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_virtio_net.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 #define VHOST_LOG_PAGE  4096
  52
  53 static inline void __attribute__((always_inline))
  54 vhost_log_page(uint8_t *log_base, uint64_t page)
  55 {
  56         log_base[page / 8] |= 1 << (page % 8);
  57 }
  58
  59 static inline void __attribute__((always_inline))
  60 vhost_log_write(struct virtio_net *dev, uint64_t addr, uint64_t len)
  61 {
  62         uint64_t page;
  63
  64         if (likely(((dev->features & (1ULL << VHOST_F_LOG_ALL)) == 0) ||
  65                    !dev->log_base || !len))
  66                 return;
  67
  68         if (unlikely(dev->log_size <= ((addr + len - 1) / VHOST_LOG_PAGE / 8)))
  69                 return;
  70
  71         /* To make sure guest memory updates are committed before logging */
  72         rte_smp_wmb();
  73
  74         page = addr / VHOST_LOG_PAGE;
  75         while (page * VHOST_LOG_PAGE < addr + len) {
  76                 vhost_log_page((uint8_t *)(uintptr_t)dev->log_base, page);
  77                 page += 1;
  78         }
  79 }
  80
  81 static inline void __attribute__((always_inline))
  82 vhost_log_used_vring(struct virtio_net *dev, struct vhost_virtqueue *vq,
  83                      uint64_t offset, uint64_t len)
  84 {
  85         vhost_log_write(dev, vq->log_guest_addr + offset, len);
  86 }
  87
  88 static bool
  89 is_valid_virt_queue_idx(uint32_t idx, int is_tx, uint32_t qp_nb)
  90 {
  91         return (is_tx ^ (idx & 1)) == 0 && idx < qp_nb * VIRTIO_QNUM;
  92 }
  93
  94 static void
  95 virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
  96 {
  97         if (m_buf->ol_flags & PKT_TX_L4_MASK) {
  98                 net_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
  99                 net_hdr->csum_start = m_buf->l2_len + m_buf->l3_len;
 100
 101                 switch (m_buf->ol_flags & PKT_TX_L4_MASK) {
 102                 case PKT_TX_TCP_CKSUM:
 103                         net_hdr->csum_offset = (offsetof(struct tcp_hdr,
 104                                                 cksum));
 105                         break;
 106                 case PKT_TX_UDP_CKSUM:
 107                         net_hdr->csum_offset = (offsetof(struct udp_hdr,
 108                                                 dgram_cksum));
 109                         break;
 110                 case PKT_TX_SCTP_CKSUM:
 111                         net_hdr->csum_offset = (offsetof(struct sctp_hdr,
 112                                                 cksum));
 113                         break;
 114                 }
 115         }
 116
 117         if (m_buf->ol_flags & PKT_TX_TCP_SEG) {
 118                 if (m_buf->ol_flags & PKT_TX_IPV4)
 119                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 120                 else
 121                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 122                 net_hdr->gso_size = m_buf->tso_segsz;
 123                 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len
 124                                         + m_buf->l4_len;
 125         }
 126 }
 127
 128 static inline void
 129 copy_virtio_net_hdr(struct virtio_net *dev, uint64_t desc_addr,
 130                     struct virtio_net_hdr_mrg_rxbuf hdr)
 131 {
 132         if (dev->vhost_hlen == sizeof(struct virtio_net_hdr_mrg_rxbuf))
 133                 *(struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)desc_addr = hdr;
 134         else
 135                 *(struct virtio_net_hdr *)(uintptr_t)desc_addr = hdr.hdr;
 136 }
 137
 138 static inline int __attribute__((always_inline))
 139 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 140                   struct rte_mbuf *m, uint16_t desc_idx)
 141 {
 142         uint32_t desc_avail, desc_offset;
 143         uint32_t mbuf_avail, mbuf_offset;
 144         uint32_t cpy_len;
 145         struct vring_desc *desc;
 146         uint64_t desc_addr;
 147         struct virtio_net_hdr_mrg_rxbuf virtio_hdr = {{0, 0, 0, 0, 0, 0}, 0};
 148
 149         desc = &vq->desc[desc_idx];
 150         desc_addr = gpa_to_vva(dev, desc->addr);
 151         /*
 152          * Checking of 'desc_addr' placed outside of 'unlikely' macro to avoid
 153          * performance issue with some versions of gcc (4.8.4 and 5.3.0) which
 154          * otherwise stores offset on the stack instead of in a register.
