lib/vhost/virtio_net.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2016 Intel Corporation
   3  */
   4
   5 #include <stdint.h>
   6 #include <stdbool.h>
   7 #include <linux/virtio_net.h>
   8
   9 #include <rte_mbuf.h>
  10 #include <rte_memcpy.h>
  11 #include <rte_net.h>
  12 #include <rte_ether.h>
  13 #include <rte_ip.h>
  14 #include <rte_dmadev.h>
  15 #include <rte_vhost.h>
  16 #include <rte_tcp.h>
  17 #include <rte_udp.h>
  18 #include <rte_sctp.h>
  19 #include <rte_arp.h>
  20 #include <rte_spinlock.h>
  21 #include <rte_malloc.h>
  22 #include <rte_vhost_async.h>
  23
  24 #include "iotlb.h"
  25 #include "vhost.h"
  26
  27 #define MAX_BATCH_LEN 256
  28
  29 /* DMA device copy operation tracking array. */
  30 struct async_dma_info dma_copy_track[RTE_DMADEV_DEFAULT_MAX];
  31
  32 static  __rte_always_inline bool
  33 rxvq_is_mergeable(struct virtio_net *dev)
  34 {
  35         return dev->features & (1ULL << VIRTIO_NET_F_MRG_RXBUF);
  36 }
  37
  38 static  __rte_always_inline bool
  39 virtio_net_is_inorder(struct virtio_net *dev)
  40 {
  41         return dev->features & (1ULL << VIRTIO_F_IN_ORDER);
  42 }
  43
  44 static bool
  45 is_valid_virt_queue_idx(uint32_t idx, int is_tx, uint32_t nr_vring)
  46 {
  47         return (is_tx ^ (idx & 1)) == 0 && idx < nr_vring;
  48 }
  49
  50 /*
  51  * This function must be called with virtqueue's access_lock taken.
  52  */
  53 static inline void
  54 vhost_queue_stats_update(struct virtio_net *dev, struct vhost_virtqueue *vq,
  55                 struct rte_mbuf **pkts, uint16_t count)
  56 {
  57         struct virtqueue_stats *stats = &vq->stats;
  58         int i;
  59
  60         if (!(dev->flags & VIRTIO_DEV_STATS_ENABLED))
  61                 return;
  62
  63         for (i = 0; i < count; i++) {
  64                 struct rte_ether_addr *ea;
  65                 struct rte_mbuf *pkt = pkts[i];
  66                 uint32_t pkt_len = rte_pktmbuf_pkt_len(pkt);
  67
  68                 stats->packets++;
  69                 stats->bytes += pkt_len;
  70
  71                 if (pkt_len == 64) {
  72                         stats->size_bins[1]++;
  73                 } else if (pkt_len > 64 && pkt_len < 1024) {
  74                         uint32_t bin;
  75
  76                         /* count zeros, and offset into correct bin */
  77                         bin = (sizeof(pkt_len) * 8) - __builtin_clz(pkt_len) - 5;
  78                         stats->size_bins[bin]++;
  79                 } else {
  80                         if (pkt_len < 64)
  81                                 stats->size_bins[0]++;
  82                         else if (pkt_len < 1519)
  83                                 stats->size_bins[6]++;
  84                         else
  85                                 stats->size_bins[7]++;
  86                 }
  87
  88                 ea = rte_pktmbuf_mtod(pkt, struct rte_ether_addr *);
  89                 if (rte_is_multicast_ether_addr(ea)) {
  90                         if (rte_is_broadcast_ether_addr(ea))
  91                                 stats->broadcast++;
  92                         else
  93                                 stats->multicast++;
  94                 }
  95         }
  96 }
  97
  98 static __rte_always_inline int64_t
  99 vhost_async_dma_transfer_one(struct virtio_net *dev, struct vhost_virtqueue *vq,
 100                 int16_t dma_id, uint16_t vchan_id, uint16_t flag_idx,
 101                 struct vhost_iov_iter *pkt)
 102 {
 103         struct async_dma_vchan_info *dma_info = &dma_copy_track[dma_id].vchans[vchan_id];
 104         uint16_t ring_mask = dma_info->ring_mask;
 105         static bool vhost_async_dma_copy_log;
 106
 107
 108         struct vhost_iovec *iov = pkt->iov;
 109         int copy_idx = 0;
 110         uint32_t nr_segs = pkt->nr_segs;
 111         uint16_t i;
 112
 113         if (rte_dma_burst_capacity(dma_id, vchan_id) < nr_segs)
 114                 return -1;
 115
 116         for (i = 0; i < nr_segs; i++) {
 117                 copy_idx = rte_dma_copy(dma_id, vchan_id, (rte_iova_t)iov[i].src_addr,
 118                                 (rte_iova_t)iov[i].dst_addr, iov[i].len, RTE_DMA_OP_FLAG_LLC);
 119                 /**
 120                  * Since all memory is pinned and DMA vChannel
 121                  * ring has enough space, failure should be a
 122                  * rare case. If failure happens, it means DMA
 123                  * device encounters serious errors; in this
 124                  * case, please stop async data-path and check
 125                  * what has happened to DMA device.
 126                  */
 127                 if (unlikely(copy_idx < 0)) {
 128                         if (!vhost_async_dma_copy_log) {
 129                                 VHOST_LOG_DATA(ERR, "(%s) DMA copy failed for channel %d:%u\n",
 130                                                 dev->ifname, dma_id, vchan_id);
 131                                 vhost_async_dma_copy_log = true;
 132                         }
 133                         return -1;
 134                 }
 135         }
 136
 137         /**
 138          * Only store packet completion flag address in the last copy's
 139          * slot, and other slots are set to NULL.
 140          */
 141         dma_info->pkts_cmpl_flag_addr[copy_idx & ring_mask] = &vq->async->pkts_cmpl_flag[flag_idx];
 142
 143         return nr_segs;
 144 }
 145
 146 static __rte_always_inline uint16_t
 147 vhost_async_dma_transfer(struct virtio_net *dev, struct vhost_virtqueue *vq,
 148                 int16_t dma_id, uint16_t vchan_id, uint16_t head_idx,
 149                 struct vhost_iov_iter *pkts, uint16_t nr_pkts)
 150 {
 151         struct async_dma_vchan_info *dma_info = &dma_copy_track[dma_id].vchans[vchan_id];
 152         int64_t ret, nr_copies = 0;
 153         uint16_t pkt_idx;
 154
 155         rte_spinlock_lock(&dma_info->dma_lock);
 156
 157         for (pkt_idx = 0; pkt_idx < nr_pkts; pkt_idx++) {
 158                 ret = vhost_async_dma_transfer_one(dev, vq, dma_id, vchan_id, head_idx,
 159                                 &pkts[pkt_idx]);
 160                 if (unlikely(ret < 0))
 161                         break;
 162
 163                 nr_copies += ret;
 164                 head_idx++;
 165                 if (head_idx >= vq->size)
 166                         head_idx -= vq->size;
 167         }
 168
 169         if (likely(nr_copies > 0))
 170                 rte_dma_submit(dma_id, vchan_id);
 171
 172         rte_spinlock_unlock(&dma_info->dma_lock);
 173
 174         return pkt_idx;
 175 }
 176
 177 static __rte_always_inline uint16_t
 178 vhost_async_dma_check_completed(struct virtio_net *dev, int16_t dma_id, uint16_t vchan_id,
 179                 uint16_t max_pkts)
 180 {
 181         struct async_dma_vchan_info *dma_info = &dma_copy_track[dma_id].vchans[vchan_id];
 182         uint16_t ring_mask = dma_info->ring_mask;
 183         uint16_t last_idx = 0;
 184         uint16_t nr_copies;
 185         uint16_t copy_idx;
 186         uint16_t i;
 187         bool has_error = false;
 188         static bool vhost_async_dma_complete_log;
 189
 190         rte_spinlock_lock(&dma_info->dma_lock);
 191
 192         /**
 193          * Print error log for debugging, if DMA reports error during
 194          * DMA transfer. We do not handle error in vhost level.
 195          */
 196         nr_copies = rte_dma_completed(dma_id, vchan_id, max_pkts, &last_idx, &has_error);
 197         if (unlikely(!vhost_async_dma_complete_log && has_error)) {
 198                 VHOST_LOG_DATA(ERR, "(%s) DMA completion failure on channel %d:%u\n", dev->ifname,
 199                                 dma_id, vchan_id);
 200                 vhost_async_dma_complete_log = true;
 201         } else if (nr_copies == 0) {
 202                 goto out;
 203         }
 204
 205         copy_idx = last_idx - nr_copies + 1;
 206         for (i = 0; i < nr_copies; i++) {
 207                 bool *flag;
 208
 209                 flag = dma_info->pkts_cmpl_flag_addr[copy_idx & ring_mask];
 210                 if (flag) {
 211                         /**
 212                          * Mark the packet flag as received. The flag
 213                          * could belong to another virtqueue but write
 214                          * is atomic.
 215                          */
 216                         *flag = true;
 217                         dma_info->pkts_cmpl_flag_addr[copy_idx & ring_mask] = NULL;
 218                 }
 219                 copy_idx++;
 220         }
 221
 222 out:
 223         rte_spinlock_unlock(&dma_info->dma_lock);
 224         return nr_copies;
 225 }
 226
 227 static inline void
 228 do_data_copy_enqueue(struct virtio_net *dev, struct vhost_virtqueue *vq)
 229 {
 230         struct batch_copy_elem *elem = vq->batch_copy_elems;
 231         uint16_t count = vq->batch_copy_nb_elems;
 232         int i;
 233
 234         for (i = 0; i < count; i++) {
 235                 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
 236                 vhost_log_cache_write_iova(dev, vq, elem[i].log_addr,
 237                                            elem[i].len);
 238                 PRINT_PACKET(dev, (uintptr_t)elem[i].dst, elem[i].len, 0);
 239         }
 240
 241         vq->batch_copy_nb_elems = 0;
 242 }
 243
 244 static inline void
 245 do_data_copy_dequeue(struct vhost_virtqueue *vq)
 246 {
 247         struct batch_copy_elem *elem = vq->batch_copy_elems;
 248         uint16_t count = vq->batch_copy_nb_elems;
 249         int i;
 250
 251         for (i = 0; i < count; i++)
 252                 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
 253
 254         vq->batch_copy_nb_elems = 0;
 255 }
 256
 257 static __rte_always_inline void
 258 do_flush_shadow_used_ring_split(struct virtio_net *dev,
 259                         struct vhost_virtqueue *vq,
 260                         uint16_t to, uint16_t from, uint16_t size)
 261 {
 262         rte_memcpy(&vq->used->ring[to],
 263                         &vq->shadow_used_split[from],
 264                         size * sizeof(struct vring_used_elem));
 265         vhost_log_cache_used_vring(dev, vq,
 266                         offsetof(struct vring_used, ring[to]),
 267                         size * sizeof(struct vring_used_elem));
 268 }
 269
 270 static __rte_always_inline void
 271 flush_shadow_used_ring_split(struct virtio_net *dev, struct vhost_virtqueue *vq)
 272 {
 273         uint16_t used_idx = vq->last_used_idx & (vq->size - 1);
 274
 275         if (used_idx + vq->shadow_used_idx <= vq->size) {
 276                 do_flush_shadow_used_ring_split(dev, vq, used_idx, 0,
 277                                           vq->shadow_used_idx);
 278         } else {
 279                 uint16_t size;
 280
 281                 /* update used ring interval [used_idx, vq->size] */
 282                 size = vq->size - used_idx;
 283                 do_flush_shadow_used_ring_split(dev, vq, used_idx, 0, size);
 284
 285                 /* update the left half used ring interval [0, left_size] */
 286                 do_flush_shadow_used_ring_split(dev, vq, 0, size,
 287                                           vq->shadow_used_idx - size);
 288         }
 289         vq->last_used_idx += vq->shadow_used_idx;
 290
 291         vhost_log_cache_sync(dev, vq);
 292
 293         __atomic_add_fetch(&vq->used->idx, vq->shadow_used_idx,
 294                            __ATOMIC_RELEASE);
 295         vq->shadow_used_idx = 0;
 296         vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
 297                 sizeof(vq->used->idx));
 298 }
 299
 300 static __rte_always_inline void
 301 update_shadow_used_ring_split(struct vhost_virtqueue *vq,
 302                          uint16_t desc_idx, uint32_t len)
 303 {
 304         uint16_t i = vq->shadow_used_idx++;
 305
 306         vq->shadow_used_split[i].id  = desc_idx;
 307         vq->shadow_used_split[i].len = len;
 308 }
 309
 310 static __rte_always_inline void
 311 vhost_flush_enqueue_shadow_packed(struct virtio_net *dev,
 312                                   struct vhost_virtqueue *vq)
 313 {
 314         int i;
 315         uint16_t used_idx = vq->last_used_idx;
 316         uint16_t head_idx = vq->last_used_idx;
 317         uint16_t head_flags = 0;
 318
 319         /* Split loop in two to save memory barriers */
 320         for (i = 0; i < vq->shadow_used_idx; i++) {
 321                 vq->desc_packed[used_idx].id = vq->shadow_used_packed[i].id;
 322                 vq->desc_packed[used_idx].len = vq->shadow_used_packed[i].len;
 323
 324                 used_idx += vq->shadow_used_packed[i].count;
 325                 if (used_idx >= vq->size)
 326                         used_idx -= vq->size;
 327         }
 328
 329         /* The ordering for storing desc flags needs to be enforced. */
 330         rte_atomic_thread_fence(__ATOMIC_RELEASE);
 331
 332         for (i = 0; i < vq->shadow_used_idx; i++) {
 333                 uint16_t flags;
 334
 335                 if (vq->shadow_used_packed[i].len)
 336                         flags = VRING_DESC_F_WRITE;
 337                 else
 338                         flags = 0;
 339
 340                 if (vq->used_wrap_counter) {
 341                         flags |= VRING_DESC_F_USED;
 342                         flags |= VRING_DESC_F_AVAIL;
 343                 } else {
 344                         flags &= ~VRING_DESC_F_USED;
 345                         flags &= ~VRING_DESC_F_AVAIL;
 346                 }
 347
 348                 if (i > 0) {
 349                         vq->desc_packed[vq->last_used_idx].flags = flags;
 350
 351                         vhost_log_cache_used_vring(dev, vq,
 352                                         vq->last_used_idx *
 353                                         sizeof(struct vring_packed_desc),
 354                                         sizeof(struct vring_packed_desc));
 355                 } else {
 356                         head_idx = vq->last_used_idx;
 357                         head_flags = flags;
 358                 }
 359
 360                 vq_inc_last_used_packed(vq, vq->shadow_used_packed[i].count);
 361         }
 362
 363         vq->desc_packed[head_idx].flags = head_flags;
 364
 365         vhost_log_cache_used_vring(dev, vq,
 366                                 head_idx *
 367                                 sizeof(struct vring_packed_desc),
 368                                 sizeof(struct vring_packed_desc));
 369
 370         vq->shadow_used_idx = 0;
 371         vhost_log_cache_sync(dev, vq);
 372 }
 373
 374 static __rte_always_inline void
 375 vhost_flush_dequeue_shadow_packed(struct virtio_net *dev,
 376                                   struct vhost_virtqueue *vq)
 377 {
 378         struct vring_used_elem_packed *used_elem = &vq->shadow_used_packed[0];
 379
 380         vq->desc_packed[vq->shadow_last_used_idx].id = used_elem->id;
 381         /* desc flags is the synchronization point for virtio packed vring */
 382         __atomic_store_n(&vq->desc_packed[vq->shadow_last_used_idx].flags,
 383                          used_elem->flags, __ATOMIC_RELEASE);
 384
 385         vhost_log_cache_used_vring(dev, vq, vq->shadow_last_used_idx *
 386                                    sizeof(struct vring_packed_desc),
 387                                    sizeof(struct vring_packed_desc));
 388         vq->shadow_used_idx = 0;
 389         vhost_log_cache_sync(dev, vq);
 390 }
 391
 392 static __rte_always_inline void
 393 vhost_flush_enqueue_batch_packed(struct virtio_net *dev,
 394                                  struct vhost_virtqueue *vq,
 395                                  uint64_t *lens,
 396                                  uint16_t *ids)
 397 {
 398         uint16_t i;
 399         uint16_t flags;
 400         uint16_t last_used_idx;
 401         struct vring_packed_desc *desc_base;
 402
 403         last_used_idx = vq->last_used_idx;
 404         desc_base = &vq->desc_packed[last_used_idx];
 405
 406         flags = PACKED_DESC_ENQUEUE_USED_FLAG(vq->used_wrap_counter);
 407
 408         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
 409                 desc_base[i].id = ids[i];
 410                 desc_base[i].len = lens[i];
 411         }
 412
 413         rte_atomic_thread_fence(__ATOMIC_RELEASE);
 414
 415         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
 416                 desc_base[i].flags = flags;
 417         }
 418
 419         vhost_log_cache_used_vring(dev, vq, last_used_idx *
 420                                    sizeof(struct vring_packed_desc),
 421                                    sizeof(struct vring_packed_desc) *
 422                                    PACKED_BATCH_SIZE);
 423         vhost_log_cache_sync(dev, vq);
 424
 425         vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
 426 }
 427
 428 static __rte_always_inline void
 429 vhost_shadow_dequeue_batch_packed_inorder(struct vhost_virtqueue *vq,
 430                                           uint16_t id)
 431 {
 432         vq->shadow_used_packed[0].id = id;
 433
 434         if (!vq->shadow_used_idx) {
 435                 vq->shadow_last_used_idx = vq->last_used_idx;
 436                 vq->shadow_used_packed[0].flags =
 437                         PACKED_DESC_DEQUEUE_USED_FLAG(vq->used_wrap_counter);
 438                 vq->shadow_used_packed[0].len = 0;
 439                 vq->shadow_used_packed[0].count = 1;
 440                 vq->shadow_used_idx++;
 441         }
 442
 443         vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
 444 }
 445
 446 static __rte_always_inline void
 447 vhost_shadow_dequeue_batch_packed(struct virtio_net *dev,
 448                                   struct vhost_virtqueue *vq,
 449                                   uint16_t *ids)
 450 {
 451         uint16_t flags;
 452         uint16_t i;
 453         uint16_t begin;
