lib/librte_vhost/virtio_net.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2016 Intel Corporation. All rights reserved.
   5  *   All rights reserved.
   6  *
   7  *   Redistribution and use in source and binary forms, with or without
   8  *   modification, are permitted provided that the following conditions
   9  *   are met:
  10  *
  11  *     * Redistributions of source code must retain the above copyright
  12  *       notice, this list of conditions and the following disclaimer.
  13  *     * Redistributions in binary form must reproduce the above copyright
  14  *       notice, this list of conditions and the following disclaimer in
  15  *       the documentation and/or other materials provided with the
  16  *       distribution.
  17  *     * Neither the name of Intel Corporation nor the names of its
  18  *       contributors may be used to endorse or promote products derived
  19  *       from this software without specific prior written permission.
  20  *
  21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  32  */
  33
  34 #include <stdint.h>
  35 #include <stdbool.h>
  36 #include <linux/virtio_net.h>
  37
  38 #include <rte_mbuf.h>
  39 #include <rte_memcpy.h>
  40 #include <rte_ether.h>
  41 #include <rte_ip.h>
  42 #include <rte_vhost.h>
  43 #include <rte_tcp.h>
  44 #include <rte_udp.h>
  45 #include <rte_sctp.h>
  46 #include <rte_arp.h>
  47
  48 #include "iotlb.h"
  49 #include "vhost.h"
  50
  51 #define MAX_PKT_BURST 32
  52
  53 #define MAX_BATCH_LEN 256
  54
  55 static bool
  56 is_valid_virt_queue_idx(uint32_t idx, int is_tx, uint32_t nr_vring)
  57 {
  58         return (is_tx ^ (idx & 1)) == 0 && idx < nr_vring;
  59 }
  60
  61 static __rte_always_inline void
  62 do_flush_shadow_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq,
  63                           uint16_t to, uint16_t from, uint16_t size)
  64 {
  65         rte_memcpy(&vq->used->ring[to],
  66                         &vq->shadow_used_ring[from],
  67                         size * sizeof(struct vring_used_elem));
  68         vhost_log_used_vring(dev, vq,
  69                         offsetof(struct vring_used, ring[to]),
  70                         size * sizeof(struct vring_used_elem));
  71 }
  72
  73 static __rte_always_inline void
  74 flush_shadow_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq)
  75 {
  76         uint16_t used_idx = vq->last_used_idx & (vq->size - 1);
  77
  78         if (used_idx + vq->shadow_used_idx <= vq->size) {
  79                 do_flush_shadow_used_ring(dev, vq, used_idx, 0,
  80                                           vq->shadow_used_idx);
  81         } else {
  82                 uint16_t size;
  83
  84                 /* update used ring interval [used_idx, vq->size] */
  85                 size = vq->size - used_idx;
  86                 do_flush_shadow_used_ring(dev, vq, used_idx, 0, size);
  87
  88                 /* update the left half used ring interval [0, left_size] */
  89                 do_flush_shadow_used_ring(dev, vq, 0, size,
  90                                           vq->shadow_used_idx - size);
  91         }
  92         vq->last_used_idx += vq->shadow_used_idx;
  93
  94         rte_smp_wmb();
  95
  96         *(volatile uint16_t *)&vq->used->idx += vq->shadow_used_idx;
  97         vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
  98                 sizeof(vq->used->idx));
  99 }
 100
 101 static __rte_always_inline void
 102 update_shadow_used_ring(struct vhost_virtqueue *vq,
 103                          uint16_t desc_idx, uint16_t len)
 104 {
 105         uint16_t i = vq->shadow_used_idx++;
 106
 107         vq->shadow_used_ring[i].id  = desc_idx;
 108         vq->shadow_used_ring[i].len = len;
 109 }
 110
 111 static inline void
 112 do_data_copy_enqueue(struct virtio_net *dev, struct vhost_virtqueue *vq)
 113 {
 114         struct batch_copy_elem *elem = vq->batch_copy_elems;
 115         uint16_t count = vq->batch_copy_nb_elems;
 116         int i;
 117
 118         for (i = 0; i < count; i++) {
 119                 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
 120                 vhost_log_write(dev, elem[i].log_addr, elem[i].len);
 121                 PRINT_PACKET(dev, (uintptr_t)elem[i].dst, elem[i].len, 0);
 122         }
 123 }
 124
 125 static inline void
 126 do_data_copy_dequeue(struct vhost_virtqueue *vq)
 127 {
 128         struct batch_copy_elem *elem = vq->batch_copy_elems;
 129         uint16_t count = vq->batch_copy_nb_elems;
 130         int i;
 131
 132         for (i = 0; i < count; i++)
 133                 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
 134 }
 135
 136 /* avoid write operation when necessary, to lessen cache issues */
 137 #define ASSIGN_UNLESS_EQUAL(var, val) do {      \
 138         if ((var) != (val))                     \
 139                 (var) = (val);                  \
 140 } while (0)
 141
 142 static void
 143 virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
 144 {
 145         uint64_t csum_l4 = m_buf->ol_flags & PKT_TX_L4_MASK;
 146
 147         if (m_buf->ol_flags & PKT_TX_TCP_SEG)
 148                 csum_l4 |= PKT_TX_TCP_CKSUM;
 149
 150         if (csum_l4) {
 151                 net_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 152                 net_hdr->csum_start = m_buf->l2_len + m_buf->l3_len;
 153
 154                 switch (csum_l4) {
 155                 case PKT_TX_TCP_CKSUM:
 156                         net_hdr->csum_offset = (offsetof(struct tcp_hdr,
 157                                                 cksum));
 158                         break;
 159                 case PKT_TX_UDP_CKSUM:
 160                         net_hdr->csum_offset = (offsetof(struct udp_hdr,
 161                                                 dgram_cksum));
 162                         break;
 163                 case PKT_TX_SCTP_CKSUM:
 164                         net_hdr->csum_offset = (offsetof(struct sctp_hdr,
 165                                                 cksum));
 166                         break;
 167                 }
 168         } else {
 169                 ASSIGN_UNLESS_EQUAL(net_hdr->csum_start, 0);
 170                 ASSIGN_UNLESS_EQUAL(net_hdr->csum_offset, 0);
 171                 ASSIGN_UNLESS_EQUAL(net_hdr->flags, 0);
 172         }