 155          */
 156         if (unlikely(desc->len < dev->vhost_hlen) || !desc_addr)
 157                 return -1;
 158
 159         rte_prefetch0((void *)(uintptr_t)desc_addr);
 160
 161         virtio_enqueue_offload(m, &virtio_hdr.hdr);
 162         copy_virtio_net_hdr(dev, desc_addr, virtio_hdr);
 163         vhost_log_write(dev, desc->addr, dev->vhost_hlen);
 164         PRINT_PACKET(dev, (uintptr_t)desc_addr, dev->vhost_hlen, 0);
 165
 166         desc_offset = dev->vhost_hlen;
 167         desc_avail  = desc->len - dev->vhost_hlen;
 168
 169         mbuf_avail  = rte_pktmbuf_data_len(m);
 170         mbuf_offset = 0;
 171         while (mbuf_avail != 0 || m->next != NULL) {
 172                 /* done with current mbuf, fetch next */
 173                 if (mbuf_avail == 0) {
 174                         m = m->next;
 175
 176                         mbuf_offset = 0;
 177                         mbuf_avail  = rte_pktmbuf_data_len(m);
 178                 }
 179
 180                 /* done with current desc buf, fetch next */
 181                 if (desc_avail == 0) {
 182                         if ((desc->flags & VRING_DESC_F_NEXT) == 0) {
 183                                 /* Room in vring buffer is not enough */
 184                                 return -1;
 185                         }
 186                         if (unlikely(desc->next >= vq->size))
 187                                 return -1;
 188
 189                         desc = &vq->desc[desc->next];
 190                         desc_addr = gpa_to_vva(dev, desc->addr);
 191                         if (unlikely(!desc_addr))
 192                                 return -1;
 193
 194                         desc_offset = 0;
 195                         desc_avail  = desc->len;
 196                 }
 197
 198                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 199                 rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)),
 200                         rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 201                         cpy_len);
 202                 vhost_log_write(dev, desc->addr + desc_offset, cpy_len);
 203                 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
 204                              cpy_len, 0);
 205
 206                 mbuf_avail  -= cpy_len;
 207                 mbuf_offset += cpy_len;
 208                 desc_avail  -= cpy_len;
 209                 desc_offset += cpy_len;
 210         }
 211
 212         return 0;
 213 }
 214
 215 /**
 216  * This function adds buffers to the virtio devices RX virtqueue. Buffers can
 217  * be received from the physical port or from another virtio device. A packet
 218  * count is returned to indicate the number of packets that are succesfully
 219  * added to the RX queue. This function works when the mbuf is scattered, but
 220  * it doesn't support the mergeable feature.
 221  */
 222 static inline uint32_t __attribute__((always_inline))
 223 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
 224               struct rte_mbuf **pkts, uint32_t count)
 225 {
 226         struct vhost_virtqueue *vq;
 227         uint16_t avail_idx, free_entries, start_idx;
 228         uint16_t desc_indexes[MAX_PKT_BURST];
 229         uint16_t used_idx;
 230         uint32_t i;
 231
 232         LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
 233         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->virt_qp_nb))) {
 234                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
 235                         dev->vid, __func__, queue_id);
 236                 return 0;
 237         }
 238
 239         vq = dev->virtqueue[queue_id];
 240         if (unlikely(vq->enabled == 0))
 241                 return 0;
 242
 243         avail_idx = *((volatile uint16_t *)&vq->avail->idx);
 244         start_idx = vq->last_used_idx;
 245         free_entries = avail_idx - start_idx;
 246         count = RTE_MIN(count, free_entries);
 247         count = RTE_MIN(count, (uint32_t)MAX_PKT_BURST);
 248         if (count == 0)
 249                 return 0;
 250
 251         LOG_DEBUG(VHOST_DATA, "(%d) start_idx %d | end_idx %d\n",
 252                 dev->vid, start_idx, start_idx + count);
 253
 254         /* Retrieve all of the desc indexes first to avoid caching issues. */
 255         rte_prefetch0(&vq->avail->ring[start_idx & (vq->size - 1)]);
 256         for (i = 0; i < count; i++) {
 257                 used_idx = (start_idx + i) & (vq->size - 1);
 258                 desc_indexes[i] = vq->avail->ring[used_idx];
 259                 vq->used->ring[used_idx].id = desc_indexes[i];
 260                 vq->used->ring[used_idx].len = pkts[i]->pkt_len +
 261                                                dev->vhost_hlen;
 262                 vhost_log_used_vring(dev, vq,
 263                         offsetof(struct vring_used, ring[used_idx]),
 264                         sizeof(vq->used->ring[used_idx]));
 265         }
 266
 267         rte_prefetch0(&vq->desc[desc_indexes[0]]);
 268         for (i = 0; i < count; i++) {
 269                 uint16_t desc_idx = desc_indexes[i];
 270                 int err;
 271
 272                 err = copy_mbuf_to_desc(dev, vq, pkts[i], desc_idx);
 273                 if (unlikely(err)) {
 274                         used_idx = (start_idx + i) & (vq->size - 1);
 275                         vq->used->ring[used_idx].len = dev->vhost_hlen;
 276                         vhost_log_used_vring(dev, vq,
