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                             struct buf_vector *buf_vec, uint16_t *num_buffers)
 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         *num_buffers = cur_idx - vq->last_used_idx;
 374         return 0;
 375 }
 376
 377 static inline int __attribute__((always_inline))
 378 copy_mbuf_to_desc_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
 379                             struct rte_mbuf *m, struct buf_vector *buf_vec,
 380                             uint16_t num_buffers)
 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 cur_idx = vq->last_used_idx;
 385         uint64_t desc_addr;
 386         uint32_t desc_chain_head;
 387         uint32_t desc_chain_len;
 388         uint32_t mbuf_offset, mbuf_avail;
 389         uint32_t desc_offset, desc_avail;
 390         uint32_t cpy_len;
 391         uint16_t desc_idx, used_idx;
 392         uint64_t hdr_addr, hdr_phys_addr;
 393         struct rte_mbuf *hdr_mbuf;
 394
 395         if (unlikely(m == NULL))
 396                 return -1;
 397
 398         LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
 399                 dev->vid, cur_idx, cur_idx + num_buffers);
 400
 401         desc_addr = gpa_to_vva(dev, buf_vec[vec_idx].buf_addr);
 402         if (buf_vec[vec_idx].buf_len < dev->vhost_hlen || !desc_addr)
 403                 return -1;
 404
 405         hdr_mbuf = m;
 406         hdr_addr = desc_addr;
 407         hdr_phys_addr = buf_vec[vec_idx].buf_addr;
 408         rte_prefetch0((void *)(uintptr_t)hdr_addr);
 409
 410         virtio_hdr.num_buffers = num_buffers;
 411         LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
 412                 dev->vid, virtio_hdr.num_buffers);
 413
 414         desc_avail  = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
 415         desc_offset = dev->vhost_hlen;
 416         desc_chain_head = buf_vec[vec_idx].desc_idx;
 417         desc_chain_len = desc_offset;
 418
 419         mbuf_avail  = rte_pktmbuf_data_len(m);
 420         mbuf_offset = 0;
 421         while (mbuf_avail != 0 || m->next != NULL) {
 422                 /* done with current desc buf, get the next one */
 423                 if (desc_avail == 0) {
 424                         desc_idx = buf_vec[vec_idx].desc_idx;
 425                         vec_idx++;
 426
 427                         if (!(vq->desc[desc_idx].flags & VRING_DESC_F_NEXT)) {
 428                                 /* Update used ring with desc information */
 429                                 used_idx = cur_idx++ & (vq->size - 1);
 430                                 vq->used->ring[used_idx].id = desc_chain_head;
 431                                 vq->used->ring[used_idx].len = desc_chain_len;
 432                                 vhost_log_used_vring(dev, vq,
 433                                         offsetof(struct vring_used,
 434                                                  ring[used_idx]),
 435                                         sizeof(vq->used->ring[used_idx]));
 436                                 desc_chain_head = buf_vec[vec_idx].desc_idx;
 437                                 desc_chain_len = 0;
 438                         }
 439
 440                         desc_addr = gpa_to_vva(dev, buf_vec[vec_idx].buf_addr);
 441                         if (unlikely(!desc_addr))
 442                                 return -1;
 443
 444                         /* Prefetch buffer address. */
 445                         rte_prefetch0((void *)(uintptr_t)desc_addr);
 446                         desc_offset = 0;
 447                         desc_avail  = buf_vec[vec_idx].buf_len;
 448                 }
 449
 450                 /* done with current mbuf, get the next one */
 451                 if (mbuf_avail == 0) {
 452                         m = m->next;
 453
 454                         mbuf_offset = 0;
 455                         mbuf_avail  = rte_pktmbuf_data_len(m);
 456                 }
 457
 458                 if (hdr_addr) {
 459                         virtio_enqueue_offload(hdr_mbuf, &virtio_hdr.hdr);
 460                         copy_virtio_net_hdr(dev, hdr_addr, virtio_hdr);
 461                         vhost_log_write(dev, hdr_phys_addr, dev->vhost_hlen);
 462                         PRINT_PACKET(dev, (uintptr_t)hdr_addr,
 463                                      dev->vhost_hlen, 0);
