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                         used_idx = (start_idx + i) & (vq->size - 1);
 378                         vq->used->ring[used_idx].len = dev->vhost_hlen;
 379                         vhost_log_used_vring(dev, vq,
 380                                 offsetof(struct vring_used, ring[used_idx]),
 381                                 sizeof(vq->used->ring[used_idx]));
 382                 }
 383
 384                 if (i + 1 < count)
 385                         rte_prefetch0(&vq->desc[desc_indexes[i+1]]);
 386         }
 387
 388         do_data_copy_enqueue(dev, vq);
 389
 390         rte_smp_wmb();
 391
 392         *(volatile uint16_t *)&vq->used->idx += count;
 393         vq->last_used_idx += count;
 394         vhost_log_used_vring(dev, vq,
 395                 offsetof(struct vring_used, idx),
 396                 sizeof(vq->used->idx));
 397
 398         /* flush used->idx update before we read avail->flags. */
 399         rte_mb();
 400
 401         /* Kick the guest if necessary. */
 402         if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
 403                         && (vq->callfd >= 0))
 404                 eventfd_write(vq->callfd, (eventfd_t)1);
 405         return count;
 406 }
 407
 408 static __rte_always_inline int
 409 fill_vec_buf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 410                          uint32_t avail_idx, uint32_t *vec_idx,
 411                          struct buf_vector *buf_vec, uint16_t *desc_chain_head,
 412                          uint16_t *desc_chain_len)
 413 {
 414         uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
 415         uint32_t vec_id = *vec_idx;
 416         uint32_t len    = 0;
 417         struct vring_desc *descs = vq->desc;
 418
 419         *desc_chain_head = idx;
 420
 421         if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
 422                 descs = (struct vring_desc *)(uintptr_t)
 423                         rte_vhost_gpa_to_vva(dev->mem, vq->desc[idx].addr);
 424                 if (unlikely(!descs))
 425                         return -1;
 426
 427                 idx = 0;
 428         }
 429
 430         while (1) {
 431                 if (unlikely(vec_id >= BUF_VECTOR_MAX || idx >= vq->size))
 432                         return -1;
 433
 434                 len += descs[idx].len;
 435                 buf_vec[vec_id].buf_addr = descs[idx].addr;
 436                 buf_vec[vec_id].buf_len  = descs[idx].len;
 437                 buf_vec[vec_id].desc_idx = idx;
 438                 vec_id++;
 439
 440                 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
 441                         break;
 442
 443                 idx = descs[idx].next;
 444         }
 445
 446         *desc_chain_len = len;
 447         *vec_idx = vec_id;
 448
 449         return 0;
 450 }
 451
 452 /*
 453  * Returns -1 on fail, 0 on success
 454  */
 455 static inline int
 456 reserve_avail_buf_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
 457                                 uint32_t size, struct buf_vector *buf_vec,
 458                                 uint16_t *num_buffers, uint16_t avail_head)
 459 {
 460         uint16_t cur_idx;
 461         uint32_t vec_idx = 0;
 462         uint16_t tries = 0;
 463
 464         uint16_t head_idx = 0;
 465         uint16_t len = 0;
 466
 467         *num_buffers = 0;
 468         cur_idx  = vq->last_avail_idx;
 469
 470         while (size > 0) {
 471                 if (unlikely(cur_idx == avail_head))
 472                         return -1;
 473
 474                 if (unlikely(fill_vec_buf(dev, vq, cur_idx, &vec_idx, buf_vec,
 475                                                 &head_idx, &len) < 0))
 476                         return -1;
 477                 len = RTE_MIN(len, size);
 478                 update_shadow_used_ring(vq, head_idx, len);
 479                 size -= len;
 480
 481                 cur_idx++;
 482                 tries++;
 483                 *num_buffers += 1;
 484
 485                 /*
 486                  * if we tried all available ring items, and still
 487                  * can't get enough buf, it means something abnormal
 488                  * happened.