 173
 174         /* IP cksum verification cannot be bypassed, then calculate here */
 175         if (m_buf->ol_flags & PKT_TX_IP_CKSUM) {
 176                 struct ipv4_hdr *ipv4_hdr;
 177
 178                 ipv4_hdr = rte_pktmbuf_mtod_offset(m_buf, struct ipv4_hdr *,
 179                                                    m_buf->l2_len);
 180                 ipv4_hdr->hdr_checksum = rte_ipv4_cksum(ipv4_hdr);
 181         }
 182
 183         if (m_buf->ol_flags & PKT_TX_TCP_SEG) {
 184                 if (m_buf->ol_flags & PKT_TX_IPV4)
 185                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 186                 else
 187                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 188                 net_hdr->gso_size = m_buf->tso_segsz;
 189                 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len
 190                                         + m_buf->l4_len;
 191         } else {
 192                 ASSIGN_UNLESS_EQUAL(net_hdr->gso_type, 0);
 193                 ASSIGN_UNLESS_EQUAL(net_hdr->gso_size, 0);
 194                 ASSIGN_UNLESS_EQUAL(net_hdr->hdr_len, 0);
 195         }
 196 }
 197
 198 static __rte_always_inline int
 199 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 200                   struct vring_desc *descs, struct rte_mbuf *m,
 201                   uint16_t desc_idx, uint32_t size)
 202 {
 203         uint32_t desc_avail, desc_offset;
 204         uint32_t mbuf_avail, mbuf_offset;
 205         uint32_t cpy_len;
 206         struct vring_desc *desc;
 207         uint64_t desc_addr;
 208         /* A counter to avoid desc dead loop chain */
 209         uint16_t nr_desc = 1;
 210         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 211         uint16_t copy_nb = vq->batch_copy_nb_elems;
 212         int error = 0;
 213
 214         desc = &descs[desc_idx];
 215         desc_addr = vhost_iova_to_vva(dev, vq, desc->addr,
 216                                         desc->len, VHOST_ACCESS_RW);
 217         /*
 218          * Checking of 'desc_addr' placed outside of 'unlikely' macro to avoid
 219          * performance issue with some versions of gcc (4.8.4 and 5.3.0) which
 220          * otherwise stores offset on the stack instead of in a register.
 221          */
 222         if (unlikely(desc->len < dev->vhost_hlen) || !desc_addr) {
 223                 error = -1;
 224                 goto out;
 225         }
 226
 227         rte_prefetch0((void *)(uintptr_t)desc_addr);
 228
 229         virtio_enqueue_offload(m, (struct virtio_net_hdr *)(uintptr_t)desc_addr);
 230         vhost_log_write(dev, desc->addr, dev->vhost_hlen);
 231         PRINT_PACKET(dev, (uintptr_t)desc_addr, dev->vhost_hlen, 0);
 232
 233         desc_offset = dev->vhost_hlen;
 234         desc_avail  = desc->len - dev->vhost_hlen;
 235
 236         mbuf_avail  = rte_pktmbuf_data_len(m);
 237         mbuf_offset = 0;
 238         while (mbuf_avail != 0 || m->next != NULL) {
 239                 /* done with current mbuf, fetch next */
 240                 if (mbuf_avail == 0) {
 241                         m = m->next;
 242
 243                         mbuf_offset = 0;
 244                         mbuf_avail  = rte_pktmbuf_data_len(m);
 245                 }
 246
 247                 /* done with current desc buf, fetch next */
 248                 if (desc_avail == 0) {
 249                         if ((desc->flags & VRING_DESC_F_NEXT) == 0) {
 250                                 /* Room in vring buffer is not enough */
 251                                 error = -1;
 252                                 goto out;
 253                         }
 254                         if (unlikely(desc->next >= size || ++nr_desc > size)) {
 255                                 error = -1;
 256                                 goto out;
 257                         }
 258
 259                         desc = &descs[desc->next];
 260                         desc_addr = vhost_iova_to_vva(dev, vq, desc->addr,
 261                                                         desc->len,
 262                                                         VHOST_ACCESS_RW);
 263                         if (unlikely(!desc_addr)) {
 264                                 error = -1;
 265                                 goto out;
 266                         }
 267
 268                         desc_offset = 0;
 269                         desc_avail  = desc->len;
 270                 }
 271
 272                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 273                 if (likely(cpy_len > MAX_BATCH_LEN || copy_nb >= vq->size)) {
 274                         rte_memcpy((void *)((uintptr_t)(desc_addr +
 275                                                         desc_offset)),
 276                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 277                                 cpy_len);
 278                         vhost_log_write(dev, desc->addr + desc_offset, cpy_len);
 279                         PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
 280                                      cpy_len, 0);
 281                 } else {
 282                         batch_copy[copy_nb].dst =
 283                                 (void *)((uintptr_t)(desc_addr + desc_offset));
 284                         batch_copy[copy_nb].src =
 285                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
 286                         batch_copy[copy_nb].log_addr = desc->addr + desc_offset;
 287                         batch_copy[copy_nb].len = cpy_len;
 288                         copy_nb++;
 289                 }
 290
 291                 mbuf_avail  -= cpy_len;
 292                 mbuf_offset += cpy_len;
 293                 desc_avail  -= cpy_len;
 294                 desc_offset += cpy_len;
 295         }
 296
 297 out:
 298         vq->batch_copy_nb_elems = copy_nb;
 299
 300         return error;
 301 }
 302
 303 /**
 304  * This function adds buffers to the virtio devices RX virtqueue. Buffers can
 305  * be received from the physical port or from another virtio device. A packet
 306  * count is returned to indicate the number of packets that are successfully
 307  * added to the RX queue. This function works when the mbuf is scattered, but
 308  * it doesn't support the mergeable feature.