 464
 465                         hdr_addr = 0;
 466                 }
 467
 468                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 469                 rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)),
 470                         rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 471                         cpy_len);
 472                 vhost_log_write(dev, buf_vec[vec_idx].buf_addr + desc_offset,
 473                         cpy_len);
 474                 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
 475                         cpy_len, 0);
 476
 477                 mbuf_avail  -= cpy_len;
 478                 mbuf_offset += cpy_len;
 479                 desc_avail  -= cpy_len;
 480                 desc_offset += cpy_len;
 481                 desc_chain_len += cpy_len;
 482         }
 483
 484         used_idx = cur_idx & (vq->size - 1);
 485         vq->used->ring[used_idx].id = desc_chain_head;
 486         vq->used->ring[used_idx].len = desc_chain_len;
 487         vhost_log_used_vring(dev, vq,
 488                 offsetof(struct vring_used, ring[used_idx]),
 489                 sizeof(vq->used->ring[used_idx]));
 490
 491         return 0;
 492 }
 493
 494 static inline uint32_t __attribute__((always_inline))
 495 virtio_dev_merge_rx(struct virtio_net *dev, uint16_t queue_id,
 496         struct rte_mbuf **pkts, uint32_t count)
 497 {
 498         struct vhost_virtqueue *vq;
 499         uint32_t pkt_idx = 0;
 500         uint16_t num_buffers;
 501         struct buf_vector buf_vec[BUF_VECTOR_MAX];
 502
 503         LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
 504         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->virt_qp_nb))) {
 505                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
 506                         dev->vid, __func__, queue_id);
 507                 return 0;
 508         }
 509
 510         vq = dev->virtqueue[queue_id];
 511         if (unlikely(vq->enabled == 0))
 512                 return 0;
 513
 514         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
 515         if (count == 0)
 516                 return 0;
 517
 518         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
 519                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
 520
 521                 if (unlikely(reserve_avail_buf_mergeable(vq, pkt_len, buf_vec,
 522                                                          &num_buffers) < 0)) {
 523                         LOG_DEBUG(VHOST_DATA,
 524                                 "(%d) failed to get enough desc from vring\n",
 525                                 dev->vid);
 526                         break;
 527                 }
 528
 529                 if (copy_mbuf_to_desc_mergeable(dev, vq, pkts[pkt_idx],
 530                                                 buf_vec, num_buffers) < 0)
 531                         break;
 532
 533                 rte_smp_wmb();
 534
 535                 *(volatile uint16_t *)&vq->used->idx += num_buffers;
 536                 vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
 537                         sizeof(vq->used->idx));
 538                 vq->last_used_idx += num_buffers;
 539         }
 540
 541         if (likely(pkt_idx)) {
 542                 /* flush used->idx update before we read avail->flags. */
 543                 rte_mb();
 544
 545                 /* Kick the guest if necessary. */
 546                 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
 547                                 && (vq->callfd >= 0))
 548                         eventfd_write(vq->callfd, (eventfd_t)1);
 549         }
 550
 551         return pkt_idx;
 552 }
 553
 554 uint16_t
 555 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
 556         struct rte_mbuf **pkts, uint16_t count)
 557 {
 558         struct virtio_net *dev = get_device(vid);
 559
 560         if (!dev)
 561                 return 0;
 562
 563         if (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF))
 564                 return virtio_dev_merge_rx(dev, queue_id, pkts, count);
 565         else
 566                 return virtio_dev_rx(dev, queue_id, pkts, count);
 567 }
 568
 569 static inline bool
 570 virtio_net_with_host_offload(struct virtio_net *dev)
 571 {
 572         if (dev->features &