 489                  */
 490                 if (unlikely(tries >= vq->size))
 491                         return -1;
 492         }
 493
 494         return 0;
 495 }
 496
 497 static __rte_always_inline int
 498 copy_mbuf_to_desc_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
 499                             struct rte_mbuf *m, struct buf_vector *buf_vec,
 500                             uint16_t num_buffers)
 501 {
 502         uint32_t vec_idx = 0;
 503         uint64_t desc_addr;
 504         uint32_t mbuf_offset, mbuf_avail;
 505         uint32_t desc_offset, desc_avail;
 506         uint32_t cpy_len;
 507         uint64_t hdr_addr, hdr_phys_addr;
 508         struct rte_mbuf *hdr_mbuf;
 509         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 510         uint16_t copy_nb = vq->batch_copy_nb_elems;
 511         int error = 0;
 512
 513         if (unlikely(m == NULL)) {
 514                 error = -1;
 515                 goto out;
 516         }
 517
 518         desc_addr = rte_vhost_gpa_to_vva(dev->mem, buf_vec[vec_idx].buf_addr);
 519         if (buf_vec[vec_idx].buf_len < dev->vhost_hlen || !desc_addr) {
 520                 error = -1;
 521                 goto out;
 522         }
 523
 524         hdr_mbuf = m;
 525         hdr_addr = desc_addr;
 526         hdr_phys_addr = buf_vec[vec_idx].buf_addr;
 527         rte_prefetch0((void *)(uintptr_t)hdr_addr);
 528
 529         LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
 530                 dev->vid, num_buffers);
 531
 532         desc_avail  = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
 533         desc_offset = dev->vhost_hlen;
 534
 535         mbuf_avail  = rte_pktmbuf_data_len(m);
 536         mbuf_offset = 0;
 537         while (mbuf_avail != 0 || m->next != NULL) {
 538                 /* done with current desc buf, get the next one */
 539                 if (desc_avail == 0) {
 540                         vec_idx++;
 541                         desc_addr = rte_vhost_gpa_to_vva(dev->mem,
 542                                         buf_vec[vec_idx].buf_addr);
 543                         if (unlikely(!desc_addr)) {
 544                                 error = -1;
 545                                 goto out;
 546                         }
 547
 548                         /* Prefetch buffer address. */
 549                         rte_prefetch0((void *)(uintptr_t)desc_addr);
 550                         desc_offset = 0;
 551                         desc_avail  = buf_vec[vec_idx].buf_len;
 552                 }
 553
 554                 /* done with current mbuf, get the next one */
 555                 if (mbuf_avail == 0) {
 556                         m = m->next;
 557
 558                         mbuf_offset = 0;
 559                         mbuf_avail  = rte_pktmbuf_data_len(m);
 560                 }
 561
 562                 if (hdr_addr) {
 563                         struct virtio_net_hdr_mrg_rxbuf *hdr;
 564
 565                         hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)
 566                                 hdr_addr;
 567                         virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
 568                         ASSIGN_UNLESS_EQUAL(hdr->num_buffers, num_buffers);
 569
 570                         vhost_log_write(dev, hdr_phys_addr, dev->vhost_hlen);
 571                         PRINT_PACKET(dev, (uintptr_t)hdr_addr,
 572                                      dev->vhost_hlen, 0);
 573
 574                         hdr_addr = 0;
 575                 }
 576
 577                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 578
 579                 if (likely(cpy_len > MAX_BATCH_LEN || copy_nb >= vq->size)) {
 580                         rte_memcpy((void *)((uintptr_t)(desc_addr +
 581                                                         desc_offset)),
 582                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 583                                 cpy_len);
 584                         vhost_log_write(dev,
 585                                 buf_vec[vec_idx].buf_addr + desc_offset,
 586                                 cpy_len);
 587                         PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
 588                                 cpy_len, 0);
 589                 } else {
 590                         batch_copy[copy_nb].dst =
 591                                 (void *)((uintptr_t)(desc_addr + desc_offset));
 592                         batch_copy[copy_nb].src =
 593                                 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
 594                         batch_copy[copy_nb].log_addr =
