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 desc_chunck_len;
 225         struct vring_desc *desc;
 226         uint64_t desc_addr, desc_gaddr;
 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         desc_chunck_len = desc->len;
 235         desc_gaddr = desc->addr;
 236         desc_addr = vhost_iova_to_vva(dev, vq, desc_gaddr,
 237                                         &desc_chunck_len, VHOST_ACCESS_RW);
 238         /*
 239          * Checking of 'desc_addr' placed outside of 'unlikely' macro to avoid
 240          * performance issue with some versions of gcc (4.8.4 and 5.3.0) which
 241          * otherwise stores offset on the stack instead of in a register.
 242          */
 243         if (unlikely(desc->len < dev->vhost_hlen) || !desc_addr) {
 244                 error = -1;
 245                 goto out;
 246         }
 247
 248         rte_prefetch0((void *)(uintptr_t)desc_addr);
 249
 250         if (likely(desc_chunck_len >= dev->vhost_hlen)) {
 251                 virtio_enqueue_offload(m,
 252                                 (struct virtio_net_hdr *)(uintptr_t)desc_addr);
 253                 PRINT_PACKET(dev, (uintptr_t)desc_addr, dev->vhost_hlen, 0);
 254                 vhost_log_write(dev, desc_gaddr, dev->vhost_hlen);
 255         } else {
 256                 struct virtio_net_hdr vnet_hdr;
 257                 uint64_t remain = dev->vhost_hlen;
 258                 uint64_t len;
 259                 uint64_t src = (uint64_t)(uintptr_t)&vnet_hdr, dst;
 260                 uint64_t guest_addr = desc_gaddr;
 261
 262                 virtio_enqueue_offload(m, &vnet_hdr);
 263
 264                 while (remain) {
 265                         len = remain;
 266                         dst = vhost_iova_to_vva(dev, vq, guest_addr,
 267                                         &len, VHOST_ACCESS_RW);
 268                         if (unlikely(!dst || !len)) {
 269                                 error = -1;
 270                                 goto out;
 271                         }
 272
 273                         rte_memcpy((void *)(uintptr_t)dst,
 274                                         (void *)(uintptr_t)src, len);
 275
 276                         PRINT_PACKET(dev, (uintptr_t)dst, (uint32_t)len, 0);
 277                         vhost_log_write(dev, guest_addr, len);
 278                         remain -= len;
 279                         guest_addr += len;
 280                         dst += len;
 281                 }
 282         }
 283
 284         desc_avail  = desc->len - dev->vhost_hlen;
 285         if (unlikely(desc_chunck_len < dev->vhost_hlen)) {
 286                 desc_chunck_len = desc_avail;
 287                 desc_gaddr = desc->addr + dev->vhost_hlen;
 288                 desc_addr = vhost_iova_to_vva(dev,
 289                                 vq, desc_gaddr,
 290                                 &desc_chunck_len,
 291                                 VHOST_ACCESS_RW);
 292                 if (unlikely(!desc_addr)) {
 293                         error = -1;
 294                         goto out;
 295                 }
 296
 297                 desc_offset = 0;
 298         } else {
 299                 desc_offset = dev->vhost_hlen;
 300                 desc_chunck_len -= dev->vhost_hlen;
 301         }
 302
 303         mbuf_avail  = rte_pktmbuf_data_len(m);
 304         mbuf_offset = 0;
 305         while (mbuf_avail != 0 || m->next != NULL) {
 306                 /* done with current mbuf, fetch next */
 307                 if (mbuf_avail == 0) {
 308                         m = m->next;
 309
 310                         mbuf_offset = 0;
 311                         mbuf_avail  = rte_pktmbuf_data_len(m);
 312                 }
 313
 314                 /* done with current desc buf, fetch next */
 315                 if (desc_avail == 0) {
 316                         if ((desc->flags & VRING_DESC_F_NEXT) == 0) {
 317                                 /* Room in vring buffer is not enough */
 318                                 error = -1;
 319                                 goto out;
 320                         }
 321                         if (unlikely(desc->next >= size || ++nr_desc > size)) {
 322                                 error = -1;
 323                                 goto out;
 324                         }
 325
 326                         desc = &descs[desc->next];
 327                         desc_chunck_len = desc->len;
 328                         desc_gaddr = desc->addr;
 329                         desc_addr = vhost_iova_to_vva(dev, vq, desc_gaddr,
 330                                                         &desc_chunck_len,
 331                                                         VHOST_ACCESS_RW);
 332                         if (unlikely(!desc_addr)) {
 333                                 error = -1;
 334                                 goto out;
 335                         }
 336
 337                         desc_offset = 0;
 338                         desc_avail  = desc->len;
 339                 } else if (unlikely(desc_chunck_len == 0)) {
 340                         desc_chunck_len = desc_avail;
 341                         desc_gaddr += desc_offset;
 342                         desc_addr = vhost_iova_to_vva(dev,
 343                                         vq, desc_gaddr,
 344                                         &desc_chunck_len, VHOST_ACCESS_RW);
 345                         if (unlikely(!desc_addr)) {
 346                                 error = -1;
 347                                 goto out;
 348                         }
 349                         desc_offset = 0;
 350                 }
 351
 352                 cpy_len = RTE_MIN(desc_chunck_len, mbuf_avail);
 353                 if (likely(cpy_len > MAX_BATCH_LEN || copy_nb >= vq->size)) {
 354                         rte_memcpy((void *)((uintptr_t)(desc_addr +
 355                                                         desc_offset)),
 356                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 357                                 cpy_len);
 358                         vhost_log_write(dev, desc_gaddr + desc_offset, cpy_len);
 359                         PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
 360                                      cpy_len, 0);
 361                 } else {
 362                         batch_copy[copy_nb].dst =
 363                                 (void *)((uintptr_t)(desc_addr + desc_offset));
 364                         batch_copy[copy_nb].src =
 365                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
 366                         batch_copy[copy_nb].log_addr = desc_gaddr + desc_offset;
 367                         batch_copy[copy_nb].len = cpy_len;
 368                         copy_nb++;
 369                 }
 370
 371                 mbuf_avail  -= cpy_len;
 372                 mbuf_offset += cpy_len;
 373                 desc_avail  -= cpy_len;
 374                 desc_offset += cpy_len;
 375                 desc_chunck_len -= cpy_len;
 376         }
 377
 378 out:
 379         vq->batch_copy_nb_elems = copy_nb;
 380
 381         return error;
 382 }
 383
 384 /**
 385  * This function adds buffers to the virtio devices RX virtqueue. Buffers can
 386  * be received from the physical port or from another virtio device. A packet
 387  * count is returned to indicate the number of packets that are successfully
 388  * added to the RX queue. This function works when the mbuf is scattered, but
 389  * it doesn't support the mergeable feature.