 573                         (VIRTIO_NET_F_CSUM | VIRTIO_NET_F_HOST_ECN |
 574                          VIRTIO_NET_F_HOST_TSO4 | VIRTIO_NET_F_HOST_TSO6 |
 575                          VIRTIO_NET_F_HOST_UFO))
 576                 return true;
 577
 578         return false;
 579 }
 580
 581 static void
 582 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
 583 {
 584         struct ipv4_hdr *ipv4_hdr;
 585         struct ipv6_hdr *ipv6_hdr;
 586         void *l3_hdr = NULL;
 587         struct ether_hdr *eth_hdr;
 588         uint16_t ethertype;
 589
 590         eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
 591
 592         m->l2_len = sizeof(struct ether_hdr);
 593         ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
 594
 595         if (ethertype == ETHER_TYPE_VLAN) {
 596                 struct vlan_hdr *vlan_hdr = (struct vlan_hdr *)(eth_hdr + 1);
 597
 598                 m->l2_len += sizeof(struct vlan_hdr);
 599                 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
 600         }
 601
 602         l3_hdr = (char *)eth_hdr + m->l2_len;
 603
 604         switch (ethertype) {
 605         case ETHER_TYPE_IPv4:
 606                 ipv4_hdr = (struct ipv4_hdr *)l3_hdr;
 607                 *l4_proto = ipv4_hdr->next_proto_id;
 608                 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
 609                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 610                 m->ol_flags |= PKT_TX_IPV4;
 611                 break;
 612         case ETHER_TYPE_IPv6:
 613                 ipv6_hdr = (struct ipv6_hdr *)l3_hdr;
 614                 *l4_proto = ipv6_hdr->proto;
 615                 m->l3_len = sizeof(struct ipv6_hdr);
 616                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 617                 m->ol_flags |= PKT_TX_IPV6;
 618                 break;
 619         default:
 620                 m->l3_len = 0;
 621                 *l4_proto = 0;
 622                 break;
 623         }
 624 }
 625
 626 static inline void __attribute__((always_inline))
 627 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
 628 {
 629         uint16_t l4_proto = 0;
 630         void *l4_hdr = NULL;
 631         struct tcp_hdr *tcp_hdr = NULL;
 632
 633         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
 634                 return;
 635
 636         parse_ethernet(m, &l4_proto, &l4_hdr);
 637         if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 638                 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
 639                         switch (hdr->csum_offset) {
 640                         case (offsetof(struct tcp_hdr, cksum)):
 641                                 if (l4_proto == IPPROTO_TCP)
 642                                         m->ol_flags |= PKT_TX_TCP_CKSUM;
 643                                 break;
 644                         case (offsetof(struct udp_hdr, dgram_cksum)):
 645                                 if (l4_proto == IPPROTO_UDP)
 646                                         m->ol_flags |= PKT_TX_UDP_CKSUM;
 647                                 break;
 648                         case (offsetof(struct sctp_hdr, cksum)):
 649                                 if (l4_proto == IPPROTO_SCTP)
 650                                         m->ol_flags |= PKT_TX_SCTP_CKSUM;
 651                                 break;
 652                         default:
 653                                 break;
 654                         }
 655                 }
 656         }
 657
 658         if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 659                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 660                 case VIRTIO_NET_HDR_GSO_TCPV4:
 661                 case VIRTIO_NET_HDR_GSO_TCPV6:
 662                         tcp_hdr = (struct tcp_hdr *)l4_hdr;
 663                         m->ol_flags |= PKT_TX_TCP_SEG;
 664                         m->tso_segsz = hdr->gso_size;
 665                         m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
 666                         break;
 667                 default:
 668                         RTE_LOG(WARNING, VHOST_DATA,
 669                                 "unsupported gso type %u.\n", hdr->gso_type);
 670                         break;
 671                 }
 672         }
 673 }
 674
 675 #define RARP_PKT_SIZE   64
 676
 677 static int