 309  */
 310 static __rte_always_inline uint32_t
 311 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
 312               struct rte_mbuf **pkts, uint32_t count)
 313 {
 314         struct vhost_virtqueue *vq;
 315         uint16_t avail_idx, free_entries, start_idx;
 316         uint16_t desc_indexes[MAX_PKT_BURST];
 317         struct vring_desc *descs;
 318         uint16_t used_idx;
 319         uint32_t i, sz;
 320
 321         LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
 322         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
 323                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
 324                         dev->vid, __func__, queue_id);
 325                 return 0;
 326         }
 327
 328         vq = dev->virtqueue[queue_id];
 329         if (unlikely(vq->enabled == 0))
 330                 return 0;
 331
 332         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 333                 vhost_user_iotlb_rd_lock(vq);
 334
 335         if (unlikely(vq->access_ok == 0)) {
 336                 if (unlikely(vring_translate(dev, vq) < 0)) {
 337                         count = 0;
 338                         goto out;
 339                 }
 340         }
 341
 342         avail_idx = *((volatile uint16_t *)&vq->avail->idx);
 343         start_idx = vq->last_used_idx;
 344         free_entries = avail_idx - start_idx;
 345         count = RTE_MIN(count, free_entries);
 346         count = RTE_MIN(count, (uint32_t)MAX_PKT_BURST);
 347         if (count == 0)
 348                 goto out;
 349
 350         LOG_DEBUG(VHOST_DATA, "(%d) start_idx %d | end_idx %d\n",
 351                 dev->vid, start_idx, start_idx + count);
 352
 353         vq->batch_copy_nb_elems = 0;
 354
 355         /* Retrieve all of the desc indexes first to avoid caching issues. */
 356         rte_prefetch0(&vq->avail->ring[start_idx & (vq->size - 1)]);
 357         for (i = 0; i < count; i++) {
 358                 used_idx = (start_idx + i) & (vq->size - 1);
 359                 desc_indexes[i] = vq->avail->ring[used_idx];
 360                 vq->used->ring[used_idx].id = desc_indexes[i];
 361                 vq->used->ring[used_idx].len = pkts[i]->pkt_len +
 362                                                dev->vhost_hlen;
 363                 vhost_log_used_vring(dev, vq,
 364                         offsetof(struct vring_used, ring[used_idx]),
 365                         sizeof(vq->used->ring[used_idx]));
 366         }
 367
 368         rte_prefetch0(&vq->desc[desc_indexes[0]]);
 369         for (i = 0; i < count; i++) {
 370                 uint16_t desc_idx = desc_indexes[i];
 371                 int err;
 372
 373                 if (vq->desc[desc_idx].flags & VRING_DESC_F_INDIRECT) {
 374                         descs = (struct vring_desc *)(uintptr_t)
 375                                 vhost_iova_to_vva(dev,
 376                                                 vq, vq->desc[desc_idx].addr,
 377                                                 vq->desc[desc_idx].len,
 378                                                 VHOST_ACCESS_RO);
 379                         if (unlikely(!descs)) {
 380                                 count = i;
 381                                 break;
 382                         }
 383
 384                         desc_idx = 0;
 385                         sz = vq->desc[desc_idx].len / sizeof(*descs);
 386                 } else {
 387                         descs = vq->desc;
 388                         sz = vq->size;
 389                 }
 390
 391                 err = copy_mbuf_to_desc(dev, vq, descs, pkts[i], desc_idx, sz);
 392                 if (unlikely(err)) {
 393                         count = i;
 394                         break;
 395                 }
 396
 397                 if (i + 1 < count)
 398                         rte_prefetch0(&vq->desc[desc_indexes[i+1]]);
 399         }
 400
 401         do_data_copy_enqueue(dev, vq);
 402
 403         rte_smp_wmb();
 404
 405         *(volatile uint16_t *)&vq->used->idx += count;
 406         vq->last_used_idx += count;
 407         vhost_log_used_vring(dev, vq,
 408                 offsetof(struct vring_used, idx),
 409                 sizeof(vq->used->idx));
 410
 411         /* flush used->idx update before we read avail->flags. */
 412         rte_mb();
 413
 414         /* Kick the guest if necessary. */
 415         if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
 416                         && (vq->callfd >= 0))
 417                 eventfd_write(vq->callfd, (eventfd_t)1);
 418 out:
 419         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 420                 vhost_user_iotlb_rd_unlock(vq);
 421
 422         return count;
 423 }
 424
 425 static __rte_always_inline int
 426 fill_vec_buf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 427                          uint32_t avail_idx, uint32_t *vec_idx,
 428                          struct buf_vector *buf_vec, uint16_t *desc_chain_head,
 429                          uint16_t *desc_chain_len)
 430 {
 431         uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
 432         uint32_t vec_id = *vec_idx;
 433         uint32_t len    = 0;
 434         struct vring_desc *descs = vq->desc;
 435
 436         *desc_chain_head = idx;
 437
 438         if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
 439                 descs = (struct vring_desc *)(uintptr_t)
 440                         vhost_iova_to_vva(dev, vq, vq->desc[idx].addr,
 441                                                 vq->desc[idx].len,
 442                                                 VHOST_ACCESS_RO);
 443                 if (unlikely(!descs))
 444                         return -1;
 445
 446                 idx = 0;
 447         }
 448
 449         while (1) {
 450                 if (unlikely(vec_id >= BUF_VECTOR_MAX || idx >= vq->size))
 451                         return -1;
 452
 453                 len += descs[idx].len;
 454                 buf_vec[vec_id].buf_addr = descs[idx].addr;
 455                 buf_vec[vec_id].buf_len  = descs[idx].len;
 456                 buf_vec[vec_id].desc_idx = idx;
 457                 vec_id++;
 458
 459                 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
 460                         break;
 461
 462                 idx = descs[idx].next;
 463         }
 464
 465         *desc_chain_len = len;
 466         *vec_idx = vec_id;
 467
 468         return 0;
 469 }
 470
 471 /*
 472  * Returns -1 on fail, 0 on success
 473  */
 474 static inline int
 475 reserve_avail_buf_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
 476                                 uint32_t size, struct buf_vector *buf_vec,
 477                                 uint16_t *num_buffers, uint16_t avail_head)
 478 {
 479         uint16_t cur_idx;
 480         uint32_t vec_idx = 0;
 481         uint16_t tries = 0;
 482
 483         uint16_t head_idx = 0;
 484         uint16_t len = 0;
 485
 486         *num_buffers = 0;
 487         cur_idx  = vq->last_avail_idx;
 488
 489         while (size > 0) {
 490                 if (unlikely(cur_idx == avail_head))
 491                         return -1;
 492
 493                 if (unlikely(fill_vec_buf(dev, vq, cur_idx, &vec_idx, buf_vec,
 494                                                 &head_idx, &len) < 0))
 495                         return -1;
 496                 len = RTE_MIN(len, size);
 497                 update_shadow_used_ring(vq, head_idx, len);
 498                 size -= len;
 499
 500                 cur_idx++;
 501                 tries++;
 502                 *num_buffers += 1;
 503
 504                 /*
 505                  * if we tried all available ring items, and still
 506                  * can't get enough buf, it means something abnormal
 507                  * happened.
