lib/librte_vhost/virtio_net.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2016 Intel Corporation
   3  */
   4
   5 #include <stdint.h>
   6 #include <stdbool.h>
   7 #include <linux/virtio_net.h>
   8
   9 #include <rte_mbuf.h>
  10 #include <rte_memcpy.h>
  11 #include <rte_ether.h>
  12 #include <rte_ip.h>
  13 #include <rte_vhost.h>
  14 #include <rte_tcp.h>
  15 #include <rte_udp.h>
  16 #include <rte_sctp.h>
  17 #include <rte_arp.h>
  18 #include <rte_spinlock.h>
  19 #include <rte_malloc.h>
  20
  21 #include "iotlb.h"
  22 #include "vhost.h"
  23
  24 #define MAX_PKT_BURST 32
  25
  26 #define MAX_BATCH_LEN 256
  27
  28 static bool
  29 is_valid_virt_queue_idx(uint32_t idx, int is_tx, uint32_t nr_vring)
  30 {
  31         return (is_tx ^ (idx & 1)) == 0 && idx < nr_vring;
  32 }
  33
  34 static __rte_always_inline struct vring_desc *
  35 alloc_copy_ind_table(struct virtio_net *dev, struct vhost_virtqueue *vq,
  36                                          struct vring_desc *desc)
  37 {
  38         struct vring_desc *idesc;
  39         uint64_t src, dst;
  40         uint64_t len, remain = desc->len;
  41         uint64_t desc_addr = desc->addr;
  42
  43         idesc = rte_malloc(__func__, desc->len, 0);
  44         if (unlikely(!idesc))
  45                 return 0;
  46
  47         dst = (uint64_t)(uintptr_t)idesc;
  48
  49         while (remain) {
  50                 len = remain;
  51                 src = vhost_iova_to_vva(dev, vq, desc_addr, &len,
  52                                 VHOST_ACCESS_RO);
  53                 if (unlikely(!src || !len)) {
  54                         rte_free(idesc);
  55                         return 0;
  56                 }
  57
  58                 rte_memcpy((void *)(uintptr_t)dst, (void *)(uintptr_t)src, len);
  59
  60                 remain -= len;
  61                 dst += len;
  62                 desc_addr += len;
  63         }
  64
  65         return idesc;
  66 }
  67
  68 static __rte_always_inline void
  69 free_ind_table(struct vring_desc *idesc)
  70 {
  71         rte_free(idesc);
  72 }
  73
  74 static __rte_always_inline void
  75 do_flush_shadow_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq,
  76                           uint16_t to, uint16_t from, uint16_t size)
  77 {
  78         rte_memcpy(&vq->used->ring[to],
  79                         &vq->shadow_used_ring[from],
  80                         size * sizeof(struct vring_used_elem));
  81         vhost_log_used_vring(dev, vq,
  82                         offsetof(struct vring_used, ring[to]),
  83                         size * sizeof(struct vring_used_elem));
  84 }
  85
  86 static __rte_always_inline void
  87 flush_shadow_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq)
  88 {
  89         uint16_t used_idx = vq->last_used_idx & (vq->size - 1);
  90
  91         if (used_idx + vq->shadow_used_idx <= vq->size) {
  92                 do_flush_shadow_used_ring(dev, vq, used_idx, 0,
  93                                           vq->shadow_used_idx);
  94         } else {
  95                 uint16_t size;
  96
  97                 /* update used ring interval [used_idx, vq->size] */
  98                 size = vq->size - used_idx;
  99                 do_flush_shadow_used_ring(dev, vq, used_idx, 0, size);
 100
 101                 /* update the left half used ring interval [0, left_size] */
 102                 do_flush_shadow_used_ring(dev, vq, 0, size,
 103                                           vq->shadow_used_idx - size);
 104         }
 105         vq->last_used_idx += vq->shadow_used_idx;
 106
 107         rte_smp_wmb();
 108
 109         *(volatile uint16_t *)&vq->used->idx += vq->shadow_used_idx;
 110         vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
 111                 sizeof(vq->used->idx));
 112 }
 113
 114 static __rte_always_inline void
 115 update_shadow_used_ring(struct vhost_virtqueue *vq,
 116                          uint16_t desc_idx, uint16_t len)
 117 {
 118         uint16_t i = vq->shadow_used_idx++;
 119
 120         vq->shadow_used_ring[i].id  = desc_idx;
 121         vq->shadow_used_ring[i].len = len;
 122 }
 123
 124 static inline void
 125 do_data_copy_enqueue(struct virtio_net *dev, struct vhost_virtqueue *vq)
 126 {
 127         struct batch_copy_elem *elem = vq->batch_copy_elems;
 128         uint16_t count = vq->batch_copy_nb_elems;
 129         int i;
 130
 131         for (i = 0; i < count; i++) {
 132                 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
 133                 vhost_log_write(dev, elem[i].log_addr, elem[i].len);
 134                 PRINT_PACKET(dev, (uintptr_t)elem[i].dst, elem[i].len, 0);
 135         }
 136 }
 137
 138 static inline void
 139 do_data_copy_dequeue(struct vhost_virtqueue *vq)
 140 {
 141         struct batch_copy_elem *elem = vq->batch_copy_elems;
 142         uint16_t count = vq->batch_copy_nb_elems;
 143         int i;
 144
 145         for (i = 0; i < count; i++)
 146                 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
 147 }
 148
 149 /* avoid write operation when necessary, to lessen cache issues */
 150 #define ASSIGN_UNLESS_EQUAL(var, val) do {      \
 151         if ((var) != (val))                     \
 152                 (var) = (val);                  \
 153 } while (0)
 154
 155 static void
 156 virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
 157 {
 158         uint64_t csum_l4 = m_buf->ol_flags & PKT_TX_L4_MASK;
 159
 160         if (m_buf->ol_flags & PKT_TX_TCP_SEG)
 161                 csum_l4 |= PKT_TX_TCP_CKSUM;
 162
 163         if (csum_l4) {
 164                 net_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 165                 net_hdr->csum_start = m_buf->l2_len + m_buf->l3_len;
 166
 167                 switch (csum_l4) {
 168                 case PKT_TX_TCP_CKSUM:
 169                         net_hdr->csum_offset = (offsetof(struct tcp_hdr,
 170                                                 cksum));
 171                         break;
 172                 case PKT_TX_UDP_CKSUM:
 173                         net_hdr->csum_offset = (offsetof(struct udp_hdr,
 174                                                 dgram_cksum));
 175                         break;
 176                 case PKT_TX_SCTP_CKSUM:
 177                         net_hdr->csum_offset = (offsetof(struct sctp_hdr,
 178                                                 cksum));
 179                         break;
 180                 }
 181         } else {
 182                 ASSIGN_UNLESS_EQUAL(net_hdr->csum_start, 0);
 183                 ASSIGN_UNLESS_EQUAL(net_hdr->csum_offset, 0);
 184                 ASSIGN_UNLESS_EQUAL(net_hdr->flags, 0);
 185         }
 186