 454
 455         flags = PACKED_DESC_DEQUEUE_USED_FLAG(vq->used_wrap_counter);
 456
 457         if (!vq->shadow_used_idx) {
 458                 vq->shadow_last_used_idx = vq->last_used_idx;
 459                 vq->shadow_used_packed[0].id  = ids[0];
 460                 vq->shadow_used_packed[0].len = 0;
 461                 vq->shadow_used_packed[0].count = 1;
 462                 vq->shadow_used_packed[0].flags = flags;
 463                 vq->shadow_used_idx++;
 464                 begin = 1;
 465         } else
 466                 begin = 0;
 467
 468         vhost_for_each_try_unroll(i, begin, PACKED_BATCH_SIZE) {
 469                 vq->desc_packed[vq->last_used_idx + i].id = ids[i];
 470                 vq->desc_packed[vq->last_used_idx + i].len = 0;
 471         }
 472
 473         rte_atomic_thread_fence(__ATOMIC_RELEASE);
 474         vhost_for_each_try_unroll(i, begin, PACKED_BATCH_SIZE)
 475                 vq->desc_packed[vq->last_used_idx + i].flags = flags;
 476
 477         vhost_log_cache_used_vring(dev, vq, vq->last_used_idx *
 478                                    sizeof(struct vring_packed_desc),
 479                                    sizeof(struct vring_packed_desc) *
 480                                    PACKED_BATCH_SIZE);
 481         vhost_log_cache_sync(dev, vq);
 482
 483         vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
 484 }
 485
 486 static __rte_always_inline void
 487 vhost_shadow_dequeue_single_packed(struct vhost_virtqueue *vq,
 488                                    uint16_t buf_id,
 489                                    uint16_t count)
 490 {
 491         uint16_t flags;
 492
 493         flags = vq->desc_packed[vq->last_used_idx].flags;
 494         if (vq->used_wrap_counter) {
 495                 flags |= VRING_DESC_F_USED;
 496                 flags |= VRING_DESC_F_AVAIL;
 497         } else {
 498                 flags &= ~VRING_DESC_F_USED;
 499                 flags &= ~VRING_DESC_F_AVAIL;
 500         }
 501
 502         if (!vq->shadow_used_idx) {
 503                 vq->shadow_last_used_idx = vq->last_used_idx;
 504
 505                 vq->shadow_used_packed[0].id  = buf_id;
 506                 vq->shadow_used_packed[0].len = 0;
 507                 vq->shadow_used_packed[0].flags = flags;
 508                 vq->shadow_used_idx++;
 509         } else {
 510                 vq->desc_packed[vq->last_used_idx].id = buf_id;
 511                 vq->desc_packed[vq->last_used_idx].len = 0;
 512                 vq->desc_packed[vq->last_used_idx].flags = flags;
 513         }
 514
 515         vq_inc_last_used_packed(vq, count);
 516 }
 517
 518 static __rte_always_inline void
 519 vhost_shadow_dequeue_single_packed_inorder(struct vhost_virtqueue *vq,
 520                                            uint16_t buf_id,
 521                                            uint16_t count)
 522 {
 523         uint16_t flags;
 524
 525         vq->shadow_used_packed[0].id = buf_id;
 526
 527         flags = vq->desc_packed[vq->last_used_idx].flags;
 528         if (vq->used_wrap_counter) {
 529                 flags |= VRING_DESC_F_USED;
 530                 flags |= VRING_DESC_F_AVAIL;
 531         } else {
 532                 flags &= ~VRING_DESC_F_USED;
 533                 flags &= ~VRING_DESC_F_AVAIL;
 534         }
 535
 536         if (!vq->shadow_used_idx) {
 537                 vq->shadow_last_used_idx = vq->last_used_idx;
 538                 vq->shadow_used_packed[0].len = 0;
 539                 vq->shadow_used_packed[0].flags = flags;
 540                 vq->shadow_used_idx++;
 541         }
 542
 543         vq_inc_last_used_packed(vq, count);
 544 }
 545
 546 static __rte_always_inline void
 547 vhost_shadow_enqueue_packed(struct vhost_virtqueue *vq,
 548                                    uint32_t *len,
 549                                    uint16_t *id,
 550                                    uint16_t *count,
 551                                    uint16_t num_buffers)
 552 {
 553         uint16_t i;
 554
 555         for (i = 0; i < num_buffers; i++) {
 556                 /* enqueue shadow flush action aligned with batch num */
 557                 if (!vq->shadow_used_idx)
 558                         vq->shadow_aligned_idx = vq->last_used_idx &
 559                                 PACKED_BATCH_MASK;
 560                 vq->shadow_used_packed[vq->shadow_used_idx].id  = id[i];
 561                 vq->shadow_used_packed[vq->shadow_used_idx].len = len[i];
 562                 vq->shadow_used_packed[vq->shadow_used_idx].count = count[i];
 563                 vq->shadow_aligned_idx += count[i];
 564                 vq->shadow_used_idx++;
 565         }
 566 }
 567
 568 static __rte_always_inline void
 569 vhost_shadow_enqueue_single_packed(struct virtio_net *dev,
 570                                    struct vhost_virtqueue *vq,
 571                                    uint32_t *len,
 572                                    uint16_t *id,
 573                                    uint16_t *count,
 574                                    uint16_t num_buffers)
 575 {
 576         vhost_shadow_enqueue_packed(vq, len, id, count, num_buffers);
 577
 578         if (vq->shadow_aligned_idx >= PACKED_BATCH_SIZE) {
 579                 do_data_copy_enqueue(dev, vq);
 580                 vhost_flush_enqueue_shadow_packed(dev, vq);
 581         }
 582 }
 583
 584 /* avoid write operation when necessary, to lessen cache issues */
 585 #define ASSIGN_UNLESS_EQUAL(var, val) do {      \
 586         if ((var) != (val))                     \
 587                 (var) = (val);                  \
 588 } while (0)
 589
 590 static __rte_always_inline void
 591 virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
 592 {
 593         uint64_t csum_l4 = m_buf->ol_flags & RTE_MBUF_F_TX_L4_MASK;
 594
 595         if (m_buf->ol_flags & RTE_MBUF_F_TX_TCP_SEG)
 596                 csum_l4 |= RTE_MBUF_F_TX_TCP_CKSUM;
 597
 598         if (csum_l4) {
 599                 net_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 600                 net_hdr->csum_start = m_buf->l2_len + m_buf->l3_len;
 601
 602                 switch (csum_l4) {
 603                 case RTE_MBUF_F_TX_TCP_CKSUM:
 604                         net_hdr->csum_offset = (offsetof(struct rte_tcp_hdr,
 605                                                 cksum));
 606                         break;
 607                 case RTE_MBUF_F_TX_UDP_CKSUM:
 608                         net_hdr->csum_offset = (offsetof(struct rte_udp_hdr,
 609                                                 dgram_cksum));
 610                         break;
 611                 case RTE_MBUF_F_TX_SCTP_CKSUM:
 612                         net_hdr->csum_offset = (offsetof(struct rte_sctp_hdr,
 613                                                 cksum));
 614                         break;
 615                 }
 616         } else {
 617                 ASSIGN_UNLESS_EQUAL(net_hdr->csum_start, 0);
 618                 ASSIGN_UNLESS_EQUAL(net_hdr->csum_offset, 0);
 619                 ASSIGN_UNLESS_EQUAL(net_hdr->flags, 0);
 620         }
 621
 622         /* IP cksum verification cannot be bypassed, then calculate here */
 623         if (m_buf->ol_flags & RTE_MBUF_F_TX_IP_CKSUM) {
 624                 struct rte_ipv4_hdr *ipv4_hdr;
 625
 626                 ipv4_hdr = rte_pktmbuf_mtod_offset(m_buf, struct rte_ipv4_hdr *,
 627                                                    m_buf->l2_len);
 628                 ipv4_hdr->hdr_checksum = 0;
 629                 ipv4_hdr->hdr_checksum = rte_ipv4_cksum(ipv4_hdr);
 630         }
 631
 632         if (m_buf->ol_flags & RTE_MBUF_F_TX_TCP_SEG) {
 633                 if (m_buf->ol_flags & RTE_MBUF_F_TX_IPV4)
 634                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 635                 else
 636                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 637                 net_hdr->gso_size = m_buf->tso_segsz;
 638                 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len
 639                                         + m_buf->l4_len;
 640         } else if (m_buf->ol_flags & RTE_MBUF_F_TX_UDP_SEG) {
 641                 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
 642                 net_hdr->gso_size = m_buf->tso_segsz;
 643                 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len +
 644                         m_buf->l4_len;
 645         } else {
 646                 ASSIGN_UNLESS_EQUAL(net_hdr->gso_type, 0);
 647                 ASSIGN_UNLESS_EQUAL(net_hdr->gso_size, 0);
 648                 ASSIGN_UNLESS_EQUAL(net_hdr->hdr_len, 0);
 649         }
 650 }
 651
 652 static __rte_always_inline int
 653 map_one_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 654                 struct buf_vector *buf_vec, uint16_t *vec_idx,
 655                 uint64_t desc_iova, uint64_t desc_len, uint8_t perm)
 656 {
 657         uint16_t vec_id = *vec_idx;
 658
 659         while (desc_len) {
 660                 uint64_t desc_addr;
 661                 uint64_t desc_chunck_len = desc_len;
 662
 663                 if (unlikely(vec_id >= BUF_VECTOR_MAX))
 664                         return -1;
 665
 666                 desc_addr = vhost_iova_to_vva(dev, vq,
 667                                 desc_iova,
 668                                 &desc_chunck_len,
 669                                 perm);
 670                 if (unlikely(!desc_addr))
 671                         return -1;
 672
 673                 rte_prefetch0((void *)(uintptr_t)desc_addr);
 674
 675                 buf_vec[vec_id].buf_iova = desc_iova;
 676                 buf_vec[vec_id].buf_addr = desc_addr;
 677                 buf_vec[vec_id].buf_len  = desc_chunck_len;
 678
 679                 desc_len -= desc_chunck_len;
 680                 desc_iova += desc_chunck_len;
 681                 vec_id++;
 682         }
 683         *vec_idx = vec_id;
 684
 685         return 0;
 686 }
 687
 688 static __rte_always_inline int
 689 fill_vec_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 690                          uint32_t avail_idx, uint16_t *vec_idx,
 691                          struct buf_vector *buf_vec, uint16_t *desc_chain_head,
 692                          uint32_t *desc_chain_len, uint8_t perm)
 693 {
 694         uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
 695         uint16_t vec_id = *vec_idx;
 696         uint32_t len    = 0;
 697         uint64_t dlen;
 698         uint32_t nr_descs = vq->size;
 699         uint32_t cnt    = 0;
 700         struct vring_desc *descs = vq->desc;
 701         struct vring_desc *idesc = NULL;
 702
 703         if (unlikely(idx >= vq->size))
 704                 return -1;
 705
 706         *desc_chain_head = idx;
 707
 708         if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
 709                 dlen = vq->desc[idx].len;
 710                 nr_descs = dlen / sizeof(struct vring_desc);
 711                 if (unlikely(nr_descs > vq->size))
 712                         return -1;
 713
 714                 descs = (struct vring_desc *)(uintptr_t)
 715                         vhost_iova_to_vva(dev, vq, vq->desc[idx].addr,
 716                                                 &dlen,
 717                                                 VHOST_ACCESS_RO);
 718                 if (unlikely(!descs))
 719                         return -1;
 720
 721                 if (unlikely(dlen < vq->desc[idx].len)) {
 722                         /*
 723                          * The indirect desc table is not contiguous
 724                          * in process VA space, we have to copy it.
 725                          */
 726                         idesc = vhost_alloc_copy_ind_table(dev, vq,
 727                                         vq->desc[idx].addr, vq->desc[idx].len);
 728                         if (unlikely(!idesc))
 729                                 return -1;
 730
 731                         descs = idesc;
 732                 }
 733
 734                 idx = 0;
 735         }
 736
 737         while (1) {
 738                 if (unlikely(idx >= nr_descs || cnt++ >= nr_descs)) {
 739                         free_ind_table(idesc);
 740                         return -1;
 741                 }
 742
 743                 dlen = descs[idx].len;
 744                 len += dlen;
 745
 746                 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 747                                                 descs[idx].addr, dlen,
 748                                                 perm))) {
 749                         free_ind_table(idesc);
 750                         return -1;
 751                 }
 752
 753                 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
 754                         break;
 755
 756                 idx = descs[idx].next;
 757         }
 758
 759         *desc_chain_len = len;
 760         *vec_idx = vec_id;
 761
 762         if (unlikely(!!idesc))
 763                 free_ind_table(idesc);
 764
 765         return 0;
 766 }
 767
 768 /*
 769  * Returns -1 on fail, 0 on success
 770  */
 771 static inline int
 772 reserve_avail_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 773                                 uint32_t size, struct buf_vector *buf_vec,
 774                                 uint16_t *num_buffers, uint16_t avail_head,
 775                                 uint16_t *nr_vec)
 776 {
 777         uint16_t cur_idx;
 778         uint16_t vec_idx = 0;
 779         uint16_t max_tries, tries = 0;
 780
 781         uint16_t head_idx = 0;
 782         uint32_t len = 0;
 783
 784         *num_buffers = 0;
 785         cur_idx  = vq->last_avail_idx;
 786
 787         if (rxvq_is_mergeable(dev))
 788                 max_tries = vq->size - 1;
 789         else
 790                 max_tries = 1;
 791
 792         while (size > 0) {
 793                 if (unlikely(cur_idx == avail_head))
 794                         return -1;
 795                 /*
 796                  * if we tried all available ring items, and still
 797                  * can't get enough buf, it means something abnormal
 798                  * happened.
 799                  */
 800                 if (unlikely(++tries > max_tries))
 801                         return -1;
 802
 803                 if (unlikely(fill_vec_buf_split(dev, vq, cur_idx,
 804                                                 &vec_idx, buf_vec,
 805                                                 &head_idx, &len,
 806                                                 VHOST_ACCESS_RW) < 0))
 807                         return -1;
 808                 len = RTE_MIN(len, size);
 809                 update_shadow_used_ring_split(vq, head_idx, len);
 810                 size -= len;
 811
 812                 cur_idx++;
 813                 *num_buffers += 1;
 814         }
 815
 816         *nr_vec = vec_idx;
 817
 818         return 0;
 819 }
 820
 821 static __rte_always_inline int
 822 fill_vec_buf_packed_indirect(struct virtio_net *dev,
 823                         struct vhost_virtqueue *vq,
 824                         struct vring_packed_desc *desc, uint16_t *vec_idx,
 825                         struct buf_vector *buf_vec, uint32_t *len, uint8_t perm)
 826 {
 827         uint16_t i;
 828         uint32_t nr_descs;
 829         uint16_t vec_id = *vec_idx;
 830         uint64_t dlen;
 831         struct vring_packed_desc *descs, *idescs = NULL;
 832
 833         dlen = desc->len;
 834         descs = (struct vring_packed_desc *)(uintptr_t)
 835                 vhost_iova_to_vva(dev, vq, desc->addr, &dlen, VHOST_ACCESS_RO);
 836         if (unlikely(!descs))
 837                 return -1;
 838
 839         if (unlikely(dlen < desc->len)) {
 840                 /*
 841                  * The indirect desc table is not contiguous
 842                  * in process VA space, we have to copy it.
 843                  */
 844                 idescs = vhost_alloc_copy_ind_table(dev,
 845                                 vq, desc->addr, desc->len);
 846                 if (unlikely(!idescs))
 847                         return -1;
 848
 849                 descs = idescs;
 850         }
 851
 852         nr_descs =  desc->len / sizeof(struct vring_packed_desc);
 853         if (unlikely(nr_descs >= vq->size)) {
 854                 free_ind_table(idescs);
 855                 return -1;
 856         }
 857
 858         for (i = 0; i < nr_descs; i++) {
 859                 if (unlikely(vec_id >= BUF_VECTOR_MAX)) {
 860                         free_ind_table(idescs);
 861                         return -1;
 862                 }
 863
 864                 dlen = descs[i].len;
 865                 *len += dlen;
 866                 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 867                                                 descs[i].addr, dlen,
 868                                                 perm)))
 869                         return -1;
 870         }
 871         *vec_idx = vec_id;
 872
 873         if (unlikely(!!idescs))
 874                 free_ind_table(idescs);
 875
 876         return 0;
 877 }
 878
 879 static __rte_always_inline int
 880 fill_vec_buf_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
 881                                 uint16_t avail_idx, uint16_t *desc_count,
 882                                 struct buf_vector *buf_vec, uint16_t *vec_idx,
 883                                 uint16_t *buf_id, uint32_t *len, uint8_t perm)
 884 {
 885         bool wrap_counter = vq->avail_wrap_counter;
 886         struct vring_packed_desc *descs = vq->desc_packed;
 887         uint16_t vec_id = *vec_idx;
 888         uint64_t dlen;
 889
 890         if (avail_idx < vq->last_avail_idx)
 891                 wrap_counter ^= 1;
 892
 893         /*
 894          * Perform a load-acquire barrier in desc_is_avail to
 895          * enforce the ordering between desc flags and desc
 896          * content.