 508                  */
 509                 if (unlikely(tries >= vq->size))
 510                         return -1;
 511         }
 512
 513         return 0;
 514 }
 515
 516 static __rte_always_inline int
 517 copy_mbuf_to_desc_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
 518                             struct rte_mbuf *m, struct buf_vector *buf_vec,
 519                             uint16_t num_buffers)
 520 {
 521         uint32_t vec_idx = 0;
 522         uint64_t desc_addr;
 523         uint32_t mbuf_offset, mbuf_avail;
 524         uint32_t desc_offset, desc_avail;
 525         uint32_t cpy_len;
 526         uint64_t hdr_addr, hdr_phys_addr;
 527         struct rte_mbuf *hdr_mbuf;
 528         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 529         uint16_t copy_nb = vq->batch_copy_nb_elems;
 530         int error = 0;
 531
 532         if (unlikely(m == NULL)) {
 533                 error = -1;
 534                 goto out;
 535         }
 536
 537         desc_addr = vhost_iova_to_vva(dev, vq, buf_vec[vec_idx].buf_addr,
 538                                                 buf_vec[vec_idx].buf_len,
 539                                                 VHOST_ACCESS_RW);
 540         if (buf_vec[vec_idx].buf_len < dev->vhost_hlen || !desc_addr) {
 541                 error = -1;
 542                 goto out;
 543         }
 544
 545         hdr_mbuf = m;
 546         hdr_addr = desc_addr;
 547         hdr_phys_addr = buf_vec[vec_idx].buf_addr;
 548         rte_prefetch0((void *)(uintptr_t)hdr_addr);
 549
 550         LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
 551                 dev->vid, num_buffers);
 552
 553         desc_avail  = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
 554         desc_offset = dev->vhost_hlen;
 555
 556         mbuf_avail  = rte_pktmbuf_data_len(m);
 557         mbuf_offset = 0;
 558         while (mbuf_avail != 0 || m->next != NULL) {
 559                 /* done with current desc buf, get the next one */
 560                 if (desc_avail == 0) {
 561                         vec_idx++;
 562                         desc_addr =
 563                                 vhost_iova_to_vva(dev, vq,
 564                                         buf_vec[vec_idx].buf_addr,
 565                                         buf_vec[vec_idx].buf_len,
 566                                         VHOST_ACCESS_RW);
 567                         if (unlikely(!desc_addr)) {
 568                                 error = -1;
 569                                 goto out;
 570                         }
 571
 572                         /* Prefetch buffer address. */
 573                         rte_prefetch0((void *)(uintptr_t)desc_addr);
 574                         desc_offset = 0;
 575                         desc_avail  = buf_vec[vec_idx].buf_len;
 576                 }
 577
 578                 /* done with current mbuf, get the next one */
 579                 if (mbuf_avail == 0) {
 580                         m = m->next;
 581
 582                         mbuf_offset = 0;
 583                         mbuf_avail  = rte_pktmbuf_data_len(m);
 584                 }
 585
 586                 if (hdr_addr) {
 587                         struct virtio_net_hdr_mrg_rxbuf *hdr;
 588
 589                         hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)
 590                                 hdr_addr;
 591                         virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
 592                         ASSIGN_UNLESS_EQUAL(hdr->num_buffers, num_buffers);
 593
 594                         vhost_log_write(dev, hdr_phys_addr, dev->vhost_hlen);
 595                         PRINT_PACKET(dev, (uintptr_t)hdr_addr,
 596                                      dev->vhost_hlen, 0);
 597
 598                         hdr_addr = 0;
 599                 }
 600
 601                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 602
 603                 if (likely(cpy_len > MAX_BATCH_LEN || copy_nb >= vq->size)) {
 604                         rte_memcpy((void *)((uintptr_t)(desc_addr +
 605                                                         desc_offset)),
 606                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 607                                 cpy_len);
 608                         vhost_log_write(dev,
 609                                 buf_vec[vec_idx].buf_addr + desc_offset,
 610                                 cpy_len);
 611                         PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
 612                                 cpy_len, 0);
 613                 } else {
 614                         batch_copy[copy_nb].dst =
 615                                 (void *)((uintptr_t)(desc_addr + desc_offset));
 616                         batch_copy[copy_nb].src =
 617                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
 618                         batch_copy[copy_nb].log_addr =
 619                                 buf_vec[vec_idx].buf_addr + desc_offset;
 620                         batch_copy[copy_nb].len = cpy_len;
 621                         copy_nb++;
 622                 }
 623
 624                 mbuf_avail  -= cpy_len;
 625                 mbuf_offset += cpy_len;
 626                 desc_avail  -= cpy_len;
 627                 desc_offset += cpy_len;
 628         }
 629
 630 out:
 631         vq->batch_copy_nb_elems = copy_nb;
 632
 633         return error;
 634 }
 635
 636 static __rte_always_inline uint32_t
 637 virtio_dev_merge_rx(struct virtio_net *dev, uint16_t queue_id,
 638         struct rte_mbuf **pkts, uint32_t count)
 639 {
 640         struct vhost_virtqueue *vq;
 641         uint32_t pkt_idx = 0;
 642         uint16_t num_buffers;
 643         struct buf_vector buf_vec[BUF_VECTOR_MAX];
 644         uint16_t avail_head;
 645
 646         LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
 647         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
 648                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
 649                         dev->vid, __func__, queue_id);
 650                 return 0;
 651         }
 652
 653         vq = dev->virtqueue[queue_id];
 654         if (unlikely(vq->enabled == 0))
 655                 return 0;
 656
 657         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 658                 vhost_user_iotlb_rd_lock(vq);
 659
 660         if (unlikely(vq->access_ok == 0))
 661                 if (unlikely(vring_translate(dev, vq) < 0))
 662                         goto out;
 663
 664         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
 665         if (count == 0)
 666                 goto out;
 667
 668         vq->batch_copy_nb_elems = 0;
 669
 670         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
 671
 672         vq->shadow_used_idx = 0;
 673         avail_head = *((volatile uint16_t *)&vq->avail->idx);
 674         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
 675                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
 676
 677                 if (unlikely(reserve_avail_buf_mergeable(dev, vq,
 678                                                 pkt_len, buf_vec, &num_buffers,
 679                                                 avail_head) < 0)) {
 680                         LOG_DEBUG(VHOST_DATA,
 681                                 "(%d) failed to get enough desc from vring\n",
 682                                 dev->vid);
 683                         vq->shadow_used_idx -= num_buffers;
 684                         break;
 685                 }
 686
 687                 LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