 595                                 buf_vec[vec_idx].buf_addr + desc_offset;
 596                         batch_copy[copy_nb].len = cpy_len;
 597                         copy_nb++;
 598                 }
 599
 600                 mbuf_avail  -= cpy_len;
 601                 mbuf_offset += cpy_len;
 602                 desc_avail  -= cpy_len;
 603                 desc_offset += cpy_len;
 604         }
 605
 606 out:
 607         vq->batch_copy_nb_elems = copy_nb;
 608
 609         return error;
 610 }
 611
 612 static __rte_always_inline uint32_t
 613 virtio_dev_merge_rx(struct virtio_net *dev, uint16_t queue_id,
 614         struct rte_mbuf **pkts, uint32_t count)
 615 {
 616         struct vhost_virtqueue *vq;
 617         uint32_t pkt_idx = 0;
 618         uint16_t num_buffers;
 619         struct buf_vector buf_vec[BUF_VECTOR_MAX];
 620         uint16_t avail_head;
 621
 622         LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
 623         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
 624                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
 625                         dev->vid, __func__, queue_id);
 626                 return 0;
 627         }
 628
 629         vq = dev->virtqueue[queue_id];
 630         if (unlikely(vq->enabled == 0))
 631                 return 0;
 632
 633         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
 634         if (count == 0)
 635                 return 0;
 636
 637         vq->batch_copy_nb_elems = 0;
 638
 639         rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
 640
 641         vq->shadow_used_idx = 0;
 642         avail_head = *((volatile uint16_t *)&vq->avail->idx);
 643         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
 644                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
 645
 646                 if (unlikely(reserve_avail_buf_mergeable(dev, vq,
 647                                                 pkt_len, buf_vec, &num_buffers,
 648                                                 avail_head) < 0)) {
 649                         LOG_DEBUG(VHOST_DATA,
 650                                 "(%d) failed to get enough desc from vring\n",
 651                                 dev->vid);
 652                         vq->shadow_used_idx -= num_buffers;
 653                         break;
 654                 }
 655
 656                 LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
 657                         dev->vid, vq->last_avail_idx,
 658                         vq->last_avail_idx + num_buffers);
 659
 660                 if (copy_mbuf_to_desc_mergeable(dev, vq, pkts[pkt_idx],
 661                                                 buf_vec, num_buffers) < 0) {
 662                         vq->shadow_used_idx -= num_buffers;
 663                         break;
 664                 }
 665
 666                 vq->last_avail_idx += num_buffers;
 667         }
 668
 669         do_data_copy_enqueue(dev, vq);
 670
 671         if (likely(vq->shadow_used_idx)) {
 672                 flush_shadow_used_ring(dev, vq);
 673
 674                 /* flush used->idx update before we read avail->flags. */
 675                 rte_mb();
 676
 677                 /* Kick the guest if necessary. */
 678                 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
 679                                 && (vq->callfd >= 0))
 680                         eventfd_write(vq->callfd, (eventfd_t)1);
 681         }
 682
 683         return pkt_idx;
 684 }
 685
 686 uint16_t
 687 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
 688         struct rte_mbuf **pkts, uint16_t count)
 689 {
 690         struct virtio_net *dev = get_device(vid);
 691
 692         if (!dev)
 693                 return 0;
 694
 695         if (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF))
 696                 return virtio_dev_merge_rx(dev, queue_id, pkts, count);
 697         else
 698                 return virtio_dev_rx(dev, queue_id, pkts, count);
 699 }
 700
 701 static inline bool
 702 virtio_net_with_host_offload(struct virtio_net *dev)
 703 {
 704         if (dev->features &
 705                         ((1ULL << VIRTIO_NET_F_CSUM) |
 706                          (1ULL << VIRTIO_NET_F_HOST_ECN) |
 707                          (1ULL << VIRTIO_NET_F_HOST_TSO4) |
 708                          (1ULL << VIRTIO_NET_F_HOST_TSO6) |
 709                          (1ULL << VIRTIO_NET_F_HOST_UFO)))
 710                 return true;
 711
 712         return false;
 713 }
 714
 715 static void
 716 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
 717 {
 718         struct ipv4_hdr *ipv4_hdr;
