1940
1941         err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, *pkts,
1942                                 mbuf_pool);
1943         if (unlikely(err)) {
1944                 rte_pktmbuf_free(*pkts);
1945                 return -1;
1946         }
1947
1948         return 0;
1949 }
1950
1951 static __rte_always_inline int
1952 virtio_dev_tx_single_packed(struct virtio_net *dev,
1953                             struct vhost_virtqueue *vq,
1954                             struct rte_mempool *mbuf_pool,
1955                             struct rte_mbuf **pkts)
1956 {
1957
1958         uint16_t buf_id, desc_count;
1959
1960         if (vhost_dequeue_single_packed(dev, vq, mbuf_pool, pkts, &buf_id,
1961                                         &desc_count))
1962                 return -1;
1963
1964         if (virtio_net_is_inorder(dev))
1965                 vhost_shadow_dequeue_single_packed_inorder(vq, buf_id,
1966                                                            desc_count);
1967         else
1968                 vhost_shadow_dequeue_single_packed(vq, buf_id, desc_count);
1969
1970         vq_inc_last_avail_packed(vq, desc_count);
1971
1972         return 0;
1973 }
1974
1975 static __rte_always_inline int
1976 virtio_dev_tx_batch_packed_zmbuf(struct virtio_net *dev,
1977                                  struct vhost_virtqueue *vq,
1978                                  struct rte_mempool *mbuf_pool,
1979                                  struct rte_mbuf **pkts)
1980 {
1981         struct zcopy_mbuf *zmbufs[PACKED_BATCH_SIZE];
1982         uintptr_t desc_addrs[PACKED_BATCH_SIZE];
1983         uint16_t ids[PACKED_BATCH_SIZE];
1984         uint16_t i;
1985
1986         uint16_t avail_idx = vq->last_avail_idx;
1987
1988         if (vhost_reserve_avail_batch_packed(dev, vq, mbuf_pool, pkts,
1989                                              avail_idx, desc_addrs, ids))
1990                 return -1;
1991
1992         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1993                 zmbufs[i] = get_zmbuf(vq);
1994
1995         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1996                 if (!zmbufs[i])
1997                         goto free_pkt;
1998         }
1999
2000         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2001                 zmbufs[i]->mbuf = pkts[i];
2002                 zmbufs[i]->desc_idx = avail_idx + i;
2003                 zmbufs[i]->desc_count = 1;
2004         }
2005
2006         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2007                 rte_mbuf_refcnt_update(pkts[i], 1);
2008
2009         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2010                 TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbufs[i], next);
2011
2012         vq->nr_zmbuf += PACKED_BATCH_SIZE;
2013         vq_inc_last_avail_packed(vq, PACKED_BATCH_SIZE);
2014
2015         return 0;
2016
2017 free_pkt:
2018         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2019                 rte_pktmbuf_free(pkts[i]);
2020
2021         return -1;
2022 }
2023
2024 static __rte_always_inline int
2025 virtio_dev_tx_single_packed_zmbuf(struct virtio_net *dev,
2026                                   struct vhost_virtqueue *vq,
2027                                   struct rte_mempool *mbuf_pool,
2028                                   struct rte_mbuf **pkts)
2029 {
2030         uint16_t buf_id, desc_count;
2031         struct zcopy_mbuf *zmbuf;
2032
2033         if (vhost_dequeue_single_packed(dev, vq, mbuf_pool, pkts, &buf_id,
2034                                         &desc_count))
2035                 return -1;
2036
2037         zmbuf = get_zmbuf(vq);
2038         if (!zmbuf) {
2039                 rte_pktmbuf_free(*pkts);
2040                 return -1;
2041         }
2042         zmbuf->mbuf = *pkts;
2043         zmbuf->desc_idx = vq->last_avail_idx;
2044         zmbuf->desc_count = desc_count;
2045
2046         rte_mbuf_refcnt_update(*pkts, 1);
2047
2048         vq->nr_zmbuf += 1;
2049         TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
2050
2051         vq_inc_last_avail_packed(vq, desc_count);
2052         return 0;
2053 }
2054
2055 static __rte_always_inline void
2056 free_zmbuf(struct vhost_virtqueue *vq)
2057 {
2058         struct zcopy_mbuf *next = NULL;