 688                         dev->vid, vq->last_avail_idx,
 689                         vq->last_avail_idx + num_buffers);
 690
 691                 if (copy_mbuf_to_desc_mergeable(dev, vq, pkts[pkt_idx],
 692                                                 buf_vec, num_buffers) < 0) {
 693                         vq->shadow_used_idx -= num_buffers;
 694                         break;
 695                 }
 696
 697                 vq->last_avail_idx += num_buffers;
 698         }
 699
 700         do_data_copy_enqueue(dev, vq);
 701
 702         if (likely(vq->shadow_used_idx)) {
 703                 flush_shadow_used_ring(dev, vq);
 704
 705                 /* flush used->idx update before we read avail->flags. */
 706                 rte_mb();
 707
 708                 /* Kick the guest if necessary. */
 709                 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
 710                                 && (vq->callfd >= 0))
 711                         eventfd_write(vq->callfd, (eventfd_t)1);
 712         }
 713
 714 out:
 715         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 716                 vhost_user_iotlb_rd_unlock(vq);
 717
 718         return pkt_idx;
 719 }
 720
 721 uint16_t
 722 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
 723         struct rte_mbuf **pkts, uint16_t count)
 724 {
 725         struct virtio_net *dev = get_device(vid);
 726
 727         if (!dev)
 728                 return 0;
 729
 730         if (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF))
 731                 return virtio_dev_merge_rx(dev, queue_id, pkts, count);
 732         else
 733                 return virtio_dev_rx(dev, queue_id, pkts, count);
 734 }
 735
 736 static inline bool
 737 virtio_net_with_host_offload(struct virtio_net *dev)
 738 {
 739         if (dev->features &
 740                         ((1ULL << VIRTIO_NET_F_CSUM) |
 741                          (1ULL << VIRTIO_NET_F_HOST_ECN) |
 742                          (1ULL << VIRTIO_NET_F_HOST_TSO4) |
 743                          (1ULL << VIRTIO_NET_F_HOST_TSO6) |
 744                          (1ULL << VIRTIO_NET_F_HOST_UFO)))
 745                 return true;
 746
 747         return false;
 748 }
 749
 750 static void
 751 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
 752 {
 753         struct ipv4_hdr *ipv4_hdr;
 754         struct ipv6_hdr *ipv6_hdr;
 755         void *l3_hdr = NULL;
 756         struct ether_hdr *eth_hdr;
 757         uint16_t ethertype;
 758
 759         eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
 760
 761         m->l2_len = sizeof(struct ether_hdr);
 762         ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
 763
 764         if (ethertype == ETHER_TYPE_VLAN) {
 765                 struct vlan_hdr *vlan_hdr = (struct vlan_hdr *)(eth_hdr + 1);
 766
 767                 m->l2_len += sizeof(struct vlan_hdr);
 768                 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
 769         }
 770
 771         l3_hdr = (char *)eth_hdr + m->l2_len;
 772
 773         switch (ethertype) {
 774         case ETHER_TYPE_IPv4:
 775                 ipv4_hdr = l3_hdr;
 776                 *l4_proto = ipv4_hdr->next_proto_id;
 777                 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
 778                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 779                 m->ol_flags |= PKT_TX_IPV4;
 780                 break;
 781         case ETHER_TYPE_IPv6:
 782                 ipv6_hdr = l3_hdr;
 783                 *l4_proto = ipv6_hdr->proto;
 784                 m->l3_len = sizeof(struct ipv6_hdr);
 785                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 786                 m->ol_flags |= PKT_TX_IPV6;
 787                 break;
 788         default:
 789                 m->l3_len = 0;
 790                 *l4_proto = 0;
 791                 *l4_hdr = NULL;
 792                 break;
 793         }
 794 }
 795
 796 static __rte_always_inline void
 797 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
 798 {
 799         uint16_t l4_proto = 0;
 800         void *l4_hdr = NULL;
 801         struct tcp_hdr *tcp_hdr = NULL;
 802
 803         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
 804                 return;
 805
 806         parse_ethernet(m, &l4_proto, &l4_hdr);
 807         if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 808                 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
 809                         switch (hdr->csum_offset) {
 810                         case (offsetof(struct tcp_hdr, cksum)):
 811                                 if (l4_proto == IPPROTO_TCP)
 812                                         m->ol_flags |= PKT_TX_TCP_CKSUM;
 813                                 break;
 814                         case (offsetof(struct udp_hdr, dgram_cksum)):
 815                                 if (l4_proto == IPPROTO_UDP)
 816                                         m->ol_flags |= PKT_TX_UDP_CKSUM;
 817                                 break;
 818                         case (offsetof(struct sctp_hdr, cksum)):
 819                                 if (l4_proto == IPPROTO_SCTP)
 820                                         m->ol_flags |= PKT_TX_SCTP_CKSUM;
 821                                 break;
 822                         default:
 823                                 break;
 824                         }
 825                 }
 826         }
 827
 828         if (l4_hdr && hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 829                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 830                 case VIRTIO_NET_HDR_GSO_TCPV4:
 831                 case VIRTIO_NET_HDR_GSO_TCPV6:
 832                         tcp_hdr = l4_hdr;
 833                         m->ol_flags |= PKT_TX_TCP_SEG;
 834                         m->tso_segsz = hdr->gso_size;
 835                         m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
 836                         break;
 837                 default:
 838                         RTE_LOG(WARNING, VHOST_DATA,
 839                                 "unsupported gso type %u.\n", hdr->gso_type);
 840                         break;
 841                 }
 842         }
 843 }
 844
 845 #define RARP_PKT_SIZE   64
 846
 847 static int
 848 make_rarp_packet(struct rte_mbuf *rarp_mbuf, const struct ether_addr *mac)
 849 {
 850         struct ether_hdr *eth_hdr;
 851         struct arp_hdr  *rarp;
 852
 853         if (rarp_mbuf->buf_len < 64) {
 854                 RTE_LOG(WARNING, VHOST_DATA,
 855                         "failed to make RARP; mbuf size too small %u (< %d)\n",
 856                         rarp_mbuf->buf_len, RARP_PKT_SIZE);
 857                 return -1;
 858         }
 859
 860         /* Ethernet header. */
 861         eth_hdr = rte_pktmbuf_mtod_offset(rarp_mbuf, struct ether_hdr *, 0);
 862         memset(eth_hdr->d_addr.addr_bytes, 0xff, ETHER_ADDR_LEN);
 863         ether_addr_copy(mac, &eth_hdr->s_addr);
 864         eth_hdr->ether_type = htons(ETHER_TYPE_RARP);
 865
 866         /* RARP header. */
 867         rarp = (struct arp_hdr *)(eth_hdr + 1);
 868         rarp->arp_hrd = htons(ARP_HRD_ETHER);
 869         rarp->arp_pro = htons(ETHER_TYPE_IPv4);
 870         rarp->arp_hln = ETHER_ADDR_LEN;
 871         rarp->arp_pln = 4;
 872         rarp->arp_op  = htons(ARP_OP_REVREQUEST);
 873
 874         ether_addr_copy(mac, &rarp->arp_data.arp_sha);
 875         ether_addr_copy(mac, &rarp->arp_data.arp_tha);
 876         memset(&rarp->arp_data.arp_sip, 0x00, 4);
 877         memset(&rarp->arp_data.arp_tip, 0x00, 4);