 897          */
 898         if (unlikely(!desc_is_avail(&descs[avail_idx], wrap_counter)))
 899                 return -1;
 900
 901         *desc_count = 0;
 902         *len = 0;
 903
 904         while (1) {
 905                 if (unlikely(vec_id >= BUF_VECTOR_MAX))
 906                         return -1;
 907
 908                 if (unlikely(*desc_count >= vq->size))
 909                         return -1;
 910
 911                 *desc_count += 1;
 912                 *buf_id = descs[avail_idx].id;
 913
 914                 if (descs[avail_idx].flags & VRING_DESC_F_INDIRECT) {
 915                         if (unlikely(fill_vec_buf_packed_indirect(dev, vq,
 916                                                         &descs[avail_idx],
 917                                                         &vec_id, buf_vec,
 918                                                         len, perm) < 0))
 919                                 return -1;
 920                 } else {
 921                         dlen = descs[avail_idx].len;
 922                         *len += dlen;
 923
 924                         if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
 925                                                         descs[avail_idx].addr,
 926                                                         dlen,
 927                                                         perm)))
 928                                 return -1;
 929                 }
 930
 931                 if ((descs[avail_idx].flags & VRING_DESC_F_NEXT) == 0)
 932                         break;
 933
 934                 if (++avail_idx >= vq->size) {
 935                         avail_idx -= vq->size;
 936                         wrap_counter ^= 1;
 937                 }
 938         }
 939
 940         *vec_idx = vec_id;
 941
 942         return 0;
 943 }
 944
 945 static __rte_noinline void
 946 copy_vnet_hdr_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 947                 struct buf_vector *buf_vec,
 948                 struct virtio_net_hdr_mrg_rxbuf *hdr)
 949 {
 950         uint64_t len;
 951         uint64_t remain = dev->vhost_hlen;
 952         uint64_t src = (uint64_t)(uintptr_t)hdr, dst;
 953         uint64_t iova = buf_vec->buf_iova;
 954
 955         while (remain) {
 956                 len = RTE_MIN(remain,
 957                                 buf_vec->buf_len);
 958                 dst = buf_vec->buf_addr;
 959                 rte_memcpy((void *)(uintptr_t)dst,
 960                                 (void *)(uintptr_t)src,
 961                                 len);
 962
 963                 PRINT_PACKET(dev, (uintptr_t)dst,
 964                                 (uint32_t)len, 0);
 965                 vhost_log_cache_write_iova(dev, vq,
 966                                 iova, len);
 967
 968                 remain -= len;
 969                 iova += len;
 970                 src += len;
 971                 buf_vec++;
 972         }
 973 }
 974
 975 static __rte_always_inline int
 976 async_iter_initialize(struct virtio_net *dev, struct vhost_async *async)
 977 {
 978         struct vhost_iov_iter *iter;
 979
 980         if (unlikely(async->iovec_idx >= VHOST_MAX_ASYNC_VEC)) {
 981                 VHOST_LOG_DATA(ERR, "(%s) no more async iovec available\n", dev->ifname);
 982                 return -1;
 983         }
 984
 985         iter = async->iov_iter + async->iter_idx;
 986         iter->iov = async->iovec + async->iovec_idx;
 987         iter->nr_segs = 0;
 988
 989         return 0;
 990 }
 991
 992 static __rte_always_inline int
 993 async_iter_add_iovec(struct virtio_net *dev, struct vhost_async *async,
 994                 void *src, void *dst, size_t len)
 995 {
 996         struct vhost_iov_iter *iter;
 997         struct vhost_iovec *iovec;
 998
 999         if (unlikely(async->iovec_idx >= VHOST_MAX_ASYNC_VEC)) {
1000                 static bool vhost_max_async_vec_log;
1001
1002                 if (!vhost_max_async_vec_log) {
1003                         VHOST_LOG_DATA(ERR, "(%s) no more async iovec available\n", dev->ifname);
1004                         vhost_max_async_vec_log = true;
1005                 }
1006
1007                 return -1;
1008         }
1009
1010         iter = async->iov_iter + async->iter_idx;
1011         iovec = async->iovec + async->iovec_idx;
1012
1013         iovec->src_addr = src;
1014         iovec->dst_addr = dst;
1015         iovec->len = len;
1016
1017         iter->nr_segs++;
1018         async->iovec_idx++;
1019
1020         return 0;
1021 }
1022
1023 static __rte_always_inline void
1024 async_iter_finalize(struct vhost_async *async)
1025 {
1026         async->iter_idx++;
1027 }
1028
1029 static __rte_always_inline void
1030 async_iter_cancel(struct vhost_async *async)
1031 {
1032         struct vhost_iov_iter *iter;
1033
1034         iter = async->iov_iter + async->iter_idx;
1035         async->iovec_idx -= iter->nr_segs;
1036         iter->nr_segs = 0;
1037         iter->iov = NULL;
1038 }
1039
1040 static __rte_always_inline void
1041 async_iter_reset(struct vhost_async *async)
1042 {
1043         async->iter_idx = 0;
1044         async->iovec_idx = 0;
1045 }
1046
1047 static __rte_always_inline int
1048 async_fill_seg(struct virtio_net *dev, struct vhost_virtqueue *vq,
1049                 struct rte_mbuf *m, uint32_t mbuf_offset,
1050                 uint64_t buf_iova, uint32_t cpy_len, bool to_desc)
1051 {
1052         struct vhost_async *async = vq->async;
1053         uint64_t mapped_len;
1054         uint32_t buf_offset = 0;
1055         void *src, *dst;
1056         void *host_iova;
1057
1058         while (cpy_len) {
1059                 host_iova = (void *)(uintptr_t)gpa_to_first_hpa(dev,
1060                                 buf_iova + buf_offset, cpy_len, &mapped_len);
1061                 if (unlikely(!host_iova)) {
1062                         VHOST_LOG_DATA(ERR, "(%s) %s: failed to get host iova.\n",
1063                                        dev->ifname, __func__);
1064                         return -1;
1065                 }
1066
1067                 if (to_desc) {
1068                         src = (void *)(uintptr_t)rte_pktmbuf_iova_offset(m, mbuf_offset);
1069                         dst = host_iova;
1070                 } else {
1071                         src = host_iova;
1072                         dst = (void *)(uintptr_t)rte_pktmbuf_iova_offset(m, mbuf_offset);
1073                 }
1074
1075                 if (unlikely(async_iter_add_iovec(dev, async, src, dst, (size_t)mapped_len)))
1076                         return -1;
1077
1078                 cpy_len -= (uint32_t)mapped_len;
1079                 mbuf_offset += (uint32_t)mapped_len;
1080                 buf_offset += (uint32_t)mapped_len;
1081         }
1082
1083         return 0;
1084 }
1085
1086 static __rte_always_inline void
1087 sync_fill_seg(struct virtio_net *dev, struct vhost_virtqueue *vq,
1088                 struct rte_mbuf *m, uint32_t mbuf_offset,
1089                 uint64_t buf_addr, uint64_t buf_iova, uint32_t cpy_len, bool to_desc)
1090 {
1091         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
1092
1093         if (likely(cpy_len > MAX_BATCH_LEN || vq->batch_copy_nb_elems >= vq->size)) {
1094                 if (to_desc) {
1095                         rte_memcpy((void *)((uintptr_t)(buf_addr)),
1096                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
1097                                 cpy_len);
1098                 } else {
1099                         rte_memcpy(rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
1100                                 (void *)((uintptr_t)(buf_addr)),
1101                                 cpy_len);
1102                 }
1103                 vhost_log_cache_write_iova(dev, vq, buf_iova, cpy_len);
1104                 PRINT_PACKET(dev, (uintptr_t)(buf_addr), cpy_len, 0);
1105         } else {
1106                 if (to_desc) {
1107                         batch_copy[vq->batch_copy_nb_elems].dst =
1108                                 (void *)((uintptr_t)(buf_addr));
1109                         batch_copy[vq->batch_copy_nb_elems].src =
1110                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
1111                 } else {
1112                         batch_copy[vq->batch_copy_nb_elems].dst =
1113                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
1114                         batch_copy[vq->batch_copy_nb_elems].src =
1115                                 (void *)((uintptr_t)(buf_addr));
1116                 }
1117                 batch_copy[vq->batch_copy_nb_elems].log_addr = buf_iova;
1118                 batch_copy[vq->batch_copy_nb_elems].len = cpy_len;
1119                 vq->batch_copy_nb_elems++;
1120         }
1121 }
1122
1123 static __rte_always_inline int
1124 mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
1125                 struct rte_mbuf *m, struct buf_vector *buf_vec,
1126                 uint16_t nr_vec, uint16_t num_buffers, bool is_async)
1127 {
1128         uint32_t vec_idx = 0;
1129         uint32_t mbuf_offset, mbuf_avail;
1130         uint32_t buf_offset, buf_avail;
1131         uint64_t buf_addr, buf_iova, buf_len;
1132         uint32_t cpy_len;
1133         uint64_t hdr_addr;
1134         struct rte_mbuf *hdr_mbuf;
1135         struct virtio_net_hdr_mrg_rxbuf tmp_hdr, *hdr = NULL;
1136         struct vhost_async *async = vq->async;
1137
1138         if (unlikely(m == NULL))
1139                 return -1;
1140
1141         buf_addr = buf_vec[vec_idx].buf_addr;
1142         buf_iova = buf_vec[vec_idx].buf_iova;
1143         buf_len = buf_vec[vec_idx].buf_len;
1144
1145         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1))
1146                 return -1;
1147
1148         hdr_mbuf = m;
1149         hdr_addr = buf_addr;
1150         if (unlikely(buf_len < dev->vhost_hlen)) {
1151                 memset(&tmp_hdr, 0, sizeof(struct virtio_net_hdr_mrg_rxbuf));
1152                 hdr = &tmp_hdr;
1153         } else
1154                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)hdr_addr;
1155
1156         VHOST_LOG_DATA(DEBUG, "(%s) RX: num merge buffers %d\n",
1157                 dev->ifname, num_buffers);
1158
1159         if (unlikely(buf_len < dev->vhost_hlen)) {
1160                 buf_offset = dev->vhost_hlen - buf_len;
1161                 vec_idx++;
1162                 buf_addr = buf_vec[vec_idx].buf_addr;
1163                 buf_iova = buf_vec[vec_idx].buf_iova;
1164                 buf_len = buf_vec[vec_idx].buf_len;
1165                 buf_avail = buf_len - buf_offset;
1166         } else {
1167                 buf_offset = dev->vhost_hlen;
1168                 buf_avail = buf_len - dev->vhost_hlen;
1169         }
1170
1171         mbuf_avail  = rte_pktmbuf_data_len(m);
1172         mbuf_offset = 0;
1173
1174         if (is_async) {
1175                 if (async_iter_initialize(dev, async))
1176                         return -1;
1177         }
1178
1179         while (mbuf_avail != 0 || m->next != NULL) {
1180                 /* done with current buf, get the next one */
1181                 if (buf_avail == 0) {
1182                         vec_idx++;
1183                         if (unlikely(vec_idx >= nr_vec))
1184                                 goto error;
1185
1186                         buf_addr = buf_vec[vec_idx].buf_addr;
1187                         buf_iova = buf_vec[vec_idx].buf_iova;
1188                         buf_len = buf_vec[vec_idx].buf_len;
1189
1190                         buf_offset = 0;
1191                         buf_avail  = buf_len;
1192                 }
1193
1194                 /* done with current mbuf, get the next one */
1195                 if (mbuf_avail == 0) {
1196                         m = m->next;
1197
1198                         mbuf_offset = 0;
1199                         mbuf_avail  = rte_pktmbuf_data_len(m);
1200                 }
1201
1202                 if (hdr_addr) {
1203                         virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
1204                         if (rxvq_is_mergeable(dev))
1205                                 ASSIGN_UNLESS_EQUAL(hdr->num_buffers,
1206                                                 num_buffers);
1207
1208                         if (unlikely(hdr == &tmp_hdr)) {
1209                                 copy_vnet_hdr_to_desc(dev, vq, buf_vec, hdr);
1210                         } else {
1211                                 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
1212                                                 dev->vhost_hlen, 0);
1213                                 vhost_log_cache_write_iova(dev, vq,
1214                                                 buf_vec[0].buf_iova,
1215                                                 dev->vhost_hlen);
1216                         }
1217
1218                         hdr_addr = 0;
1219                 }
1220
1221                 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
1222
1223                 if (is_async) {
1224                         if (async_fill_seg(dev, vq, m, mbuf_offset,
1225                                            buf_iova + buf_offset, cpy_len, true) < 0)
1226                                 goto error;
1227                 } else {
1228                         sync_fill_seg(dev, vq, m, mbuf_offset,
1229                                       buf_addr + buf_offset,
1230                                       buf_iova + buf_offset, cpy_len, true);
1231                 }
1232
1233                 mbuf_avail  -= cpy_len;
1234                 mbuf_offset += cpy_len;
1235                 buf_avail  -= cpy_len;
1236                 buf_offset += cpy_len;
1237         }
1238
1239         if (is_async)
1240                 async_iter_finalize(async);
1241
1242         return 0;
1243 error:
1244         if (is_async)
1245                 async_iter_cancel(async);
1246
1247         return -1;
1248 }
1249
1250 static __rte_always_inline int
1251 vhost_enqueue_single_packed(struct virtio_net *dev,
1252                             struct vhost_virtqueue *vq,
1253                             struct rte_mbuf *pkt,
1254                             struct buf_vector *buf_vec,
1255                             uint16_t *nr_descs)
1256 {
1257         uint16_t nr_vec = 0;
1258         uint16_t avail_idx = vq->last_avail_idx;
1259         uint16_t max_tries, tries = 0;
1260         uint16_t buf_id = 0;
1261         uint32_t len = 0;
1262         uint16_t desc_count;
1263         uint32_t size = pkt->pkt_len + sizeof(struct virtio_net_hdr_mrg_rxbuf);
1264         uint16_t num_buffers = 0;
1265         uint32_t buffer_len[vq->size];
1266         uint16_t buffer_buf_id[vq->size];
1267         uint16_t buffer_desc_count[vq->size];
1268
1269         if (rxvq_is_mergeable(dev))
1270                 max_tries = vq->size - 1;
1271         else
1272                 max_tries = 1;
1273
1274         while (size > 0) {
1275                 /*
1276                  * if we tried all available ring items, and still
1277                  * can't get enough buf, it means something abnormal
1278                  * happened.
1279                  */
1280                 if (unlikely(++tries > max_tries))
1281                         return -1;
1282
1283                 if (unlikely(fill_vec_buf_packed(dev, vq,
1284                                                 avail_idx, &desc_count,
1285                                                 buf_vec, &nr_vec,
1286                                                 &buf_id, &len,
1287                                                 VHOST_ACCESS_RW) < 0))
1288                         return -1;
1289
1290                 len = RTE_MIN(len, size);
1291                 size -= len;
1292
1293                 buffer_len[num_buffers] = len;
1294                 buffer_buf_id[num_buffers] = buf_id;
1295                 buffer_desc_count[num_buffers] = desc_count;
1296                 num_buffers += 1;
1297
1298                 *nr_descs += desc_count;
1299                 avail_idx += desc_count;
1300                 if (avail_idx >= vq->size)
1301                         avail_idx -= vq->size;
1302         }
1303
1304         if (mbuf_to_desc(dev, vq, pkt, buf_vec, nr_vec, num_buffers, false) < 0)
1305                 return -1;
1306
1307         vhost_shadow_enqueue_single_packed(dev, vq, buffer_len, buffer_buf_id,
1308                                            buffer_desc_count, num_buffers);
1309
1310         return 0;
1311 }
1312
1313 static __rte_noinline uint32_t
1314 virtio_dev_rx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
1315         struct rte_mbuf **pkts, uint32_t count)
1316 {
1317         uint32_t pkt_idx = 0;
1318         uint16_t num_buffers;
1319         struct buf_vector buf_vec[BUF_VECTOR_MAX];
1320         uint16_t avail_head;
1321
1322         /*
1323          * The ordering between avail index and
1324          * desc reads needs to be enforced.