 390  */
 391 static __rte_always_inline uint32_t
 392 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
 393               struct rte_mbuf **pkts, uint32_t count)
 394 {
 395         struct vhost_virtqueue *vq;
 396         uint16_t avail_idx, free_entries, start_idx;
 397         uint16_t desc_indexes[MAX_PKT_BURST];
 398         struct vring_desc *descs;
 399         uint16_t used_idx;
 400         uint32_t i, sz;
 401
 402         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
 403         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
 404                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
 405                         dev->vid, __func__, queue_id);
 406                 return 0;
 407         }
 408
 409         vq = dev->virtqueue[queue_id];
 410
 411         rte_spinlock_lock(&vq->access_lock);
 412
 413         if (unlikely(vq->enabled == 0))
 414                 goto out_access_unlock;
 415
 416         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 417                 vhost_user_iotlb_rd_lock(vq);
 418
 419         if (unlikely(vq->access_ok == 0)) {
 420                 if (unlikely(vring_translate(dev, vq) < 0)) {
 421                         count = 0;
 422                         goto out;
 423                 }
 424         }
 425
 426         avail_idx = *((volatile uint16_t *)&vq->avail->idx);
 427         start_idx = vq->last_used_idx;
 428         free_entries = avail_idx - start_idx;
 429         count = RTE_MIN(count, free_entries);
 430         count = RTE_MIN(count, (uint32_t)MAX_PKT_BURST);
 431         if (count == 0)
 432                 goto out;
 433
 434         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) start_idx %d | end_idx %d\n",
 435                 dev->vid, start_idx, start_idx + count);
 436
 437         vq->batch_copy_nb_elems = 0;
 438
 439         /* Retrieve all of the desc indexes first to avoid caching issues. */
 440         rte_prefetch0(&vq->avail->ring[start_idx & (vq->size - 1)]);
 441         for (i = 0; i < count; i++) {
 442                 used_idx = (start_idx + i) & (vq->size - 1);
 443                 desc_indexes[i] = vq->avail->ring[used_idx];
 444                 vq->used->ring[used_idx].id = desc_indexes[i];
 445                 vq->used->ring[used_idx].len = pkts[i]->pkt_len +
 446                                                dev->vhost_hlen;
 447                 vhost_log_used_vring(dev, vq,
 448                         offsetof(struct vring_used, ring[used_idx]),
 449                         sizeof(vq->used->ring[used_idx]));
 450         }
 451
 452         rte_prefetch0(&vq->desc[desc_indexes[0]]);
 453         for (i = 0; i < count; i++) {
 454                 struct vring_desc *idesc = NULL;
 455                 uint16_t desc_idx = desc_indexes[i];
 456                 int err;
 457
 458                 if (vq->desc[desc_idx].flags & VRING_DESC_F_INDIRECT) {
 459                         uint64_t dlen = vq->desc[desc_idx].len;
 460                         descs = (struct vring_desc *)(uintptr_t)
 461                                 vhost_iova_to_vva(dev,
 462                                                 vq, vq->desc[desc_idx].addr,
 463                                                 &dlen, VHOST_ACCESS_RO);
 464                         if (unlikely(!descs)) {
 465                                 count = i;
 466                                 break;
 467                         }
 468
 469                         if (unlikely(dlen < vq->desc[desc_idx].len)) {
 470                                 /*
 471                                  * The indirect desc table is not contiguous
 472                                  * in process VA space, we have to copy it.
 473                                  */
 474                                 idesc = alloc_copy_ind_table(dev, vq,
 475                                                         &vq->desc[desc_idx]);
 476                                 if (unlikely(!idesc))
 477                                         break;
 478
 479                                 descs = idesc;
 480                         }
 481
 482                         desc_idx = 0;
 483                         sz = vq->desc[desc_idx].len / sizeof(*descs);
 484                 } else {
 485                         descs = vq->desc;
 486                         sz = vq->size;
 487                 }
 488
 489                 err = copy_mbuf_to_desc(dev, vq, descs, pkts[i], desc_idx, sz);
 490                 if (unlikely(err)) {
 491                         count = i;
 492                         free_ind_table(idesc);
 493                         break;
 494                 }
 495
 496                 if (i + 1 < count)
 497                         rte_prefetch0(&vq->desc[desc_indexes[i+1]]);
 498
 499                 if (unlikely(!!idesc))
 500                         free_ind_table(idesc);
 501         }
 502
 503         do_data_copy_enqueue(dev, vq);
 504
 505         rte_smp_wmb();
 506
 507         *(volatile uint16_t *)&vq->used->idx += count;
 508         vq->last_used_idx += count;
 509         vhost_log_used_vring(dev, vq,
 510                 offsetof(struct vring_used, idx),
 511                 sizeof(vq->used->idx));
 512
 513         vhost_vring_call(dev, vq);
 514 out:
 515         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 516                 vhost_user_iotlb_rd_unlock(vq);
 517
 518 out_access_unlock:
 519         rte_spinlock_unlock(&vq->access_lock);
 520
 521         return count;
 522 }
 523
 524 static __rte_always_inline int
 525 fill_vec_buf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 526                          uint32_t avail_idx, uint32_t *vec_idx,
 527                          struct buf_vector *buf_vec, uint16_t *desc_chain_head,
 528                          uint16_t *desc_chain_len)
 529 {
 530         uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
 531         uint32_t vec_id = *vec_idx;
 532         uint32_t len    = 0;
 533         uint64_t dlen;
 534         struct vring_desc *descs = vq->desc;
 535         struct vring_desc *idesc = NULL;
 536
 537         *desc_chain_head = idx;
 538
 539         if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
 540                 dlen = vq->desc[idx].len;
 541                 descs = (struct vring_desc *)(uintptr_t)
 542                         vhost_iova_to_vva(dev, vq, vq->desc[idx].addr,
 543                                                 &dlen,
 544                                                 VHOST_ACCESS_RO);
 545                 if (unlikely(!descs))
 546                         return -1;
 547
 548                 if (unlikely(dlen < vq->desc[idx].len)) {
 549                         /*
 550                          * The indirect desc table is not contiguous
 551                          * in process VA space, we have to copy it.