 277                                 offsetof(struct vring_used, ring[used_idx]),
 278                                 sizeof(vq->used->ring[used_idx]));
 279                 }
 280
 281                 if (i + 1 < count)
 282                         rte_prefetch0(&vq->desc[desc_indexes[i+1]]);
 283         }
 284
 285         rte_smp_wmb();
 286
 287         *(volatile uint16_t *)&vq->used->idx += count;
 288         vq->last_used_idx += count;
 289         vhost_log_used_vring(dev, vq,
 290                 offsetof(struct vring_used, idx),
 291                 sizeof(vq->used->idx));
 292
 293         /* flush used->idx update before we read avail->flags. */
 294         rte_mb();
 295
 296         /* Kick the guest if necessary. */
 297         if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
 298                         && (vq->callfd >= 0))
 299                 eventfd_write(vq->callfd, (eventfd_t)1);
 300         return count;
 301 }
 302
 303 static inline int
 304 fill_vec_buf(struct vhost_virtqueue *vq, uint32_t avail_idx,
 305              uint32_t *allocated, uint32_t *vec_idx,
 306              struct buf_vector *buf_vec)
 307 {
 308         uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
 309         uint32_t vec_id = *vec_idx;
 310         uint32_t len    = *allocated;
 311
 312         while (1) {
 313                 if (unlikely(vec_id >= BUF_VECTOR_MAX || idx >= vq->size))
 314                         return -1;
 315
 316                 len += vq->desc[idx].len;
 317                 buf_vec[vec_id].buf_addr = vq->desc[idx].addr;
 318                 buf_vec[vec_id].buf_len  = vq->desc[idx].len;
 319                 buf_vec[vec_id].desc_idx = idx;
 320                 vec_id++;
 321
 322                 if ((vq->desc[idx].flags & VRING_DESC_F_NEXT) == 0)
 323                         break;
 324
 325                 idx = vq->desc[idx].next;
 326         }
 327
 328         *allocated = len;
 329         *vec_idx   = vec_id;
 330
 331         return 0;
 332 }
 333
 334 /*
 335  * Returns -1 on fail, 0 on success
 336  */
 337 static inline int
 338 reserve_avail_buf_mergeable(struct vhost_virtqueue *vq, uint32_t size,
 339                             uint16_t *end, struct buf_vector *buf_vec)
 340 {
 341         uint16_t cur_idx;
 342         uint16_t avail_idx;
 343         uint32_t allocated = 0;
 344         uint32_t vec_idx = 0;
 345         uint16_t tries = 0;
 346
 347         cur_idx  = vq->last_used_idx;
 348
 349         while (1) {
 350                 avail_idx = *((volatile uint16_t *)&vq->avail->idx);
 351                 if (unlikely(cur_idx == avail_idx))
 352                         return -1;
 353
 354                 if (unlikely(fill_vec_buf(vq, cur_idx, &allocated,
 355                                           &vec_idx, buf_vec) < 0))
 356                         return -1;
 357
 358                 cur_idx++;
 359                 tries++;
 360
 361                 if (allocated >= size)
 362                         break;
 363
 364                 /*
 365                  * if we tried all available ring items, and still
 366                  * can't get enough buf, it means something abnormal
 367                  * happened.
 368                  */
 369                 if (unlikely(tries >= vq->size))
 370                         return -1;
 371         }
 372
 373         *end = cur_idx;
 374         return 0;
 375 }
 376
 377 static inline uint32_t __attribute__((always_inline))
 378 copy_mbuf_to_desc_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
 379                             uint16_t end_idx, struct rte_mbuf *m,
 380                             struct buf_vector *buf_vec)
 381 {
 382         struct virtio_net_hdr_mrg_rxbuf virtio_hdr = {{0, 0, 0, 0, 0, 0}, 0};
 383         uint32_t vec_idx = 0;
 384         uint16_t start_idx = vq->last_used_idx;
 385         uint16_t cur_idx = start_idx;
 386         uint64_t desc_addr;
 387         uint32_t desc_chain_head;
 388         uint32_t desc_chain_len;
 389         uint32_t mbuf_offset, mbuf_avail;
 390         uint32_t desc_offset, desc_avail;
 391         uint32_t cpy_len;
 392         uint16_t desc_idx, used_idx;
 393
 394         if (unlikely(m == NULL))
 395                 return 0;
 396
 397         LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
 398                 dev->vid, cur_idx, end_idx);
 399
 400         desc_addr = gpa_to_vva(dev, buf_vec[vec_idx].buf_addr);
 401         if (buf_vec[vec_idx].buf_len < dev->vhost_hlen || !desc_addr)
 402                 return 0;
 403
 404         rte_prefetch0((void *)(uintptr_t)desc_addr);
 405
 406         virtio_hdr.num_buffers = end_idx - start_idx;
 407         LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
 408                 dev->vid, virtio_hdr.num_buffers);
 409
 410         virtio_enqueue_offload(m, &virtio_hdr.hdr);
 411         copy_virtio_net_hdr(dev, desc_addr, virtio_hdr);
 412         vhost_log_write(dev, buf_vec[vec_idx].buf_addr, dev->vhost_hlen);
 413         PRINT_PACKET(dev, (uintptr_t)desc_addr, dev->vhost_hlen, 0);
 414
 415         desc_avail  = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
 416         desc_offset = dev->vhost_hlen;
 417         desc_chain_head = buf_vec[vec_idx].desc_idx;
 418         desc_chain_len = desc_offset;
 419
 420         mbuf_avail  = rte_pktmbuf_data_len(m);
 421         mbuf_offset = 0;
 422         while (mbuf_avail != 0 || m->next != NULL) {
 423                 /* done with current desc buf, get the next one */