1325          */
1326         avail_head = __atomic_load_n(&vq->avail->idx, __ATOMIC_ACQUIRE);
1327
1328         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
1329
1330         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
1331                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
1332                 uint16_t nr_vec = 0;
1333
1334                 if (unlikely(reserve_avail_buf_split(dev, vq,
1335                                                 pkt_len, buf_vec, &num_buffers,
1336                                                 avail_head, &nr_vec) < 0)) {
1337                         VHOST_LOG_DATA(DEBUG,
1338                                 "(%s) failed to get enough desc from vring\n",
1339                                 dev->ifname);
1340                         vq->shadow_used_idx -= num_buffers;
1341                         break;
1342                 }
1343
1344                 VHOST_LOG_DATA(DEBUG, "(%s) current index %d | end index %d\n",
1345                         dev->ifname, vq->last_avail_idx,
1346                         vq->last_avail_idx + num_buffers);
1347
1348                 if (mbuf_to_desc(dev, vq, pkts[pkt_idx], buf_vec, nr_vec,
1349                                         num_buffers, false) < 0) {
1350                         vq->shadow_used_idx -= num_buffers;
1351                         break;
1352                 }
1353
1354                 vq->last_avail_idx += num_buffers;
1355         }
1356
1357         do_data_copy_enqueue(dev, vq);
1358
1359         if (likely(vq->shadow_used_idx)) {
1360                 flush_shadow_used_ring_split(dev, vq);
1361                 vhost_vring_call_split(dev, vq);
1362         }
1363
1364         return pkt_idx;
1365 }
1366
1367 static __rte_always_inline int
1368 virtio_dev_rx_sync_batch_check(struct virtio_net *dev,
1369                            struct vhost_virtqueue *vq,
1370                            struct rte_mbuf **pkts,
1371                            uint64_t *desc_addrs,
1372                            uint64_t *lens)
1373 {
1374         bool wrap_counter = vq->avail_wrap_counter;
1375         struct vring_packed_desc *descs = vq->desc_packed;
1376         uint16_t avail_idx = vq->last_avail_idx;
1377         uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1378         uint16_t i;
1379
1380         if (unlikely(avail_idx & PACKED_BATCH_MASK))
1381                 return -1;
1382
1383         if (unlikely((avail_idx + PACKED_BATCH_SIZE) > vq->size))
1384                 return -1;
1385
1386         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1387                 if (unlikely(pkts[i]->next != NULL))
1388                         return -1;
1389                 if (unlikely(!desc_is_avail(&descs[avail_idx + i],
1390                                             wrap_counter)))
1391                         return -1;
1392         }
1393
1394         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1395                 lens[i] = descs[avail_idx + i].len;
1396
1397         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1398                 if (unlikely(pkts[i]->pkt_len > (lens[i] - buf_offset)))
1399                         return -1;
1400         }
1401
1402         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1403                 desc_addrs[i] = vhost_iova_to_vva(dev, vq,
1404                                                   descs[avail_idx + i].addr,
1405                                                   &lens[i],
1406                                                   VHOST_ACCESS_RW);
1407
1408         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1409                 if (unlikely(!desc_addrs[i]))
1410                         return -1;
1411                 if (unlikely(lens[i] != descs[avail_idx + i].len))
1412                         return -1;
1413         }
1414
1415         return 0;
1416 }
1417
1418 static __rte_always_inline void
1419 virtio_dev_rx_batch_packed_copy(struct virtio_net *dev,
1420                            struct vhost_virtqueue *vq,
1421                            struct rte_mbuf **pkts,
1422                            uint64_t *desc_addrs,
1423                            uint64_t *lens)
1424 {
1425         uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1426         struct virtio_net_hdr_mrg_rxbuf *hdrs[PACKED_BATCH_SIZE];
1427         struct vring_packed_desc *descs = vq->desc_packed;
1428         uint16_t avail_idx = vq->last_avail_idx;
1429         uint16_t ids[PACKED_BATCH_SIZE];
1430         uint16_t i;
1431
1432         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1433                 rte_prefetch0((void *)(uintptr_t)desc_addrs[i]);
1434                 hdrs[i] = (struct virtio_net_hdr_mrg_rxbuf *)
1435                                         (uintptr_t)desc_addrs[i];
1436                 lens[i] = pkts[i]->pkt_len +
1437                         sizeof(struct virtio_net_hdr_mrg_rxbuf);
1438         }
1439
1440         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1441                 virtio_enqueue_offload(pkts[i], &hdrs[i]->hdr);
1442
1443         vq_inc_last_avail_packed(vq, PACKED_BATCH_SIZE);
1444
1445         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1446                 rte_memcpy((void *)(uintptr_t)(desc_addrs[i] + buf_offset),
1447                            rte_pktmbuf_mtod_offset(pkts[i], void *, 0),
1448                            pkts[i]->pkt_len);
1449         }
1450
1451         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1452                 vhost_log_cache_write_iova(dev, vq, descs[avail_idx + i].addr,
1453                                            lens[i]);
1454
1455         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1456                 ids[i] = descs[avail_idx + i].id;
1457
1458         vhost_flush_enqueue_batch_packed(dev, vq, lens, ids);
1459 }
1460
1461 static __rte_always_inline int
1462 virtio_dev_rx_sync_batch_packed(struct virtio_net *dev,
1463                            struct vhost_virtqueue *vq,
1464                            struct rte_mbuf **pkts)
1465 {
1466         uint64_t desc_addrs[PACKED_BATCH_SIZE];
1467         uint64_t lens[PACKED_BATCH_SIZE];
1468
1469         if (virtio_dev_rx_sync_batch_check(dev, vq, pkts, desc_addrs, lens) == -1)
1470                 return -1;
1471
1472         if (vq->shadow_used_idx) {
1473                 do_data_copy_enqueue(dev, vq);
1474                 vhost_flush_enqueue_shadow_packed(dev, vq);
1475         }
1476
1477         virtio_dev_rx_batch_packed_copy(dev, vq, pkts, desc_addrs, lens);
1478
1479         return 0;
1480 }
1481
1482 static __rte_always_inline int16_t
1483 virtio_dev_rx_single_packed(struct virtio_net *dev,
1484                             struct vhost_virtqueue *vq,
1485                             struct rte_mbuf *pkt)
1486 {
1487         struct buf_vector buf_vec[BUF_VECTOR_MAX];
1488         uint16_t nr_descs = 0;
1489
1490         if (unlikely(vhost_enqueue_single_packed(dev, vq, pkt, buf_vec,
1491                                                  &nr_descs) < 0)) {
1492                 VHOST_LOG_DATA(DEBUG, "(%s) failed to get enough desc from vring\n",
1493                                 dev->ifname);
1494                 return -1;
1495         }
1496
1497         VHOST_LOG_DATA(DEBUG, "(%s) current index %d | end index %d\n",
1498                         dev->ifname, vq->last_avail_idx,
1499                         vq->last_avail_idx + nr_descs);
1500
1501         vq_inc_last_avail_packed(vq, nr_descs);
1502
1503         return 0;
1504 }
1505
1506 static __rte_noinline uint32_t
1507 virtio_dev_rx_packed(struct virtio_net *dev,
1508                      struct vhost_virtqueue *__rte_restrict vq,
1509                      struct rte_mbuf **__rte_restrict pkts,
1510                      uint32_t count)
1511 {
1512         uint32_t pkt_idx = 0;
1513
1514         do {
1515                 rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
1516
1517                 if (count - pkt_idx >= PACKED_BATCH_SIZE) {
1518                         if (!virtio_dev_rx_sync_batch_packed(dev, vq,
1519                                                         &pkts[pkt_idx])) {
1520                                 pkt_idx += PACKED_BATCH_SIZE;
1521                                 continue;
1522                         }
1523                 }
1524
1525                 if (virtio_dev_rx_single_packed(dev, vq, pkts[pkt_idx]))
1526                         break;
1527                 pkt_idx++;
1528
1529         } while (pkt_idx < count);
1530
1531         if (vq->shadow_used_idx) {
1532                 do_data_copy_enqueue(dev, vq);
1533                 vhost_flush_enqueue_shadow_packed(dev, vq);
1534         }
1535
1536         if (pkt_idx)
1537                 vhost_vring_call_packed(dev, vq);
1538
1539         return pkt_idx;
1540 }
1541
1542 static __rte_always_inline uint32_t
1543 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
1544         struct rte_mbuf **pkts, uint32_t count)
1545 {
1546         struct vhost_virtqueue *vq;
1547         uint32_t nb_tx = 0;
1548
1549         VHOST_LOG_DATA(DEBUG, "(%s) %s\n", dev->ifname, __func__);
1550         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
1551                 VHOST_LOG_DATA(ERR, "(%s) %s: invalid virtqueue idx %d.\n",
1552                         dev->ifname, __func__, queue_id);
1553                 return 0;
1554         }
1555
1556         vq = dev->virtqueue[queue_id];
1557
1558         rte_spinlock_lock(&vq->access_lock);
1559
1560         if (unlikely(!vq->enabled))
1561                 goto out_access_unlock;
1562
1563         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1564                 vhost_user_iotlb_rd_lock(vq);
1565
1566         if (unlikely(!vq->access_ok))
1567                 if (unlikely(vring_translate(dev, vq) < 0))
1568                         goto out;
1569
1570         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
1571         if (count == 0)
1572                 goto out;
1573
1574         if (vq_is_packed(dev))
1575                 nb_tx = virtio_dev_rx_packed(dev, vq, pkts, count);
1576         else
1577                 nb_tx = virtio_dev_rx_split(dev, vq, pkts, count);
1578
1579         vhost_queue_stats_update(dev, vq, pkts, nb_tx);
1580
1581 out:
1582         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1583                 vhost_user_iotlb_rd_unlock(vq);
1584
1585 out_access_unlock:
1586         rte_spinlock_unlock(&vq->access_lock);
1587
1588         return nb_tx;
1589 }
1590
1591 uint16_t
1592 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
1593         struct rte_mbuf **__rte_restrict pkts, uint16_t count)
1594 {
1595         struct virtio_net *dev = get_device(vid);
1596
1597         if (!dev)
1598                 return 0;
1599
1600         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
1601                 VHOST_LOG_DATA(ERR, "(%s) %s: built-in vhost net backend is disabled.\n",
1602                         dev->ifname, __func__);
1603                 return 0;
1604         }
1605
1606         return virtio_dev_rx(dev, queue_id, pkts, count);
1607 }
1608
1609 static __rte_always_inline uint16_t
1610 async_get_first_inflight_pkt_idx(struct vhost_virtqueue *vq)
1611 {
1612         struct vhost_async *async = vq->async;
1613
1614         if (async->pkts_idx >= async->pkts_inflight_n)
1615                 return async->pkts_idx - async->pkts_inflight_n;
1616         else
1617                 return vq->size - async->pkts_inflight_n + async->pkts_idx;
1618 }
1619
1620 static __rte_always_inline void
1621 store_dma_desc_info_split(struct vring_used_elem *s_ring, struct vring_used_elem *d_ring,
1622                 uint16_t ring_size, uint16_t s_idx, uint16_t d_idx, uint16_t count)
1623 {
1624         size_t elem_size = sizeof(struct vring_used_elem);
1625
1626         if (d_idx + count <= ring_size) {
1627                 rte_memcpy(d_ring + d_idx, s_ring + s_idx, count * elem_size);
1628         } else {
1629                 uint16_t size = ring_size - d_idx;
1630
1631                 rte_memcpy(d_ring + d_idx, s_ring + s_idx, size * elem_size);
1632                 rte_memcpy(d_ring, s_ring + s_idx + size, (count - size) * elem_size);
1633         }
1634 }
1635
1636 static __rte_always_inline void
1637 store_dma_desc_info_packed(struct vring_used_elem_packed *s_ring,
1638                 struct vring_used_elem_packed *d_ring,
1639                 uint16_t ring_size, uint16_t s_idx, uint16_t d_idx, uint16_t count)
1640 {
1641         size_t elem_size = sizeof(struct vring_used_elem_packed);
1642
1643         if (d_idx + count <= ring_size) {
1644                 rte_memcpy(d_ring + d_idx, s_ring + s_idx, count * elem_size);
1645         } else {
1646                 uint16_t size = ring_size - d_idx;
1647
1648                 rte_memcpy(d_ring + d_idx, s_ring + s_idx, size * elem_size);
1649                 rte_memcpy(d_ring, s_ring + s_idx + size, (count - size) * elem_size);
1650         }
1651 }
1652
1653 static __rte_noinline uint32_t
1654 virtio_dev_rx_async_submit_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
1655                 uint16_t queue_id, struct rte_mbuf **pkts, uint32_t count,
1656                 int16_t dma_id, uint16_t vchan_id)
1657 {
1658         struct buf_vector buf_vec[BUF_VECTOR_MAX];
1659         uint32_t pkt_idx = 0;
1660         uint16_t num_buffers;
1661         uint16_t avail_head;
1662
1663         struct vhost_async *async = vq->async;
1664         struct async_inflight_info *pkts_info = async->pkts_info;
1665         uint32_t pkt_err = 0;
1666         uint16_t n_xfer;
1667         uint16_t slot_idx = 0;
1668
1669         /*
1670          * The ordering between avail index and desc reads need to be enforced.
1671          */
1672         avail_head = __atomic_load_n(&vq->avail->idx, __ATOMIC_ACQUIRE);
1673
1674         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
1675
1676         async_iter_reset(async);
1677
1678         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
1679                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
1680                 uint16_t nr_vec = 0;
1681
1682                 if (unlikely(reserve_avail_buf_split(dev, vq, pkt_len, buf_vec,
1683                                                 &num_buffers, avail_head, &nr_vec) < 0)) {
1684                         VHOST_LOG_DATA(DEBUG, "(%s) failed to get enough desc from vring\n",
1685                                         dev->ifname);
1686                         vq->shadow_used_idx -= num_buffers;
1687                         break;
1688                 }
1689
1690                 VHOST_LOG_DATA(DEBUG, "(%s) current index %d | end index %d\n",
1691                         dev->ifname, vq->last_avail_idx, vq->last_avail_idx + num_buffers);
1692
1693                 if (mbuf_to_desc(dev, vq, pkts[pkt_idx], buf_vec, nr_vec, num_buffers, true) < 0) {
1694                         vq->shadow_used_idx -= num_buffers;
1695                         break;
1696                 }
1697
1698                 slot_idx = (async->pkts_idx + pkt_idx) & (vq->size - 1);
1699                 pkts_info[slot_idx].descs = num_buffers;
1700                 pkts_info[slot_idx].mbuf = pkts[pkt_idx];
1701
1702                 vq->last_avail_idx += num_buffers;
1703         }
1704
1705         if (unlikely(pkt_idx == 0))
1706                 return 0;
1707
1708         n_xfer = vhost_async_dma_transfer(dev, vq, dma_id, vchan_id, async->pkts_idx,
1709                         async->iov_iter, pkt_idx);
1710
1711         pkt_err = pkt_idx - n_xfer;
1712         if (unlikely(pkt_err)) {
1713                 uint16_t num_descs = 0;
1714
1715                 VHOST_LOG_DATA(DEBUG, "(%s) %s: failed to transfer %u packets for queue %u.\n",
1716                                 dev->ifname, __func__, pkt_err, queue_id);
1717
1718                 /* update number of completed packets */
1719                 pkt_idx = n_xfer;
1720
1721                 /* calculate the sum of descriptors to revert */
1722                 while (pkt_err-- > 0) {
1723                         num_descs += pkts_info[slot_idx & (vq->size - 1)].descs;
1724                         slot_idx--;
1725                 }
1726
1727                 /* recover shadow used ring and available ring */
1728                 vq->shadow_used_idx -= num_descs;
1729                 vq->last_avail_idx -= num_descs;
1730         }
1731
1732         /* keep used descriptors */
1733         if (likely(vq->shadow_used_idx)) {
1734                 uint16_t to = async->desc_idx_split & (vq->size - 1);
1735
1736                 store_dma_desc_info_split(vq->shadow_used_split,
1737                                 async->descs_split, vq->size, 0, to,
1738                                 vq->shadow_used_idx);
1739
1740                 async->desc_idx_split += vq->shadow_used_idx;
1741
1742                 async->pkts_idx += pkt_idx;
1743                 if (async->pkts_idx >= vq->size)
1744                         async->pkts_idx -= vq->size;
1745
1746                 async->pkts_inflight_n += pkt_idx;
1747                 vq->shadow_used_idx = 0;
1748         }
1749
1750         return pkt_idx;
1751 }
1752
1753
1754 static __rte_always_inline int
1755 vhost_enqueue_async_packed(struct virtio_net *dev,
1756                             struct vhost_virtqueue *vq,
1757                             struct rte_mbuf *pkt,
1758                             struct buf_vector *buf_vec,
1759                             uint16_t *nr_descs,
1760                             uint16_t *nr_buffers)
1761 {
1762         uint16_t nr_vec = 0;
1763         uint16_t avail_idx = vq->last_avail_idx;
1764         uint16_t max_tries, tries = 0;
1765         uint16_t buf_id = 0;
1766         uint32_t len = 0;
1767         uint16_t desc_count = 0;
1768         uint32_t size = pkt->pkt_len + sizeof(struct virtio_net_hdr_mrg_rxbuf);
1769         uint32_t buffer_len[vq->size];
1770         uint16_t buffer_buf_id[vq->size];
1771         uint16_t buffer_desc_count[vq->size];
1772
1773         if (rxvq_is_mergeable(dev))
1774                 max_tries = vq->size - 1;
1775         else
1776                 max_tries = 1;
1777
1778         while (size > 0) {
1779                 /*
1780                  * if we tried all available ring items, and still
1781                  * can't get enough buf, it means something abnormal
1782                  * happened.