 678 make_rarp_packet(struct rte_mbuf *rarp_mbuf, const struct ether_addr *mac)
 679 {
 680         struct ether_hdr *eth_hdr;
 681         struct arp_hdr  *rarp;
 682
 683         if (rarp_mbuf->buf_len < 64) {
 684                 RTE_LOG(WARNING, VHOST_DATA,
 685                         "failed to make RARP; mbuf size too small %u (< %d)\n",
 686                         rarp_mbuf->buf_len, RARP_PKT_SIZE);
 687                 return -1;
 688         }
 689
 690         /* Ethernet header. */
 691         eth_hdr = rte_pktmbuf_mtod_offset(rarp_mbuf, struct ether_hdr *, 0);
 692         memset(eth_hdr->d_addr.addr_bytes, 0xff, ETHER_ADDR_LEN);
 693         ether_addr_copy(mac, &eth_hdr->s_addr);
 694         eth_hdr->ether_type = htons(ETHER_TYPE_RARP);
 695
 696         /* RARP header. */
 697         rarp = (struct arp_hdr *)(eth_hdr + 1);
 698         rarp->arp_hrd = htons(ARP_HRD_ETHER);
 699         rarp->arp_pro = htons(ETHER_TYPE_IPv4);
 700         rarp->arp_hln = ETHER_ADDR_LEN;
 701         rarp->arp_pln = 4;
 702         rarp->arp_op  = htons(ARP_OP_REVREQUEST);
 703
 704         ether_addr_copy(mac, &rarp->arp_data.arp_sha);
 705         ether_addr_copy(mac, &rarp->arp_data.arp_tha);
 706         memset(&rarp->arp_data.arp_sip, 0x00, 4);
 707         memset(&rarp->arp_data.arp_tip, 0x00, 4);
 708
 709         rarp_mbuf->pkt_len  = rarp_mbuf->data_len = RARP_PKT_SIZE;
 710
 711         return 0;
 712 }
 713
 714 static inline void __attribute__((always_inline))
 715 put_zmbuf(struct zcopy_mbuf *zmbuf)
 716 {
 717         zmbuf->in_use = 0;
 718 }
 719
 720 static inline int __attribute__((always_inline))
 721 copy_desc_to_mbuf(struct virtio_net *dev, struct vring_desc *descs,
 722                   uint16_t max_desc, struct rte_mbuf *m, uint16_t desc_idx,
 723                   struct rte_mempool *mbuf_pool)
 724 {
 725         struct vring_desc *desc;
 726         uint64_t desc_addr;
 727         uint32_t desc_avail, desc_offset;
 728         uint32_t mbuf_avail, mbuf_offset;
 729         uint32_t cpy_len;
 730         struct rte_mbuf *cur = m, *prev = m;
 731         struct virtio_net_hdr *hdr = NULL;
 732         /* A counter to avoid desc dead loop chain */
 733         uint32_t nr_desc = 1;
 734
 735         desc = &descs[desc_idx];
 736         if (unlikely((desc->len < dev->vhost_hlen)) ||
 737                         (desc->flags & VRING_DESC_F_INDIRECT))
 738                 return -1;
 739
 740         desc_addr = gpa_to_vva(dev, desc->addr);
 741         if (unlikely(!desc_addr))
 742                 return -1;
 743
 744         if (virtio_net_with_host_offload(dev)) {
 745                 hdr = (struct virtio_net_hdr *)((uintptr_t)desc_addr);
 746                 rte_prefetch0(hdr);
 747         }
 748
 749         /*
 750          * A virtio driver normally uses at least 2 desc buffers
 751          * for Tx: the first for storing the header, and others
 752          * for storing the data.
 753          */
 754         if (likely((desc->len == dev->vhost_hlen) &&
 755                    (desc->flags & VRING_DESC_F_NEXT) != 0)) {
 756                 desc = &descs[desc->next];
 757                 if (unlikely(desc->flags & VRING_DESC_F_INDIRECT))
 758                         return -1;
 759
 760                 desc_addr = gpa_to_vva(dev, desc->addr);
 761                 if (unlikely(!desc_addr))
 762                         return -1;
 763
 764                 desc_offset = 0;
 765                 desc_avail  = desc->len;
 766                 nr_desc    += 1;
 767         } else {
 768                 desc_avail  = desc->len - dev->vhost_hlen;
 769                 desc_offset = dev->vhost_hlen;
 770         }
 771
 772         rte_prefetch0((void *)(uintptr_t)(desc_addr + desc_offset));
 773
 774         PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset), desc_avail, 0);
 775
 776         mbuf_offset = 0;
 777         mbuf_avail  = m->buf_len - RTE_PKTMBUF_HEADROOM;
 778         while (1) {
 779                 uint64_t hpa;
 780
 781                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 782
 783                 /*
 784                  * A desc buf might across two host physical pages that are
 785                  * not continuous. In such case (gpa_to_hpa returns 0), data
 786                  * will be copied even though zero copy is enabled.