 552                          */
 553                         idesc = alloc_copy_ind_table(dev, vq, &vq->desc[idx]);
 554                         if (unlikely(!idesc))
 555                                 return -1;
 556
 557                         descs = idesc;
 558                 }
 559
 560                 idx = 0;
 561         }
 562
 563         while (1) {
 564                 if (unlikely(vec_id >= BUF_VECTOR_MAX || idx >= vq->size)) {
 565                         free_ind_table(idesc);
 566                         return -1;
 567                 }
 568
 569                 len += descs[idx].len;
 570                 buf_vec[vec_id].buf_addr = descs[idx].addr;
 571                 buf_vec[vec_id].buf_len  = descs[idx].len;
 572                 buf_vec[vec_id].desc_idx = idx;
 573                 vec_id++;
 574
 575                 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
 576                         break;
 577
 578                 idx = descs[idx].next;
 579         }
 580
 581         *desc_chain_len = len;
 582         *vec_idx = vec_id;
 583
 584         if (unlikely(!!idesc))
 585                 free_ind_table(idesc);
 586
 587         return 0;
 588 }
 589
 590 /*
 591  * Returns -1 on fail, 0 on success
 592  */
 593 static inline int
 594 reserve_avail_buf_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
 595                                 uint32_t size, struct buf_vector *buf_vec,
 596                                 uint16_t *num_buffers, uint16_t avail_head)
 597 {
 598         uint16_t cur_idx;
 599         uint32_t vec_idx = 0;
 600         uint16_t tries = 0;
 601
 602         uint16_t head_idx = 0;
 603         uint16_t len = 0;
 604
 605         *num_buffers = 0;
 606         cur_idx  = vq->last_avail_idx;
 607
 608         while (size > 0) {
 609                 if (unlikely(cur_idx == avail_head))
 610                         return -1;
 611
 612                 if (unlikely(fill_vec_buf(dev, vq, cur_idx, &vec_idx, buf_vec,
 613                                                 &head_idx, &len) < 0))
 614                         return -1;
 615                 len = RTE_MIN(len, size);
 616                 update_shadow_used_ring(vq, head_idx, len);
 617                 size -= len;
 618
 619                 cur_idx++;
 620                 tries++;
 621                 *num_buffers += 1;
 622
 623                 /*
 624                  * if we tried all available ring items, and still
 625                  * can't get enough buf, it means something abnormal
 626                  * happened.
 627                  */
 628                 if (unlikely(tries >= vq->size))
 629                         return -1;
 630         }
 631
 632         return 0;
 633 }
 634
 635 static __rte_always_inline int
 636 copy_mbuf_to_desc_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
 637                             struct rte_mbuf *m, struct buf_vector *buf_vec,
 638                             uint16_t num_buffers)
 639 {
 640         uint32_t vec_idx = 0;
 641         uint64_t desc_addr;
 642         uint32_t mbuf_offset, mbuf_avail;
 643         uint32_t desc_offset, desc_avail;
 644         uint32_t cpy_len;
 645         uint64_t dlen;
 646         uint64_t hdr_addr, hdr_phys_addr;
 647         struct rte_mbuf *hdr_mbuf;
 648         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 649         uint16_t copy_nb = vq->batch_copy_nb_elems;
 650         int error = 0;
 651
 652         if (unlikely(m == NULL)) {
 653                 error = -1;
 654                 goto out;
 655         }
 656
 657         dlen = buf_vec[vec_idx].buf_len;
 658         desc_addr = vhost_iova_to_vva(dev, vq, buf_vec[vec_idx].buf_addr,
 659                                                 &dlen, VHOST_ACCESS_RW);
 660         if (dlen != buf_vec[vec_idx].buf_len ||
 661                         buf_vec[vec_idx].buf_len < dev->vhost_hlen ||
 662                         !desc_addr) {
 663                 error = -1;
 664                 goto out;
 665         }
 666
 667         hdr_mbuf = m;
 668         hdr_addr = desc_addr;
 669         hdr_phys_addr = buf_vec[vec_idx].buf_addr;
 670         rte_prefetch0((void *)(uintptr_t)hdr_addr);
 671
 672         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
 673                 dev->vid, num_buffers);
 674
 675         desc_avail  = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
 676         desc_offset = dev->vhost_hlen;
 677
 678         mbuf_avail  = rte_pktmbuf_data_len(m);
 679         mbuf_offset = 0;
 680         while (mbuf_avail != 0 || m->next != NULL) {
 681                 /* done with current desc buf, get the next one */
 682                 if (desc_avail == 0) {
 683                         vec_idx++;
 684                         dlen = buf_vec[vec_idx].buf_len;
 685                         desc_addr =
 686                                 vhost_iova_to_vva(dev, vq,
 687                                         buf_vec[vec_idx].buf_addr,
 688                                         &dlen,
 689                                         VHOST_ACCESS_RW);
 690                         if (unlikely(!desc_addr ||
 691                                         dlen != buf_vec[vec_idx].buf_len)) {
 692                                 error = -1;
 693                                 goto out;
 694                         }
 695
 696                         /* Prefetch buffer address. */
 697                         rte_prefetch0((void *)(uintptr_t)desc_addr);
 698                         desc_offset = 0;
 699                         desc_avail  = buf_vec[vec_idx].buf_len;
 700                 }
 701
 702                 /* done with current mbuf, get the next one */
 703                 if (mbuf_avail == 0) {
 704                         m = m->next;
 705
 706                         mbuf_offset = 0;
 707                         mbuf_avail  = rte_pktmbuf_data_len(m);
 708                 }
 709
 710                 if (hdr_addr) {
 711                         struct virtio_net_hdr_mrg_rxbuf *hdr;
 712
 713                         hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)
 714                                 hdr_addr;
 715                         virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
 716                         ASSIGN_UNLESS_EQUAL(hdr->num_buffers, num_buffers);
 717
 718                         vhost_log_write(dev, hdr_phys_addr, dev->vhost_hlen);
 719                         PRINT_PACKET(dev, (uintptr_t)hdr_addr,
 720                                      dev->vhost_hlen, 0);
 721
 722                         hdr_addr = 0;
 723                 }
 724
 725                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 726
 727                 if (likely(cpy_len > MAX_BATCH_LEN || copy_nb >= vq->size)) {
 728                         rte_memcpy((void *)((uintptr_t)(desc_addr +
 729                                                         desc_offset)),