 719         struct ipv6_hdr *ipv6_hdr;
 720         void *l3_hdr = NULL;
 721         struct ether_hdr *eth_hdr;
 722         uint16_t ethertype;
 723
 724         eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
 725
 726         m->l2_len = sizeof(struct ether_hdr);
 727         ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
 728
 729         if (ethertype == ETHER_TYPE_VLAN) {
 730                 struct vlan_hdr *vlan_hdr = (struct vlan_hdr *)(eth_hdr + 1);
 731
 732                 m->l2_len += sizeof(struct vlan_hdr);
 733                 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
 734         }
 735
 736         l3_hdr = (char *)eth_hdr + m->l2_len;
 737
 738         switch (ethertype) {
 739         case ETHER_TYPE_IPv4:
 740                 ipv4_hdr = l3_hdr;
 741                 *l4_proto = ipv4_hdr->next_proto_id;
 742                 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
 743                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 744                 m->ol_flags |= PKT_TX_IPV4;
 745                 break;
 746         case ETHER_TYPE_IPv6:
 747                 ipv6_hdr = l3_hdr;
 748                 *l4_proto = ipv6_hdr->proto;
 749                 m->l3_len = sizeof(struct ipv6_hdr);
 750                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 751                 m->ol_flags |= PKT_TX_IPV6;
 752                 break;
 753         default:
 754                 m->l3_len = 0;
 755                 *l4_proto = 0;
 756                 *l4_hdr = NULL;
 757                 break;
 758         }
 759 }
 760
 761 static __rte_always_inline void
 762 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
 763 {
 764         uint16_t l4_proto = 0;
 765         void *l4_hdr = NULL;
 766         struct tcp_hdr *tcp_hdr = NULL;
 767
 768         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
 769                 return;
 770
 771         parse_ethernet(m, &l4_proto, &l4_hdr);
 772         if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 773                 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
 774                         switch (hdr->csum_offset) {
 775                         case (offsetof(struct tcp_hdr, cksum)):
 776                                 if (l4_proto == IPPROTO_TCP)
 777                                         m->ol_flags |= PKT_TX_TCP_CKSUM;
 778                                 break;
 779                         case (offsetof(struct udp_hdr, dgram_cksum)):
 780                                 if (l4_proto == IPPROTO_UDP)
 781                                         m->ol_flags |= PKT_TX_UDP_CKSUM;
 782                                 break;
 783                         case (offsetof(struct sctp_hdr, cksum)):
 784                                 if (l4_proto == IPPROTO_SCTP)
 785                                         m->ol_flags |= PKT_TX_SCTP_CKSUM;
 786                                 break;
 787                         default:
 788                                 break;
 789                         }
 790                 }
 791         }
 792
 793         if (l4_hdr && hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 794                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 795                 case VIRTIO_NET_HDR_GSO_TCPV4:
 796                 case VIRTIO_NET_HDR_GSO_TCPV6:
 797                         tcp_hdr = l4_hdr;
 798                         m->ol_flags |= PKT_TX_TCP_SEG;
 799                         m->tso_segsz = hdr->gso_size;
 800                         m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
 801                         break;
 802                 default:
 803                         RTE_LOG(WARNING, VHOST_DATA,
 804                                 "unsupported gso type %u.\n", hdr->gso_type);
 805                         break;
 806                 }
 807         }
 808 }
 809
 810 #define RARP_PKT_SIZE   64
 811
 812 static int
 813 make_rarp_packet(struct rte_mbuf *rarp_mbuf, const struct ether_addr *mac)
 814 {
 815         struct ether_hdr *eth_hdr;
 816         struct arp_hdr  *rarp;
 817
 818         if (rarp_mbuf->buf_len < 64) {
 819                 RTE_LOG(WARNING, VHOST_DATA,
 820                         "failed to make RARP; mbuf size too small %u (< %d)\n",
 821                         rarp_mbuf->buf_len, RARP_PKT_SIZE);
 822                 return -1;
 823         }
 824
 825         /* Ethernet header. */
 826         eth_hdr = rte_pktmbuf_mtod_offset(rarp_mbuf, struct ether_hdr *, 0);
 827         memset(eth_hdr->d_addr.addr_bytes, 0xff, ETHER_ADDR_LEN);
 828         ether_addr_copy(mac, &eth_hdr->s_addr);