 187         /* IP cksum verification cannot be bypassed, then calculate here */
 188         if (m_buf->ol_flags & PKT_TX_IP_CKSUM) {
 189                 struct ipv4_hdr *ipv4_hdr;
 190
 191                 ipv4_hdr = rte_pktmbuf_mtod_offset(m_buf, struct ipv4_hdr *,
 192                                                    m_buf->l2_len);
 193                 ipv4_hdr->hdr_checksum = rte_ipv4_cksum(ipv4_hdr);
 194         }
 195
 196         if (m_buf->ol_flags & PKT_TX_TCP_SEG) {
 197                 if (m_buf->ol_flags & PKT_TX_IPV4)
 198                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 199                 else
 200                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 201                 net_hdr->gso_size = m_buf->tso_segsz;
 202                 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len
 203                                         + m_buf->l4_len;
 204         } else if (m_buf->ol_flags & PKT_TX_UDP_SEG) {
 205                 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
 206                 net_hdr->gso_size = m_buf->tso_segsz;
 207                 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len +
 208                         m_buf->l4_len;
 209         } else {
 210                 ASSIGN_UNLESS_EQUAL(net_hdr->gso_type, 0);
 211                 ASSIGN_UNLESS_EQUAL(net_hdr->gso_size, 0);
 212                 ASSIGN_UNLESS_EQUAL(net_hdr->hdr_len, 0);
 213         }
 214 }
 215
 216 static __rte_always_inline int
 217 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 218                   struct vring_desc *descs, struct rte_mbuf *m,
 219                   uint16_t desc_idx, uint32_t size)
 220 {
 221         uint32_t desc_avail, desc_offset;
 222         uint32_t mbuf_avail, mbuf_offset;
 223         uint32_t cpy_len;
 224         uint64_t dlen;
 225         struct vring_desc *desc;
 226         uint64_t desc_addr;
 227         /* A counter to avoid desc dead loop chain */
 228         uint16_t nr_desc = 1;
 229         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 230         uint16_t copy_nb = vq->batch_copy_nb_elems;
 231         int error = 0;
 232
 233         desc = &descs[desc_idx];
 234         dlen = desc->len;
 235         desc_addr = vhost_iova_to_vva(dev, vq, desc->addr,
 236                                         &dlen, VHOST_ACCESS_RW);
 237         /*
 238          * Checking of 'desc_addr' placed outside of 'unlikely' macro to avoid
 239          * performance issue with some versions of gcc (4.8.4 and 5.3.0) which
 240          * otherwise stores offset on the stack instead of in a register.
 241          */
 242         if (unlikely(dlen != desc->len || desc->len < dev->vhost_hlen) ||
 243                         !desc_addr) {
 244                 error = -1;
 245                 goto out;
 246         }
 247
 248         rte_prefetch0((void *)(uintptr_t)desc_addr);
 249
 250         virtio_enqueue_offload(m, (struct virtio_net_hdr *)(uintptr_t)desc_addr);
 251         vhost_log_write(dev, desc->addr, dev->vhost_hlen);
 252         PRINT_PACKET(dev, (uintptr_t)desc_addr, dev->vhost_hlen, 0);
 253
 254         desc_offset = dev->vhost_hlen;
 255         desc_avail  = desc->len - dev->vhost_hlen;
 256
 257         mbuf_avail  = rte_pktmbuf_data_len(m);
 258         mbuf_offset = 0;
 259         while (mbuf_avail != 0 || m->next != NULL) {
 260                 /* done with current mbuf, fetch next */
 261                 if (mbuf_avail == 0) {
 262                         m = m->next;
 263
 264                         mbuf_offset = 0;
 265                         mbuf_avail  = rte_pktmbuf_data_len(m);
 266                 }
 267
 268                 /* done with current desc buf, fetch next */
 269                 if (desc_avail == 0) {
 270                         if ((desc->flags & VRING_DESC_F_NEXT) == 0) {
 271                                 /* Room in vring buffer is not enough */
 272                                 error = -1;
 273                                 goto out;
 274                         }
 275                         if (unlikely(desc->next >= size || ++nr_desc > size)) {
 276                                 error = -1;
 277                                 goto out;
 278                         }
 279
 280                         desc = &descs[desc->next];
 281                         dlen = desc->len;
 282                         desc_addr = vhost_iova_to_vva(dev, vq, desc->addr,
 283                                                         &dlen,
 284                                                         VHOST_ACCESS_RW);
 285                         if (unlikely(!desc_addr || dlen != desc->len)) {
 286                                 error = -1;
 287                                 goto out;
 288                         }
 289
 290                         desc_offset = 0;
 291                         desc_avail  = desc->len;
 292                 }
 293
 294                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 295                 if (likely(cpy_len > MAX_BATCH_LEN || copy_nb >= vq->size)) {
 296                         rte_memcpy((void *)((uintptr_t)(desc_addr +
 297                                                         desc_offset)),
 298                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 299                                 cpy_len);
 300                         vhost_log_write(dev, desc->addr + desc_offset, cpy_len);
 301                         PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
 302                                      cpy_len, 0);
 303                 } else {
 304                         batch_copy[copy_nb].dst =
 305                                 (void *)((uintptr_t)(desc_addr + desc_offset));
 306                         batch_copy[copy_nb].src =
 307                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
 308                         batch_copy[copy_nb].log_addr = desc->addr + desc_offset;
 309                         batch_copy[copy_nb].len = cpy_len;
 310                         copy_nb++;
 311                 }
 312
 313                 mbuf_avail  -= cpy_len;
 314                 mbuf_offset += cpy_len;
 315                 desc_avail  -= cpy_len;
 316                 desc_offset += cpy_len;
 317         }
 318
 319 out:
 320         vq->batch_copy_nb_elems = copy_nb;
 321
 322         return error;
 323 }
 324
 325 /**
 326  * This function adds buffers to the virtio devices RX virtqueue. Buffers can
 327  * be received from the physical port or from another virtio device. A packet
 328  * count is returned to indicate the number of packets that are successfully
 329  * added to the RX queue. This function works when the mbuf is scattered, but
 330  * it doesn't support the mergeable feature.