 730                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 731                                 cpy_len);
 732                         vhost_log_write(dev,
 733                                 buf_vec[vec_idx].buf_addr + desc_offset,
 734                                 cpy_len);
 735                         PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
 736                                 cpy_len, 0);
 737                 } else {
 738                         batch_copy[copy_nb].dst =
 739                                 (void *)((uintptr_t)(desc_addr + desc_offset));
 740                         batch_copy[copy_nb].src =
 741                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
 742                         batch_copy[copy_nb].log_addr =
 743                                 buf_vec[vec_idx].buf_addr + desc_offset;
 744                         batch_copy[copy_nb].len = cpy_len;
 745                         copy_nb++;
 746                 }
 747
 748                 mbuf_avail  -= cpy_len;
 749                 mbuf_offset += cpy_len;
 750                 desc_avail  -= cpy_len;
 751                 desc_offset += cpy_len;
 752         }
 753
 754 out:
 755         vq->batch_copy_nb_elems = copy_nb;
 756
 757         return error;
 758 }
 759
 760 static __rte_always_inline uint32_t
 761 virtio_dev_merge_rx(struct virtio_net *dev, uint16_t queue_id,
 762         struct rte_mbuf **pkts, uint32_t count)
 763 {
 764         struct vhost_virtqueue *vq;
 765         uint32_t pkt_idx = 0;
 766         uint16_t num_buffers;
 767         struct buf_vector buf_vec[BUF_VECTOR_MAX];
 768         uint16_t avail_head;
 769
 770         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
 771         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
 772                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
 773                         dev->vid, __func__, queue_id);
 774                 return 0;
 775         }
 776
 777         vq = dev->virtqueue[queue_id];
 778
 779         rte_spinlock_lock(&vq->access_lock);
 780
 781         if (unlikely(vq->enabled == 0))
 782                 goto out_access_unlock;
 783
 784         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 785                 vhost_user_iotlb_rd_lock(vq);
 786
 787         if (unlikely(vq->access_ok == 0))
 788                 if (unlikely(vring_translate(dev, vq) < 0))
 789                         goto out;
 790
 791         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
 792         if (count == 0)
 793                 goto out;
 794
 795         vq->batch_copy_nb_elems = 0;
 796
 797         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
 798
 799         vq->shadow_used_idx = 0;
 800         avail_head = *((volatile uint16_t *)&vq->avail->idx);
 801         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
 802                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
 803
 804                 if (unlikely(reserve_avail_buf_mergeable(dev, vq,
 805                                                 pkt_len, buf_vec, &num_buffers,
 806                                                 avail_head) < 0)) {
 807                         VHOST_LOG_DEBUG(VHOST_DATA,
 808                                 "(%d) failed to get enough desc from vring\n",
 809                                 dev->vid);
 810                         vq->shadow_used_idx -= num_buffers;
 811                         break;
 812                 }
 813
 814                 VHOST_LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
 815                         dev->vid, vq->last_avail_idx,
 816                         vq->last_avail_idx + num_buffers);
 817
 818                 if (copy_mbuf_to_desc_mergeable(dev, vq, pkts[pkt_idx],
 819                                                 buf_vec, num_buffers) < 0) {
 820                         vq->shadow_used_idx -= num_buffers;
 821                         break;
 822                 }
 823
 824                 vq->last_avail_idx += num_buffers;
 825         }
 826
 827         do_data_copy_enqueue(dev, vq);
 828
 829         if (likely(vq->shadow_used_idx)) {
 830                 flush_shadow_used_ring(dev, vq);
 831                 vhost_vring_call(dev, vq);
 832         }
 833
 834 out:
 835         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 836                 vhost_user_iotlb_rd_unlock(vq);
 837
 838 out_access_unlock:
 839         rte_spinlock_unlock(&vq->access_lock);
 840
 841         return pkt_idx;
 842 }
 843
 844 uint16_t
 845 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
 846         struct rte_mbuf **pkts, uint16_t count)
 847 {
 848         struct virtio_net *dev = get_device(vid);
 849
 850         if (!dev)
 851                 return 0;
 852
 853         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
 854                 RTE_LOG(ERR, VHOST_DATA,
 855                         "(%d) %s: built-in vhost net backend is disabled.\n",
 856                         dev->vid, __func__);
 857                 return 0;
 858         }
 859
 860         if (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF))
 861                 return virtio_dev_merge_rx(dev, queue_id, pkts, count);
 862         else
 863                 return virtio_dev_rx(dev, queue_id, pkts, count);
 864 }
 865
 866 static inline bool
 867 virtio_net_with_host_offload(struct virtio_net *dev)
 868 {
 869         if (dev->features &
 870                         ((1ULL << VIRTIO_NET_F_CSUM) |
 871                          (1ULL << VIRTIO_NET_F_HOST_ECN) |
 872                          (1ULL << VIRTIO_NET_F_HOST_TSO4) |
 873                          (1ULL << VIRTIO_NET_F_HOST_TSO6) |
 874                          (1ULL << VIRTIO_NET_F_HOST_UFO)))
 875                 return true;
 876
 877         return false;
 878 }
 879
 880 static void
 881 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
 882 {
 883         struct ipv4_hdr *ipv4_hdr;
 884         struct ipv6_hdr *ipv6_hdr;
 885         void *l3_hdr = NULL;
 886         struct ether_hdr *eth_hdr;
 887         uint16_t ethertype;
 888
 889         eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
 890
 891         m->l2_len = sizeof(struct ether_hdr);
 892         ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
 893
 894         if (ethertype == ETHER_TYPE_VLAN) {
 895                 struct vlan_hdr *vlan_hdr = (struct vlan_hdr *)(eth_hdr + 1);
 896
 897                 m->l2_len += sizeof(struct vlan_hdr);
 898                 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
 899         }
 900
 901         l3_hdr = (char *)eth_hdr + m->l2_len;
 902
 903         switch (ethertype) {
 904         case ETHER_TYPE_IPv4:
 905                 ipv4_hdr = l3_hdr;
 906                 *l4_proto = ipv4_hdr->next_proto_id;
 907                 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