 829         eth_hdr->ether_type = htons(ETHER_TYPE_RARP);
 830
 831         /* RARP header. */
 832         rarp = (struct arp_hdr *)(eth_hdr + 1);
 833         rarp->arp_hrd = htons(ARP_HRD_ETHER);
 834         rarp->arp_pro = htons(ETHER_TYPE_IPv4);
 835         rarp->arp_hln = ETHER_ADDR_LEN;
 836         rarp->arp_pln = 4;
 837         rarp->arp_op  = htons(ARP_OP_REVREQUEST);
 838
 839         ether_addr_copy(mac, &rarp->arp_data.arp_sha);
 840         ether_addr_copy(mac, &rarp->arp_data.arp_tha);
 841         memset(&rarp->arp_data.arp_sip, 0x00, 4);
 842         memset(&rarp->arp_data.arp_tip, 0x00, 4);
 843
 844         rarp_mbuf->pkt_len  = rarp_mbuf->data_len = RARP_PKT_SIZE;
 845
 846         return 0;
 847 }
 848
 849 static __rte_always_inline void
 850 put_zmbuf(struct zcopy_mbuf *zmbuf)
 851 {
 852         zmbuf->in_use = 0;
 853 }
 854
 855 static __rte_always_inline int
 856 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 857                   struct vring_desc *descs, uint16_t max_desc,
 858                   struct rte_mbuf *m, uint16_t desc_idx,
 859                   struct rte_mempool *mbuf_pool)
 860 {
 861         struct vring_desc *desc;
 862         uint64_t desc_addr;
 863         uint32_t desc_avail, desc_offset;
 864         uint32_t mbuf_avail, mbuf_offset;
 865         uint32_t cpy_len;
 866         struct rte_mbuf *cur = m, *prev = m;
 867         struct virtio_net_hdr *hdr = NULL;
 868         /* A counter to avoid desc dead loop chain */
 869         uint32_t nr_desc = 1;
 870         struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 871         uint16_t copy_nb = vq->batch_copy_nb_elems;
 872         int error = 0;
 873
 874         desc = &descs[desc_idx];
 875         if (unlikely((desc->len < dev->vhost_hlen)) ||
 876                         (desc->flags & VRING_DESC_F_INDIRECT)) {
 877                 error = -1;
 878                 goto out;
 879         }
 880
 881         desc_addr = rte_vhost_gpa_to_vva(dev->mem, desc->addr);
 882         if (unlikely(!desc_addr)) {
 883                 error = -1;
 884                 goto out;
 885         }
 886
 887         if (virtio_net_with_host_offload(dev)) {
 888                 hdr = (struct virtio_net_hdr *)((uintptr_t)desc_addr);
 889                 rte_prefetch0(hdr);
 890         }
 891
 892         /*
 893          * A virtio driver normally uses at least 2 desc buffers
 894          * for Tx: the first for storing the header, and others
 895          * for storing the data.
 896          */
 897         if (likely((desc->len == dev->vhost_hlen) &&
 898                    (desc->flags & VRING_DESC_F_NEXT) != 0)) {
 899                 desc = &descs[desc->next];
 900                 if (unlikely(desc->flags & VRING_DESC_F_INDIRECT)) {
 901                         error = -1;
 902                         goto out;
 903                 }
 904
 905                 desc_addr = rte_vhost_gpa_to_vva(dev->mem, desc->addr);
 906                 if (unlikely(!desc_addr)) {
 907                         error = -1;
 908                         goto out;
 909                 }
 910
 911                 desc_offset = 0;
 912                 desc_avail  = desc->len;
 913                 nr_desc    += 1;
 914         } else {
 915                 desc_avail  = desc->len - dev->vhost_hlen;
 916                 desc_offset = dev->vhost_hlen;
 917         }
 918
 919         rte_prefetch0((void *)(uintptr_t)(desc_addr + desc_offset));
 920
 921         PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset), desc_avail, 0);
 922
 923         mbuf_offset = 0;
 924         mbuf_avail  = m->buf_len - RTE_PKTMBUF_HEADROOM;
 925         while (1) {
 926                 uint64_t hpa;
 927
 928                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 929
 930                 /*
 931                  * A desc buf might across two host physical pages that are
 932                  * not continuous. In such case (gpa_to_hpa returns 0), data
 933                  * will be copied even though zero copy is enabled.
 934                  */
 935                 if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
 936                                         desc->addr + desc_offset, cpy_len)))) {
 937                         cur->data_len = cpy_len;
 938                         cur->data_off = 0;
 939                         cur->buf_addr = (void *)(uintptr_t)desc_addr;
 940                         cur->buf_physaddr = hpa;
 941
 942                         /*
 943                          * In zero copy mode, one mbuf can only reference data
 944                          * for one or partial of one desc buff.