1783                  */
1784                 if (unlikely(++tries > max_tries))
1785                         return -1;
1786
1787                 if (unlikely(fill_vec_buf_packed(dev, vq,
1788                                                 avail_idx, &desc_count,
1789                                                 buf_vec, &nr_vec,
1790                                                 &buf_id, &len,
1791                                                 VHOST_ACCESS_RW) < 0))
1792                         return -1;
1793
1794                 len = RTE_MIN(len, size);
1795                 size -= len;
1796
1797                 buffer_len[*nr_buffers] = len;
1798                 buffer_buf_id[*nr_buffers] = buf_id;
1799                 buffer_desc_count[*nr_buffers] = desc_count;
1800                 *nr_buffers += 1;
1801                 *nr_descs += desc_count;
1802                 avail_idx += desc_count;
1803                 if (avail_idx >= vq->size)
1804                         avail_idx -= vq->size;
1805         }
1806
1807         if (unlikely(mbuf_to_desc(dev, vq, pkt, buf_vec, nr_vec, *nr_buffers, true) < 0))
1808                 return -1;
1809
1810         vhost_shadow_enqueue_packed(vq, buffer_len, buffer_buf_id, buffer_desc_count, *nr_buffers);
1811
1812         return 0;
1813 }
1814
1815 static __rte_always_inline int16_t
1816 virtio_dev_rx_async_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
1817                             struct rte_mbuf *pkt, uint16_t *nr_descs, uint16_t *nr_buffers)
1818 {
1819         struct buf_vector buf_vec[BUF_VECTOR_MAX];
1820
1821         if (unlikely(vhost_enqueue_async_packed(dev, vq, pkt, buf_vec,
1822                                         nr_descs, nr_buffers) < 0)) {
1823                 VHOST_LOG_DATA(DEBUG, "(%s) failed to get enough desc from vring\n", dev->ifname);
1824                 return -1;
1825         }
1826
1827         VHOST_LOG_DATA(DEBUG, "(%s) current index %d | end index %d\n",
1828                         dev->ifname, vq->last_avail_idx, vq->last_avail_idx + *nr_descs);
1829
1830         return 0;
1831 }
1832
1833 static __rte_always_inline void
1834 dma_error_handler_packed(struct vhost_virtqueue *vq, uint16_t slot_idx,
1835                         uint32_t nr_err, uint32_t *pkt_idx)
1836 {
1837         uint16_t descs_err = 0;
1838         uint16_t buffers_err = 0;
1839         struct async_inflight_info *pkts_info = vq->async->pkts_info;
1840
1841         *pkt_idx -= nr_err;
1842         /* calculate the sum of buffers and descs of DMA-error packets. */
1843         while (nr_err-- > 0) {
1844                 descs_err += pkts_info[slot_idx % vq->size].descs;
1845                 buffers_err += pkts_info[slot_idx % vq->size].nr_buffers;
1846                 slot_idx--;
1847         }
1848
1849         if (vq->last_avail_idx >= descs_err) {
1850                 vq->last_avail_idx -= descs_err;
1851         } else {
1852                 vq->last_avail_idx = vq->last_avail_idx + vq->size - descs_err;
1853                 vq->avail_wrap_counter ^= 1;
1854         }
1855
1856         vq->shadow_used_idx -= buffers_err;
1857 }
1858
1859 static __rte_noinline uint32_t
1860 virtio_dev_rx_async_submit_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
1861                 uint16_t queue_id, struct rte_mbuf **pkts, uint32_t count,
1862                 int16_t dma_id, uint16_t vchan_id)
1863 {
1864         uint32_t pkt_idx = 0;
1865         uint32_t remained = count;
1866         uint16_t n_xfer;
1867         uint16_t num_buffers;
1868         uint16_t num_descs;
1869
1870         struct vhost_async *async = vq->async;
1871         struct async_inflight_info *pkts_info = async->pkts_info;
1872         uint32_t pkt_err = 0;
1873         uint16_t slot_idx = 0;
1874
1875         do {
1876                 rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
1877
1878                 num_buffers = 0;
1879                 num_descs = 0;
1880                 if (unlikely(virtio_dev_rx_async_packed(dev, vq, pkts[pkt_idx],
1881                                                 &num_descs, &num_buffers) < 0))
1882                         break;
1883
1884                 slot_idx = (async->pkts_idx + pkt_idx) % vq->size;
1885
1886                 pkts_info[slot_idx].descs = num_descs;
1887                 pkts_info[slot_idx].nr_buffers = num_buffers;
1888                 pkts_info[slot_idx].mbuf = pkts[pkt_idx];
1889
1890                 pkt_idx++;
1891                 remained--;
1892                 vq_inc_last_avail_packed(vq, num_descs);
1893         } while (pkt_idx < count);
1894
1895         if (unlikely(pkt_idx == 0))
1896                 return 0;
1897
1898         n_xfer = vhost_async_dma_transfer(dev, vq, dma_id, vchan_id, async->pkts_idx,
1899                         async->iov_iter, pkt_idx);
1900
1901         async_iter_reset(async);
1902
1903         pkt_err = pkt_idx - n_xfer;
1904         if (unlikely(pkt_err)) {
1905                 VHOST_LOG_DATA(DEBUG, "(%s) %s: failed to transfer %u packets for queue %u.\n",
1906                                 dev->ifname, __func__, pkt_err, queue_id);
1907                 dma_error_handler_packed(vq, slot_idx, pkt_err, &pkt_idx);
1908         }
1909
1910         if (likely(vq->shadow_used_idx)) {
1911                 /* keep used descriptors. */
1912                 store_dma_desc_info_packed(vq->shadow_used_packed, async->buffers_packed,
1913                                         vq->size, 0, async->buffer_idx_packed,
1914                                         vq->shadow_used_idx);
1915
1916                 async->buffer_idx_packed += vq->shadow_used_idx;
1917                 if (async->buffer_idx_packed >= vq->size)
1918                         async->buffer_idx_packed -= vq->size;
1919
1920                 async->pkts_idx += pkt_idx;
1921                 if (async->pkts_idx >= vq->size)
1922                         async->pkts_idx -= vq->size;
1923
1924                 vq->shadow_used_idx = 0;
1925                 async->pkts_inflight_n += pkt_idx;
1926         }
1927
1928         return pkt_idx;
1929 }
1930
1931 static __rte_always_inline void
1932 write_back_completed_descs_split(struct vhost_virtqueue *vq, uint16_t n_descs)
1933 {
1934         struct vhost_async *async = vq->async;
1935         uint16_t nr_left = n_descs;
1936         uint16_t nr_copy;
1937         uint16_t to, from;
1938
1939         do {
1940                 from = async->last_desc_idx_split & (vq->size - 1);
1941                 nr_copy = nr_left + from <= vq->size ? nr_left : vq->size - from;
1942                 to = vq->last_used_idx & (vq->size - 1);
1943
1944                 if (to + nr_copy <= vq->size) {
1945                         rte_memcpy(&vq->used->ring[to], &async->descs_split[from],
1946                                         nr_copy * sizeof(struct vring_used_elem));
1947                 } else {
1948                         uint16_t size = vq->size - to;
1949
1950                         rte_memcpy(&vq->used->ring[to], &async->descs_split[from],
1951                                         size * sizeof(struct vring_used_elem));
1952                         rte_memcpy(&vq->used->ring[0], &async->descs_split[from + size],
1953                                         (nr_copy - size) * sizeof(struct vring_used_elem));
1954                 }
1955
1956                 async->last_desc_idx_split += nr_copy;
1957                 vq->last_used_idx += nr_copy;
1958                 nr_left -= nr_copy;
1959         } while (nr_left > 0);
1960 }
1961
1962 static __rte_always_inline void
1963 write_back_completed_descs_packed(struct vhost_virtqueue *vq,
1964                                 uint16_t n_buffers)
1965 {
1966         struct vhost_async *async = vq->async;
1967         uint16_t from = async->last_buffer_idx_packed;
1968         uint16_t used_idx = vq->last_used_idx;
1969         uint16_t head_idx = vq->last_used_idx;
1970         uint16_t head_flags = 0;
1971         uint16_t i;
1972
1973         /* Split loop in two to save memory barriers */
1974         for (i = 0; i < n_buffers; i++) {
1975                 vq->desc_packed[used_idx].id = async->buffers_packed[from].id;
1976                 vq->desc_packed[used_idx].len = async->buffers_packed[from].len;
1977
1978                 used_idx += async->buffers_packed[from].count;
1979                 if (used_idx >= vq->size)
1980                         used_idx -= vq->size;
1981
1982                 from++;
1983                 if (from >= vq->size)
1984                         from = 0;
1985         }
1986
1987         /* The ordering for storing desc flags needs to be enforced. */
1988         rte_atomic_thread_fence(__ATOMIC_RELEASE);
1989
1990         from = async->last_buffer_idx_packed;
1991
1992         for (i = 0; i < n_buffers; i++) {
1993                 uint16_t flags;
1994
1995                 if (async->buffers_packed[from].len)
1996                         flags = VRING_DESC_F_WRITE;
1997                 else
1998                         flags = 0;
1999
2000                 if (vq->used_wrap_counter) {
2001                         flags |= VRING_DESC_F_USED;
2002                         flags |= VRING_DESC_F_AVAIL;
2003                 } else {
2004                         flags &= ~VRING_DESC_F_USED;
2005                         flags &= ~VRING_DESC_F_AVAIL;
2006                 }
2007
2008                 if (i > 0) {
2009                         vq->desc_packed[vq->last_used_idx].flags = flags;
2010                 } else {
2011                         head_idx = vq->last_used_idx;
2012                         head_flags = flags;
2013                 }
2014
2015                 vq_inc_last_used_packed(vq, async->buffers_packed[from].count);
2016
2017                 from++;
2018                 if (from == vq->size)
2019                         from = 0;
2020         }
2021
2022         vq->desc_packed[head_idx].flags = head_flags;
2023         async->last_buffer_idx_packed = from;
2024 }
2025
2026 static __rte_always_inline uint16_t
2027 vhost_poll_enqueue_completed(struct virtio_net *dev, uint16_t queue_id,
2028                 struct rte_mbuf **pkts, uint16_t count, int16_t dma_id,
2029                 uint16_t vchan_id)
2030 {
2031         struct vhost_virtqueue *vq = dev->virtqueue[queue_id];
2032         struct vhost_async *async = vq->async;
2033         struct async_inflight_info *pkts_info = async->pkts_info;
2034         uint16_t nr_cpl_pkts = 0;
2035         uint16_t n_descs = 0, n_buffers = 0;
2036         uint16_t start_idx, from, i;
2037
2038         /* Check completed copies for the given DMA vChannel */
2039         vhost_async_dma_check_completed(dev, dma_id, vchan_id, VHOST_DMA_MAX_COPY_COMPLETE);
2040
2041         start_idx = async_get_first_inflight_pkt_idx(vq);
2042         /**
2043          * Calculate the number of copy completed packets.
2044          * Note that there may be completed packets even if
2045          * no copies are reported done by the given DMA vChannel,
2046          * as it's possible that a virtqueue uses multiple DMA
2047          * vChannels.
2048          */
2049         from = start_idx;
2050         while (vq->async->pkts_cmpl_flag[from] && count--) {
2051                 vq->async->pkts_cmpl_flag[from] = false;
2052                 from++;
2053                 if (from >= vq->size)
2054                         from -= vq->size;
2055                 nr_cpl_pkts++;
2056         }
2057
2058         if (nr_cpl_pkts == 0)
2059                 return 0;
2060
2061         for (i = 0; i < nr_cpl_pkts; i++) {
2062                 from = (start_idx + i) % vq->size;
2063                 /* Only used with packed ring */
2064                 n_buffers += pkts_info[from].nr_buffers;
2065                 /* Only used with split ring */
2066                 n_descs += pkts_info[from].descs;
2067                 pkts[i] = pkts_info[from].mbuf;
2068         }
2069
2070         async->pkts_inflight_n -= nr_cpl_pkts;
2071
2072         if (likely(vq->enabled && vq->access_ok)) {
2073                 if (vq_is_packed(dev)) {
2074                         write_back_completed_descs_packed(vq, n_buffers);
2075                         vhost_vring_call_packed(dev, vq);
2076                 } else {
2077                         write_back_completed_descs_split(vq, n_descs);
2078                         __atomic_add_fetch(&vq->used->idx, n_descs, __ATOMIC_RELEASE);
2079                         vhost_vring_call_split(dev, vq);
2080                 }
2081         } else {
2082                 if (vq_is_packed(dev)) {
2083                         async->last_buffer_idx_packed += n_buffers;
2084                         if (async->last_buffer_idx_packed >= vq->size)
2085                                 async->last_buffer_idx_packed -= vq->size;
2086                 } else {
2087                         async->last_desc_idx_split += n_descs;
2088                 }
2089         }
2090
2091         return nr_cpl_pkts;
2092 }
2093
2094 uint16_t
2095 rte_vhost_poll_enqueue_completed(int vid, uint16_t queue_id,
2096                 struct rte_mbuf **pkts, uint16_t count, int16_t dma_id,
2097                 uint16_t vchan_id)
2098 {
2099         struct virtio_net *dev = get_device(vid);
2100         struct vhost_virtqueue *vq;
2101         uint16_t n_pkts_cpl = 0;
2102
2103         if (unlikely(!dev))
2104                 return 0;
2105
2106         VHOST_LOG_DATA(DEBUG, "(%s) %s\n", dev->ifname, __func__);
2107         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
2108                 VHOST_LOG_DATA(ERR, "(%s) %s: invalid virtqueue idx %d.\n",
2109                         dev->ifname, __func__, queue_id);
2110                 return 0;
2111         }
2112
2113         if (unlikely(!dma_copy_track[dma_id].vchans ||
2114                                 !dma_copy_track[dma_id].vchans[vchan_id].pkts_cmpl_flag_addr)) {
2115                 VHOST_LOG_DATA(ERR, "(%s) %s: invalid channel %d:%u.\n", dev->ifname, __func__,
2116                                dma_id, vchan_id);
2117                 return 0;
2118         }
2119
2120         vq = dev->virtqueue[queue_id];
2121
2122         if (!rte_spinlock_trylock(&vq->access_lock)) {
2123                 VHOST_LOG_DATA(DEBUG, "(%s) %s: virtqueue %u is busy.\n", dev->ifname, __func__,
2124                                 queue_id);
2125                 return 0;
2126         }
2127
2128         if (unlikely(!vq->async)) {
2129                 VHOST_LOG_DATA(ERR, "(%s) %s: async not registered for virtqueue %d.\n",
2130                                 dev->ifname, __func__, queue_id);
2131                 goto out;
2132         }
2133
2134         n_pkts_cpl = vhost_poll_enqueue_completed(dev, queue_id, pkts, count, dma_id, vchan_id);
2135
2136         vhost_queue_stats_update(dev, vq, pkts, n_pkts_cpl);
2137         vq->stats.inflight_completed += n_pkts_cpl;
2138
2139 out:
2140         rte_spinlock_unlock(&vq->access_lock);
2141
2142         return n_pkts_cpl;
2143 }
2144
2145 uint16_t
2146 rte_vhost_clear_queue_thread_unsafe(int vid, uint16_t queue_id,
2147                 struct rte_mbuf **pkts, uint16_t count, int16_t dma_id,
2148                 uint16_t vchan_id)
2149 {
2150         struct virtio_net *dev = get_device(vid);
2151         struct vhost_virtqueue *vq;
2152         uint16_t n_pkts_cpl = 0;
2153
2154         if (!dev)
2155                 return 0;
2156
2157         VHOST_LOG_DATA(DEBUG, "(%s) %s\n", dev->ifname, __func__);
2158         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
2159                 VHOST_LOG_DATA(ERR, "(%s) %s: invalid virtqueue idx %d.\n",
2160                         dev->ifname, __func__, queue_id);
2161                 return 0;
2162         }
2163
2164         vq = dev->virtqueue[queue_id];
2165
2166         if (unlikely(!rte_spinlock_is_locked(&vq->access_lock))) {
2167                 VHOST_LOG_DATA(ERR, "(%s) %s() called without access lock taken.\n",
2168                                 dev->ifname, __func__);
2169                 return -1;
2170         }
2171
2172         if (unlikely(!vq->async)) {
2173                 VHOST_LOG_DATA(ERR, "(%s) %s: async not registered for queue id %d.\n",
2174                         dev->ifname, __func__, queue_id);
2175                 return 0;
2176         }
2177
2178         if (unlikely(!dma_copy_track[dma_id].vchans ||
2179                                 !dma_copy_track[dma_id].vchans[vchan_id].pkts_cmpl_flag_addr)) {
2180                 VHOST_LOG_DATA(ERR, "(%s) %s: invalid channel %d:%u.\n", dev->ifname, __func__,
2181                                 dma_id, vchan_id);
2182                 return 0;
2183         }
2184
2185         n_pkts_cpl = vhost_poll_enqueue_completed(dev, queue_id, pkts, count, dma_id, vchan_id);
2186
2187         vhost_queue_stats_update(dev, vq, pkts, n_pkts_cpl);
2188         vq->stats.inflight_completed += n_pkts_cpl;
2189
2190         return n_pkts_cpl;
2191 }
2192
2193 static __rte_always_inline uint32_t
2194 virtio_dev_rx_async_submit(struct virtio_net *dev, uint16_t queue_id,
2195         struct rte_mbuf **pkts, uint32_t count, int16_t dma_id, uint16_t vchan_id)
2196 {
2197         struct vhost_virtqueue *vq;
2198         uint32_t nb_tx = 0;
2199
2200         VHOST_LOG_DATA(DEBUG, "(%s) %s\n", dev->ifname, __func__);
2201         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
2202                 VHOST_LOG_DATA(ERR, "(%s) %s: invalid virtqueue idx %d.\n",
2203                         dev->ifname, __func__, queue_id);
2204                 return 0;
2205         }
2206
2207         if (unlikely(!dma_copy_track[dma_id].vchans ||
2208                                 !dma_copy_track[dma_id].vchans[vchan_id].pkts_cmpl_flag_addr)) {
2209                 VHOST_LOG_DATA(ERR, "(%s) %s: invalid channel %d:%u.\n", dev->ifname, __func__,
2210                                dma_id, vchan_id);
2211                 return 0;
2212         }
2213
2214         vq = dev->virtqueue[queue_id];
2215
2216         rte_spinlock_lock(&vq->access_lock);
2217
2218         if (unlikely(!vq->enabled || !vq->async))
2219                 goto out_access_unlock;
2220
2221         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
2222                 vhost_user_iotlb_rd_lock(vq);
2223
2224         if (unlikely(!vq->access_ok))
2225                 if (unlikely(vring_translate(dev, vq) < 0))
2226                         goto out;
2227
2228         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
2229         if (count == 0)
2230                 goto out;
2231
2232         if (vq_is_packed(dev))
2233                 nb_tx = virtio_dev_rx_async_submit_packed(dev, vq, queue_id,
