lib/librte_vhost/vhost_rxtx.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2014 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-net.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         if (unlikely(desc->len < dev->vhost_hlen))
 151                 return -1;
 152
 153         desc_addr = gpa_to_vva(dev, desc->addr);
 154         rte_prefetch0((void *)(uintptr_t)desc_addr);
 155
 156         virtio_enqueue_offload(m, &virtio_hdr.hdr);
 157         copy_virtio_net_hdr(dev, desc_addr, virtio_hdr);
 158         vhost_log_write(dev, desc->addr, dev->vhost_hlen);
 159         PRINT_PACKET(dev, (uintptr_t)desc_addr, dev->vhost_hlen, 0);
 160
 161         desc_offset = dev->vhost_hlen;
 162         desc_avail  = desc->len - dev->vhost_hlen;
 163
 164         mbuf_avail  = rte_pktmbuf_data_len(m);
 165         mbuf_offset = 0;
 166         while (mbuf_avail != 0 || m->next != NULL) {
 167                 /* done with current mbuf, fetch next */
 168                 if (mbuf_avail == 0) {
 169                         m = m->next;
 170
 171                         mbuf_offset = 0;
 172                         mbuf_avail  = rte_pktmbuf_data_len(m);
 173                 }
 174
 175                 /* done with current desc buf, fetch next */
 176                 if (desc_avail == 0) {
 177                         if ((desc->flags & VRING_DESC_F_NEXT) == 0) {
 178                                 /* Room in vring buffer is not enough */
 179                                 return -1;
 180                         }
 181                         if (unlikely(desc->next >= vq->size))
 182                                 return -1;
 183
 184                         desc = &vq->desc[desc->next];
 185                         desc_addr   = gpa_to_vva(dev, desc->addr);
 186                         desc_offset = 0;
 187                         desc_avail  = desc->len;
 188                 }
 189
 190                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 191                 rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)),
 192                         rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 193                         cpy_len);
 194                 vhost_log_write(dev, desc->addr + desc_offset, cpy_len);
 195                 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
 196                              cpy_len, 0);
 197
 198                 mbuf_avail  -= cpy_len;
 199                 mbuf_offset += cpy_len;
 200                 desc_avail  -= cpy_len;
 201                 desc_offset += cpy_len;
 202         }
 203
 204         return 0;
 205 }
 206
 207 /*
 208  * As many data cores may want to access available buffers
 209  * they need to be reserved.
 210  */
 211 static inline uint32_t
 212 reserve_avail_buf(struct vhost_virtqueue *vq, uint32_t count,
 213                   uint16_t *start, uint16_t *end)
 214 {
 215         uint16_t res_start_idx;
 216         uint16_t res_end_idx;
 217         uint16_t avail_idx;
 218         uint16_t free_entries;
 219         int success;
 220
 221         count = RTE_MIN(count, (uint32_t)MAX_PKT_BURST);
 222
 223 again:
 224         res_start_idx = vq->last_used_idx_res;
 225         avail_idx = *((volatile uint16_t *)&vq->avail->idx);
 226
 227         free_entries = avail_idx - res_start_idx;
 228         count = RTE_MIN(count, free_entries);
 229         if (count == 0)
 230                 return 0;
 231
 232         res_end_idx = res_start_idx + count;
 233
 234         /*
 235          * update vq->last_used_idx_res atomically; try again if failed.
 236          *
 237          * TODO: Allow to disable cmpset if no concurrency in application.
 238          */
 239         success = rte_atomic16_cmpset(&vq->last_used_idx_res,
 240                                       res_start_idx, res_end_idx);
 241         if (unlikely(!success))
 242                 goto again;
 243
 244         *start = res_start_idx;
 245         *end   = res_end_idx;
 246
 247         return count;
 248 }
 249
 250 /**
 251  * This function adds buffers to the virtio devices RX virtqueue. Buffers can
 252  * be received from the physical port or from another virtio device. A packet
 253  * count is returned to indicate the number of packets that are succesfully
 254  * added to the RX queue. This function works when the mbuf is scattered, but
 255  * it doesn't support the mergeable feature.