 331  */
 332 static __rte_always_inline uint32_t
 333 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
 334               struct rte_mbuf **pkts, uint32_t count)
 335 {
 336         struct vhost_virtqueue *vq;
 337         uint16_t avail_idx, free_entries, start_idx;
 338         uint16_t desc_indexes[MAX_PKT_BURST];
 339         struct vring_desc *descs;
 340         uint16_t used_idx;
 341         uint32_t i, sz;
 342
 343         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
 344         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
 345                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
 346                         dev->vid, __func__, queue_id);
 347                 return 0;
 348         }
 349
 350         vq = dev->virtqueue[queue_id];
 351
 352         rte_spinlock_lock(&vq->access_lock);
 353
 354         if (unlikely(vq->enabled == 0))
 355                 goto out_access_unlock;
 356
 357         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 358                 vhost_user_iotlb_rd_lock(vq);
 359
 360         if (unlikely(vq->access_ok == 0)) {
 361                 if (unlikely(vring_translate(dev, vq) < 0)) {
 362                         count = 0;
 363                         goto out;
 364                 }
 365         }
 366
 367         avail_idx = *((volatile uint16_t *)&vq->avail->idx);
 368         start_idx = vq->last_used_idx;
 369         free_entries = avail_idx - start_idx;
 370         count = RTE_MIN(count, free_entries);
 371         count = RTE_MIN(count, (uint32_t)MAX_PKT_BURST);
 372         if (count == 0)
 373                 goto out;
 374
 375         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) start_idx %d | end_idx %d\n",
 376                 dev->vid, start_idx, start_idx + count);
 377
 378         vq->batch_copy_nb_elems = 0;
 379
 380         /* Retrieve all of the desc indexes first to avoid caching issues. */
 381         rte_prefetch0(&vq->avail->ring[start_idx & (vq->size - 1)]);
 382         for (i = 0; i < count; i++) {
 383                 used_idx = (start_idx + i) & (vq->size - 1);
 384                 desc_indexes[i] = vq->avail->ring[used_idx];
 385                 vq->used->ring[used_idx].id = desc_indexes[i];
 386                 vq->used->ring[used_idx].len = pkts[i]->pkt_len +
 387                                                dev->vhost_hlen;
 388                 vhost_log_used_vring(dev, vq,
 389                         offsetof(struct vring_used, ring[used_idx]),
 390                         sizeof(vq->used->ring[used_idx]));
 391         }
 392
 393         rte_prefetch0(&vq->desc[desc_indexes[0]]);
 394         for (i = 0; i < count; i++) {
 395                 struct vring_desc *idesc = NULL;
 396                 uint16_t desc_idx = desc_indexes[i];
 397                 int err;
 398
 399                 if (vq->desc[desc_idx].flags & VRING_DESC_F_INDIRECT) {
 400                         uint64_t dlen = vq->desc[desc_idx].len;
 401                         descs = (struct vring_desc *)(uintptr_t)
 402                                 vhost_iova_to_vva(dev,
 403                                                 vq, vq->desc[desc_idx].addr,
 404                                                 &dlen, VHOST_ACCESS_RO);
 405                         if (unlikely(!descs)) {
 406                                 count = i;
 407                                 break;
 408                         }
 409
 410                         if (unlikely(dlen < vq->desc[desc_idx].len)) {
 411                                 /*
 412                                  * The indirect desc table is not contiguous
 413                                  * in process VA space, we have to copy it.
 414                                  */
 415                                 idesc = alloc_copy_ind_table(dev, vq,
 416                                                         &vq->desc[desc_idx]);
 417                                 if (unlikely(!idesc))
 418                                         break;
 419
 420                                 descs = idesc;
 421                         }
 422
 423                         desc_idx = 0;
 424                         sz = vq->desc[desc_idx].len / sizeof(*descs);
 425                 } else {
 426                         descs = vq->desc;
 427                         sz = vq->size;
 428                 }
 429
 430                 err = copy_mbuf_to_desc(dev, vq, descs, pkts[i], desc_idx, sz);
 431                 if (unlikely(err)) {
 432                         count = i;
 433                         free_ind_table(idesc);
 434                         break;
 435                 }
 436
 437                 if (i + 1 < count)
 438                         rte_prefetch0(&vq->desc[desc_indexes[i+1]]);
 439
 440                 if (unlikely(!!idesc))
 441                         free_ind_table(idesc);
 442         }
 443
 444         do_data_copy_enqueue(dev, vq);
 445
 446         rte_smp_wmb();
 447
 448         *(volatile uint16_t *)&vq->used->idx += count;
 449         vq->last_used_idx += count;
 450         vhost_log_used_vring(dev, vq,
 451                 offsetof(struct vring_used, idx),
 452                 sizeof(vq->used->idx));
 453
 454         vhost_vring_call(dev, vq);
 455 out:
 456         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 457                 vhost_user_iotlb_rd_unlock(vq);
 458
 459 out_access_unlock:
 460         rte_spinlock_unlock(&vq->access_lock);
 461
 462         return count;
 463 }
 464
 465 static __rte_always_inline int
 466 fill_vec_buf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 467                          uint32_t avail_idx, uint32_t *vec_idx,
 468                          struct buf_vector *buf_vec, uint16_t *desc_chain_head,
 469                          uint16_t *desc_chain_len)
 470 {
 471         uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
 472         uint32_t vec_id = *vec_idx;
 473         uint32_t len    = 0;
 474         uint64_t dlen;
 475         struct vring_desc *descs = vq->desc;
 476         struct vring_desc *idesc = NULL;
 477
 478         *desc_chain_head = idx;
 479
 480         if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
 481                 dlen = vq->desc[idx].len;
 482                 descs = (struct vring_desc *)(uintptr_t)
 483                         vhost_iova_to_vva(dev, vq, vq->desc[idx].addr,
 484                                                 &dlen,
 485                                                 VHOST_ACCESS_RO);
 486                 if (unlikely(!descs))
 487                         return -1;
 488
 489                 if (unlikely(dlen < vq->desc[idx].len)) {
 490                         /*
 491                          * The indirect desc table is not contiguous
 492                          * in process VA space, we have to copy it.