 945                          */
 946                         mbuf_avail = cpy_len;
 947                 } else {
 948                         if (likely(cpy_len > MAX_BATCH_LEN ||
 949                                    copy_nb >= vq->size)) {
 950                                 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
 951                                                                    mbuf_offset),
 952                                            (void *)((uintptr_t)(desc_addr +
 953                                                                 desc_offset)),
 954                                            cpy_len);
 955                         } else {
 956                                 batch_copy[copy_nb].dst =
 957                                         rte_pktmbuf_mtod_offset(cur, void *,
 958                                                                 mbuf_offset);
 959                                 batch_copy[copy_nb].src =
 960                                         (void *)((uintptr_t)(desc_addr +
 961                                                              desc_offset));
 962                                 batch_copy[copy_nb].len = cpy_len;
 963                                 copy_nb++;
 964                         }
 965                 }
 966
 967                 mbuf_avail  -= cpy_len;
 968                 mbuf_offset += cpy_len;
 969                 desc_avail  -= cpy_len;
 970                 desc_offset += cpy_len;
 971
 972                 /* This desc reaches to its end, get the next one */
 973                 if (desc_avail == 0) {
 974                         if ((desc->flags & VRING_DESC_F_NEXT) == 0)
 975                                 break;
 976
 977                         if (unlikely(desc->next >= max_desc ||
 978                                      ++nr_desc > max_desc)) {
 979                                 error = -1;
 980                                 goto out;
 981                         }
 982                         desc = &descs[desc->next];
 983                         if (unlikely(desc->flags & VRING_DESC_F_INDIRECT)) {
 984                                 error = -1;
 985                                 goto out;
 986                         }
 987
 988                         desc_addr = rte_vhost_gpa_to_vva(dev->mem, desc->addr);
 989                         if (unlikely(!desc_addr)) {
 990                                 error = -1;
 991                                 goto out;
 992                         }
 993
 994                         rte_prefetch0((void *)(uintptr_t)desc_addr);
 995
 996                         desc_offset = 0;
 997                         desc_avail  = desc->len;
 998
 999                         PRINT_PACKET(dev, (uintptr_t)desc_addr, desc->len, 0);
1000                 }
1001
1002                 /*
1003                  * This mbuf reaches to its end, get a new one
1004                  * to hold more data.
1005                  */
1006                 if (mbuf_avail == 0) {
1007                         cur = rte_pktmbuf_alloc(mbuf_pool);
1008                         if (unlikely(cur == NULL)) {
1009                                 RTE_LOG(ERR, VHOST_DATA, "Failed to "
1010                                         "allocate memory for mbuf.\n");
1011                                 error = -1;
1012                                 goto out;
1013                         }
1014                         if (unlikely(dev->dequeue_zero_copy))
1015                                 rte_mbuf_refcnt_update(cur, 1);
1016
1017                         prev->next = cur;
1018                         prev->data_len = mbuf_offset;
1019                         m->nb_segs += 1;
1020                         m->pkt_len += mbuf_offset;
1021                         prev = cur;
1022
1023                         mbuf_offset = 0;
1024                         mbuf_avail  = cur->buf_len - RTE_PKTMBUF_HEADROOM;
1025                 }
1026         }
1027
1028         prev->data_len = mbuf_offset;
1029         m->pkt_len    += mbuf_offset;
1030
1031         if (hdr)
1032                 vhost_dequeue_offload(hdr, m);
1033
1034 out:
1035         vq->batch_copy_nb_elems = copy_nb;
1036
1037         return error;
1038 }
1039
1040 static __rte_always_inline void
1041 update_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq,
1042                  uint32_t used_idx, uint32_t desc_idx)
1043 {
1044         vq->used->ring[used_idx].id  = desc_idx;
1045         vq->used->ring[used_idx].len = 0;
1046         vhost_log_used_vring(dev, vq,
1047                         offsetof(struct vring_used, ring[used_idx]),
1048                         sizeof(vq->used->ring[used_idx]));
1049 }
1050