2059         struct zcopy_mbuf *zmbuf;
2060
2061         for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
2062              zmbuf != NULL; zmbuf = next) {
2063                 next = TAILQ_NEXT(zmbuf, next);
2064
2065                 uint16_t last_used_idx = vq->last_used_idx;
2066
2067                 if (mbuf_is_consumed(zmbuf->mbuf)) {
2068                         uint16_t flags;
2069                         flags = vq->desc_packed[last_used_idx].flags;
2070                         if (vq->used_wrap_counter) {
2071                                 flags |= VRING_DESC_F_USED;
2072                                 flags |= VRING_DESC_F_AVAIL;
2073                         } else {
2074                                 flags &= ~VRING_DESC_F_USED;
2075                                 flags &= ~VRING_DESC_F_AVAIL;
2076                         }
2077
2078                         vq->desc_packed[last_used_idx].id = zmbuf->desc_idx;
2079                         vq->desc_packed[last_used_idx].len = 0;
2080
2081                         rte_smp_wmb();
2082                         vq->desc_packed[last_used_idx].flags = flags;
2083
2084                         vq_inc_last_used_packed(vq, zmbuf->desc_count);
2085
2086                         TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
2087                         restore_mbuf(zmbuf->mbuf);
2088                         rte_pktmbuf_free(zmbuf->mbuf);
2089                         put_zmbuf(zmbuf);
2090                         vq->nr_zmbuf -= 1;
2091                 }
2092         }
2093 }
2094
2095 static __rte_noinline uint16_t
2096 virtio_dev_tx_packed_zmbuf(struct virtio_net *dev,
2097                            struct vhost_virtqueue *vq,
2098                            struct rte_mempool *mbuf_pool,
2099                            struct rte_mbuf **pkts,
2100                            uint32_t count)
2101 {
2102         uint32_t pkt_idx = 0;
2103         uint32_t remained = count;
2104
2105         free_zmbuf(vq);
2106
2107         do {
2108                 if (remained >= PACKED_BATCH_SIZE) {
2109                         if (!virtio_dev_tx_batch_packed_zmbuf(dev, vq,
2110                                 mbuf_pool, &pkts[pkt_idx])) {
2111                                 pkt_idx += PACKED_BATCH_SIZE;
2112                                 remained -= PACKED_BATCH_SIZE;
2113                                 continue;
2114                         }
2115                 }
2116
2117                 if (virtio_dev_tx_single_packed_zmbuf(dev, vq, mbuf_pool,
2118                                                       &pkts[pkt_idx]))
2119                         break;
2120                 pkt_idx++;
2121                 remained--;
2122
2123         } while (remained);
2124
2125         if (pkt_idx)
2126                 vhost_vring_call_packed(dev, vq);
2127
2128         return pkt_idx;
2129 }
2130
2131 static __rte_noinline uint16_t
2132 virtio_dev_tx_packed(struct virtio_net *dev,
2133                      struct vhost_virtqueue *vq,
2134                      struct rte_mempool *mbuf_pool,
2135                      struct rte_mbuf **pkts,
2136                      uint32_t count)
2137 {
2138         uint32_t pkt_idx = 0;
2139         uint32_t remained = count;
2140
2141         do {
2142                 rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
2143
2144                 if (remained >= PACKED_BATCH_SIZE) {
2145                         if (!virtio_dev_tx_batch_packed(dev, vq, mbuf_pool,
2146                                                         &pkts[pkt_idx])) {
2147                                 vhost_flush_dequeue_packed(dev, vq);
2148                                 pkt_idx += PACKED_BATCH_SIZE;
2149                                 remained -= PACKED_BATCH_SIZE;
2150                                 continue;
2151                         }
2152                 }
2153
2154                 if (virtio_dev_tx_single_packed(dev, vq, mbuf_pool,
2155                                                 &pkts[pkt_idx]))
2156                         break;
2157                 vhost_flush_dequeue_packed(dev, vq);
2158                 pkt_idx++;
2159                 remained--;
2160
2161         } while (remained);
2162
2163         if (vq->shadow_used_idx)