 493                          */
 494                         idesc = alloc_copy_ind_table(dev, vq, &vq->desc[idx]);
 495                         if (unlikely(!idesc))
 496                                 return -1;
 497
 498                         descs = idesc;
 499                 }
 500
 501                 idx = 0;
 502         }
 503
 504         while (1) {
 505                 if (unlikely(vec_id >= BUF_VECTOR_MAX || idx >= vq->size)) {
 506                         free_ind_table(idesc);
 507                         return -1;
 508                 }
 509
 510                 len += descs[idx].len;
 511                 buf_vec[vec_id].buf_addr = descs[idx].addr;
 512                 buf_vec[vec_id].buf_len  = descs[idx].len;
 513                 buf_vec[vec_id].desc_idx = idx;
 514                 vec_id++;
 515
 516                 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
 517                         break;
 518
 519                 idx = descs[idx].next;
 520         }
 521
 522         *desc_chain_len = len;
 523         *vec_idx = vec_id;
 524
 525         if (unlikely(!!idesc))
 526                 free_ind_table(idesc);
 527
 528         return 0;
 529 }
 530
 531 /*
 532  * Returns -1 on fail, 0 on success
 533  */
 534 static inline int
 535 reserve_avail_buf_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
 536                                 uint32_t size, struct buf_vector *buf_vec,
 537                                 uint16_t *num_buffers, uint16_t avail_head)
 538 {
 539         uint16_t cur_idx;
 540         uint32_t vec_idx = 0;
 541         uint16_t tries = 0;
 542
 543         uint16_t head_idx = 0;
 544         uint16_t len = 0;
 545
 546         *num_buffers = 0;
 547         cur_idx  = vq->last_avail_idx;
 548
 549         while (size > 0) {
 550                 if (unlikely(cur_idx == avail_head))
 551                         return -1;
 552
 553                 if (unlikely(fill_vec_buf(dev, vq, cur_idx, &vec_idx, buf_vec,
 554                                                 &head_idx, &len) < 0))
 555                         return -1;
 556                 len = RTE_MIN(len, size);
 557                 update_shadow_used_ring(vq, head_idx, len);
 558                 size -= len;
 559
 560                 cur_idx++;
 561                 tries++;
 562                 *num_buffers += 1;
 563
 564                 /*
 565                  * if we tried all available ring items, and still
 566                  * can't get enough buf, it means something abnormal
 567                  * happened.
 568                  */
 569                 if (unlikely(tries >= vq->size))
 570                         return -1;
 571         }
 572
 573         return 0;
 574 }
 575
 576 static __rte_always_inline int
 577 copy_mbuf_to_desc_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
 578                             struct rte_mbuf *m, struct buf_vector *buf_vec,
 579                             uint16_t num_buffers)
 580 {
 581         uint32_t vec_idx = 0;
 582         uint64_t desc_addr;
 583         uint32_t mbuf_offset, mbuf_avail;
 584         uint32_t desc_offset, desc_avail;
 585         uint32_t cpy_len;
 586         uint64_t dlen;
 587         uint64_t hdr_addr, hdr_phys_addr;
 588         struct rte_mbuf *hdr_mbuf;
 589         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 590         uint16_t copy_nb = vq->batch_copy_nb_elems;
 591         int error = 0;
 592
 593         if (unlikely(m == NULL)) {
 594                 error = -1;
 595                 goto out;
 596         }
 597
 598         dlen = buf_vec[vec_idx].buf_len;
 599         desc_addr = vhost_iova_to_vva(dev, vq, buf_vec[vec_idx].buf_addr,
 600                                                 &dlen, VHOST_ACCESS_RW);
 601         if (dlen != buf_vec[vec_idx].buf_len ||
 602                         buf_vec[vec_idx].buf_len < dev->vhost_hlen ||
 603                         !desc_addr) {
 604                 error = -1;
 605                 goto out;
 606         }
 607
 608         hdr_mbuf = m;
 609         hdr_addr = desc_addr;
 610         hdr_phys_addr = buf_vec[vec_idx].buf_addr;
 611         rte_prefetch0((void *)(uintptr_t)hdr_addr);
 612
 613         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
 614                 dev->vid, num_buffers);
 615
 616         desc_avail  = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
 617         desc_offset = dev->vhost_hlen;
 618
 619         mbuf_avail  = rte_pktmbuf_data_len(m);
 620         mbuf_offset = 0;
 621         while (mbuf_avail != 0 || m->next != NULL) {
 622                 /* done with current desc buf, get the next one */
 623                 if (desc_avail == 0) {
 624                         vec_idx++;
 625                         dlen = buf_vec[vec_idx].buf_len;
 626                         desc_addr =
 627                                 vhost_iova_to_vva(dev, vq,
 628                                         buf_vec[vec_idx].buf_addr,
 629                                         &dlen,
 630                                         VHOST_ACCESS_RW);
 631                         if (unlikely(!desc_addr ||
 632                                         dlen != buf_vec[vec_idx].buf_len)) {
 633                                 error = -1;
 634                                 goto out;
 635                         }
 636
 637                         /* Prefetch buffer address. */
 638                         rte_prefetch0((void *)(uintptr_t)desc_addr);
 639                         desc_offset = 0;
 640                         desc_avail  = buf_vec[vec_idx].buf_len;
 641                 }
 642
 643                 /* done with current mbuf, get the next one */
 644                 if (mbuf_avail == 0) {
 645                         m = m->next;
 646
 647                         mbuf_offset = 0;
 648                         mbuf_avail  = rte_pktmbuf_data_len(m);
 649                 }
 650
 651                 if (hdr_addr) {
 652                         struct virtio_net_hdr_mrg_rxbuf *hdr;
 653
 654                         hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)
 655                                 hdr_addr;
 656                         virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
 657                         ASSIGN_UNLESS_EQUAL(hdr->num_buffers, num_buffers);
 658
 659                         vhost_log_write(dev, hdr_phys_addr, dev->vhost_hlen);
 660                         PRINT_PACKET(dev, (uintptr_t)hdr_addr,
 661                                      dev->vhost_hlen, 0);
 662
 663                         hdr_addr = 0;
 664                 }
 665
 666                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 667
 668                 if (likely(cpy_len > MAX_BATCH_LEN || copy_nb >= vq->size)) {
 669                         rte_memcpy((void *)((uintptr_t)(desc_addr +
 670                                                         desc_offset)),
 671                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 672                                 cpy_len);
 673                         vhost_log_write(dev,
 674                                 buf_vec[vec_idx].buf_addr + desc_offset,
 675                                 cpy_len);
 676                         PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
 677                                 cpy_len, 0);
 678                 } else {
 679                         batch_copy[copy_nb].dst =
 680                                 (void *)((uintptr_t)(desc_addr + desc_offset));
 681                         batch_copy[copy_nb].src =
 682                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
 683                         batch_copy[copy_nb].log_addr =
 684                                 buf_vec[vec_idx].buf_addr + desc_offset;
 685                         batch_copy[copy_nb].len = cpy_len;
 686                         copy_nb++;
 687                 }
 688
 689                 mbuf_avail  -= cpy_len;
 690                 mbuf_offset += cpy_len;
 691                 desc_avail  -= cpy_len;
 692                 desc_offset += cpy_len;
 693         }
 694
 695 out:
 696         vq->batch_copy_nb_elems = copy_nb;
 697
 698         return error;
 699 }
 700