 908                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 909                 m->ol_flags |= PKT_TX_IPV4;
 910                 break;
 911         case ETHER_TYPE_IPv6:
 912                 ipv6_hdr = l3_hdr;
 913                 *l4_proto = ipv6_hdr->proto;
 914                 m->l3_len = sizeof(struct ipv6_hdr);
 915                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 916                 m->ol_flags |= PKT_TX_IPV6;
 917                 break;
 918         default:
 919                 m->l3_len = 0;
 920                 *l4_proto = 0;
 921                 *l4_hdr = NULL;
 922                 break;
 923         }
 924 }
 925
 926 static __rte_always_inline void
 927 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
 928 {
 929         uint16_t l4_proto = 0;
 930         void *l4_hdr = NULL;
 931         struct tcp_hdr *tcp_hdr = NULL;
 932
 933         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
 934                 return;
 935
 936         parse_ethernet(m, &l4_proto, &l4_hdr);
 937         if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 938                 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
 939                         switch (hdr->csum_offset) {
 940                         case (offsetof(struct tcp_hdr, cksum)):
 941                                 if (l4_proto == IPPROTO_TCP)
 942                                         m->ol_flags |= PKT_TX_TCP_CKSUM;
 943                                 break;
 944                         case (offsetof(struct udp_hdr, dgram_cksum)):
 945                                 if (l4_proto == IPPROTO_UDP)
 946                                         m->ol_flags |= PKT_TX_UDP_CKSUM;
 947                                 break;
 948                         case (offsetof(struct sctp_hdr, cksum)):
 949                                 if (l4_proto == IPPROTO_SCTP)
 950                                         m->ol_flags |= PKT_TX_SCTP_CKSUM;
 951                                 break;
 952                         default:
 953                                 break;
 954                         }
 955                 }
 956         }
 957
 958         if (l4_hdr && hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 959                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 960                 case VIRTIO_NET_HDR_GSO_TCPV4:
 961                 case VIRTIO_NET_HDR_GSO_TCPV6:
 962                         tcp_hdr = l4_hdr;
 963                         m->ol_flags |= PKT_TX_TCP_SEG;
 964                         m->tso_segsz = hdr->gso_size;
 965                         m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
 966                         break;
 967                 case VIRTIO_NET_HDR_GSO_UDP:
 968                         m->ol_flags |= PKT_TX_UDP_SEG;
 969                         m->tso_segsz = hdr->gso_size;
 970                         m->l4_len = sizeof(struct udp_hdr);
 971                         break;
 972                 default:
 973                         RTE_LOG(WARNING, VHOST_DATA,
 974                                 "unsupported gso type %u.\n", hdr->gso_type);
 975                         break;
 976                 }
 977         }
 978 }
 979
 980 static __rte_always_inline void
 981 put_zmbuf(struct zcopy_mbuf *zmbuf)
 982 {
 983         zmbuf->in_use = 0;
 984 }
 985
 986 static __rte_always_inline int
 987 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 988                   struct vring_desc *descs, uint16_t max_desc,
 989                   struct rte_mbuf *m, uint16_t desc_idx,
 990                   struct rte_mempool *mbuf_pool)
 991 {
 992         struct vring_desc *desc;
 993         uint64_t desc_addr, desc_gaddr;
 994         uint32_t desc_avail, desc_offset;
 995         uint32_t mbuf_avail, mbuf_offset;
 996         uint32_t cpy_len;
 997         uint64_t desc_chunck_len;
 998         struct rte_mbuf *cur = m, *prev = m;
 999         struct virtio_net_hdr tmp_hdr;
1000         struct virtio_net_hdr *hdr = NULL;
1001         /* A counter to avoid desc dead loop chain */
1002         uint32_t nr_desc = 1;
1003         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
1004         uint16_t copy_nb = vq->batch_copy_nb_elems;
1005         int error = 0;
1006
1007         desc = &descs[desc_idx];
1008         if (unlikely((desc->len < dev->vhost_hlen)) ||
1009                         (desc->flags & VRING_DESC_F_INDIRECT)) {
1010                 error = -1;
1011                 goto out;
1012         }
1013
1014         desc_chunck_len = desc->len;
1015         desc_gaddr = desc->addr;
1016         desc_addr = vhost_iova_to_vva(dev,
1017                                         vq, desc_gaddr,
1018                                         &desc_chunck_len,
1019                                         VHOST_ACCESS_RO);
1020         if (unlikely(!desc_addr)) {
1021                 error = -1;
1022                 goto out;
1023         }
1024
1025         if (virtio_net_with_host_offload(dev)) {
1026                 if (unlikely(desc_chunck_len < sizeof(struct virtio_net_hdr))) {
1027                         uint64_t len = desc_chunck_len;
1028                         uint64_t remain = sizeof(struct virtio_net_hdr);
1029                         uint64_t src = desc_addr;
1030                         uint64_t dst = (uint64_t)(uintptr_t)&tmp_hdr;
1031                         uint64_t guest_addr = desc_gaddr;
1032
1033                         /*
1034                          * No luck, the virtio-net header doesn't fit
1035                          * in a contiguous virtual area.
1036                          */
1037                         while (remain) {
1038                                 len = remain;
1039                                 src = vhost_iova_to_vva(dev, vq,
1040                                                 guest_addr, &len,
1041                                                 VHOST_ACCESS_RO);
1042                                 if (unlikely(!src || !len)) {
1043                                         error = -1;
1044                                         goto out;
1045                                 }
1046
1047                                 rte_memcpy((void *)(uintptr_t)dst,
1048                                                    (void *)(uintptr_t)src, len);
1049
1050                                 guest_addr += len;
1051                                 remain -= len;
1052                                 dst += len;
1053                         }
1054
1055                         hdr = &tmp_hdr;
1056                 } else {
1057                         hdr = (struct virtio_net_hdr *)((uintptr_t)desc_addr);
1058                         rte_prefetch0(hdr);
1059                 }
1060         }
1061
1062         /*
1063          * A virtio driver normally uses at least 2 desc buffers
1064          * for Tx: the first for storing the header, and others
1065          * for storing the data.