 787                  */
 788                 if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
 789                                         desc->addr + desc_offset, cpy_len)))) {
 790                         cur->data_len = cpy_len;
 791                         cur->data_off = 0;
 792                         cur->buf_addr = (void *)(uintptr_t)desc_addr;
 793                         cur->buf_physaddr = hpa;
 794
 795                         /*
 796                          * In zero copy mode, one mbuf can only reference data
 797                          * for one or partial of one desc buff.
 798                          */
 799                         mbuf_avail = cpy_len;
 800                 } else {
 801                         rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
 802                                                            mbuf_offset),
 803                                 (void *)((uintptr_t)(desc_addr + desc_offset)),
 804                                 cpy_len);
 805                 }
 806
 807                 mbuf_avail  -= cpy_len;
 808                 mbuf_offset += cpy_len;
 809                 desc_avail  -= cpy_len;
 810                 desc_offset += cpy_len;
 811
 812                 /* This desc reaches to its end, get the next one */
 813                 if (desc_avail == 0) {
 814                         if ((desc->flags & VRING_DESC_F_NEXT) == 0)
 815                                 break;
 816
 817                         if (unlikely(desc->next >= max_desc ||
 818                                      ++nr_desc > max_desc))
 819                                 return -1;
 820                         desc = &descs[desc->next];
 821                         if (unlikely(desc->flags & VRING_DESC_F_INDIRECT))
 822                                 return -1;
 823
 824                         desc_addr = gpa_to_vva(dev, desc->addr);
 825                         if (unlikely(!desc_addr))
 826                                 return -1;
 827
 828                         rte_prefetch0((void *)(uintptr_t)desc_addr);
 829
 830                         desc_offset = 0;
 831                         desc_avail  = desc->len;
 832
 833                         PRINT_PACKET(dev, (uintptr_t)desc_addr, desc->len, 0);
 834                 }
 835
 836                 /*
 837                  * This mbuf reaches to its end, get a new one
 838                  * to hold more data.
 839                  */
 840                 if (mbuf_avail == 0) {
 841                         cur = rte_pktmbuf_alloc(mbuf_pool);
 842                         if (unlikely(cur == NULL)) {
 843                                 RTE_LOG(ERR, VHOST_DATA, "Failed to "
 844                                         "allocate memory for mbuf.\n");
 845                                 return -1;
 846                         }
 847
 848                         prev->next = cur;
 849                         prev->data_len = mbuf_offset;
 850                         m->nb_segs += 1;
 851                         m->pkt_len += mbuf_offset;
 852                         prev = cur;
 853
 854                         mbuf_offset = 0;
 855                         mbuf_avail  = cur->buf_len - RTE_PKTMBUF_HEADROOM;
 856                 }
 857         }
 858
 859         prev->data_len = mbuf_offset;
 860         m->pkt_len    += mbuf_offset;
 861
 862         if (hdr)
 863                 vhost_dequeue_offload(hdr, m);
 864
 865         return 0;
 866 }
 867
 868 static inline void __attribute__((always_inline))
 869 update_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq,
 870                  uint32_t used_idx, uint32_t desc_idx)
 871 {
 872         vq->used->ring[used_idx].id  = desc_idx;
 873         vq->used->ring[used_idx].len = 0;
 874         vhost_log_used_vring(dev, vq,
 875                         offsetof(struct vring_used, ring[used_idx]),
 876                         sizeof(vq->used->ring[used_idx]));
 877 }
 878
 879 static inline void __attribute__((always_inline))
 880 update_used_idx(struct virtio_net *dev, struct vhost_virtqueue *vq,