1051 static __rte_always_inline void
1052 update_used_idx(struct virtio_net *dev, struct vhost_virtqueue *vq,
1053                 uint32_t count)
1054 {
1055         if (unlikely(count == 0))
1056                 return;
1057
1058         rte_smp_wmb();
1059         rte_smp_rmb();
1060
1061         vq->used->idx += count;
1062         vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
1063                         sizeof(vq->used->idx));
1064
1065         /* Kick guest if required. */
1066         if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
1067                         && (vq->callfd >= 0))
1068                 eventfd_write(vq->callfd, (eventfd_t)1);
1069 }
1070
1071 static __rte_always_inline struct zcopy_mbuf *
1072 get_zmbuf(struct vhost_virtqueue *vq)
1073 {
1074         uint16_t i;
1075         uint16_t last;
1076         int tries = 0;
1077
1078         /* search [last_zmbuf_idx, zmbuf_size) */
1079         i = vq->last_zmbuf_idx;
1080         last = vq->zmbuf_size;
1081
1082 again:
1083         for (; i < last; i++) {
1084                 if (vq->zmbufs[i].in_use == 0) {
1085                         vq->last_zmbuf_idx = i + 1;
1086                         vq->zmbufs[i].in_use = 1;
1087                         return &vq->zmbufs[i];
1088                 }
1089         }
1090
1091         tries++;
1092         if (tries == 1) {
1093                 /* search [0, last_zmbuf_idx) */
1094                 i = 0;
1095                 last = vq->last_zmbuf_idx;
1096                 goto again;
1097         }
1098
1099         return NULL;
1100 }
1101
1102 static __rte_always_inline bool
1103 mbuf_is_consumed(struct rte_mbuf *m)
1104 {
1105         while (m) {
1106                 if (rte_mbuf_refcnt_read(m) > 1)
1107                         return false;
1108                 m = m->next;
1109         }
1110
1111         return true;
1112 }
1113
1114 uint16_t
1115 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
1116         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
1117 {
1118         struct virtio_net *dev;
1119         struct rte_mbuf *rarp_mbuf = NULL;
1120         struct vhost_virtqueue *vq;
1121         uint32_t desc_indexes[MAX_PKT_BURST];
1122         uint32_t used_idx;
1123         uint32_t i = 0;
1124         uint16_t free_entries;
1125         uint16_t avail_idx;
1126
1127         dev = get_device(vid);
1128         if (!dev)
1129                 return 0;
1130
1131         if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
1132                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
1133                         dev->vid, __func__, queue_id);
1134                 return 0;
1135         }
1136
1137         vq = dev->virtqueue[queue_id];
1138         if (unlikely(vq->enabled == 0))
1139                 return 0;
1140
1141         vq->batch_copy_nb_elems = 0;
1142
1143         if (unlikely(dev->dequeue_zero_copy)) {
1144                 struct zcopy_mbuf *zmbuf, *next;
1145                 int nr_updated = 0;
1146
1147                 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1148                      zmbuf != NULL; zmbuf = next) {
1149                         next = TAILQ_NEXT(zmbuf, next);
1150
1151                         if (mbuf_is_consumed(zmbuf->mbuf)) {
1152                                 used_idx = vq->last_used_idx++ & (vq->size - 1);
1153                                 update_used_ring(dev, vq, used_idx,
1154                                                  zmbuf->desc_idx);
1155                                 nr_updated += 1;
1156
1157                                 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1158                                 rte_pktmbuf_free(zmbuf->mbuf);
1159                                 put_zmbuf(zmbuf);
1160                                 vq->nr_zmbuf -= 1;
1161                         }
1162                 }
1163
1164                 update_used_idx(dev, vq, nr_updated);
1165         }
1166
1167         /*
1168          * Construct a RARP broadcast packet, and inject it to the "pkts"
1169          * array, to looks like that guest actually send such packet.
1170          *
1171          * Check user_send_rarp() for more information.
1172          *
1173          * broadcast_rarp shares a cacheline in the virtio_net structure
1174          * with some fields that are accessed during enqueue and
1175          * rte_atomic16_cmpset() causes a write if using cmpxchg. This could
1176          * result in false sharing between enqueue and dequeue.
1177          *
1178          * Prevent unnecessary false sharing by reading broadcast_rarp first
1179          * and only performing cmpset if the read indicates it is likely to
1180          * be set.