 424                 if (desc_avail == 0) {
 425                         desc_idx = buf_vec[vec_idx].desc_idx;
 426                         vec_idx++;
 427
 428                         if (!(vq->desc[desc_idx].flags & VRING_DESC_F_NEXT)) {
 429                                 /* Update used ring with desc information */
 430                                 used_idx = cur_idx++ & (vq->size - 1);
 431                                 vq->used->ring[used_idx].id = desc_chain_head;
 432                                 vq->used->ring[used_idx].len = desc_chain_len;
 433                                 vhost_log_used_vring(dev, vq,
 434                                         offsetof(struct vring_used,
 435                                                  ring[used_idx]),
 436                                         sizeof(vq->used->ring[used_idx]));
 437                                 desc_chain_head = buf_vec[vec_idx].desc_idx;
 438                                 desc_chain_len = 0;
 439                         }
 440
 441                         desc_addr = gpa_to_vva(dev, buf_vec[vec_idx].buf_addr);
 442                         if (unlikely(!desc_addr))
 443                                 return 0;
 444
 445                         /* Prefetch buffer address. */
 446                         rte_prefetch0((void *)(uintptr_t)desc_addr);
 447                         desc_offset = 0;
 448                         desc_avail  = buf_vec[vec_idx].buf_len;
 449                 }
 450
 451                 /* done with current mbuf, get the next one */
 452                 if (mbuf_avail == 0) {
 453                         m = m->next;
 454
 455                         mbuf_offset = 0;
 456                         mbuf_avail  = rte_pktmbuf_data_len(m);
 457                 }
 458
 459                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 460                 rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)),
 461                         rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 462                         cpy_len);
 463                 vhost_log_write(dev, buf_vec[vec_idx].buf_addr + desc_offset,
 464                         cpy_len);
 465                 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
 466                         cpy_len, 0);
 467
 468                 mbuf_avail  -= cpy_len;
 469                 mbuf_offset += cpy_len;
 470                 desc_avail  -= cpy_len;
 471                 desc_offset += cpy_len;
 472                 desc_chain_len += cpy_len;
 473         }
 474
 475         used_idx = cur_idx & (vq->size - 1);
 476         vq->used->ring[used_idx].id = desc_chain_head;
 477         vq->used->ring[used_idx].len = desc_chain_len;
 478         vhost_log_used_vring(dev, vq,
 479                 offsetof(struct vring_used, ring[used_idx]),
 480                 sizeof(vq->used->ring[used_idx]));
 481
 482         return end_idx - start_idx;
 483 }
 484
 485 static inline uint32_t __attribute__((always_inline))
 486 virtio_dev_merge_rx(struct virtio_net *dev, uint16_t queue_id,
 487         struct rte_mbuf **pkts, uint32_t count)
 488 {
 489         struct vhost_virtqueue *vq;
 490         uint32_t pkt_idx = 0, nr_used = 0;
 491         uint16_t end;
 492         struct buf_vector buf_vec[BUF_VECTOR_MAX];
 493
 494         LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
 495         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->virt_qp_nb))) {
 496                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
 497                         dev->vid, __func__, queue_id);
 498                 return 0;
 499         }
 500
 501         vq = dev->virtqueue[queue_id];
 502         if (unlikely(vq->enabled == 0))
 503                 return 0;
 504
 505         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
 506         if (count == 0)
 507                 return 0;
 508
 509         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
 510                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
 511
 512                 if (unlikely(reserve_avail_buf_mergeable(vq, pkt_len,
 513                                                          &end, buf_vec) < 0)) {
 514                         LOG_DEBUG(VHOST_DATA,
 515                                 "(%d) failed to get enough desc from vring\n",
 516                                 dev->vid);
 517                         break;
 518                 }
 519
 520                 nr_used = copy_mbuf_to_desc_mergeable(dev, vq, end,
 521                                                       pkts[pkt_idx], buf_vec);
 522                 rte_smp_wmb();
 523
 524                 *(volatile uint16_t *)&vq->used->idx += nr_used;
 525                 vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
 526                         sizeof(vq->used->idx));
 527                 vq->last_used_idx += nr_used;
 528         }
 529
 530         if (likely(pkt_idx)) {
 531                 /* flush used->idx update before we read avail->flags. */
 532                 rte_mb();
 533
 534                 /* Kick the guest if necessary. */
 535                 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
 536                                 && (vq->callfd >= 0))
 537                         eventfd_write(vq->callfd, (eventfd_t)1);
 538         }
 539
 540         return pkt_idx;
 541 }
 542
 543 uint16_t
 544 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
 545         struct rte_mbuf **pkts, uint16_t count)
 546 {