1066          */
1067         if (likely((desc->len == dev->vhost_hlen) &&
1068                    (desc->flags & VRING_DESC_F_NEXT) != 0)) {
1069                 desc = &descs[desc->next];
1070                 if (unlikely(desc->flags & VRING_DESC_F_INDIRECT)) {
1071                         error = -1;
1072                         goto out;
1073                 }
1074
1075                 desc_chunck_len = desc->len;
1076                 desc_gaddr = desc->addr;
1077                 desc_addr = vhost_iova_to_vva(dev,
1078                                                         vq, desc_gaddr,
1079                                                         &desc_chunck_len,
1080                                                         VHOST_ACCESS_RO);
1081                 if (unlikely(!desc_addr)) {
1082                         error = -1;
1083                         goto out;
1084                 }
1085
1086                 desc_offset = 0;
1087                 desc_avail  = desc->len;
1088                 nr_desc    += 1;
1089         } else {
1090                 desc_avail  = desc->len - dev->vhost_hlen;
1091
1092                 if (unlikely(desc_chunck_len < dev->vhost_hlen)) {
1093                         desc_chunck_len = desc_avail;
1094                         desc_gaddr += dev->vhost_hlen;
1095                         desc_addr = vhost_iova_to_vva(dev,
1096                                         vq, desc_gaddr,
1097                                         &desc_chunck_len,
1098                                         VHOST_ACCESS_RO);
1099                         if (unlikely(!desc_addr)) {
1100                                 error = -1;
1101                                 goto out;
1102                         }
1103
1104                         desc_offset = 0;
1105                 } else {
1106                         desc_offset = dev->vhost_hlen;
1107                         desc_chunck_len -= dev->vhost_hlen;
1108                 }
1109         }
1110
1111         rte_prefetch0((void *)(uintptr_t)(desc_addr + desc_offset));
1112
1113         PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
1114                         (uint32_t)desc_chunck_len, 0);
1115
1116         mbuf_offset = 0;
1117         mbuf_avail  = m->buf_len - RTE_PKTMBUF_HEADROOM;
1118         while (1) {
1119                 uint64_t hpa;
1120
1121                 cpy_len = RTE_MIN(desc_chunck_len, mbuf_avail);
1122
1123                 /*
1124                  * A desc buf might across two host physical pages that are
1125                  * not continuous. In such case (gpa_to_hpa returns 0), data
1126                  * will be copied even though zero copy is enabled.
1127                  */
1128                 if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
1129                                         desc_gaddr + desc_offset, cpy_len)))) {
1130                         cur->data_len = cpy_len;
1131                         cur->data_off = 0;
1132                         cur->buf_addr = (void *)(uintptr_t)(desc_addr
1133                                 + desc_offset);
1134                         cur->buf_iova = hpa;
1135
1136                         /*
1137                          * In zero copy mode, one mbuf can only reference data
1138                          * for one or partial of one desc buff.
1139                          */
1140                         mbuf_avail = cpy_len;
1141                 } else {
1142                         if (likely(cpy_len > MAX_BATCH_LEN ||
1143                                    copy_nb >= vq->size ||
1144                                    (hdr && cur == m) ||
1145                                    desc->len != desc_chunck_len)) {
1146                                 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
1147                                                                    mbuf_offset),
1148                                            (void *)((uintptr_t)(desc_addr +
1149                                                                 desc_offset)),
1150                                            cpy_len);
1151                         } else {
1152                                 batch_copy[copy_nb].dst =
1153                                         rte_pktmbuf_mtod_offset(cur, void *,
1154                                                                 mbuf_offset);
1155                                 batch_copy[copy_nb].src =
1156                                         (void *)((uintptr_t)(desc_addr +
1157                                                              desc_offset));
1158                                 batch_copy[copy_nb].len = cpy_len;
1159                                 copy_nb++;
1160                         }
1161                 }
1162
1163                 mbuf_avail  -= cpy_len;
1164                 mbuf_offset += cpy_len;
1165                 desc_avail  -= cpy_len;
1166                 desc_chunck_len -= cpy_len;
1167                 desc_offset += cpy_len;
1168
1169                 /* This desc reaches to its end, get the next one */
1170                 if (desc_avail == 0) {
1171                         if ((desc->flags & VRING_DESC_F_NEXT) == 0)
1172                                 break;
1173
1174                         if (unlikely(desc->next >= max_desc ||
1175                                      ++nr_desc > max_desc)) {
1176                                 error = -1;
1177                                 goto out;
1178                         }
1179                         desc = &descs[desc->next];
1180                         if (unlikely(desc->flags & VRING_DESC_F_INDIRECT)) {
1181                                 error = -1;
1182                                 goto out;
1183                         }
1184
1185                         desc_chunck_len = desc->len;
1186                         desc_gaddr = desc->addr;
1187                         desc_addr = vhost_iova_to_vva(dev,
1188                                                         vq, desc_gaddr,
1189                                                         &desc_chunck_len,
1190                                                         VHOST_ACCESS_RO);
1191                         if (unlikely(!desc_addr)) {
1192                                 error = -1;
1193                                 goto out;
1194                         }
1195
1196                         rte_prefetch0((void *)(uintptr_t)desc_addr);
1197
1198                         desc_offset = 0;
1199                         desc_avail  = desc->len;
1200
1201                         PRINT_PACKET(dev, (uintptr_t)desc_addr,
1202                                         (uint32_t)desc_chunck_len, 0);
1203                 } else if (unlikely(desc_chunck_len == 0)) {
1204                         desc_chunck_len = desc_avail;
1205                         desc_gaddr += desc_offset;
1206                         desc_addr = vhost_iova_to_vva(dev, vq,
1207                                         desc_gaddr,
1208                                         &desc_chunck_len,
1209                                         VHOST_ACCESS_RO);
1210                         if (unlikely(!desc_addr)) {
1211                                 error = -1;
1212                                 goto out;
1213                         }
1214                         desc_offset = 0;
1215
1216                         PRINT_PACKET(dev, (uintptr_t)desc_addr,
1217                                         (uint32_t)desc_chunck_len, 0);
1218                 }
1219
1220                 /*
1221                  * This mbuf reaches to its end, get a new one
1222                  * to hold more data.