 878
 879         rarp_mbuf->pkt_len  = rarp_mbuf->data_len = RARP_PKT_SIZE;
 880
 881         return 0;
 882 }
 883
 884 static __rte_always_inline void
 885 put_zmbuf(struct zcopy_mbuf *zmbuf)
 886 {
 887         zmbuf->in_use = 0;
 888 }
 889
 890 static __rte_always_inline int
 891 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 892                   struct vring_desc *descs, uint16_t max_desc,
 893                   struct rte_mbuf *m, uint16_t desc_idx,
 894                   struct rte_mempool *mbuf_pool)
 895 {
 896         struct vring_desc *desc;
 897         uint64_t desc_addr;
 898         uint32_t desc_avail, desc_offset;
 899         uint32_t mbuf_avail, mbuf_offset;
 900         uint32_t cpy_len;
 901         struct rte_mbuf *cur = m, *prev = m;
 902         struct virtio_net_hdr *hdr = NULL;
 903         /* A counter to avoid desc dead loop chain */
 904         uint32_t nr_desc = 1;
 905         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 906         uint16_t copy_nb = vq->batch_copy_nb_elems;
 907         int error = 0;
 908
 909         desc = &descs[desc_idx];
 910         if (unlikely((desc->len < dev->vhost_hlen)) ||
 911                         (desc->flags & VRING_DESC_F_INDIRECT)) {
 912                 error = -1;
 913                 goto out;
 914         }
 915
 916         desc_addr = vhost_iova_to_vva(dev,
 917                                         vq, desc->addr,
 918                                         desc->len,
 919                                         VHOST_ACCESS_RO);
 920         if (unlikely(!desc_addr)) {
 921                 error = -1;
 922                 goto out;
 923         }
 924
 925         if (virtio_net_with_host_offload(dev)) {
 926                 hdr = (struct virtio_net_hdr *)((uintptr_t)desc_addr);
 927                 rte_prefetch0(hdr);
 928         }
 929
 930         /*
 931          * A virtio driver normally uses at least 2 desc buffers
 932          * for Tx: the first for storing the header, and others
 933          * for storing the data.
 934          */
 935         if (likely((desc->len == dev->vhost_hlen) &&
 936                    (desc->flags & VRING_DESC_F_NEXT) != 0)) {
 937                 desc = &descs[desc->next];
 938                 if (unlikely(desc->flags & VRING_DESC_F_INDIRECT)) {
 939                         error = -1;
 940                         goto out;
 941                 }
 942
 943                 desc_addr = vhost_iova_to_vva(dev,
 944                                                         vq, desc->addr,
 945                                                         desc->len,
 946                                                         VHOST_ACCESS_RO);
 947                 if (unlikely(!desc_addr)) {
 948                         error = -1;
 949                         goto out;
 950                 }
 951
 952                 desc_offset = 0;
 953                 desc_avail  = desc->len;
 954                 nr_desc    += 1;
 955         } else {
 956                 desc_avail  = desc->len - dev->vhost_hlen;
 957                 desc_offset = dev->vhost_hlen;
 958         }
 959
 960         rte_prefetch0((void *)(uintptr_t)(desc_addr + desc_offset));
 961
 962         PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset), desc_avail, 0);
 963
 964         mbuf_offset = 0;
 965         mbuf_avail  = m->buf_len - RTE_PKTMBUF_HEADROOM;
 966         while (1) {
 967                 uint64_t hpa;
 968
 969                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 970
 971                 /*
 972                  * A desc buf might across two host physical pages that are
 973                  * not continuous. In such case (gpa_to_hpa returns 0), data
 974                  * will be copied even though zero copy is enabled.
 975                  */
 976                 if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
 977                                         desc->addr + desc_offset, cpy_len)))) {
 978                         cur->data_len = cpy_len;
 979                         cur->data_off = 0;
 980                         cur->buf_addr = (void *)(uintptr_t)desc_addr;
 981                         cur->buf_physaddr = hpa;
 982
 983                         /*
 984                          * In zero copy mode, one mbuf can only reference data
 985                          * for one or partial of one desc buff.
 986                          */
 987                         mbuf_avail = cpy_len;
 988                 } else {
 989                         if (likely(cpy_len > MAX_BATCH_LEN ||
 990                                    copy_nb >= vq->size ||
 991                                    (hdr && cur == m))) {
 992                                 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
 993                                                                    mbuf_offset),
 994                                            (void *)((uintptr_t)(desc_addr +
 995                                                                 desc_offset)),
 996                                            cpy_len);
 997                         } else {
 998                                 batch_copy[copy_nb].dst =
 999                                         rte_pktmbuf_mtod_offset(cur, void *,
1000                                                                 mbuf_offset);
1001                                 batch_copy[copy_nb].src =
1002                                         (void *)((uintptr_t)(desc_addr +
1003                                                              desc_offset));
1004                                 batch_copy[copy_nb].len = cpy_len;
1005                                 copy_nb++;
1006                         }
1007                 }
1008
1009                 mbuf_avail  -= cpy_len;
1010                 mbuf_offset += cpy_len;
1011                 desc_avail  -= cpy_len;
1012                 desc_offset += cpy_len;
1013
1014                 /* This desc reaches to its end, get the next one */
1015                 if (desc_avail == 0) {
1016                         if ((desc->flags & VRING_DESC_F_NEXT) == 0)
1017                                 break;
1018
1019                         if (unlikely(desc->next >= max_desc ||
1020                                      ++nr_desc > max_desc)) {
1021                                 error = -1;
1022                                 goto out;
1023                         }
1024                         desc = &descs[desc->next];
1025                         if (unlikely(desc->flags & VRING_DESC_F_INDIRECT)) {
1026                                 error = -1;
1027                                 goto out;
1028                         }
1029
1030                         desc_addr = vhost_iova_to_vva(dev,
1031                                                         vq, desc->addr,
1032                                                         desc->len,
1033                                                         VHOST_ACCESS_RO);
1034                         if (unlikely(!desc_addr)) {
1035                                 error = -1;
1036                                 goto out;
1037                         }
1038
1039                         rte_prefetch0((void *)(uintptr_t)desc_addr);
1040
1041                         desc_offset = 0;
1042                         desc_avail  = desc->len;
1043
1044                         PRINT_PACKET(dev, (uintptr_t)desc_addr, desc->len, 0);
1045                 }
1046
1047                 /*
1048                  * This mbuf reaches to its end, get a new one
1049                  * to hold more data.