2234                                 pkts, count, dma_id, vchan_id);
2235         else
2236                 nb_tx = virtio_dev_rx_async_submit_split(dev, vq, queue_id,
2237                                 pkts, count, dma_id, vchan_id);
2238
2239         vq->stats.inflight_submitted += nb_tx;
2240
2241 out:
2242         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
2243                 vhost_user_iotlb_rd_unlock(vq);
2244
2245 out_access_unlock:
2246         rte_spinlock_unlock(&vq->access_lock);
2247
2248         return nb_tx;
2249 }
2250
2251 uint16_t
2252 rte_vhost_submit_enqueue_burst(int vid, uint16_t queue_id,
2253                 struct rte_mbuf **pkts, uint16_t count, int16_t dma_id,
2254                 uint16_t vchan_id)
2255 {
2256         struct virtio_net *dev = get_device(vid);
2257
2258         if (!dev)
2259                 return 0;
2260
2261         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
2262                 VHOST_LOG_DATA(ERR, "(%s) %s: built-in vhost net backend is disabled.\n",
2263                         dev->ifname, __func__);
2264                 return 0;
2265         }
2266
2267         return virtio_dev_rx_async_submit(dev, queue_id, pkts, count, dma_id, vchan_id);
2268 }
2269
2270 static inline bool
2271 virtio_net_with_host_offload(struct virtio_net *dev)
2272 {
2273         if (dev->features &
2274                         ((1ULL << VIRTIO_NET_F_CSUM) |
2275                          (1ULL << VIRTIO_NET_F_HOST_ECN) |
2276                          (1ULL << VIRTIO_NET_F_HOST_TSO4) |
2277                          (1ULL << VIRTIO_NET_F_HOST_TSO6) |
2278                          (1ULL << VIRTIO_NET_F_HOST_UFO)))
2279                 return true;
2280
2281         return false;
2282 }
2283
2284 static int
2285 parse_headers(struct rte_mbuf *m, uint8_t *l4_proto)
2286 {
2287         struct rte_ipv4_hdr *ipv4_hdr;
2288         struct rte_ipv6_hdr *ipv6_hdr;
2289         struct rte_ether_hdr *eth_hdr;
2290         uint16_t ethertype;
2291         uint16_t data_len = rte_pktmbuf_data_len(m);
2292
2293         if (data_len < sizeof(struct rte_ether_hdr))
2294                 return -EINVAL;
2295
2296         eth_hdr = rte_pktmbuf_mtod(m, struct rte_ether_hdr *);
2297
2298         m->l2_len = sizeof(struct rte_ether_hdr);
2299         ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
2300
2301         if (ethertype == RTE_ETHER_TYPE_VLAN) {
2302                 if (data_len < sizeof(struct rte_ether_hdr) +
2303                                 sizeof(struct rte_vlan_hdr))
2304                         goto error;
2305
2306                 struct rte_vlan_hdr *vlan_hdr =
2307                         (struct rte_vlan_hdr *)(eth_hdr + 1);
2308
2309                 m->l2_len += sizeof(struct rte_vlan_hdr);
2310                 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
2311         }
2312
2313         switch (ethertype) {
2314         case RTE_ETHER_TYPE_IPV4:
2315                 if (data_len < m->l2_len + sizeof(struct rte_ipv4_hdr))
2316                         goto error;
2317                 ipv4_hdr = rte_pktmbuf_mtod_offset(m, struct rte_ipv4_hdr *,
2318                                 m->l2_len);
2319                 m->l3_len = rte_ipv4_hdr_len(ipv4_hdr);
2320                 if (data_len < m->l2_len + m->l3_len)
2321                         goto error;
2322                 m->ol_flags |= RTE_MBUF_F_TX_IPV4;
2323                 *l4_proto = ipv4_hdr->next_proto_id;
2324                 break;
2325         case RTE_ETHER_TYPE_IPV6:
2326                 if (data_len < m->l2_len + sizeof(struct rte_ipv6_hdr))
2327                         goto error;
2328                 ipv6_hdr = rte_pktmbuf_mtod_offset(m, struct rte_ipv6_hdr *,
2329                                 m->l2_len);
2330                 m->l3_len = sizeof(struct rte_ipv6_hdr);
2331                 m->ol_flags |= RTE_MBUF_F_TX_IPV6;
2332                 *l4_proto = ipv6_hdr->proto;
2333                 break;
2334         default:
2335                 /* a valid L3 header is needed for further L4 parsing */
2336                 goto error;
2337         }
2338
2339         /* both CSUM and GSO need a valid L4 header */
2340         switch (*l4_proto) {
2341         case IPPROTO_TCP:
2342                 if (data_len < m->l2_len + m->l3_len +
2343                                 sizeof(struct rte_tcp_hdr))
2344                         goto error;
2345                 break;
2346         case IPPROTO_UDP:
2347                 if (data_len < m->l2_len + m->l3_len +
2348                                 sizeof(struct rte_udp_hdr))
2349                         goto error;
2350                 break;
2351         case IPPROTO_SCTP:
2352                 if (data_len < m->l2_len + m->l3_len +
2353                                 sizeof(struct rte_sctp_hdr))
2354                         goto error;
2355                 break;
2356         default:
2357                 goto error;
2358         }
2359
2360         return 0;
2361
2362 error:
2363         m->l2_len = 0;
2364         m->l3_len = 0;
2365         m->ol_flags = 0;
2366         return -EINVAL;
2367 }
2368
2369 static __rte_always_inline void
2370 vhost_dequeue_offload_legacy(struct virtio_net *dev, struct virtio_net_hdr *hdr,
2371                 struct rte_mbuf *m)
2372 {
2373         uint8_t l4_proto = 0;
2374         struct rte_tcp_hdr *tcp_hdr = NULL;
2375         uint16_t tcp_len;
2376         uint16_t data_len = rte_pktmbuf_data_len(m);
2377
2378         if (parse_headers(m, &l4_proto) < 0)
2379                 return;
2380
2381         if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
2382                 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
2383                         switch (hdr->csum_offset) {
2384                         case (offsetof(struct rte_tcp_hdr, cksum)):
2385                                 if (l4_proto != IPPROTO_TCP)
2386                                         goto error;
2387                                 m->ol_flags |= RTE_MBUF_F_TX_TCP_CKSUM;
2388                                 break;
2389                         case (offsetof(struct rte_udp_hdr, dgram_cksum)):
2390                                 if (l4_proto != IPPROTO_UDP)
2391                                         goto error;
2392                                 m->ol_flags |= RTE_MBUF_F_TX_UDP_CKSUM;
2393                                 break;
2394                         case (offsetof(struct rte_sctp_hdr, cksum)):
2395                                 if (l4_proto != IPPROTO_SCTP)
2396                                         goto error;
2397                                 m->ol_flags |= RTE_MBUF_F_TX_SCTP_CKSUM;
2398                                 break;
2399                         default:
2400                                 goto error;
2401                         }
2402                 } else {
2403                         goto error;
2404                 }
2405         }
2406
2407         if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
2408                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
2409                 case VIRTIO_NET_HDR_GSO_TCPV4:
2410                 case VIRTIO_NET_HDR_GSO_TCPV6:
2411                         if (l4_proto != IPPROTO_TCP)
2412                                 goto error;
2413                         tcp_hdr = rte_pktmbuf_mtod_offset(m,
2414                                         struct rte_tcp_hdr *,
2415                                         m->l2_len + m->l3_len);
2416                         tcp_len = (tcp_hdr->data_off & 0xf0) >> 2;
2417                         if (data_len < m->l2_len + m->l3_len + tcp_len)
2418                                 goto error;
2419                         m->ol_flags |= RTE_MBUF_F_TX_TCP_SEG;
2420                         m->tso_segsz = hdr->gso_size;
2421                         m->l4_len = tcp_len;
2422                         break;
2423                 case VIRTIO_NET_HDR_GSO_UDP:
2424                         if (l4_proto != IPPROTO_UDP)
2425                                 goto error;
2426                         m->ol_flags |= RTE_MBUF_F_TX_UDP_SEG;
2427                         m->tso_segsz = hdr->gso_size;
2428                         m->l4_len = sizeof(struct rte_udp_hdr);
2429                         break;
2430                 default:
2431                         VHOST_LOG_DATA(WARNING, "(%s) unsupported gso type %u.\n",
2432                                         dev->ifname, hdr->gso_type);
2433                         goto error;
2434                 }
2435         }
2436         return;
2437
2438 error:
2439         m->l2_len = 0;
2440         m->l3_len = 0;
2441         m->ol_flags = 0;
2442 }
2443
2444 static __rte_always_inline void
2445 vhost_dequeue_offload(struct virtio_net *dev, struct virtio_net_hdr *hdr,
2446                 struct rte_mbuf *m, bool legacy_ol_flags)
2447 {
2448         struct rte_net_hdr_lens hdr_lens;
2449         int l4_supported = 0;
2450         uint32_t ptype;
2451
2452         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
2453                 return;
2454
2455         if (legacy_ol_flags) {
2456                 vhost_dequeue_offload_legacy(dev, hdr, m);
2457                 return;
2458         }
2459
2460         m->ol_flags |= RTE_MBUF_F_RX_IP_CKSUM_UNKNOWN;
2461
2462         ptype = rte_net_get_ptype(m, &hdr_lens, RTE_PTYPE_ALL_MASK);
2463         m->packet_type = ptype;
2464         if ((ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_TCP ||
2465             (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_UDP ||
2466             (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_SCTP)
2467                 l4_supported = 1;
2468
2469         /* According to Virtio 1.1 spec, the device only needs to look at
2470          * VIRTIO_NET_HDR_F_NEEDS_CSUM in the packet transmission path.
2471          * This differs from the processing incoming packets path where the
2472          * driver could rely on VIRTIO_NET_HDR_F_DATA_VALID flag set by the
2473          * device.
2474          *
2475          * 5.1.6.2.1 Driver Requirements: Packet Transmission
2476          * The driver MUST NOT set the VIRTIO_NET_HDR_F_DATA_VALID and
2477          * VIRTIO_NET_HDR_F_RSC_INFO bits in flags.
2478          *
2479          * 5.1.6.2.2 Device Requirements: Packet Transmission
2480          * The device MUST ignore flag bits that it does not recognize.
2481          */
2482         if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
2483                 uint32_t hdrlen;
2484
2485                 hdrlen = hdr_lens.l2_len + hdr_lens.l3_len + hdr_lens.l4_len;
2486                 if (hdr->csum_start <= hdrlen && l4_supported != 0) {
2487                         m->ol_flags |= RTE_MBUF_F_RX_L4_CKSUM_NONE;
2488                 } else {
2489                         /* Unknown proto or tunnel, do sw cksum. We can assume
2490                          * the cksum field is in the first segment since the
2491                          * buffers we provided to the host are large enough.
2492                          * In case of SCTP, this will be wrong since it's a CRC
2493                          * but there's nothing we can do.
2494                          */
2495                         uint16_t csum = 0, off;
2496
2497                         if (rte_raw_cksum_mbuf(m, hdr->csum_start,
2498                                         rte_pktmbuf_pkt_len(m) - hdr->csum_start, &csum) < 0)
2499                                 return;
2500                         if (likely(csum != 0xffff))
2501                                 csum = ~csum;
2502                         off = hdr->csum_offset + hdr->csum_start;
2503                         if (rte_pktmbuf_data_len(m) >= off + 1)
2504                                 *rte_pktmbuf_mtod_offset(m, uint16_t *, off) = csum;
2505                 }
2506         }
2507
2508         if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
2509                 if (hdr->gso_size == 0)
2510                         return;
2511
2512                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
2513                 case VIRTIO_NET_HDR_GSO_TCPV4:
2514                 case VIRTIO_NET_HDR_GSO_TCPV6:
2515                         if ((ptype & RTE_PTYPE_L4_MASK) != RTE_PTYPE_L4_TCP)
2516                                 break;
2517                         m->ol_flags |= RTE_MBUF_F_RX_LRO | RTE_MBUF_F_RX_L4_CKSUM_NONE;
2518                         m->tso_segsz = hdr->gso_size;
2519                         break;
2520                 case VIRTIO_NET_HDR_GSO_UDP:
2521                         if ((ptype & RTE_PTYPE_L4_MASK) != RTE_PTYPE_L4_UDP)
2522                                 break;
2523                         m->ol_flags |= RTE_MBUF_F_RX_LRO | RTE_MBUF_F_RX_L4_CKSUM_NONE;
2524                         m->tso_segsz = hdr->gso_size;
2525                         break;
2526                 default:
2527                         break;
2528                 }
2529         }
2530 }
2531
2532 static __rte_noinline void
2533 copy_vnet_hdr_from_desc(struct virtio_net_hdr *hdr,
2534                 struct buf_vector *buf_vec)
2535 {
2536         uint64_t len;
2537         uint64_t remain = sizeof(struct virtio_net_hdr);
2538         uint64_t src;
2539         uint64_t dst = (uint64_t)(uintptr_t)hdr;
2540
2541         while (remain) {
2542                 len = RTE_MIN(remain, buf_vec->buf_len);
2543                 src = buf_vec->buf_addr;
2544                 rte_memcpy((void *)(uintptr_t)dst,
2545                                 (void *)(uintptr_t)src, len);
2546
2547                 remain -= len;
2548                 dst += len;
2549                 buf_vec++;
2550         }
2551 }
2552
2553 static __rte_always_inline int
2554 desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
2555                   struct buf_vector *buf_vec, uint16_t nr_vec,
2556                   struct rte_mbuf *m, struct rte_mempool *mbuf_pool,
2557                   bool legacy_ol_flags, uint16_t slot_idx, bool is_async)
2558 {
2559         uint32_t buf_avail, buf_offset, buf_len;
2560         uint64_t buf_addr, buf_iova;
2561         uint32_t mbuf_avail, mbuf_offset;
2562         uint32_t cpy_len;
2563         struct rte_mbuf *cur = m, *prev = m;
2564         struct virtio_net_hdr tmp_hdr;
2565         struct virtio_net_hdr *hdr = NULL;
2566         /* A counter to avoid desc dead loop chain */
2567         uint16_t vec_idx = 0;
2568         struct vhost_async *async = vq->async;
2569         struct async_inflight_info *pkts_info;
2570
2571         buf_addr = buf_vec[vec_idx].buf_addr;
2572         buf_iova = buf_vec[vec_idx].buf_iova;
2573         buf_len = buf_vec[vec_idx].buf_len;
2574
2575         if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1))
2576                 return -1;
2577
2578         if (virtio_net_with_host_offload(dev)) {
2579                 if (unlikely(buf_len < sizeof(struct virtio_net_hdr))) {
2580                         /*
2581                          * No luck, the virtio-net header doesn't fit
2582                          * in a contiguous virtual area.
2583                          */
2584                         copy_vnet_hdr_from_desc(&tmp_hdr, buf_vec);
2585                         hdr = &tmp_hdr;
2586                 } else {
2587                         hdr = (struct virtio_net_hdr *)((uintptr_t)buf_addr);
2588                 }
2589         }
2590
2591         /*
2592          * A virtio driver normally uses at least 2 desc buffers
2593          * for Tx: the first for storing the header, and others
2594          * for storing the data.
2595          */
2596         if (unlikely(buf_len < dev->vhost_hlen)) {
2597                 buf_offset = dev->vhost_hlen - buf_len;
2598                 vec_idx++;
2599                 buf_addr = buf_vec[vec_idx].buf_addr;
2600                 buf_iova = buf_vec[vec_idx].buf_iova;
2601                 buf_len = buf_vec[vec_idx].buf_len;
2602                 buf_avail  = buf_len - buf_offset;
2603         } else if (buf_len == dev->vhost_hlen) {
2604                 if (unlikely(++vec_idx >= nr_vec))
2605                         goto error;
2606                 buf_addr = buf_vec[vec_idx].buf_addr;
2607                 buf_iova = buf_vec[vec_idx].buf_iova;
2608                 buf_len = buf_vec[vec_idx].buf_len;
2609
2610                 buf_offset = 0;
2611                 buf_avail = buf_len;
2612         } else {
2613                 buf_offset = dev->vhost_hlen;
2614                 buf_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
2615         }
2616
2617         PRINT_PACKET(dev,
2618                         (uintptr_t)(buf_addr + buf_offset),
2619                         (uint32_t)buf_avail, 0);
2620
2621         mbuf_offset = 0;
2622         mbuf_avail  = m->buf_len - RTE_PKTMBUF_HEADROOM;
2623
2624         if (is_async) {
2625                 pkts_info = async->pkts_info;
2626                 if (async_iter_initialize(dev, async))
2627                         return -1;
2628         }
2629
2630         while (1) {
2631                 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
2632
2633                 if (is_async) {
2634                         if (async_fill_seg(dev, vq, cur, mbuf_offset,
2635                                            buf_iova + buf_offset, cpy_len, false) < 0)
2636                                 goto error;
2637                 } else {
2638                         sync_fill_seg(dev, vq, cur, mbuf_offset,
2639                                       buf_addr + buf_offset,
2640                                       buf_iova + buf_offset, cpy_len, false);
2641                 }
2642
2643                 mbuf_avail  -= cpy_len;
2644                 mbuf_offset += cpy_len;
2645                 buf_avail -= cpy_len;
2646                 buf_offset += cpy_len;
2647
2648                 /* This buf reaches to its end, get the next one */
2649                 if (buf_avail == 0) {
2650                         if (++vec_idx >= nr_vec)
2651                                 break;
2652
2653                         buf_addr = buf_vec[vec_idx].buf_addr;
2654                         buf_iova = buf_vec[vec_idx].buf_iova;
2655                         buf_len = buf_vec[vec_idx].buf_len;
2656
2657                         buf_offset = 0;
2658                         buf_avail  = buf_len;
2659
2660                         PRINT_PACKET(dev, (uintptr_t)buf_addr,
2661                                         (uint32_t)buf_avail, 0);
2662                 }
2663
2664                 /*
2665                  * This mbuf reaches to its end, get a new one
2666                  * to hold more data.
