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