 701 static __rte_always_inline uint32_t
 702 virtio_dev_merge_rx(struct virtio_net *dev, uint16_t queue_id,
 703         struct rte_mbuf **pkts, uint32_t count)
 704 {
 705         struct vhost_virtqueue *vq;
 706         uint32_t pkt_idx = 0;
 707         uint16_t num_buffers;
 708         struct buf_vector buf_vec[BUF_VECTOR_MAX];
 709         uint16_t avail_head;
 710
 711         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
 712         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
 713                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
 714                         dev->vid, __func__, queue_id);
 715                 return 0;
 716         }
 717
 718         vq = dev->virtqueue[queue_id];
 719
 720         rte_spinlock_lock(&vq->access_lock);
 721
 722         if (unlikely(vq->enabled == 0))
 723                 goto out_access_unlock;
 724
 725         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 726                 vhost_user_iotlb_rd_lock(vq);
 727
 728         if (unlikely(vq->access_ok == 0))
 729                 if (unlikely(vring_translate(dev, vq) < 0))
 730                         goto out;
 731
 732         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
 733         if (count == 0)
 734                 goto out;
 735
 736         vq->batch_copy_nb_elems = 0;
 737
 738         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
 739
 740         vq->shadow_used_idx = 0;
 741         avail_head = *((volatile uint16_t *)&vq->avail->idx);
 742         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
 743                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
 744
 745                 if (unlikely(reserve_avail_buf_mergeable(dev, vq,
 746                                                 pkt_len, buf_vec, &num_buffers,
 747                                                 avail_head) < 0)) {
 748                         VHOST_LOG_DEBUG(VHOST_DATA,
 749                                 "(%d) failed to get enough desc from vring\n",
 750                                 dev->vid);
 751                         vq->shadow_used_idx -= num_buffers;
 752                         break;
 753                 }
 754
 755                 VHOST_LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
 756                         dev->vid, vq->last_avail_idx,
 757                         vq->last_avail_idx + num_buffers);
 758
 759                 if (copy_mbuf_to_desc_mergeable(dev, vq, pkts[pkt_idx],
 760                                                 buf_vec, num_buffers) < 0) {
 761                         vq->shadow_used_idx -= num_buffers;
 762                         break;
 763                 }
 764
 765                 vq->last_avail_idx += num_buffers;
 766         }
 767
 768         do_data_copy_enqueue(dev, vq);
 769
 770         if (likely(vq->shadow_used_idx)) {
 771                 flush_shadow_used_ring(dev, vq);
 772                 vhost_vring_call(dev, vq);
 773         }
 774
 775 out:
 776         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 777                 vhost_user_iotlb_rd_unlock(vq);
 778
 779 out_access_unlock:
 780         rte_spinlock_unlock(&vq->access_lock);
 781
 782         return pkt_idx;
 783 }
 784
 785 uint16_t
 786 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
 787         struct rte_mbuf **pkts, uint16_t count)
 788 {
 789         struct virtio_net *dev = get_device(vid);
 790
 791         if (!dev)
 792                 return 0;
 793
 794         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
 795                 RTE_LOG(ERR, VHOST_DATA,
 796                         "(%d) %s: built-in vhost net backend is disabled.\n",
 797                         dev->vid, __func__);
 798                 return 0;
 799         }
 800
 801         if (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF))
 802                 return virtio_dev_merge_rx(dev, queue_id, pkts, count);
 803         else
 804                 return virtio_dev_rx(dev, queue_id, pkts, count);
 805 }
 806
 807 static inline bool
 808 virtio_net_with_host_offload(struct virtio_net *dev)
 809 {
 810         if (dev->features &
 811                         ((1ULL << VIRTIO_NET_F_CSUM) |
 812                          (1ULL << VIRTIO_NET_F_HOST_ECN) |
 813                          (1ULL << VIRTIO_NET_F_HOST_TSO4) |
 814                          (1ULL << VIRTIO_NET_F_HOST_TSO6) |
 815                          (1ULL << VIRTIO_NET_F_HOST_UFO)))
 816                 return true;
 817
 818         return false;
 819 }
 820
 821 static void
 822 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
 823 {
 824         struct ipv4_hdr *ipv4_hdr;
 825         struct ipv6_hdr *ipv6_hdr;
 826         void *l3_hdr = NULL;
 827         struct ether_hdr *eth_hdr;
 828         uint16_t ethertype;
 829
 830         eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
 831
 832         m->l2_len = sizeof(struct ether_hdr);
 833         ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
 834
 835         if (ethertype == ETHER_TYPE_VLAN) {
 836                 struct vlan_hdr *vlan_hdr = (struct vlan_hdr *)(eth_hdr + 1);
 837
 838                 m->l2_len += sizeof(struct vlan_hdr);
 839                 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
 840         }
 841
 842         l3_hdr = (char *)eth_hdr + m->l2_len;
 843
 844         switch (ethertype) {
 845         case ETHER_TYPE_IPv4:
 846                 ipv4_hdr = l3_hdr;
 847                 *l4_proto = ipv4_hdr->next_proto_id;
 848                 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
 849                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 850                 m->ol_flags |= PKT_TX_IPV4;
 851                 break;
 852         case ETHER_TYPE_IPv6:
 853                 ipv6_hdr = l3_hdr;
 854                 *l4_proto = ipv6_hdr->proto;
 855                 m->l3_len = sizeof(struct ipv6_hdr);
 856                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 857                 m->ol_flags |= PKT_TX_IPV6;
 858                 break;
 859         default:
 860                 m->l3_len = 0;
 861                 *l4_proto = 0;
 862                 *l4_hdr = NULL;
 863                 break;
 864         }
 865 }
 866
 867 static __rte_always_inline void
 868 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
 869 {
 870         uint16_t l4_proto = 0;
 871         void *l4_hdr = NULL;
 872         struct tcp_hdr *tcp_hdr = NULL;
 873
 874         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
 875                 return;
 876
 877         parse_ethernet(m, &l4_proto, &l4_hdr);
 878         if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 879                 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
 880                         switch (hdr->csum_offset) {
 881                         case (offsetof(struct tcp_hdr, cksum)):
 882                                 if (l4_proto == IPPROTO_TCP)
 883                                         m->ol_flags |= PKT_TX_TCP_CKSUM;
 884                                 break;
 885                         case (offsetof(struct udp_hdr, dgram_cksum)):
 886                                 if (l4_proto == IPPROTO_UDP)
 887                                         m->ol_flags |= PKT_TX_UDP_CKSUM;
 888                                 break;
 889                         case (offsetof(struct sctp_hdr, cksum)):
 890                                 if (l4_proto == IPPROTO_SCTP)
 891                                         m->ol_flags |= PKT_TX_SCTP_CKSUM;
 892                                 break;
 893                         default:
 894                                 break;
 895                         }
 896                 }
 897         }
 898
 899         if (l4_hdr && hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 900                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 901                 case VIRTIO_NET_HDR_GSO_TCPV4:
 902                 case VIRTIO_NET_HDR_GSO_TCPV6:
 903                         tcp_hdr = l4_hdr;
 904                         m->ol_flags |= PKT_TX_TCP_SEG;
 905                         m->tso_segsz = hdr->gso_size;
 906                         m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