2164                 do_data_copy_dequeue(vq);
2165
2166         return pkt_idx;
2167 }
2168
2169 uint16_t
2170 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
2171         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
2172 {
2173         struct virtio_net *dev;
2174         struct rte_mbuf *rarp_mbuf = NULL;
2175         struct vhost_virtqueue *vq;
2176
2177         dev = get_device(vid);
2178         if (!dev)
2179                 return 0;
2180
2181         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
2182                 RTE_LOG(ERR, VHOST_DATA,
2183                         "(%d) %s: built-in vhost net backend is disabled.\n",
2184                         dev->vid, __func__);
2185                 return 0;
2186         }
2187
2188         if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
2189                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
2190                         dev->vid, __func__, queue_id);
2191                 return 0;
2192         }
2193
2194         vq = dev->virtqueue[queue_id];
2195
2196         if (unlikely(rte_spinlock_trylock(&vq->access_lock) == 0))
2197                 return 0;
2198
2199         if (unlikely(vq->enabled == 0)) {
2200                 count = 0;
2201                 goto out_access_unlock;
2202         }
2203
2204         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
2205                 vhost_user_iotlb_rd_lock(vq);
2206
2207         if (unlikely(vq->access_ok == 0))
2208                 if (unlikely(vring_translate(dev, vq) < 0)) {
2209                         count = 0;
2210                         goto out;
2211                 }
2212
2213         /*
2214          * Construct a RARP broadcast packet, and inject it to the "pkts"
2215          * array, to looks like that guest actually send such packet.
2216          *
2217          * Check user_send_rarp() for more information.
2218          *
2219          * broadcast_rarp shares a cacheline in the virtio_net structure
2220          * with some fields that are accessed during enqueue and
2221          * rte_atomic16_cmpset() causes a write if using cmpxchg. This could
2222          * result in false sharing between enqueue and dequeue.
2223          *
2224          * Prevent unnecessary false sharing by reading broadcast_rarp first
2225          * and only performing cmpset if the read indicates it is likely to
2226          * be set.
2227          */
2228         if (unlikely(rte_atomic16_read(&dev->broadcast_rarp) &&
2229                         rte_atomic16_cmpset((volatile uint16_t *)
2230                                 &dev->broadcast_rarp.cnt, 1, 0))) {
2231
2232                 rarp_mbuf = rte_net_make_rarp_packet(mbuf_pool, &dev->mac);
2233                 if (rarp_mbuf == NULL) {
2234                         RTE_LOG(ERR, VHOST_DATA,
2235                                 "Failed to make RARP packet.\n");
2236                         count = 0;
2237                         goto out;
2238                 }
2239                 count -= 1;
2240         }
2241
2242         if (vq_is_packed(dev)) {
2243                 if (unlikely(dev->dequeue_zero_copy))
2244                         count = virtio_dev_tx_packed_zmbuf(dev, vq, mbuf_pool,
2245                                                            pkts, count);
2246                 else
2247                         count = virtio_dev_tx_packed(dev, vq, mbuf_pool, pkts,
2248                                                      count);
2249         } else
2250                 count = virtio_dev_tx_split(dev, vq, mbuf_pool, pkts, count);
2251
2252 out:
2253         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
2254                 vhost_user_iotlb_rd_unlock(vq);
2255
2256 out_access_unlock:
2257         rte_spinlock_unlock(&vq->access_lock);
2258
2259         if (unlikely(rarp_mbuf != NULL)) {
2260                 /*
2261                  * Inject it to the head of "pkts" array, so that switch's mac
2262                  * learning table will get updated first.
2263                  */
2264                 memmove(&pkts[1], pkts, count * sizeof(struct rte_mbuf *));
2265                 pkts[0] = rarp_mbuf;
2266                 count += 1;
2267         }
2268
2269         return count;
2270 }