1223                  */
1224                 if (mbuf_avail == 0) {
1225                         cur = rte_pktmbuf_alloc(mbuf_pool);
1226                         if (unlikely(cur == NULL)) {
1227                                 RTE_LOG(ERR, VHOST_DATA, "Failed to "
1228                                         "allocate memory for mbuf.\n");
1229                                 error = -1;
1230                                 goto out;
1231                         }
1232                         if (unlikely(dev->dequeue_zero_copy))
1233                                 rte_mbuf_refcnt_update(cur, 1);
1234
1235                         prev->next = cur;
1236                         prev->data_len = mbuf_offset;
1237                         m->nb_segs += 1;
1238                         m->pkt_len += mbuf_offset;
1239                         prev = cur;
1240
1241                         mbuf_offset = 0;
1242                         mbuf_avail  = cur->buf_len - RTE_PKTMBUF_HEADROOM;
1243                 }
1244         }
1245
1246         prev->data_len = mbuf_offset;
1247         m->pkt_len    += mbuf_offset;
1248
1249         if (hdr)
1250                 vhost_dequeue_offload(hdr, m);
1251
1252 out:
1253         vq->batch_copy_nb_elems = copy_nb;
1254
1255         return error;
1256 }
1257
1258 static __rte_always_inline void
1259 update_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq,
1260                  uint32_t used_idx, uint32_t desc_idx)
1261 {
1262         vq->used->ring[used_idx].id  = desc_idx;
1263         vq->used->ring[used_idx].len = 0;
1264         vhost_log_used_vring(dev, vq,
1265                         offsetof(struct vring_used, ring[used_idx]),
1266                         sizeof(vq->used->ring[used_idx]));
1267 }
1268
1269 static __rte_always_inline void
1270 update_used_idx(struct virtio_net *dev, struct vhost_virtqueue *vq,
1271                 uint32_t count)
1272 {
1273         if (unlikely(count == 0))
1274                 return;
1275
1276         rte_smp_wmb();
1277         rte_smp_rmb();
1278
1279         vq->used->idx += count;
1280         vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
1281                         sizeof(vq->used->idx));
1282         vhost_vring_call(dev, vq);
1283 }
1284
1285 static __rte_always_inline struct zcopy_mbuf *
1286 get_zmbuf(struct vhost_virtqueue *vq)
1287 {
1288         uint16_t i;
1289         uint16_t last;
1290         int tries = 0;
1291
1292         /* search [last_zmbuf_idx, zmbuf_size) */
1293         i = vq->last_zmbuf_idx;
1294         last = vq->zmbuf_size;
1295
1296 again:
1297         for (; i < last; i++) {
1298                 if (vq->zmbufs[i].in_use == 0) {
1299                         vq->last_zmbuf_idx = i + 1;
1300                         vq->zmbufs[i].in_use = 1;
1301                         return &vq->zmbufs[i];
1302                 }
1303         }
1304
1305         tries++;
1306         if (tries == 1) {
1307                 /* search [0, last_zmbuf_idx) */
1308                 i = 0;
1309                 last = vq->last_zmbuf_idx;
1310                 goto again;
1311         }
1312
1313         return NULL;
1314 }
1315
1316 static __rte_always_inline bool
1317 mbuf_is_consumed(struct rte_mbuf *m)
1318 {
1319         while (m) {
1320                 if (rte_mbuf_refcnt_read(m) > 1)
1321                         return false;
1322                 m = m->next;
1323         }
1324
1325         return true;
1326 }
1327
1328 static __rte_always_inline void
1329 restore_mbuf(struct rte_mbuf *m)
1330 {
1331         uint32_t mbuf_size, priv_size;
1332
1333         while (m) {
1334                 priv_size = rte_pktmbuf_priv_size(m->pool);
1335                 mbuf_size = sizeof(struct rte_mbuf) + priv_size;
1336                 /* start of buffer is after mbuf structure and priv data */
1337
1338                 m->buf_addr = (char *)m + mbuf_size;
1339                 m->buf_iova = rte_mempool_virt2iova(m) + mbuf_size;
1340                 m = m->next;
1341         }
1342 }
1343
1344 uint16_t
1345 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
1346         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1347 {
1348         struct virtio_net *dev;
1349         struct rte_mbuf *rarp_mbuf = NULL;
1350         struct vhost_virtqueue *vq;
1351         uint32_t desc_indexes[MAX_PKT_BURST];
1352         uint32_t used_idx;
1353         uint32_t i = 0;
1354         uint16_t free_entries;
1355         uint16_t avail_idx;
1356
1357         dev = get_device(vid);
1358         if (!dev)
1359                 return 0;
1360
1361         if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
1362                 RTE_LOG(ERR, VHOST_DATA,
1363                         "(%d) %s: built-in vhost net backend is disabled.\n",
1364                         dev->vid, __func__);
1365                 return 0;
1366         }
1367
1368         if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
1369                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
1370                         dev->vid, __func__, queue_id);
1371                 return 0;
1372         }
1373
1374         vq = dev->virtqueue[queue_id];
1375
1376         if (unlikely(rte_spinlock_trylock(&vq->access_lock) == 0))
1377                 return 0;
1378
1379         if (unlikely(vq->enabled == 0))
1380                 goto out_access_unlock;
1381
1382         vq->batch_copy_nb_elems = 0;
1383
1384         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1385                 vhost_user_iotlb_rd_lock(vq);
1386
1387         if (unlikely(vq->access_ok == 0))
1388                 if (unlikely(vring_translate(dev, vq) < 0))
1389                         goto out;
1390
1391         if (unlikely(dev->dequeue_zero_copy)) {
1392                 struct zcopy_mbuf *zmbuf, *next;
1393                 int nr_updated = 0;
1394
1395                 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1396                      zmbuf != NULL; zmbuf = next) {
1397                         next = TAILQ_NEXT(zmbuf, next);
1398
1399                         if (mbuf_is_consumed(zmbuf->mbuf)) {
1400                                 used_idx = vq->last_used_idx++ & (vq->size - 1);
1401                                 update_used_ring(dev, vq, used_idx,
1402                                                  zmbuf->desc_idx);
1403                                 nr_updated += 1;
1404
1405                                 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1406                                 restore_mbuf(zmbuf->mbuf);
1407                                 rte_pktmbuf_free(zmbuf->mbuf);
1408                                 put_zmbuf(zmbuf);
1409                                 vq->nr_zmbuf -= 1;
1410                         }
1411                 }
1412
1413                 update_used_idx(dev, vq, nr_updated);
1414         }
1415
1416         /*
1417          * Construct a RARP broadcast packet, and inject it to the "pkts"
1418          * array, to looks like that guest actually send such packet.
1419          *
1420          * Check user_send_rarp() for more information.
1421          *
1422          * broadcast_rarp shares a cacheline in the virtio_net structure
1423          * with some fields that are accessed during enqueue and
1424          * rte_atomic16_cmpset() causes a write if using cmpxchg. This could
1425          * result in false sharing between enqueue and dequeue.