 547         struct virtio_net *dev = get_device(vid);
 548
 549         if (!dev)
 550                 return 0;
 551
 552         if (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF))
 553                 return virtio_dev_merge_rx(dev, queue_id, pkts, count);
 554         else
 555                 return virtio_dev_rx(dev, queue_id, pkts, count);
 556 }
 557
 558 static void
 559 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
 560 {
 561         struct ipv4_hdr *ipv4_hdr;
 562         struct ipv6_hdr *ipv6_hdr;
 563         void *l3_hdr = NULL;
 564         struct ether_hdr *eth_hdr;
 565         uint16_t ethertype;
 566
 567         eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
 568
 569         m->l2_len = sizeof(struct ether_hdr);
 570         ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
 571
 572         if (ethertype == ETHER_TYPE_VLAN) {
 573                 struct vlan_hdr *vlan_hdr = (struct vlan_hdr *)(eth_hdr + 1);
 574
 575                 m->l2_len += sizeof(struct vlan_hdr);
 576                 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
 577         }
 578
 579         l3_hdr = (char *)eth_hdr + m->l2_len;
 580
 581         switch (ethertype) {
 582         case ETHER_TYPE_IPv4:
 583                 ipv4_hdr = (struct ipv4_hdr *)l3_hdr;
 584                 *l4_proto = ipv4_hdr->next_proto_id;
 585                 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
 586                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 587                 m->ol_flags |= PKT_TX_IPV4;
 588                 break;
 589         case ETHER_TYPE_IPv6:
 590                 ipv6_hdr = (struct ipv6_hdr *)l3_hdr;
 591                 *l4_proto = ipv6_hdr->proto;
 592                 m->l3_len = sizeof(struct ipv6_hdr);
 593                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 594                 m->ol_flags |= PKT_TX_IPV6;
 595                 break;
 596         default:
 597                 m->l3_len = 0;
 598                 *l4_proto = 0;
 599                 break;
 600         }
 601 }
 602
 603 static inline void __attribute__((always_inline))
 604 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
 605 {
 606         uint16_t l4_proto = 0;
 607         void *l4_hdr = NULL;
 608         struct tcp_hdr *tcp_hdr = NULL;
 609
 610         parse_ethernet(m, &l4_proto, &l4_hdr);
 611         if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 612                 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
 613                         switch (hdr->csum_offset) {
 614                         case (offsetof(struct tcp_hdr, cksum)):
 615                                 if (l4_proto == IPPROTO_TCP)
 616                                         m->ol_flags |= PKT_TX_TCP_CKSUM;
 617                                 break;
 618                         case (offsetof(struct udp_hdr, dgram_cksum)):
 619                                 if (l4_proto == IPPROTO_UDP)
 620                                         m->ol_flags |= PKT_TX_UDP_CKSUM;
 621                                 break;
 622                         case (offsetof(struct sctp_hdr, cksum)):
 623                                 if (l4_proto == IPPROTO_SCTP)
 624                                         m->ol_flags |= PKT_TX_SCTP_CKSUM;
 625                                 break;
 626                         default:
 627                                 break;
 628                         }
 629                 }
 630         }
 631
 632         if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 633                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 634                 case VIRTIO_NET_HDR_GSO_TCPV4:
 635                 case VIRTIO_NET_HDR_GSO_TCPV6:
 636                         tcp_hdr = (struct tcp_hdr *)l4_hdr;
 637                         m->ol_flags |= PKT_TX_TCP_SEG;
 638                         m->tso_segsz = hdr->gso_size;
 639                         m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
 640                         break;
 641                 default:
 642                         RTE_LOG(WARNING, VHOST_DATA,
 643                                 "unsupported gso type %u.\n", hdr->gso_type);
 644                         break;
 645                 }
 646         }
 647 }
 648
 649 #define RARP_PKT_SIZE   64
 650
 651 static int
 652 make_rarp_packet(struct rte_mbuf *rarp_mbuf, const struct ether_addr *mac)
 653 {
 654         struct ether_hdr *eth_hdr;
 655         struct arp_hdr  *rarp;
 656
 657         if (rarp_mbuf->buf_len < 64) {
 658                 RTE_LOG(WARNING, VHOST_DATA,
 659                         "failed to make RARP; mbuf size too small %u (< %d)\n",
 660                         rarp_mbuf->buf_len, RARP_PKT_SIZE);
 661                 return -1;
 662         }
 663
 664         /* Ethernet header. */
 665         eth_hdr = rte_pktmbuf_mtod_offset(rarp_mbuf, struct ether_hdr *, 0);
 666         memset(eth_hdr->d_addr.addr_bytes, 0xff, ETHER_ADDR_LEN);
 667         ether_addr_copy(mac, &eth_hdr->s_addr);
 668         eth_hdr->ether_type = htons(ETHER_TYPE_RARP);
 669
 670         /* RARP header. */
 671         rarp = (struct arp_hdr *)(eth_hdr + 1);
 672         rarp->arp_hrd = htons(ARP_HRD_ETHER);
 673         rarp->arp_pro = htons(ETHER_TYPE_IPv4);
 674         rarp->arp_hln = ETHER_ADDR_LEN;
 675         rarp->arp_pln = 4;
 676         rarp->arp_op  = htons(ARP_OP_REVREQUEST);
 677
 678         ether_addr_copy(mac, &rarp->arp_data.arp_sha);