2667                  */
2668                 if (mbuf_avail == 0) {
2669                         cur = rte_pktmbuf_alloc(mbuf_pool);
2670                         if (unlikely(cur == NULL)) {
2671                                 VHOST_LOG_DATA(ERR, "(%s) failed to allocate memory for mbuf.\n",
2672                                                 dev->ifname);
2673                                 goto error;
2674                         }
2675
2676                         prev->next = cur;
2677                         prev->data_len = mbuf_offset;
2678                         m->nb_segs += 1;
2679                         m->pkt_len += mbuf_offset;
2680                         prev = cur;
2681
2682                         mbuf_offset = 0;
2683                         mbuf_avail  = cur->buf_len - RTE_PKTMBUF_HEADROOM;
2684                 }
2685         }
2686
2687         prev->data_len = mbuf_offset;
2688         m->pkt_len    += mbuf_offset;
2689
2690         if (is_async) {
2691                 async_iter_finalize(async);
2692                 if (hdr)
2693                         pkts_info[slot_idx].nethdr = *hdr;
2694         } else {
2695                 if (hdr)
2696                         vhost_dequeue_offload(dev, hdr, m, legacy_ol_flags);
2697         }
2698
2699         return 0;
2700 error:
2701         if (is_async)
2702                 async_iter_cancel(async);
2703
2704         return -1;
2705 }
2706
2707 static void
2708 virtio_dev_extbuf_free(void *addr __rte_unused, void *opaque)
2709 {
2710         rte_free(opaque);
2711 }
2712
2713 static int
2714 virtio_dev_extbuf_alloc(struct virtio_net *dev, struct rte_mbuf *pkt, uint32_t size)
2715 {
2716         struct rte_mbuf_ext_shared_info *shinfo = NULL;
2717         uint32_t total_len = RTE_PKTMBUF_HEADROOM + size;
2718         uint16_t buf_len;
2719         rte_iova_t iova;
2720         void *buf;
2721
2722         total_len += sizeof(*shinfo) + sizeof(uintptr_t);
2723         total_len = RTE_ALIGN_CEIL(total_len, sizeof(uintptr_t));
2724
2725         if (unlikely(total_len > UINT16_MAX))
2726                 return -ENOSPC;
2727
2728         buf_len = total_len;
2729         buf = rte_malloc(NULL, buf_len, RTE_CACHE_LINE_SIZE);
2730         if (unlikely(buf == NULL))
2731                 return -ENOMEM;
2732
2733         /* Initialize shinfo */
2734         shinfo = rte_pktmbuf_ext_shinfo_init_helper(buf, &buf_len,
2735                                                 virtio_dev_extbuf_free, buf);
2736         if (unlikely(shinfo == NULL)) {
2737                 rte_free(buf);
2738                 VHOST_LOG_DATA(ERR, "(%s) failed to init shinfo\n", dev->ifname);
2739                 return -1;
2740         }
2741
2742         iova = rte_malloc_virt2iova(buf);
2743         rte_pktmbuf_attach_extbuf(pkt, buf, iova, buf_len, shinfo);
2744         rte_pktmbuf_reset_headroom(pkt);
2745
2746         return 0;
2747 }
2748
2749 /*
2750  * Prepare a host supported pktmbuf.
2751  */
2752 static __rte_always_inline int
2753 virtio_dev_pktmbuf_prep(struct virtio_net *dev, struct rte_mbuf *pkt,
2754                          uint32_t data_len)
2755 {
2756         if (rte_pktmbuf_tailroom(pkt) >= data_len)
2757                 return 0;
2758
2759         /* attach an external buffer if supported */
2760         if (dev->extbuf && !virtio_dev_extbuf_alloc(dev, pkt, data_len))
2761                 return 0;
2762
2763         /* check if chained buffers are allowed */
2764         if (!dev->linearbuf)
2765                 return 0;
2766
2767         return -1;
2768 }
2769
2770 __rte_always_inline
2771 static uint16_t
2772 virtio_dev_tx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
2773         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count,
2774         bool legacy_ol_flags)
2775 {
2776         uint16_t i;
2777         uint16_t free_entries;
2778         uint16_t dropped = 0;
2779         static bool allocerr_warned;
2780
2781         /*
2782          * The ordering between avail index and
2783          * desc reads needs to be enforced.
2784          */
2785         free_entries = __atomic_load_n(&vq->avail->idx, __ATOMIC_ACQUIRE) -
2786                         vq->last_avail_idx;
2787         if (free_entries == 0)
2788                 return 0;
2789
2790         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
2791
2792         VHOST_LOG_DATA(DEBUG, "(%s) %s\n", dev->ifname, __func__);
2793
2794         count = RTE_MIN(count, MAX_PKT_BURST);
2795         count = RTE_MIN(count, free_entries);
2796         VHOST_LOG_DATA(DEBUG, "(%s) about to dequeue %u buffers\n",
2797                         dev->ifname, count);
2798
2799         if (rte_pktmbuf_alloc_bulk(mbuf_pool, pkts, count))
2800                 return 0;
2801
2802         for (i = 0; i < count; i++) {
2803                 struct buf_vector buf_vec[BUF_VECTOR_MAX];
2804                 uint16_t head_idx;
2805                 uint32_t buf_len;
2806                 uint16_t nr_vec = 0;
2807                 int err;
2808
2809                 if (unlikely(fill_vec_buf_split(dev, vq,
2810                                                 vq->last_avail_idx + i,
2811                                                 &nr_vec, buf_vec,
2812                                                 &head_idx, &buf_len,
2813                                                 VHOST_ACCESS_RO) < 0))
2814                         break;
2815
2816                 update_shadow_used_ring_split(vq, head_idx, 0);
2817
2818                 err = virtio_dev_pktmbuf_prep(dev, pkts[i], buf_len);
2819                 if (unlikely(err)) {
2820                         /*
2821                          * mbuf allocation fails for jumbo packets when external
2822                          * buffer allocation is not allowed and linear buffer
2823                          * is required. Drop this packet.
2824                          */
2825                         if (!allocerr_warned) {
2826                                 VHOST_LOG_DATA(ERR, "(%s) failed mbuf alloc of size %d from %s.\n",
2827                                         dev->ifname, buf_len, mbuf_pool->name);
2828                                 allocerr_warned = true;
2829                         }
2830                         dropped += 1;
2831                         i++;
2832                         break;
2833                 }
2834
2835                 err = desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts[i],
2836                                    mbuf_pool, legacy_ol_flags, 0, false);
2837                 if (unlikely(err)) {
2838                         if (!allocerr_warned) {
2839                                 VHOST_LOG_DATA(ERR, "(%s) failed to copy desc to mbuf.\n",
2840                                         dev->ifname);
2841                                 allocerr_warned = true;
2842                         }
2843                         dropped += 1;
2844                         i++;
2845                         break;
2846                 }
2847
2848         }
2849
2850         if (dropped)
2851                 rte_pktmbuf_free_bulk(&pkts[i - 1], count - i + 1);
2852
2853         vq->last_avail_idx += i;
2854
2855         do_data_copy_dequeue(vq);
2856         if (unlikely(i < count))
2857                 vq->shadow_used_idx = i;
2858         if (likely(vq->shadow_used_idx)) {
2859                 flush_shadow_used_ring_split(dev, vq);
2860                 vhost_vring_call_split(dev, vq);
2861         }
2862
2863         return (i - dropped);
2864 }
2865
2866 __rte_noinline
2867 static uint16_t
2868 virtio_dev_tx_split_legacy(struct virtio_net *dev,
2869         struct vhost_virtqueue *vq, struct rte_mempool *mbuf_pool,
2870         struct rte_mbuf **pkts, uint16_t count)
2871 {
2872         return virtio_dev_tx_split(dev, vq, mbuf_pool, pkts, count, true);
2873 }
2874
2875 __rte_noinline
2876 static uint16_t
2877 virtio_dev_tx_split_compliant(struct virtio_net *dev,
2878         struct vhost_virtqueue *vq, struct rte_mempool *mbuf_pool,
2879         struct rte_mbuf **pkts, uint16_t count)
2880 {
2881         return virtio_dev_tx_split(dev, vq, mbuf_pool, pkts, count, false);
2882 }
2883
2884 static __rte_always_inline int
2885 vhost_reserve_avail_batch_packed(struct virtio_net *dev,
2886                                  struct vhost_virtqueue *vq,
2887                                  struct rte_mbuf **pkts,
2888                                  uint16_t avail_idx,
2889                                  uintptr_t *desc_addrs,
2890                                  uint16_t *ids)
2891 {
2892         bool wrap = vq->avail_wrap_counter;
2893         struct vring_packed_desc *descs = vq->desc_packed;
2894         uint64_t lens[PACKED_BATCH_SIZE];
2895         uint64_t buf_lens[PACKED_BATCH_SIZE];
2896         uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2897         uint16_t flags, i;
2898
2899         if (unlikely(avail_idx & PACKED_BATCH_MASK))
2900                 return -1;
2901         if (unlikely((avail_idx + PACKED_BATCH_SIZE) > vq->size))
2902                 return -1;
2903
2904         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2905                 flags = descs[avail_idx + i].flags;
2906                 if (unlikely((wrap != !!(flags & VRING_DESC_F_AVAIL)) ||
2907                              (wrap == !!(flags & VRING_DESC_F_USED))  ||
2908                              (flags & PACKED_DESC_SINGLE_DEQUEUE_FLAG)))
2909                         return -1;
2910         }
2911
2912         rte_atomic_thread_fence(__ATOMIC_ACQUIRE);
2913
2914         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2915                 lens[i] = descs[avail_idx + i].len;
2916
2917         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2918                 desc_addrs[i] = vhost_iova_to_vva(dev, vq,
2919                                                   descs[avail_idx + i].addr,
2920                                                   &lens[i], VHOST_ACCESS_RW);
2921         }
2922
2923         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2924                 if (unlikely(!desc_addrs[i]))
2925                         return -1;
2926                 if (unlikely((lens[i] != descs[avail_idx + i].len)))
2927                         return -1;
2928         }
2929
2930         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2931                 if (virtio_dev_pktmbuf_prep(dev, pkts[i], lens[i]))
2932                         goto err;
2933         }
2934
2935         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2936                 buf_lens[i] = pkts[i]->buf_len - pkts[i]->data_off;
2937
2938         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2939                 if (unlikely(buf_lens[i] < (lens[i] - buf_offset)))
2940                         goto err;
2941         }
2942
2943         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2944                 pkts[i]->pkt_len = lens[i] - buf_offset;
2945                 pkts[i]->data_len = pkts[i]->pkt_len;
2946                 ids[i] = descs[avail_idx + i].id;
2947         }
2948
2949         return 0;
2950
2951 err:
2952         return -1;
2953 }
2954
2955 static __rte_always_inline int
2956 virtio_dev_tx_batch_packed(struct virtio_net *dev,
2957                            struct vhost_virtqueue *vq,
2958                            struct rte_mbuf **pkts,
2959                            bool legacy_ol_flags)
2960 {
2961         uint16_t avail_idx = vq->last_avail_idx;
2962         uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2963         struct virtio_net_hdr *hdr;
2964         uintptr_t desc_addrs[PACKED_BATCH_SIZE];
2965         uint16_t ids[PACKED_BATCH_SIZE];
2966         uint16_t i;
2967
2968         if (vhost_reserve_avail_batch_packed(dev, vq, pkts, avail_idx,
2969                                              desc_addrs, ids))
2970                 return -1;
2971
2972         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2973                 rte_prefetch0((void *)(uintptr_t)desc_addrs[i]);
2974
2975         vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2976                 rte_memcpy(rte_pktmbuf_mtod_offset(pkts[i], void *, 0),
2977                            (void *)(uintptr_t)(desc_addrs[i] + buf_offset),
2978                            pkts[i]->pkt_len);
2979
2980         if (virtio_net_with_host_offload(dev)) {
2981                 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2982                         hdr = (struct virtio_net_hdr *)(desc_addrs[i]);
2983                         vhost_dequeue_offload(dev, hdr, pkts[i], legacy_ol_flags);
2984                 }
2985         }
2986
2987         if (virtio_net_is_inorder(dev))
2988                 vhost_shadow_dequeue_batch_packed_inorder(vq,
2989                         ids[PACKED_BATCH_SIZE - 1]);
2990         else
2991                 vhost_shadow_dequeue_batch_packed(dev, vq, ids);
2992
2993         vq_inc_last_avail_packed(vq, PACKED_BATCH_SIZE);
2994
2995         return 0;
2996 }
2997
2998 static __rte_always_inline int
2999 vhost_dequeue_single_packed(struct virtio_net *dev,
3000                             struct vhost_virtqueue *vq,
3001                             struct rte_mempool *mbuf_pool,
3002                             struct rte_mbuf *pkts,
3003                             uint16_t *buf_id,
3004                             uint16_t *desc_count,
3005                             bool legacy_ol_flags)
3006 {
3007         struct buf_vector buf_vec[BUF_VECTOR_MAX];
3008         uint32_t buf_len;
3009         uint16_t nr_vec = 0;
3010         int err;
3011         static bool allocerr_warned;
3012
3013         if (unlikely(fill_vec_buf_packed(dev, vq,
3014                                          vq->last_avail_idx, desc_count,
3015                                          buf_vec, &nr_vec,
3016                                          buf_id, &buf_len,
3017                                          VHOST_ACCESS_RO) < 0))
3018                 return -1;
3019
3020         if (unlikely(virtio_dev_pktmbuf_prep(dev, pkts, buf_len))) {
3021                 if (!allocerr_warned) {
3022                         VHOST_LOG_DATA(ERR, "(%s) failed mbuf alloc of size %d from %s.\n",
3023                                 dev->ifname, buf_len, mbuf_pool->name);
3024                         allocerr_warned = true;
3025                 }
3026                 return -1;
3027         }
3028
3029         err = desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts,
3030                            mbuf_pool, legacy_ol_flags, 0, false);
3031         if (unlikely(err)) {
3032                 if (!allocerr_warned) {
3033                         VHOST_LOG_DATA(ERR, "(%s) failed to copy desc to mbuf.\n",
3034                                 dev->ifname);
3035                         allocerr_warned = true;
3036                 }
3037                 return -1;
3038         }
3039
3040         return 0;
3041 }
3042
3043 static __rte_always_inline int
3044 virtio_dev_tx_single_packed(struct virtio_net *dev,
3045                             struct vhost_virtqueue *vq,
3046                             struct rte_mempool *mbuf_pool,
3047                             struct rte_mbuf *pkts,
3048                             bool legacy_ol_flags)
3049 {
3050
3051         uint16_t buf_id, desc_count = 0;
3052         int ret;
3053
3054         ret = vhost_dequeue_single_packed(dev, vq, mbuf_pool, pkts, &buf_id,
3055                                         &desc_count, legacy_ol_flags);
3056
3057         if (likely(desc_count > 0)) {
3058                 if (virtio_net_is_inorder(dev))
3059                         vhost_shadow_dequeue_single_packed_inorder(vq, buf_id,
3060                                                                    desc_count);
3061                 else
3062                         vhost_shadow_dequeue_single_packed(vq, buf_id,
3063                                         desc_count);
3064
3065                 vq_inc_last_avail_packed(vq, desc_count);
3066         }
3067
3068         return ret;
3069 }
3070
3071 __rte_always_inline
3072 static uint16_t
3073 virtio_dev_tx_packed(struct virtio_net *dev,
3074                      struct vhost_virtqueue *__rte_restrict vq,
3075                      struct rte_mempool *mbuf_pool,
3076                      struct rte_mbuf **__rte_restrict pkts,
3077                      uint32_t count,
3078                      bool legacy_ol_flags)
3079 {
3080         uint32_t pkt_idx = 0;
3081
3082         if (rte_pktmbuf_alloc_bulk(mbuf_pool, pkts, count))
3083                 return 0;
3084
3085         do {
3086                 rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
3087
3088                 if (count - pkt_idx >= PACKED_BATCH_SIZE) {
3089                         if (!virtio_dev_tx_batch_packed(dev, vq,
3090                                                         &pkts[pkt_idx],
3091                                                         legacy_ol_flags)) {
3092                                 pkt_idx += PACKED_BATCH_SIZE;
3093                                 continue;
3094                         }
3095                 }
3096
3097                 if (virtio_dev_tx_single_packed(dev, vq, mbuf_pool,
3098                                                 pkts[pkt_idx],
3099                                                 legacy_ol_flags))
3100                         break;
3101                 pkt_idx++;
3102         } while (pkt_idx < count);
3103
3104         if (pkt_idx != count)
3105                 rte_pktmbuf_free_bulk(&pkts[pkt_idx], count - pkt_idx);
3106
3107         if (vq->shadow_used_idx) {
3108                 do_data_copy_dequeue(vq);
3109
3110                 vhost_flush_dequeue_shadow_packed(dev, vq);
3111                 vhost_vring_call_packed(dev, vq);
3112         }
3113
3114         return pkt_idx;
3115 }
3116
3117 __rte_noinline
3118 static uint16_t
3119 virtio_dev_tx_packed_legacy(struct virtio_net *dev,
3120         struct vhost_virtqueue *__rte_restrict vq, struct rte_mempool *mbuf_pool,
3121         struct rte_mbuf **__rte_restrict pkts, uint32_t count)
3122 {
3123         return virtio_dev_tx_packed(dev, vq, mbuf_pool, pkts, count, true);
3124 }
3125
3126 __rte_noinline
3127 static uint16_t
3128 virtio_dev_tx_packed_compliant(struct virtio_net *dev,
3129         struct vhost_virtqueue *__rte_restrict vq, struct rte_mempool *mbuf_pool,
3130         struct rte_mbuf **__rte_restrict pkts, uint32_t count)
3131 {
3132         return virtio_dev_tx_packed(dev, vq, mbuf_pool, pkts, count, false);
3133 }
3134
3135 uint16_t
3136 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
3137         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
3138 {
3139         struct virtio_net *dev;
3140         struct rte_mbuf *rarp_mbuf = NULL;
3141         struct vhost_virtqueue *vq;
3142         int16_t success = 1;
3143
3144         dev = get_device(vid);
3145         if (!dev)
3146                 return 0;
3147
3148         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
3149                 VHOST_LOG_DATA(ERR, "(%s) %s: built-in vhost net backend is disabled.\n",
3150                                 dev->ifname, __func__);
3151                 return 0;
3152         }
3153
3154         if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
3155                 VHOST_LOG_DATA(ERR, "(%s) %s: invalid virtqueue idx %d.\n",
3156                                 dev->ifname, __func__, queue_id);
3157                 return 0;
3158         }
3159
3160         vq = dev->virtqueue[queue_id];
3161
3162         if (unlikely(rte_spinlock_trylock(&vq->access_lock) == 0))
3163                 return 0;
3164
3165         if (unlikely(!vq->enabled)) {
3166                 count = 0;
3167                 goto out_access_unlock;
3168         }
3169
3170         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
3171                 vhost_user_iotlb_rd_lock(vq);
3172
3173         if (unlikely(!vq->access_ok))
3174                 if (unlikely(vring_translate(dev, vq) < 0)) {
3175                         count = 0;
3176                         goto out;
3177                 }
3178
3179         /*
3180          * Construct a RARP broadcast packet, and inject it to the "pkts"
3181          * array, to looks like that guest actually send such packet.