1181          */
1182
1183         if (unlikely(rte_atomic16_read(&dev->broadcast_rarp) &&
1184                         rte_atomic16_cmpset((volatile uint16_t *)
1185                                 &dev->broadcast_rarp.cnt, 1, 0))) {
1186
1187                 rarp_mbuf = rte_pktmbuf_alloc(mbuf_pool);
1188                 if (rarp_mbuf == NULL) {
1189                         RTE_LOG(ERR, VHOST_DATA,
1190                                 "Failed to allocate memory for mbuf.\n");
1191                         return 0;
1192                 }
1193
1194                 if (make_rarp_packet(rarp_mbuf, &dev->mac)) {
1195                         rte_pktmbuf_free(rarp_mbuf);
1196                         rarp_mbuf = NULL;
1197                 } else {
1198                         count -= 1;
1199                 }
1200         }
1201
1202         free_entries = *((volatile uint16_t *)&vq->avail->idx) -
1203                         vq->last_avail_idx;
1204         if (free_entries == 0)
1205                 goto out;
1206
1207         LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1208
1209         /* Prefetch available and used ring */
1210         avail_idx = vq->last_avail_idx & (vq->size - 1);
1211         used_idx  = vq->last_used_idx  & (vq->size - 1);
1212         rte_prefetch0(&vq->avail->ring[avail_idx]);
1213         rte_prefetch0(&vq->used->ring[used_idx]);
1214
1215         count = RTE_MIN(count, MAX_PKT_BURST);
1216         count = RTE_MIN(count, free_entries);
1217         LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1218                         dev->vid, count);
1219
1220         /* Retrieve all of the head indexes first to avoid caching issues. */
1221         for (i = 0; i < count; i++) {
1222                 avail_idx = (vq->last_avail_idx + i) & (vq->size - 1);
1223                 used_idx  = (vq->last_used_idx  + i) & (vq->size - 1);
1224                 desc_indexes[i] = vq->avail->ring[avail_idx];
1225
1226                 if (likely(dev->dequeue_zero_copy == 0))
1227                         update_used_ring(dev, vq, used_idx, desc_indexes[i]);
1228         }
1229
1230         /* Prefetch descriptor index. */
1231         rte_prefetch0(&vq->desc[desc_indexes[0]]);
1232         for (i = 0; i < count; i++) {
1233                 struct vring_desc *desc;
1234                 uint16_t sz, idx;
1235                 int err;
1236
1237                 if (likely(i + 1 < count))
1238                         rte_prefetch0(&vq->desc[desc_indexes[i + 1]]);
1239
1240                 if (vq->desc[desc_indexes[i]].flags & VRING_DESC_F_INDIRECT) {
1241                         desc = (struct vring_desc *)(uintptr_t)
1242                                 rte_vhost_gpa_to_vva(dev->mem,
1243                                         vq->desc[desc_indexes[i]].addr);
1244                         if (unlikely(!desc))
1245                                 break;
1246
1247                         rte_prefetch0(desc);
1248                         sz = vq->desc[desc_indexes[i]].len / sizeof(*desc);
1249                         idx = 0;
1250                 } else {
1251                         desc = vq->desc;
1252                         sz = vq->size;
1253                         idx = desc_indexes[i];
1254                 }
1255
1256                 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
1257                 if (unlikely(pkts[i] == NULL)) {
1258                         RTE_LOG(ERR, VHOST_DATA,
1259                                 "Failed to allocate memory for mbuf.\n");
1260                         break;
1261                 }
1262
1263                 err = copy_desc_to_mbuf(dev, vq, desc, sz, pkts[i], idx,
1264                                         mbuf_pool);
1265                 if (unlikely(err)) {
1266                         rte_pktmbuf_free(pkts[i]);
1267                         break;
1268                 }
1269
1270                 if (unlikely(dev->dequeue_zero_copy)) {
1271                         struct zcopy_mbuf *zmbuf;
1272
1273                         zmbuf = get_zmbuf(vq);
1274                         if (!zmbuf) {
1275                                 rte_pktmbuf_free(pkts[i]);
1276                                 break;
1277                         }
1278                         zmbuf->mbuf = pkts[i];
1279                         zmbuf->desc_idx = desc_indexes[i];
1280
1281                         /*
1282                          * Pin lock the mbuf; we will check later to see
1283                          * whether the mbuf is freed (when we are the last
1284                          * user) or not. If that's the case, we then could
1285                          * update the used ring safely.
1286                          */
1287                         rte_mbuf_refcnt_update(pkts[i], 1);
1288
1289                         vq->nr_zmbuf += 1;
1290                         TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1291                 }
1292         }
1293         vq->last_avail_idx += i;
1294
1295         if (likely(dev->dequeue_zero_copy == 0)) {
1296                 do_data_copy_dequeue(vq);
1297                 vq->last_used_idx += i;
1298                 update_used_idx(dev, vq, i);
1299         }
1300
1301 out:
1302         if (unlikely(rarp_mbuf != NULL)) {
1303                 /*
1304                  * Inject it to the head of "pkts" array, so that switch's mac
1305                  * learning table will get updated first.
1306                  */
1307                 memmove(&pkts[1], pkts, i * sizeof(struct rte_mbuf *));
1308                 pkts[0] = rarp_mbuf;
1309                 i += 1;
1310         }
1311
1312         return i;
1313 }