 679         ether_addr_copy(mac, &rarp->arp_data.arp_tha);
 680         memset(&rarp->arp_data.arp_sip, 0x00, 4);
 681         memset(&rarp->arp_data.arp_tip, 0x00, 4);
 682
 683         rarp_mbuf->pkt_len  = rarp_mbuf->data_len = RARP_PKT_SIZE;
 684
 685         return 0;
 686 }
 687
 688 static inline void __attribute__((always_inline))
 689 put_zmbuf(struct zcopy_mbuf *zmbuf)
 690 {
 691         zmbuf->in_use = 0;
 692 }
 693
 694 static inline int __attribute__((always_inline))
 695 copy_desc_to_mbuf(struct virtio_net *dev, struct vring_desc *descs,
 696                   uint16_t max_desc, struct rte_mbuf *m, uint16_t desc_idx,
 697                   struct rte_mempool *mbuf_pool)
 698 {
 699         struct vring_desc *desc;
 700         uint64_t desc_addr;
 701         uint32_t desc_avail, desc_offset;
 702         uint32_t mbuf_avail, mbuf_offset;
 703         uint32_t cpy_len;
 704         struct rte_mbuf *cur = m, *prev = m;
 705         struct virtio_net_hdr *hdr;
 706         /* A counter to avoid desc dead loop chain */
 707         uint32_t nr_desc = 1;
 708
 709         desc = &descs[desc_idx];
 710         if (unlikely((desc->len < dev->vhost_hlen)) ||
 711                         (desc->flags & VRING_DESC_F_INDIRECT))
 712                 return -1;
 713
 714         desc_addr = gpa_to_vva(dev, desc->addr);
 715         if (unlikely(!desc_addr))
 716                 return -1;
 717
 718         hdr = (struct virtio_net_hdr *)((uintptr_t)desc_addr);
 719         rte_prefetch0(hdr);
 720
 721         /*
 722          * A virtio driver normally uses at least 2 desc buffers
 723          * for Tx: the first for storing the header, and others
 724          * for storing the data.
 725          */
 726         if (likely((desc->len == dev->vhost_hlen) &&
 727                    (desc->flags & VRING_DESC_F_NEXT) != 0)) {
 728                 desc = &descs[desc->next];
 729                 if (unlikely(desc->flags & VRING_DESC_F_INDIRECT))
 730                         return -1;
 731
 732                 desc_addr = gpa_to_vva(dev, desc->addr);
 733                 if (unlikely(!desc_addr))
 734                         return -1;
 735
 736                 rte_prefetch0((void *)(uintptr_t)desc_addr);
 737
 738                 desc_offset = 0;
 739                 desc_avail  = desc->len;
 740                 nr_desc    += 1;
 741
 742                 PRINT_PACKET(dev, (uintptr_t)desc_addr, desc->len, 0);
 743         } else {
 744                 desc_avail  = desc->len - dev->vhost_hlen;
 745                 desc_offset = dev->vhost_hlen;
 746         }
 747
 748         mbuf_offset = 0;
 749         mbuf_avail  = m->buf_len - RTE_PKTMBUF_HEADROOM;
 750         while (1) {
 751                 uint64_t hpa;
 752
 753                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 754
 755                 /*
 756                  * A desc buf might across two host physical pages that are
 757                  * not continuous. In such case (gpa_to_hpa returns 0), data
 758                  * will be copied even though zero copy is enabled.
 759                  */
 760                 if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
 761                                         desc->addr + desc_offset, cpy_len)))) {
 762                         cur->data_len = cpy_len;
 763                         cur->data_off = 0;
 764                         cur->buf_addr = (void *)(uintptr_t)desc_addr;
 765                         cur->buf_physaddr = hpa;
 766
 767                         /*
 768                          * In zero copy mode, one mbuf can only reference data
 769                          * for one or partial of one desc buff.
 770                          */
 771                         mbuf_avail = cpy_len;
 772                 } else {
 773                         rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
 774                                                            mbuf_offset),
 775                                 (void *)((uintptr_t)(desc_addr + desc_offset)),
 776                                 cpy_len);
 777                 }
 778
 779                 mbuf_avail  -= cpy_len;
 780                 mbuf_offset += cpy_len;
 781                 desc_avail  -= cpy_len;
 782                 desc_offset += cpy_len;
 783
 784                 /* This desc reaches to its end, get the next one */
 785                 if (desc_avail == 0) {
 786                         if ((desc->flags & VRING_DESC_F_NEXT) == 0)
 787                                 break;
 788
 789                         if (unlikely(desc->next >= max_desc ||
 790                                      ++nr_desc > max_desc))
 791                                 return -1;
 792                         desc = &descs[desc->next];
 793                         if (unlikely(desc->flags & VRING_DESC_F_INDIRECT))
 794                                 return -1;
 795
 796                         desc_addr = gpa_to_vva(dev, desc->addr);
 797                         if (unlikely(!desc_addr))
 798                                 return -1;
 799
 800                         rte_prefetch0((void *)(uintptr_t)desc_addr);
 801
 802                         desc_offset = 0;
 803                         desc_avail  = desc->len;
 804
 805                         PRINT_PACKET(dev, (uintptr_t)desc_addr, desc->len, 0);
 806                 }
 807
 808                 /*
 809                  * This mbuf reaches to its end, get a new one
 810                  * to hold more data.