1050                  */
1051                 if (mbuf_avail == 0) {
1052                         cur = rte_pktmbuf_alloc(mbuf_pool);
1053                         if (unlikely(cur == NULL)) {
1054                                 RTE_LOG(ERR, VHOST_DATA, "Failed to "
1055                                         "allocate memory for mbuf.\n");
1056                                 error = -1;
1057                                 goto out;
1058                         }
1059                         if (unlikely(dev->dequeue_zero_copy))
1060                                 rte_mbuf_refcnt_update(cur, 1);
1061
1062                         prev->next = cur;
1063                         prev->data_len = mbuf_offset;
1064                         m->nb_segs += 1;
1065                         m->pkt_len += mbuf_offset;
1066                         prev = cur;
1067
1068                         mbuf_offset = 0;
1069                         mbuf_avail  = cur->buf_len - RTE_PKTMBUF_HEADROOM;
1070                 }
1071         }
1072
1073         prev->data_len = mbuf_offset;
1074         m->pkt_len    += mbuf_offset;
1075
1076         if (hdr)
1077                 vhost_dequeue_offload(hdr, m);
1078
1079 out:
1080         vq->batch_copy_nb_elems = copy_nb;
1081
1082         return error;
1083 }
1084
1085 static __rte_always_inline void
1086 update_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq,
1087                  uint32_t used_idx, uint32_t desc_idx)
1088 {
1089         vq->used->ring[used_idx].id  = desc_idx;
1090         vq->used->ring[used_idx].len = 0;
1091         vhost_log_used_vring(dev, vq,
1092                         offsetof(struct vring_used, ring[used_idx]),
1093                         sizeof(vq->used->ring[used_idx]));
1094 }
1095
1096 static __rte_always_inline void
1097 update_used_idx(struct virtio_net *dev, struct vhost_virtqueue *vq,
1098                 uint32_t count)
1099 {
1100         if (unlikely(count == 0))
1101                 return;
1102
1103         rte_smp_wmb();
1104         rte_smp_rmb();
1105
1106         vq->used->idx += count;
1107         vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
1108                         sizeof(vq->used->idx));
1109
1110         /* Kick guest if required. */
1111         if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
1112                         && (vq->callfd >= 0))
1113                 eventfd_write(vq->callfd, (eventfd_t)1);
1114 }
1115
1116 static __rte_always_inline struct zcopy_mbuf *
1117 get_zmbuf(struct vhost_virtqueue *vq)
1118 {
1119         uint16_t i;
1120         uint16_t last;
1121         int tries = 0;
1122
1123         /* search [last_zmbuf_idx, zmbuf_size) */
1124         i = vq->last_zmbuf_idx;
1125         last = vq->zmbuf_size;
1126
1127 again:
1128         for (; i < last; i++) {
1129                 if (vq->zmbufs[i].in_use == 0) {
1130                         vq->last_zmbuf_idx = i + 1;
1131                         vq->zmbufs[i].in_use = 1;
1132                         return &vq->zmbufs[i];
1133                 }
1134         }
1135
1136         tries++;
1137         if (tries == 1) {
1138                 /* search [0, last_zmbuf_idx) */
1139                 i = 0;
1140                 last = vq->last_zmbuf_idx;
1141                 goto again;
1142         }
1143
1144         return NULL;
1145 }
1146
1147 static __rte_always_inline bool
1148 mbuf_is_consumed(struct rte_mbuf *m)
1149 {
1150         while (m) {
1151                 if (rte_mbuf_refcnt_read(m) > 1)
1152                         return false;
1153                 m = m->next;
1154         }
1155
1156         return true;
1157 }
1158
1159 uint16_t
1160 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
1161         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1162 {
1163         struct virtio_net *dev;
1164         struct rte_mbuf *rarp_mbuf = NULL;
1165         struct vhost_virtqueue *vq;
1166         uint32_t desc_indexes[MAX_PKT_BURST];
1167         uint32_t used_idx;
1168         uint32_t i = 0;
1169         uint16_t free_entries;
1170         uint16_t avail_idx;
1171
1172         dev = get_device(vid);
1173         if (!dev)
1174                 return 0;
1175
1176         if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
1177                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
1178                         dev->vid, __func__, queue_id);
1179                 return 0;
1180         }
1181
1182         vq = dev->virtqueue[queue_id];
1183         if (unlikely(vq->enabled == 0))
1184                 return 0;
1185
1186         vq->batch_copy_nb_elems = 0;
1187
1188         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1189                 vhost_user_iotlb_rd_lock(vq);
1190
1191         if (unlikely(vq->access_ok == 0))
1192                 if (unlikely(vring_translate(dev, vq) < 0))
1193                         goto out;
1194
1195         if (unlikely(dev->dequeue_zero_copy)) {
1196                 struct zcopy_mbuf *zmbuf, *next;
1197                 int nr_updated = 0;
1198
1199                 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1200                      zmbuf != NULL; zmbuf = next) {
1201                         next = TAILQ_NEXT(zmbuf, next);
1202
1203                         if (mbuf_is_consumed(zmbuf->mbuf)) {
1204                                 used_idx = vq->last_used_idx++ & (vq->size - 1);
1205                                 update_used_ring(dev, vq, used_idx,
1206                                                  zmbuf->desc_idx);
1207                                 nr_updated += 1;
1208
1209                                 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1210                                 rte_pktmbuf_free(zmbuf->mbuf);
1211                                 put_zmbuf(zmbuf);
1212                                 vq->nr_zmbuf -= 1;
1213                         }
1214                 }
1215
1216                 update_used_idx(dev, vq, nr_updated);
1217         }
1218
1219         /*
1220          * Construct a RARP broadcast packet, and inject it to the "pkts"
1221          * array, to looks like that guest actually send such packet.