 907                         break;
 908                 case VIRTIO_NET_HDR_GSO_UDP:
 909                         m->ol_flags |= PKT_TX_UDP_SEG;
 910                         m->tso_segsz = hdr->gso_size;
 911                         m->l4_len = sizeof(struct udp_hdr);
 912                         break;
 913                 default:
 914                         RTE_LOG(WARNING, VHOST_DATA,
 915                                 "unsupported gso type %u.\n", hdr->gso_type);
 916                         break;
 917                 }
 918         }
 919 }
 920
 921 static __rte_always_inline void
 922 put_zmbuf(struct zcopy_mbuf *zmbuf)
 923 {
 924         zmbuf->in_use = 0;
 925 }
 926
 927 static __rte_always_inline int
 928 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 929                   struct vring_desc *descs, uint16_t max_desc,
 930                   struct rte_mbuf *m, uint16_t desc_idx,
 931                   struct rte_mempool *mbuf_pool)
 932 {
 933         struct vring_desc *desc;
 934         uint64_t desc_addr;
 935         uint32_t desc_avail, desc_offset;
 936         uint32_t mbuf_avail, mbuf_offset;
 937         uint32_t cpy_len;
 938         uint64_t dlen;
 939         struct rte_mbuf *cur = m, *prev = m;
 940         struct virtio_net_hdr *hdr = NULL;
 941         /* A counter to avoid desc dead loop chain */
 942         uint32_t nr_desc = 1;
 943         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 944         uint16_t copy_nb = vq->batch_copy_nb_elems;
 945         int error = 0;
 946
 947         desc = &descs[desc_idx];
 948         if (unlikely((desc->len < dev->vhost_hlen)) ||
 949                         (desc->flags & VRING_DESC_F_INDIRECT)) {
 950                 error = -1;
 951                 goto out;
 952         }
 953
 954         dlen = desc->len;
 955         desc_addr = vhost_iova_to_vva(dev,
 956                                         vq, desc->addr,
 957                                         &dlen,
 958                                         VHOST_ACCESS_RO);
 959         if (unlikely(!desc_addr || dlen != desc->len)) {
 960                 error = -1;
 961                 goto out;
 962         }
 963
 964         if (virtio_net_with_host_offload(dev)) {
 965                 hdr = (struct virtio_net_hdr *)((uintptr_t)desc_addr);
 966                 rte_prefetch0(hdr);
 967         }
 968
 969         /*
 970          * A virtio driver normally uses at least 2 desc buffers
 971          * for Tx: the first for storing the header, and others
 972          * for storing the data.
 973          */
 974         if (likely((desc->len == dev->vhost_hlen) &&
 975                    (desc->flags & VRING_DESC_F_NEXT) != 0)) {
 976                 desc = &descs[desc->next];
 977                 if (unlikely(desc->flags & VRING_DESC_F_INDIRECT)) {
 978                         error = -1;
 979                         goto out;
 980                 }
 981
 982                 dlen = desc->len;
 983                 desc_addr = vhost_iova_to_vva(dev,
 984                                                         vq, desc->addr,
 985                                                         &dlen,
 986                                                         VHOST_ACCESS_RO);
 987                 if (unlikely(!desc_addr || dlen != desc->len)) {
 988                         error = -1;
 989                         goto out;
 990                 }
 991
 992                 desc_offset = 0;
 993                 desc_avail  = desc->len;
 994                 nr_desc    += 1;
 995         } else {
 996                 desc_avail  = desc->len - dev->vhost_hlen;
 997                 desc_offset = dev->vhost_hlen;
 998         }
 999
1000         rte_prefetch0((void *)(uintptr_t)(desc_addr + desc_offset));
1001
1002         PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset), desc_avail, 0);
1003
1004         mbuf_offset = 0;
1005         mbuf_avail  = m->buf_len - RTE_PKTMBUF_HEADROOM;
1006         while (1) {
1007                 uint64_t hpa;
1008
1009                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
1010
1011                 /*
1012                  * A desc buf might across two host physical pages that are
1013                  * not continuous. In such case (gpa_to_hpa returns 0), data
1014                  * will be copied even though zero copy is enabled.
1015                  */
1016                 if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
1017                                         desc->addr + desc_offset, cpy_len)))) {
1018                         cur->data_len = cpy_len;
1019                         cur->data_off = 0;
1020                         cur->buf_addr = (void *)(uintptr_t)(desc_addr
1021                                 + desc_offset);
1022                         cur->buf_iova = hpa;
1023
1024                         /*
1025                          * In zero copy mode, one mbuf can only reference data
1026                          * for one or partial of one desc buff.
1027                          */
1028                         mbuf_avail = cpy_len;
1029                 } else {
1030                         if (likely(cpy_len > MAX_BATCH_LEN ||
1031                                    copy_nb >= vq->size ||
1032                                    (hdr && cur == m))) {
1033                                 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
1034                                                                    mbuf_offset),
1035                                            (void *)((uintptr_t)(desc_addr +
1036                                                                 desc_offset)),
1037                                            cpy_len);
1038                         } else {
1039                                 batch_copy[copy_nb].dst =
1040                                         rte_pktmbuf_mtod_offset(cur, void *,
1041                                                                 mbuf_offset);
1042                                 batch_copy[copy_nb].src =
1043                                         (void *)((uintptr_t)(desc_addr +
1044                                                              desc_offset));
1045                                 batch_copy[copy_nb].len = cpy_len;
1046                                 copy_nb++;
1047                         }
1048                 }
1049
1050                 mbuf_avail  -= cpy_len;
1051                 mbuf_offset += cpy_len;
1052                 desc_avail  -= cpy_len;
1053                 desc_offset += cpy_len;
1054
1055                 /* This desc reaches to its end, get the next one */
1056                 if (desc_avail == 0) {
1057                         if ((desc->flags & VRING_DESC_F_NEXT) == 0)
1058                                 break;
1059
1060                         if (unlikely(desc->next >= max_desc ||
1061                                      ++nr_desc > max_desc)) {
1062                                 error = -1;
1063                                 goto out;
1064                         }
1065                         desc = &descs[desc->next];
1066                         if (unlikely(desc->flags & VRING_DESC_F_INDIRECT)) {
1067                                 error = -1;
1068                                 goto out;
1069                         }
1070
1071                         dlen = desc->len;
1072                         desc_addr = vhost_iova_to_vva(dev,
1073                                                         vq, desc->addr,
1074                                                         &dlen, VHOST_ACCESS_RO);
1075                         if (unlikely(!desc_addr || dlen != desc->len)) {
1076                                 error = -1;
1077                                 goto out;
1078                         }
1079
1080                         rte_prefetch0((void *)(uintptr_t)desc_addr);
1081
1082                         desc_offset = 0;
1083                         desc_avail  = desc->len;
1084
1085                         PRINT_PACKET(dev, (uintptr_t)desc_addr, desc->len, 0);
1086                 }
1087
1088                 /*
1089                  * This mbuf reaches to its end, get a new one
1090                  * to hold more data.