 881                 uint32_t count)
 882 {
 883         if (unlikely(count == 0))
 884                 return;
 885
 886         rte_smp_wmb();
 887         rte_smp_rmb();
 888
 889         vq->used->idx += count;
 890         vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
 891                         sizeof(vq->used->idx));
 892
 893         /* Kick guest if required. */
 894         if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
 895                         && (vq->callfd >= 0))
 896                 eventfd_write(vq->callfd, (eventfd_t)1);
 897 }
 898
 899 static inline struct zcopy_mbuf *__attribute__((always_inline))
 900 get_zmbuf(struct vhost_virtqueue *vq)
 901 {
 902         uint16_t i;
 903         uint16_t last;
 904         int tries = 0;
 905
 906         /* search [last_zmbuf_idx, zmbuf_size) */
 907         i = vq->last_zmbuf_idx;
 908         last = vq->zmbuf_size;
 909
 910 again:
 911         for (; i < last; i++) {
 912                 if (vq->zmbufs[i].in_use == 0) {
 913                         vq->last_zmbuf_idx = i + 1;
 914                         vq->zmbufs[i].in_use = 1;
 915                         return &vq->zmbufs[i];
 916                 }
 917         }
 918
 919         tries++;
 920         if (tries == 1) {
 921                 /* search [0, last_zmbuf_idx) */
 922                 i = 0;
 923                 last = vq->last_zmbuf_idx;
 924                 goto again;
 925         }
 926
 927         return NULL;
 928 }
 929
 930 static inline bool __attribute__((always_inline))
 931 mbuf_is_consumed(struct rte_mbuf *m)
 932 {
 933         while (m) {
 934                 if (rte_mbuf_refcnt_read(m) > 1)
 935                         return false;
 936                 m = m->next;
 937         }
 938
 939         return true;
 940 }
 941
 942 uint16_t
 943 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
 944         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
 945 {
 946         struct virtio_net *dev;
 947         struct rte_mbuf *rarp_mbuf = NULL;
 948         struct vhost_virtqueue *vq;
 949         uint32_t desc_indexes[MAX_PKT_BURST];
 950         uint32_t used_idx;
 951         uint32_t i = 0;
 952         uint16_t free_entries;
 953         uint16_t avail_idx;
 954
 955         dev = get_device(vid);
 956         if (!dev)
 957                 return 0;
 958
 959         if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->virt_qp_nb))) {
 960                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
 961                         dev->vid, __func__, queue_id);
 962                 return 0;
 963         }
 964
 965         vq = dev->virtqueue[queue_id];
 966         if (unlikely(vq->enabled == 0))
 967                 return 0;
 968
 969         if (unlikely(dev->dequeue_zero_copy)) {
 970                 struct zcopy_mbuf *zmbuf, *next;
 971                 int nr_updated = 0;
 972
 973                 for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
 974                      zmbuf != NULL; zmbuf = next) {
 975                         next = TAILQ_NEXT(zmbuf, next);
 976
 977                         if (mbuf_is_consumed(zmbuf->mbuf)) {
 978                                 used_idx = vq->last_used_idx++ & (vq->size - 1);
 979                                 update_used_ring(dev, vq, used_idx,
 980                                                  zmbuf->desc_idx);
 981                                 nr_updated += 1;
 982
 983                                 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
 984                                 rte_pktmbuf_free(zmbuf->mbuf);
 985                                 put_zmbuf(zmbuf);
 986                                 vq->nr_zmbuf -= 1;
 987                         }
 988                 }
 989
 990                 update_used_idx(dev, vq, nr_updated);
 991         }
 992
 993         /*
 994          * Construct a RARP broadcast packet, and inject it to the "pkts"
 995          * array, to looks like that guest actually send such packet.
 996          *
 997          * Check user_send_rarp() for more information.