1222          *
1223          * Check user_send_rarp() for more information.
1224          *
1225          * broadcast_rarp shares a cacheline in the virtio_net structure
1226          * with some fields that are accessed during enqueue and
1227          * rte_atomic16_cmpset() causes a write if using cmpxchg. This could
1228          * result in false sharing between enqueue and dequeue.
1229          *
1230          * Prevent unnecessary false sharing by reading broadcast_rarp first
1231          * and only performing cmpset if the read indicates it is likely to
1232          * be set.
1233          */
1234
1235         if (unlikely(rte_atomic16_read(&dev->broadcast_rarp) &&
1236                         rte_atomic16_cmpset((volatile uint16_t *)
1237                                 &dev->broadcast_rarp.cnt, 1, 0))) {
1238
1239                 rarp_mbuf = rte_pktmbuf_alloc(mbuf_pool);
1240                 if (rarp_mbuf == NULL) {
1241                         RTE_LOG(ERR, VHOST_DATA,
1242                                 "Failed to allocate memory for mbuf.\n");
1243                         return 0;
1244                 }
1245
1246                 if (make_rarp_packet(rarp_mbuf, &dev->mac)) {
1247                         rte_pktmbuf_free(rarp_mbuf);
1248                         rarp_mbuf = NULL;
1249                 } else {
1250                         count -= 1;
1251                 }
1252         }
1253
1254         free_entries = *((volatile uint16_t *)&vq->avail->idx) -
1255                         vq->last_avail_idx;
1256         if (free_entries == 0)
1257                 goto out;
1258
1259         LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1260
1261         /* Prefetch available and used ring */
1262         avail_idx = vq->last_avail_idx & (vq->size - 1);
1263         used_idx  = vq->last_used_idx  & (vq->size - 1);
1264         rte_prefetch0(&vq->avail->ring[avail_idx]);
1265         rte_prefetch0(&vq->used->ring[used_idx]);
1266
1267         count = RTE_MIN(count, MAX_PKT_BURST);
1268         count = RTE_MIN(count, free_entries);
1269         LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1270                         dev->vid, count);
1271
1272         /* Retrieve all of the head indexes first to avoid caching issues. */
1273         for (i = 0; i < count; i++) {
1274                 avail_idx = (vq->last_avail_idx + i) & (vq->size - 1);
1275                 used_idx  = (vq->last_used_idx  + i) & (vq->size - 1);
1276                 desc_indexes[i] = vq->avail->ring[avail_idx];
1277
1278                 if (likely(dev->dequeue_zero_copy == 0))
1279                         update_used_ring(dev, vq, used_idx, desc_indexes[i]);
1280         }
1281
1282         /* Prefetch descriptor index. */
1283         rte_prefetch0(&vq->desc[desc_indexes[0]]);
1284         for (i = 0; i < count; i++) {
1285                 struct vring_desc *desc;
1286                 uint16_t sz, idx;
1287                 int err;
1288
1289                 if (likely(i + 1 < count))
1290                         rte_prefetch0(&vq->desc[desc_indexes[i + 1]]);
1291
1292                 if (vq->desc[desc_indexes[i]].flags & VRING_DESC_F_INDIRECT) {
1293                         desc = (struct vring_desc *)(uintptr_t)
1294                                 vhost_iova_to_vva(dev, vq,
1295                                                 vq->desc[desc_indexes[i]].addr,
1296                                                 sizeof(*desc),
1297                                                 VHOST_ACCESS_RO);
1298                         if (unlikely(!desc))
1299                                 break;
1300
1301                         rte_prefetch0(desc);
1302                         sz = vq->desc[desc_indexes[i]].len / sizeof(*desc);
1303                         idx = 0;
1304                 } else {
1305                         desc = vq->desc;
1306                         sz = vq->size;
1307                         idx = desc_indexes[i];
1308                 }
1309
1310                 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
1311                 if (unlikely(pkts[i] == NULL)) {
1312                         RTE_LOG(ERR, VHOST_DATA,
1313                                 "Failed to allocate memory for mbuf.\n");
1314                         break;
1315                 }
1316
1317                 err = copy_desc_to_mbuf(dev, vq, desc, sz, pkts[i], idx,
1318                                         mbuf_pool);
1319                 if (unlikely(err)) {
1320                         rte_pktmbuf_free(pkts[i]);
1321                         break;
1322                 }
1323
1324                 if (unlikely(dev->dequeue_zero_copy)) {
1325                         struct zcopy_mbuf *zmbuf;
1326
1327                         zmbuf = get_zmbuf(vq);
1328                         if (!zmbuf) {
1329                                 rte_pktmbuf_free(pkts[i]);
1330                                 break;
1331                         }
1332                         zmbuf->mbuf = pkts[i];
1333                         zmbuf->desc_idx = desc_indexes[i];
1334
1335                         /*
1336                          * Pin lock the mbuf; we will check later to see
1337                          * whether the mbuf is freed (when we are the last
1338                          * user) or not. If that's the case, we then could
1339                          * update the used ring safely.
1340                          */
1341                         rte_mbuf_refcnt_update(pkts[i], 1);
1342
1343                         vq->nr_zmbuf += 1;
1344                         TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1345                 }
1346         }
1347         vq->last_avail_idx += i;
1348
1349         if (likely(dev->dequeue_zero_copy == 0)) {
1350                 do_data_copy_dequeue(vq);
1351                 vq->last_used_idx += i;
1352                 update_used_idx(dev, vq, i);
1353         }
1354
1355 out:
1356         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1357                 vhost_user_iotlb_rd_unlock(vq);
1358
1359         if (unlikely(rarp_mbuf != NULL)) {
1360                 /*
1361                  * Inject it to the head of "pkts" array, so that switch's mac
1362                  * learning table will get updated first.
1363                  */
1364                 memmove(&pkts[1], pkts, i * sizeof(struct rte_mbuf *));
1365                 pkts[0] = rarp_mbuf;
1366                 i += 1;
1367         }
1368
1369         return i;
1370 }