 811                  */
 812                 if (mbuf_avail == 0) {
 813                         cur = rte_pktmbuf_alloc(mbuf_pool);
 814                         if (unlikely(cur == NULL)) {
 815                                 RTE_LOG(ERR, VHOST_DATA, "Failed to "
 816                                         "allocate memory for mbuf.\n");
 817                                 return -1;
 818                         }
 819
 820                         prev->next = cur;
 821                         prev->data_len = mbuf_offset;
 822                         m->nb_segs += 1;
 823                         m->pkt_len += mbuf_offset;
 824                         prev = cur;
 825
 826                         mbuf_offset = 0;
 827                         mbuf_avail  = cur->buf_len - RTE_PKTMBUF_HEADROOM;
 828                 }
 829         }
 830
 831         prev->data_len = mbuf_offset;
 832         m->pkt_len    += mbuf_offset;
 833
 834         if (hdr->flags != 0 || hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE)
 835                 vhost_dequeue_offload(hdr, m);
 836
 837         return 0;
 838 }
 839
 840 static inline void __attribute__((always_inline))
 841 update_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq,
 842                  uint32_t used_idx, uint32_t desc_idx)
 843 {
 844         vq->used->ring[used_idx].id  = desc_idx;
 845         vq->used->ring[used_idx].len = 0;
 846         vhost_log_used_vring(dev, vq,
 847                         offsetof(struct vring_used, ring[used_idx]),
 848                         sizeof(vq->used->ring[used_idx]));
 849 }
 850
 851 static inline void __attribute__((always_inline))
 852 update_used_idx(struct virtio_net *dev, struct vhost_virtqueue *vq,
 853                 uint32_t count)
 854 {
 855         if (unlikely(count == 0))
 856                 return;
 857
 858         rte_smp_wmb();
 859         rte_smp_rmb();
 860
 861         vq->used->idx += count;
 862         vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
 863                         sizeof(vq->used->idx));
 864
 865         /* Kick guest if required. */
 866         if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
 867                         && (vq->callfd >= 0))
 868                 eventfd_write(vq->callfd, (eventfd_t)1);
 869 }
 870
 871 static inline struct zcopy_mbuf *__attribute__((always_inline))
 872 get_zmbuf(struct vhost_virtqueue *vq)
 873 {
 874         uint16_t i;
 875         uint16_t last;
 876         int tries = 0;
 877
 878         /* search [last_zmbuf_idx, zmbuf_size) */
 879         i = vq->last_zmbuf_idx;
 880         last = vq->zmbuf_size;
 881
 882 again:
 883         for (; i < last; i++) {
 884                 if (vq->zmbufs[i].in_use == 0) {
 885                         vq->last_zmbuf_idx = i + 1;
 886                         vq->zmbufs[i].in_use = 1;
 887                         return &vq->zmbufs[i];
 888                 }
 889         }
 890
 891         tries++;
 892         if (tries == 1) {
 893                 /* search [0, last_zmbuf_idx) */
 894                 i = 0;
 895                 last = vq->last_zmbuf_idx;
 896                 goto again;
 897         }
 898
 899         return NULL;
 900 }
 901
 902 static inline bool __attribute__((always_inline))
 903 mbuf_is_consumed(struct rte_mbuf *m)
 904 {
 905         while (m) {
 906                 if (rte_mbuf_refcnt_read(m) > 1)
 907                         return false;
 908                 m = m->next;
 909         }
 910
 911         return true;
 912 }
 913
 914 uint16_t
 915 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
 916         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
 917 {
 918         struct virtio_net *dev;
 919         struct rte_mbuf *rarp_mbuf = NULL;
 920         struct vhost_virtqueue *vq;
 921         uint32_t desc_indexes[MAX_PKT_BURST];
 922         uint32_t used_idx;
 923         uint32_t i = 0;
 924         uint16_t free_entries;
 925         uint16_t avail_idx;
 926
 927         dev = get_device(vid);
 928         if (!dev)
 929                 return 0;
 930
 931         if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->virt_qp_nb))) {
 932                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
 933                         dev->vid, __func__, queue_id);
 934                 return 0;
 935         }
 936
 937         vq = dev->virtqueue[queue_id];
 938         if (unlikely(vq->enabled == 0))
 939                 return 0;
 940
 941         if (unlikely(dev->dequeue_zero_copy)) {
 942                 struct zcopy_mbuf *zmbuf, *next;
 943                 int nr_updated = 0;
 944
 945                 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
 946                      zmbuf != NULL; zmbuf = next) {
 947                         next = TAILQ_NEXT(zmbuf, next);
 948
 949                         if (mbuf_is_consumed(zmbuf->mbuf)) {
 950                                 used_idx = vq->last_used_idx++ & (vq->size - 1);
 951                                 update_used_ring(dev, vq, used_idx,
 952                                                  zmbuf->desc_idx);
 953                                 nr_updated += 1;
 954
 955                                 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
 956                                 rte_pktmbuf_free(zmbuf->mbuf);
 957                                 put_zmbuf(zmbuf);
 958                                 vq->nr_zmbuf -= 1;
 959                         }
 960                 }
 961
 962                 update_used_idx(dev, vq, nr_updated);
 963         }
 964
 965         /*
 966          * Construct a RARP broadcast packet, and inject it to the "pkts"
 967          * array, to looks like that guest actually send such packet.
 968          *
 969          * Check user_send_rarp() for more information.