1091                  */
1092                 if (mbuf_avail == 0) {
1093                         cur = rte_pktmbuf_alloc(mbuf_pool);
1094                         if (unlikely(cur == NULL)) {
1095                                 RTE_LOG(ERR, VHOST_DATA, "Failed to "
1096                                         "allocate memory for mbuf.\n");
1097                                 error = -1;
1098                                 goto out;
1099                         }
1100                         if (unlikely(dev->dequeue_zero_copy))
1101                                 rte_mbuf_refcnt_update(cur, 1);
1102
1103                         prev->next = cur;
1104                         prev->data_len = mbuf_offset;
1105                         m->nb_segs += 1;
1106                         m->pkt_len += mbuf_offset;
1107                         prev = cur;
1108
1109                         mbuf_offset = 0;
1110                         mbuf_avail  = cur->buf_len - RTE_PKTMBUF_HEADROOM;
1111                 }
1112         }
1113
1114         prev->data_len = mbuf_offset;
1115         m->pkt_len    += mbuf_offset;
1116
1117         if (hdr)
1118                 vhost_dequeue_offload(hdr, m);
1119
1120 out:
1121         vq->batch_copy_nb_elems = copy_nb;
1122
1123         return error;
1124 }
1125
1126 static __rte_always_inline void
1127 update_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq,
1128                  uint32_t used_idx, uint32_t desc_idx)
1129 {
1130         vq->used->ring[used_idx].id  = desc_idx;
1131         vq->used->ring[used_idx].len = 0;
1132         vhost_log_used_vring(dev, vq,
1133                         offsetof(struct vring_used, ring[used_idx]),
1134                         sizeof(vq->used->ring[used_idx]));
1135 }
1136
1137 static __rte_always_inline void
1138 update_used_idx(struct virtio_net *dev, struct vhost_virtqueue *vq,
1139                 uint32_t count)
1140 {
1141         if (unlikely(count == 0))
1142                 return;
1143
1144         rte_smp_wmb();
1145         rte_smp_rmb();
1146
1147         vq->used->idx += count;
1148         vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
1149                         sizeof(vq->used->idx));
1150         vhost_vring_call(dev, vq);
1151 }
1152
1153 static __rte_always_inline struct zcopy_mbuf *
1154 get_zmbuf(struct vhost_virtqueue *vq)
1155 {
1156         uint16_t i;
1157         uint16_t last;
1158         int tries = 0;
1159
1160         /* search [last_zmbuf_idx, zmbuf_size) */
1161         i = vq->last_zmbuf_idx;
1162         last = vq->zmbuf_size;
1163
1164 again:
1165         for (; i < last; i++) {
1166                 if (vq->zmbufs[i].in_use == 0) {
1167                         vq->last_zmbuf_idx = i + 1;
1168                         vq->zmbufs[i].in_use = 1;
1169                         return &vq->zmbufs[i];
1170                 }
1171         }
1172
1173         tries++;
1174         if (tries == 1) {
1175                 /* search [0, last_zmbuf_idx) */
1176                 i = 0;
1177                 last = vq->last_zmbuf_idx;
1178                 goto again;
1179         }
1180
1181         return NULL;
1182 }
1183
1184 static __rte_always_inline bool
1185 mbuf_is_consumed(struct rte_mbuf *m)
1186 {
1187         while (m) {
1188                 if (rte_mbuf_refcnt_read(m) > 1)
1189                         return false;
1190                 m = m->next;
1191         }
1192
1193         return true;
1194 }
1195
1196 static __rte_always_inline void
1197 restore_mbuf(struct rte_mbuf *m)
1198 {
1199         uint32_t mbuf_size, priv_size;
1200
1201         while (m) {
1202                 priv_size = rte_pktmbuf_priv_size(m->pool);
1203                 mbuf_size = sizeof(struct rte_mbuf) + priv_size;
1204                 /* start of buffer is after mbuf structure and priv data */
1205
1206                 m->buf_addr = (char *)m + mbuf_size;
1207                 m->buf_iova = rte_mempool_virt2iova(m) + mbuf_size;
1208                 m = m->next;
1209         }
1210 }
1211
1212 uint16_t
1213 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
1214         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1215 {
1216         struct virtio_net *dev;
1217         struct rte_mbuf *rarp_mbuf = NULL;
1218         struct vhost_virtqueue *vq;
1219         uint32_t desc_indexes[MAX_PKT_BURST];
1220         uint32_t used_idx;
1221         uint32_t i = 0;
1222         uint16_t free_entries;
1223         uint16_t avail_idx;
1224
1225         dev = get_device(vid);
1226         if (!dev)
1227                 return 0;
1228
1229         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
1230                 RTE_LOG(ERR, VHOST_DATA,
1231                         "(%d) %s: built-in vhost net backend is disabled.\n",
1232                         dev->vid, __func__);
1233                 return 0;
1234         }
1235
1236         if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
1237                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
1238                         dev->vid, __func__, queue_id);
1239                 return 0;
1240         }
1241
1242         vq = dev->virtqueue[queue_id];
1243
1244         if (unlikely(rte_spinlock_trylock(&vq->access_lock) == 0))
1245                 return 0;
1246
1247         if (unlikely(vq->enabled == 0))
1248                 goto out_access_unlock;
1249
1250         vq->batch_copy_nb_elems = 0;
1251
1252         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1253                 vhost_user_iotlb_rd_lock(vq);
1254
1255         if (unlikely(vq->access_ok == 0))
1256                 if (unlikely(vring_translate(dev, vq) < 0))
1257                         goto out;
1258
1259         if (unlikely(dev->dequeue_zero_copy)) {
1260                 struct zcopy_mbuf *zmbuf, *next;
1261                 int nr_updated = 0;
1262
1263                 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1264                      zmbuf != NULL; zmbuf = next) {
1265                         next = TAILQ_NEXT(zmbuf, next);
1266
1267                         if (mbuf_is_consumed(zmbuf->mbuf)) {
1268                                 used_idx = vq->last_used_idx++ & (vq->size - 1);
1269                                 update_used_ring(dev, vq, used_idx,
1270                                                  zmbuf->desc_idx);
1271                                 nr_updated += 1;
1272
1273                                 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1274                                 restore_mbuf(zmbuf->mbuf);
1275                                 rte_pktmbuf_free(zmbuf->mbuf);
1276                                 put_zmbuf(zmbuf);
1277                                 vq->nr_zmbuf -= 1;
1278                         }
1279                 }
1280
1281                 update_used_idx(dev, vq, nr_updated);
1282         }
1283
1284         /*
1285          * Construct a RARP broadcast packet, and inject it to the "pkts"
1286          * array, to looks like that guest actually send such packet.