 998          */
 999         if (unlikely(rte_atomic16_cmpset((volatile uint16_t *)
1000                                          &dev->broadcast_rarp.cnt, 1, 0))) {
1001                 rarp_mbuf = rte_pktmbuf_alloc(mbuf_pool);
1002                 if (rarp_mbuf == NULL) {
1003                         RTE_LOG(ERR, VHOST_DATA,
1004                                 "Failed to allocate memory for mbuf.\n");
1005                         return 0;
1006                 }
1007
1008                 if (make_rarp_packet(rarp_mbuf, &dev->mac)) {
1009                         rte_pktmbuf_free(rarp_mbuf);
1010                         rarp_mbuf = NULL;
1011                 } else {
1012                         count -= 1;
1013                 }
1014         }
1015
1016         free_entries = *((volatile uint16_t *)&vq->avail->idx) -
1017                         vq->last_avail_idx;
1018         if (free_entries == 0)
1019                 goto out;
1020
1021         LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
1022
1023         /* Prefetch available and used ring */
1024         avail_idx = vq->last_avail_idx & (vq->size - 1);
1025         used_idx  = vq->last_used_idx  & (vq->size - 1);
1026         rte_prefetch0(&vq->avail->ring[avail_idx]);
1027         rte_prefetch0(&vq->used->ring[used_idx]);
1028
1029         count = RTE_MIN(count, MAX_PKT_BURST);
1030         count = RTE_MIN(count, free_entries);
1031         LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
1032                         dev->vid, count);
1033
1034         /* Retrieve all of the head indexes first to avoid caching issues. */
1035         for (i = 0; i < count; i++) {
1036                 avail_idx = (vq->last_avail_idx + i) & (vq->size - 1);
1037                 used_idx  = (vq->last_used_idx  + i) & (vq->size - 1);
1038                 desc_indexes[i] = vq->avail->ring[avail_idx];
1039
1040                 if (likely(dev->dequeue_zero_copy == 0))
1041                         update_used_ring(dev, vq, used_idx, desc_indexes[i]);
1042         }
1043
1044         /* Prefetch descriptor index. */
1045         rte_prefetch0(&vq->desc[desc_indexes[0]]);
1046         for (i = 0; i < count; i++) {
1047                 struct vring_desc *desc;
1048                 uint16_t sz, idx;
1049                 int err;
1050
1051                 if (likely(i + 1 < count))
1052                         rte_prefetch0(&vq->desc[desc_indexes[i + 1]]);
1053
1054                 if (vq->desc[desc_indexes[i]].flags & VRING_DESC_F_INDIRECT) {
1055                         desc = (struct vring_desc *)(uintptr_t)gpa_to_vva(dev,
1056                                         vq->desc[desc_indexes[i]].addr);
1057                         if (unlikely(!desc))
1058                                 break;
1059
1060                         rte_prefetch0(desc);
1061                         sz = vq->desc[desc_indexes[i]].len / sizeof(*desc);
1062                         idx = 0;
1063                 } else {
1064                         desc = vq->desc;
1065                         sz = vq->size;
1066                         idx = desc_indexes[i];
1067                 }
1068
1069                 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
1070                 if (unlikely(pkts[i] == NULL)) {
1071                         RTE_LOG(ERR, VHOST_DATA,
1072                                 "Failed to allocate memory for mbuf.\n");
1073                         break;
1074                 }
1075
1076                 err = copy_desc_to_mbuf(dev, desc, sz, pkts[i], idx, mbuf_pool);
1077                 if (unlikely(err)) {
1078                         rte_pktmbuf_free(pkts[i]);
1079                         break;
1080                 }
1081
1082                 if (unlikely(dev->dequeue_zero_copy)) {
1083                         struct zcopy_mbuf *zmbuf;
1084
1085                         zmbuf = get_zmbuf(vq);
1086                         if (!zmbuf) {
1087                                 rte_pktmbuf_free(pkts[i]);
1088                                 break;
1089                         }
1090                         zmbuf->mbuf = pkts[i];
1091                         zmbuf->desc_idx = desc_indexes[i];
1092
1093                         /*
1094                          * Pin lock the mbuf; we will check later to see
1095                          * whether the mbuf is freed (when we are the last
1096                          * user) or not. If that's the case, we then could
1097                          * update the used ring safely.
1098                          */
1099                         rte_mbuf_refcnt_update(pkts[i], 1);
1100
1101                         vq->nr_zmbuf += 1;
1102                         TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
1103                 }
1104         }
1105         vq->last_avail_idx += i;
1106
1107         if (likely(dev->dequeue_zero_copy == 0)) {
1108                 vq->last_used_idx += i;
1109                 update_used_idx(dev, vq, i);
1110         }
1111
1112 out:
1113         if (unlikely(rarp_mbuf != NULL)) {
1114                 /*
1115                  * Inject it to the head of "pkts" array, so that switch's mac
1116                  * learning table will get updated first.
1117                  */
1118                 memmove(&pkts[1], pkts, i * sizeof(struct rte_mbuf *));
1119                 pkts[0] = rarp_mbuf;
1120                 i += 1;
1121         }
1122
1123         return i;
1124 }