3182          *
3183          * Check user_send_rarp() for more information.
3184          *
3185          * broadcast_rarp shares a cacheline in the virtio_net structure
3186          * with some fields that are accessed during enqueue and
3187          * __atomic_compare_exchange_n causes a write if performed compare
3188          * and exchange. This could result in false sharing between enqueue
3189          * and dequeue.
3190          *
3191          * Prevent unnecessary false sharing by reading broadcast_rarp first
3192          * and only performing compare and exchange if the read indicates it
3193          * is likely to be set.
3194          */
3195         if (unlikely(__atomic_load_n(&dev->broadcast_rarp, __ATOMIC_ACQUIRE) &&
3196                         __atomic_compare_exchange_n(&dev->broadcast_rarp,
3197                         &success, 0, 0, __ATOMIC_RELEASE, __ATOMIC_RELAXED))) {
3198
3199                 rarp_mbuf = rte_net_make_rarp_packet(mbuf_pool, &dev->mac);
3200                 if (rarp_mbuf == NULL) {
3201                         VHOST_LOG_DATA(ERR, "(%s) failed to make RARP packet.\n", dev->ifname);
3202                         count = 0;
3203                         goto out;
3204                 }
3205                 /*
3206                  * Inject it to the head of "pkts" array, so that switch's mac
3207                  * learning table will get updated first.
3208                  */
3209                 pkts[0] = rarp_mbuf;
3210                 vhost_queue_stats_update(dev, vq, pkts, 1);
3211                 pkts++;
3212                 count -= 1;
3213         }
3214
3215         if (vq_is_packed(dev)) {
3216                 if (dev->flags & VIRTIO_DEV_LEGACY_OL_FLAGS)
3217                         count = virtio_dev_tx_packed_legacy(dev, vq, mbuf_pool, pkts, count);
3218                 else
3219                         count = virtio_dev_tx_packed_compliant(dev, vq, mbuf_pool, pkts, count);
3220         } else {
3221                 if (dev->flags & VIRTIO_DEV_LEGACY_OL_FLAGS)
3222                         count = virtio_dev_tx_split_legacy(dev, vq, mbuf_pool, pkts, count);
3223                 else
3224                         count = virtio_dev_tx_split_compliant(dev, vq, mbuf_pool, pkts, count);
3225         }
3226
3227         vhost_queue_stats_update(dev, vq, pkts, count);
3228
3229 out:
3230         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
3231                 vhost_user_iotlb_rd_unlock(vq);
3232
3233 out_access_unlock:
3234         rte_spinlock_unlock(&vq->access_lock);
3235
3236         if (unlikely(rarp_mbuf != NULL))
3237                 count += 1;
3238
3239         return count;
3240 }
3241
3242 static __rte_always_inline uint16_t
3243 async_poll_dequeue_completed_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
3244                 struct rte_mbuf **pkts, uint16_t count, int16_t dma_id,
3245                 uint16_t vchan_id, bool legacy_ol_flags)
3246 {
3247         uint16_t start_idx, from, i;
3248         uint16_t nr_cpl_pkts = 0;
3249         struct async_inflight_info *pkts_info = vq->async->pkts_info;
3250
3251         vhost_async_dma_check_completed(dev, dma_id, vchan_id, VHOST_DMA_MAX_COPY_COMPLETE);
3252
3253         start_idx = async_get_first_inflight_pkt_idx(vq);
3254
3255         from = start_idx;
3256         while (vq->async->pkts_cmpl_flag[from] && count--) {
3257                 vq->async->pkts_cmpl_flag[from] = false;
3258                 from = (from + 1) & (vq->size - 1);
3259                 nr_cpl_pkts++;
3260         }
3261
3262         if (nr_cpl_pkts == 0)
3263                 return 0;
3264
3265         for (i = 0; i < nr_cpl_pkts; i++) {
3266                 from = (start_idx + i) & (vq->size - 1);
3267                 pkts[i] = pkts_info[from].mbuf;
3268
3269                 if (virtio_net_with_host_offload(dev))
3270                         vhost_dequeue_offload(dev, &pkts_info[from].nethdr, pkts[i],
3271                                               legacy_ol_flags);
3272         }
3273
3274         /* write back completed descs to used ring and update used idx */
3275         write_back_completed_descs_split(vq, nr_cpl_pkts);
3276         __atomic_add_fetch(&vq->used->idx, nr_cpl_pkts, __ATOMIC_RELEASE);
3277         vhost_vring_call_split(dev, vq);
3278
3279         vq->async->pkts_inflight_n -= nr_cpl_pkts;
3280
3281         return nr_cpl_pkts;
3282 }
3283
3284 static __rte_always_inline uint16_t
3285 virtio_dev_tx_async_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
3286                 struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count,
3287                 int16_t dma_id, uint16_t vchan_id, bool legacy_ol_flags)
3288 {
3289         static bool allocerr_warned;
3290         bool dropped = false;
3291         uint16_t free_entries;
3292         uint16_t pkt_idx, slot_idx = 0;
3293         uint16_t nr_done_pkts = 0;
3294         uint16_t pkt_err = 0;
3295         uint16_t n_xfer;
3296         struct vhost_async *async = vq->async;
3297         struct async_inflight_info *pkts_info = async->pkts_info;
3298         struct rte_mbuf *pkts_prealloc[MAX_PKT_BURST];
3299         uint16_t pkts_size = count;
3300
3301         /**
3302          * The ordering between avail index and
3303          * desc reads needs to be enforced.
3304          */
3305         free_entries = __atomic_load_n(&vq->avail->idx, __ATOMIC_ACQUIRE) -
3306                         vq->last_avail_idx;
3307         if (free_entries == 0)
3308                 goto out;
3309
3310         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
3311
3312         async_iter_reset(async);
3313
3314         count = RTE_MIN(count, MAX_PKT_BURST);
3315         count = RTE_MIN(count, free_entries);
3316         VHOST_LOG_DATA(DEBUG, "(%s) about to dequeue %u buffers\n",
3317                         dev->ifname, count);
3318
3319         if (rte_pktmbuf_alloc_bulk(mbuf_pool, pkts_prealloc, count))
3320                 goto out;
3321
3322         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
3323                 uint16_t head_idx = 0;
3324                 uint16_t nr_vec = 0;
3325                 uint16_t to;
3326                 uint32_t buf_len;
3327                 int err;
3328                 struct buf_vector buf_vec[BUF_VECTOR_MAX];
3329                 struct rte_mbuf *pkt = pkts_prealloc[pkt_idx];
3330
3331                 if (unlikely(fill_vec_buf_split(dev, vq, vq->last_avail_idx,
3332                                                 &nr_vec, buf_vec,
3333                                                 &head_idx, &buf_len,
3334                                                 VHOST_ACCESS_RO) < 0)) {
3335                         dropped = true;
3336                         break;
3337                 }
3338
3339                 err = virtio_dev_pktmbuf_prep(dev, pkt, buf_len);
3340                 if (unlikely(err)) {
3341                         /**
3342                          * mbuf allocation fails for jumbo packets when external
3343                          * buffer allocation is not allowed and linear buffer
3344                          * is required. Drop this packet.
3345                          */
3346                         if (!allocerr_warned) {
3347                                 VHOST_LOG_DATA(ERR,
3348                                         "(%s) %s: Failed mbuf alloc of size %d from %s\n",
3349                                         dev->ifname, __func__, buf_len, mbuf_pool->name);
3350                                 allocerr_warned = true;
3351                         }
3352                         dropped = true;
3353                         break;
3354                 }
3355
3356                 slot_idx = (async->pkts_idx + pkt_idx) & (vq->size - 1);
3357                 err = desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkt, mbuf_pool,
3358                                         legacy_ol_flags, slot_idx, true);
3359                 if (unlikely(err)) {
3360                         if (!allocerr_warned) {
3361                                 VHOST_LOG_DATA(ERR,
3362                                         "(%s) %s: Failed to offload copies to async channel.\n",
3363                                         dev->ifname, __func__);
3364                                 allocerr_warned = true;
3365                         }
3366                         dropped = true;
3367                         break;
3368                 }
3369
3370                 pkts_info[slot_idx].mbuf = pkt;
3371
3372                 /* store used descs */
3373                 to = async->desc_idx_split & (vq->size - 1);
3374                 async->descs_split[to].id = head_idx;
3375                 async->descs_split[to].len = 0;
3376                 async->desc_idx_split++;
3377
3378                 vq->last_avail_idx++;
3379         }
3380
3381         if (unlikely(dropped))
3382                 rte_pktmbuf_free_bulk(&pkts_prealloc[pkt_idx], count - pkt_idx);
3383
3384         n_xfer = vhost_async_dma_transfer(dev, vq, dma_id, vchan_id, async->pkts_idx,
3385                                           async->iov_iter, pkt_idx);
3386
3387         async->pkts_inflight_n += n_xfer;
3388
3389         pkt_err = pkt_idx - n_xfer;
3390         if (unlikely(pkt_err)) {
3391                 VHOST_LOG_DATA(DEBUG, "(%s) %s: failed to transfer data.\n",
3392                                 dev->ifname, __func__);
3393
3394                 pkt_idx = n_xfer;
3395                 /* recover available ring */
3396                 vq->last_avail_idx -= pkt_err;
3397
3398                 /**
3399                  * recover async channel copy related structures and free pktmbufs
3400                  * for error pkts.
3401                  */
3402                 async->desc_idx_split -= pkt_err;
3403                 while (pkt_err-- > 0) {
3404                         rte_pktmbuf_free(pkts_info[slot_idx & (vq->size - 1)].mbuf);
3405                         slot_idx--;
3406                 }
3407         }
3408
3409         async->pkts_idx += pkt_idx;
3410         if (async->pkts_idx >= vq->size)
3411                 async->pkts_idx -= vq->size;
3412
3413 out:
3414         /* DMA device may serve other queues, unconditionally check completed. */
3415         nr_done_pkts = async_poll_dequeue_completed_split(dev, vq, pkts, pkts_size,
3416                                                           dma_id, vchan_id, legacy_ol_flags);
3417
3418         return nr_done_pkts;
3419 }
3420
3421 __rte_noinline
3422 static uint16_t
3423 virtio_dev_tx_async_split_legacy(struct virtio_net *dev,
3424                 struct vhost_virtqueue *vq, struct rte_mempool *mbuf_pool,
3425                 struct rte_mbuf **pkts, uint16_t count,
3426                 int16_t dma_id, uint16_t vchan_id)
3427 {
3428         return virtio_dev_tx_async_split(dev, vq, mbuf_pool,
3429                                 pkts, count, dma_id, vchan_id, true);
3430 }
3431
3432 __rte_noinline
3433 static uint16_t
3434 virtio_dev_tx_async_split_compliant(struct virtio_net *dev,
3435                 struct vhost_virtqueue *vq, struct rte_mempool *mbuf_pool,
3436                 struct rte_mbuf **pkts, uint16_t count,
3437                 int16_t dma_id, uint16_t vchan_id)
3438 {
3439         return virtio_dev_tx_async_split(dev, vq, mbuf_pool,
3440                                 pkts, count, dma_id, vchan_id, false);
3441 }
3442
3443 uint16_t
3444 rte_vhost_async_try_dequeue_burst(int vid, uint16_t queue_id,
3445         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count,
3446         int *nr_inflight, int16_t dma_id, uint16_t vchan_id)
3447 {
3448         struct virtio_net *dev;
3449         struct rte_mbuf *rarp_mbuf = NULL;
3450         struct vhost_virtqueue *vq;
3451         int16_t success = 1;
3452
3453         dev = get_device(vid);
3454         if (!dev || !nr_inflight)
3455                 return 0;
3456
3457         *nr_inflight = -1;
3458
3459         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
3460                 VHOST_LOG_DATA(ERR, "(%s) %s: built-in vhost net backend is disabled.\n",
3461                                 dev->ifname, __func__);
3462                 return 0;
3463         }
3464
3465         if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
3466                 VHOST_LOG_DATA(ERR, "(%s) %s: invalid virtqueue idx %d.\n",
3467                                 dev->ifname, __func__, queue_id);
3468                 return 0;
3469         }
3470
3471         if (unlikely(dma_id < 0 || dma_id >= RTE_DMADEV_DEFAULT_MAX)) {
3472                 VHOST_LOG_DATA(ERR, "(%s) %s: invalid dma id %d.\n",
3473                                 dev->ifname, __func__, dma_id);
3474                 return 0;
3475         }
3476
3477         if (unlikely(!dma_copy_track[dma_id].vchans ||
3478                                 !dma_copy_track[dma_id].vchans[vchan_id].pkts_cmpl_flag_addr)) {
3479                 VHOST_LOG_DATA(ERR, "(%s) %s: invalid channel %d:%u.\n", dev->ifname, __func__,
3480                                 dma_id, vchan_id);
3481                 return 0;
3482         }
3483
3484         vq = dev->virtqueue[queue_id];
3485
3486         if (unlikely(rte_spinlock_trylock(&vq->access_lock) == 0))
3487                 return 0;
3488
3489         if (unlikely(vq->enabled == 0)) {
3490                 count = 0;
3491                 goto out_access_unlock;
3492         }
3493
3494         if (unlikely(!vq->async)) {
3495                 VHOST_LOG_DATA(ERR, "(%s) %s: async not registered for queue id %d.\n",
3496                                 dev->ifname, __func__, queue_id);
3497                 count = 0;
3498                 goto out_access_unlock;
3499         }
3500
3501         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
3502                 vhost_user_iotlb_rd_lock(vq);
3503
3504         if (unlikely(vq->access_ok == 0))
3505                 if (unlikely(vring_translate(dev, vq) < 0)) {
3506                         count = 0;
3507                         goto out;
3508                 }
3509
3510         /*
3511          * Construct a RARP broadcast packet, and inject it to the "pkts"
3512          * array, to looks like that guest actually send such packet.
3513          *
3514          * Check user_send_rarp() for more information.
3515          *
3516          * broadcast_rarp shares a cacheline in the virtio_net structure
3517          * with some fields that are accessed during enqueue and
3518          * __atomic_compare_exchange_n causes a write if performed compare
3519          * and exchange. This could result in false sharing between enqueue
3520          * and dequeue.
3521          *
3522          * Prevent unnecessary false sharing by reading broadcast_rarp first
3523          * and only performing compare and exchange if the read indicates it
3524          * is likely to be set.
3525          */
3526         if (unlikely(__atomic_load_n(&dev->broadcast_rarp, __ATOMIC_ACQUIRE) &&
3527                         __atomic_compare_exchange_n(&dev->broadcast_rarp,
3528                         &success, 0, 0, __ATOMIC_RELEASE, __ATOMIC_RELAXED))) {
3529
3530                 rarp_mbuf = rte_net_make_rarp_packet(mbuf_pool, &dev->mac);
3531                 if (rarp_mbuf == NULL) {
3532                         VHOST_LOG_DATA(ERR, "Failed to make RARP packet.\n");
3533                         count = 0;
3534                         goto out;
3535                 }
3536                 /*
3537                  * Inject it to the head of "pkts" array, so that switch's mac
3538                  * learning table will get updated first.
3539                  */
3540                 pkts[0] = rarp_mbuf;
3541                 pkts++;
3542                 count -= 1;
3543         }
3544
3545         if (unlikely(vq_is_packed(dev))) {
3546                 static bool not_support_pack_log;
3547                 if (!not_support_pack_log) {
3548                         VHOST_LOG_DATA(ERR,
3549                                 "(%s) %s: async dequeue does not support packed ring.\n",
3550                                 dev->ifname, __func__);
3551                         not_support_pack_log = true;
3552                 }
3553                 count = 0;
3554                 goto out;
3555         }
3556
3557         if (dev->flags & VIRTIO_DEV_LEGACY_OL_FLAGS)
3558                 count = virtio_dev_tx_async_split_legacy(dev, vq, mbuf_pool, pkts,
3559                                                          count, dma_id, vchan_id);
3560         else
3561                 count = virtio_dev_tx_async_split_compliant(dev, vq, mbuf_pool, pkts,
3562                                                             count, dma_id, vchan_id);
3563
3564         *nr_inflight = vq->async->pkts_inflight_n;
3565
3566 out:
3567         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
3568                 vhost_user_iotlb_rd_unlock(vq);
3569
3570 out_access_unlock:
3571         rte_spinlock_unlock(&vq->access_lock);
3572
3573         if (unlikely(rarp_mbuf != NULL))
3574                 count += 1;
3575
3576         return count;
3577 }