1287          *
1288          * Check user_send_rarp() for more information.
1289          *
1290          * broadcast_rarp shares a cacheline in the virtio_net structure
1291          * with some fields that are accessed during enqueue and
1292          * rte_atomic16_cmpset() causes a write if using cmpxchg. This could
1293          * result in false sharing between enqueue and dequeue.
1294          *
1295          * Prevent unnecessary false sharing by reading broadcast_rarp first
1296          * and only performing cmpset if the read indicates it is likely to
1297          * be set.
1298          */
1299
1300         if (unlikely(rte_atomic16_read(&dev->broadcast_rarp) &&
1301                         rte_atomic16_cmpset((volatile uint16_t *)
1302                                 &dev->broadcast_rarp.cnt, 1, 0))) {
1303
1304                 rarp_mbuf = rte_net_make_rarp_packet(mbuf_pool, &dev->mac);
1305                 if (rarp_mbuf == NULL) {
1306                         RTE_LOG(ERR, VHOST_DATA,
1307                                 "Failed to make RARP packet.\n");
1308                         return 0;
1309                 }
1310                 count -= 1;
1311         }
1312
1313         free_entries = *((volatile uint16_t *)&vq->avail->idx) -
1314                         vq->last_avail_idx;
1315         if (free_entries == 0)
1316                 goto out;
1317
1318         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1319
1320         /* Prefetch available and used ring */
1321         avail_idx = vq->last_avail_idx & (vq->size - 1);
1322         used_idx  = vq->last_used_idx  & (vq->size - 1);
1323         rte_prefetch0(&vq->avail->ring[avail_idx]);
1324         rte_prefetch0(&vq->used->ring[used_idx]);
1325
1326         count = RTE_MIN(count, MAX_PKT_BURST);
1327         count = RTE_MIN(count, free_entries);
1328         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1329                         dev->vid, count);
1330
1331         /* Retrieve all of the head indexes first to avoid caching issues. */
1332         for (i = 0; i < count; i++) {
1333                 avail_idx = (vq->last_avail_idx + i) & (vq->size - 1);
1334                 used_idx  = (vq->last_used_idx  + i) & (vq->size - 1);
1335                 desc_indexes[i] = vq->avail->ring[avail_idx];
1336
1337                 if (likely(dev->dequeue_zero_copy == 0))
1338                         update_used_ring(dev, vq, used_idx, desc_indexes[i]);
1339         }
1340
1341         /* Prefetch descriptor index. */
1342         rte_prefetch0(&vq->desc[desc_indexes[0]]);
1343         for (i = 0; i < count; i++) {
1344                 struct vring_desc *desc, *idesc = NULL;
1345                 uint16_t sz, idx;
1346                 uint64_t dlen;
1347                 int err;
1348
1349                 if (likely(i + 1 < count))
1350                         rte_prefetch0(&vq->desc[desc_indexes[i + 1]]);
1351
1352                 if (vq->desc[desc_indexes[i]].flags & VRING_DESC_F_INDIRECT) {
1353                         dlen = vq->desc[desc_indexes[i]].len;
1354                         desc = (struct vring_desc *)(uintptr_t)
1355                                 vhost_iova_to_vva(dev, vq,
1356                                                 vq->desc[desc_indexes[i]].addr,
1357                                                 &dlen,
1358                                                 VHOST_ACCESS_RO);
1359                         if (unlikely(!desc))
1360                                 break;
1361
1362                         if (unlikely(dlen < vq->desc[desc_indexes[i]].len)) {
1363                                 /*
1364                                  * The indirect desc table is not contiguous
1365                                  * in process VA space, we have to copy it.
1366                                  */
1367                                 idesc = alloc_copy_ind_table(dev, vq,
1368                                                 &vq->desc[desc_indexes[i]]);
1369                                 if (unlikely(!idesc))
1370                                         break;
1371
1372                                 desc = idesc;
1373                         }
1374
1375                         rte_prefetch0(desc);
1376                         sz = vq->desc[desc_indexes[i]].len / sizeof(*desc);
1377                         idx = 0;
1378                 } else {
1379                         desc = vq->desc;
1380                         sz = vq->size;
1381                         idx = desc_indexes[i];
1382                 }
1383
1384                 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
1385                 if (unlikely(pkts[i] == NULL)) {
1386                         RTE_LOG(ERR, VHOST_DATA,
1387                                 "Failed to allocate memory for mbuf.\n");
1388                         free_ind_table(idesc);
1389                         break;
1390                 }
1391
1392                 err = copy_desc_to_mbuf(dev, vq, desc, sz, pkts[i], idx,
1393                                         mbuf_pool);
1394                 if (unlikely(err)) {
1395                         rte_pktmbuf_free(pkts[i]);
1396                         free_ind_table(idesc);
1397                         break;
1398                 }
1399
1400                 if (unlikely(dev->dequeue_zero_copy)) {
1401                         struct zcopy_mbuf *zmbuf;
1402
1403                         zmbuf = get_zmbuf(vq);
1404                         if (!zmbuf) {
1405                                 rte_pktmbuf_free(pkts[i]);
1406                                 free_ind_table(idesc);
1407                                 break;
1408                         }
1409                         zmbuf->mbuf = pkts[i];
1410                         zmbuf->desc_idx = desc_indexes[i];
1411
1412                         /*
1413                          * Pin lock the mbuf; we will check later to see
1414                          * whether the mbuf is freed (when we are the last
1415                          * user) or not. If that's the case, we then could
1416                          * update the used ring safely.
1417                          */
1418                         rte_mbuf_refcnt_update(pkts[i], 1);
1419
1420                         vq->nr_zmbuf += 1;
1421                         TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1422                 }
1423
1424                 if (unlikely(!!idesc))
1425                         free_ind_table(idesc);
1426         }
1427         vq->last_avail_idx += i;
1428
1429         if (likely(dev->dequeue_zero_copy == 0)) {
1430                 do_data_copy_dequeue(vq);
1431                 vq->last_used_idx += i;
1432                 update_used_idx(dev, vq, i);
1433         }
1434
1435 out:
1436         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1437                 vhost_user_iotlb_rd_unlock(vq);
1438
1439 out_access_unlock:
1440         rte_spinlock_unlock(&vq->access_lock);
1441
1442         if (unlikely(rarp_mbuf != NULL)) {
1443                 /*
1444                  * Inject it to the head of "pkts" array, so that switch's mac
1445                  * learning table will get updated first.
1446                  */
1447                 memmove(&pkts[1], pkts, i * sizeof(struct rte_mbuf *));
1448                 pkts[0] = rarp_mbuf;
1449                 i += 1;
1450         }
1451
1452         return i;
1453 }