 970          */
 971         if (unlikely(rte_atomic16_cmpset((volatile uint16_t *)
 972                                          &dev->broadcast_rarp.cnt, 1, 0))) {
 973                 rarp_mbuf = rte_pktmbuf_alloc(mbuf_pool);
 974                 if (rarp_mbuf == NULL) {
 975                         RTE_LOG(ERR, VHOST_DATA,
 976                                 "Failed to allocate memory for mbuf.\n");
 977                         return 0;
 978                 }
 979
 980                 if (make_rarp_packet(rarp_mbuf, &dev->mac)) {
 981                         rte_pktmbuf_free(rarp_mbuf);
 982                         rarp_mbuf = NULL;
 983                 } else {
 984                         count -= 1;
 985                 }
 986         }
 987
 988         free_entries = *((volatile uint16_t *)&vq->avail->idx) -
 989                         vq->last_avail_idx;
 990         if (free_entries == 0)
 991                 goto out;
 992
 993         LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
 994
 995         /* Prefetch available and used ring */
 996         avail_idx = vq->last_avail_idx & (vq->size - 1);
 997         used_idx  = vq->last_used_idx  & (vq->size - 1);
 998         rte_prefetch0(&vq->avail->ring[avail_idx]);
 999         rte_prefetch0(&vq->used->ring[used_idx]);
1000
1001         count = RTE_MIN(count, MAX_PKT_BURST);
1002         count = RTE_MIN(count, free_entries);
1003         LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1004                         dev->vid, count);
1005
1006         /* Retrieve all of the head indexes first to avoid caching issues. */
1007         for (i = 0; i < count; i++) {
1008                 avail_idx = (vq->last_avail_idx + i) & (vq->size - 1);
1009                 used_idx  = (vq->last_used_idx  + i) & (vq->size - 1);
1010                 desc_indexes[i] = vq->avail->ring[avail_idx];
1011
1012                 if (likely(dev->dequeue_zero_copy == 0))
1013                         update_used_ring(dev, vq, used_idx, desc_indexes[i]);
1014         }
1015
1016         /* Prefetch descriptor index. */
1017         rte_prefetch0(&vq->desc[desc_indexes[0]]);
1018         for (i = 0; i < count; i++) {
1019                 struct vring_desc *desc;
1020                 uint16_t sz, idx;
1021                 int err;
1022
1023                 if (likely(i + 1 < count))
1024                         rte_prefetch0(&vq->desc[desc_indexes[i + 1]]);
1025
1026                 if (vq->desc[desc_indexes[i]].flags & VRING_DESC_F_INDIRECT) {
1027                         desc = (struct vring_desc *)(uintptr_t)gpa_to_vva(dev,
1028                                         vq->desc[desc_indexes[i]].addr);
1029                         if (unlikely(!desc))
1030                                 break;
1031
1032                         rte_prefetch0(desc);
1033                         sz = vq->desc[desc_indexes[i]].len / sizeof(*desc);
1034                         idx = 0;
1035                 } else {
1036                         desc = vq->desc;
1037                         sz = vq->size;
1038                         idx = desc_indexes[i];
1039                 }
1040
1041                 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
1042                 if (unlikely(pkts[i] == NULL)) {
1043                         RTE_LOG(ERR, VHOST_DATA,
1044                                 "Failed to allocate memory for mbuf.\n");
1045                         break;
1046                 }
1047
1048                 err = copy_desc_to_mbuf(dev, desc, sz, pkts[i], idx, mbuf_pool);
1049                 if (unlikely(err)) {
1050                         rte_pktmbuf_free(pkts[i]);
1051                         break;
1052                 }
1053
1054                 if (unlikely(dev->dequeue_zero_copy)) {
1055                         struct zcopy_mbuf *zmbuf;
1056
1057                         zmbuf = get_zmbuf(vq);
1058                         if (!zmbuf) {
1059                                 rte_pktmbuf_free(pkts[i]);
1060                                 break;
1061                         }
1062                         zmbuf->mbuf = pkts[i];
1063                         zmbuf->desc_idx = desc_indexes[i];
1064
1065                         /*
1066                          * Pin lock the mbuf; we will check later to see
1067                          * whether the mbuf is freed (when we are the last
1068                          * user) or not. If that's the case, we then could
1069                          * update the used ring safely.
1070                          */
1071                         rte_mbuf_refcnt_update(pkts[i], 1);
1072
1073                         vq->nr_zmbuf += 1;
1074                         TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1075                 }
1076         }
1077         vq->last_avail_idx += i;
1078
1079         if (likely(dev->dequeue_zero_copy == 0)) {
1080                 vq->last_used_idx += i;
1081                 update_used_idx(dev, vq, i);
1082         }
1083
1084 out:
1085         if (unlikely(rarp_mbuf != NULL)) {
1086                 /*
1087                  * Inject it to the head of "pkts" array, so that switch's mac
1088                  * learning table will get updated first.
1089                  */
1090                 memmove(&pkts[1], pkts, i * sizeof(struct rte_mbuf *));
1091                 pkts[0] = rarp_mbuf;
1092                 i += 1;
1093         }
1094
1095         return i;
1096 }