1426          *
1427          * Prevent unnecessary false sharing by reading broadcast_rarp first
1428          * and only performing cmpset if the read indicates it is likely to
1429          * be set.
1430          */
1431
1432         if (unlikely(rte_atomic16_read(&dev->broadcast_rarp) &&
1433                         rte_atomic16_cmpset((volatile uint16_t *)
1434                                 &dev->broadcast_rarp.cnt, 1, 0))) {
1435
1436                 rarp_mbuf = rte_net_make_rarp_packet(mbuf_pool, &dev->mac);
1437                 if (rarp_mbuf == NULL) {
1438                         RTE_LOG(ERR, VHOST_DATA,
1439                                 "Failed to make RARP packet.\n");
1440                         return 0;
1441                 }
1442                 count -= 1;
1443         }
1444
1445         free_entries = *((volatile uint16_t *)&vq->avail->idx) -
1446                         vq->last_avail_idx;
1447         if (free_entries == 0)
1448                 goto out;
1449
1450         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1451
1452         /* Prefetch available and used ring */
1453         avail_idx = vq->last_avail_idx & (vq->size - 1);
1454         used_idx  = vq->last_used_idx  & (vq->size - 1);
1455         rte_prefetch0(&vq->avail->ring[avail_idx]);
1456         rte_prefetch0(&vq->used->ring[used_idx]);
1457
1458         count = RTE_MIN(count, MAX_PKT_BURST);
1459         count = RTE_MIN(count, free_entries);
1460         VHOST_LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1461                         dev->vid, count);
1462
1463         /* Retrieve all of the head indexes first to avoid caching issues. */
1464         for (i = 0; i < count; i++) {
1465                 avail_idx = (vq->last_avail_idx + i) & (vq->size - 1);
1466                 used_idx  = (vq->last_used_idx  + i) & (vq->size - 1);
1467                 desc_indexes[i] = vq->avail->ring[avail_idx];
1468
1469                 if (likely(dev->dequeue_zero_copy == 0))
1470                         update_used_ring(dev, vq, used_idx, desc_indexes[i]);
1471         }
1472
1473         /* Prefetch descriptor index. */
1474         rte_prefetch0(&vq->desc[desc_indexes[0]]);
1475         for (i = 0; i < count; i++) {
1476                 struct vring_desc *desc, *idesc = NULL;
1477                 uint16_t sz, idx;
1478                 uint64_t dlen;
1479                 int err;
1480
1481                 if (likely(i + 1 < count))
1482                         rte_prefetch0(&vq->desc[desc_indexes[i + 1]]);
1483
1484                 if (vq->desc[desc_indexes[i]].flags & VRING_DESC_F_INDIRECT) {
1485                         dlen = vq->desc[desc_indexes[i]].len;
1486                         desc = (struct vring_desc *)(uintptr_t)
1487                                 vhost_iova_to_vva(dev, vq,
1488                                                 vq->desc[desc_indexes[i]].addr,
1489                                                 &dlen,
1490                                                 VHOST_ACCESS_RO);
1491                         if (unlikely(!desc))
1492                                 break;
1493
1494                         if (unlikely(dlen < vq->desc[desc_indexes[i]].len)) {
1495                                 /*
1496                                  * The indirect desc table is not contiguous
1497                                  * in process VA space, we have to copy it.
1498                                  */
1499                                 idesc = alloc_copy_ind_table(dev, vq,
1500                                                 &vq->desc[desc_indexes[i]]);
1501                                 if (unlikely(!idesc))
1502                                         break;
1503
1504                                 desc = idesc;
1505                         }
1506
1507                         rte_prefetch0(desc);
1508                         sz = vq->desc[desc_indexes[i]].len / sizeof(*desc);
1509                         idx = 0;
1510                 } else {
1511                         desc = vq->desc;
1512                         sz = vq->size;
1513                         idx = desc_indexes[i];
1514                 }
1515
1516                 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
1517                 if (unlikely(pkts[i] == NULL)) {
1518                         RTE_LOG(ERR, VHOST_DATA,
1519                                 "Failed to allocate memory for mbuf.\n");
1520                         free_ind_table(idesc);
1521                         break;
1522                 }
1523
1524                 err = copy_desc_to_mbuf(dev, vq, desc, sz, pkts[i], idx,
1525                                         mbuf_pool);
1526                 if (unlikely(err)) {
1527                         rte_pktmbuf_free(pkts[i]);
1528                         free_ind_table(idesc);
1529                         break;
1530                 }
1531
1532                 if (unlikely(dev->dequeue_zero_copy)) {
1533                         struct zcopy_mbuf *zmbuf;
1534
1535                         zmbuf = get_zmbuf(vq);
1536                         if (!zmbuf) {
1537                                 rte_pktmbuf_free(pkts[i]);
1538                                 free_ind_table(idesc);
1539                                 break;
1540                         }
1541                         zmbuf->mbuf = pkts[i];
1542                         zmbuf->desc_idx = desc_indexes[i];
1543
1544                         /*
1545                          * Pin lock the mbuf; we will check later to see
1546                          * whether the mbuf is freed (when we are the last
1547                          * user) or not. If that's the case, we then could
1548                          * update the used ring safely.
1549                          */
1550                         rte_mbuf_refcnt_update(pkts[i], 1);
1551
1552                         vq->nr_zmbuf += 1;
1553                         TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1554                 }
1555
1556                 if (unlikely(!!idesc))
1557                         free_ind_table(idesc);
1558         }
1559         vq->last_avail_idx += i;
1560
1561         if (likely(dev->dequeue_zero_copy == 0)) {
1562                 do_data_copy_dequeue(vq);
1563                 vq->last_used_idx += i;
1564                 update_used_idx(dev, vq, i);
1565         }
1566
1567 out:
1568         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1569                 vhost_user_iotlb_rd_unlock(vq);
1570
1571 out_access_unlock:
1572         rte_spinlock_unlock(&vq->access_lock);
1573
1574         if (unlikely(rarp_mbuf != NULL)) {
1575                 /*
1576                  * Inject it to the head of "pkts" array, so that switch's mac
1577                  * learning table will get updated first.
1578                  */
1579                 memmove(&pkts[1], pkts, i * sizeof(struct rte_mbuf *));
1580                 pkts[0] = rarp_mbuf;
1581                 i += 1;
1582         }
1583
1584         return i;
1585 }