 256  */
 257 static inline uint32_t __attribute__((always_inline))
 258 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
 259               struct rte_mbuf **pkts, uint32_t count)
 260 {
 261         struct vhost_virtqueue *vq;
 262         uint16_t res_start_idx, res_end_idx;
 263         uint16_t desc_indexes[MAX_PKT_BURST];
 264         uint16_t used_idx;
 265         uint32_t i;
 266
 267         LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
 268         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->virt_qp_nb))) {
 269                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
 270                         dev->vid, __func__, queue_id);
 271                 return 0;
 272         }
 273
 274         vq = dev->virtqueue[queue_id];
 275         if (unlikely(vq->enabled == 0))
 276                 return 0;
 277
 278         count = reserve_avail_buf(vq, count, &res_start_idx, &res_end_idx);
 279         if (count == 0)
 280                 return 0;
 281
 282         LOG_DEBUG(VHOST_DATA, "(%d) res_start_idx %d | res_end_idx Index %d\n",
 283                 dev->vid, res_start_idx, res_end_idx);
 284
 285         /* Retrieve all of the desc indexes first to avoid caching issues. */
 286         rte_prefetch0(&vq->avail->ring[res_start_idx & (vq->size - 1)]);
 287         for (i = 0; i < count; i++) {
 288                 used_idx = (res_start_idx + i) & (vq->size - 1);
 289                 desc_indexes[i] = vq->avail->ring[used_idx];
 290                 vq->used->ring[used_idx].id = desc_indexes[i];
 291                 vq->used->ring[used_idx].len = pkts[i]->pkt_len +
 292                                                dev->vhost_hlen;
 293                 vhost_log_used_vring(dev, vq,
 294                         offsetof(struct vring_used, ring[used_idx]),
 295                         sizeof(vq->used->ring[used_idx]));
 296         }
 297
 298         rte_prefetch0(&vq->desc[desc_indexes[0]]);
 299         for (i = 0; i < count; i++) {
 300                 uint16_t desc_idx = desc_indexes[i];
 301                 int err;
 302
 303                 err = copy_mbuf_to_desc(dev, vq, pkts[i], desc_idx);
 304                 if (unlikely(err)) {
 305                         used_idx = (res_start_idx + i) & (vq->size - 1);
 306                         vq->used->ring[used_idx].len = dev->vhost_hlen;
 307                         vhost_log_used_vring(dev, vq,
 308                                 offsetof(struct vring_used, ring[used_idx]),
 309                                 sizeof(vq->used->ring[used_idx]));
 310                 }
 311
 312                 if (i + 1 < count)
 313                         rte_prefetch0(&vq->desc[desc_indexes[i+1]]);
 314         }
 315
 316         rte_smp_wmb();
 317
 318         /* Wait until it's our turn to add our buffer to the used ring. */
 319         while (unlikely(vq->last_used_idx != res_start_idx))
 320                 rte_pause();
 321
 322         *(volatile uint16_t *)&vq->used->idx += count;
 323         vq->last_used_idx = res_end_idx;
 324         vhost_log_used_vring(dev, vq,
 325                 offsetof(struct vring_used, idx),
 326                 sizeof(vq->used->idx));
 327
 328         /* flush used->idx update before we read avail->flags. */
 329         rte_mb();
 330
 331         /* Kick the guest if necessary. */
 332         if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
 333                         && (vq->callfd >= 0))
 334                 eventfd_write(vq->callfd, (eventfd_t)1);
 335         return count;
 336 }
 337
 338 static inline int
 339 fill_vec_buf(struct vhost_virtqueue *vq, uint32_t avail_idx,
 340              uint32_t *allocated, uint32_t *vec_idx,
 341              struct buf_vector *buf_vec)
 342 {
 343         uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
 344         uint32_t vec_id = *vec_idx;
 345         uint32_t len    = *allocated;
 346
 347         while (1) {
 348                 if (unlikely(vec_id >= BUF_VECTOR_MAX || idx >= vq->size))
 349                         return -1;
 350
 351                 len += vq->desc[idx].len;
 352                 buf_vec[vec_id].buf_addr = vq->desc[idx].addr;
 353                 buf_vec[vec_id].buf_len  = vq->desc[idx].len;
 354                 buf_vec[vec_id].desc_idx = idx;
 355                 vec_id++;
 356
 357                 if ((vq->desc[idx].flags & VRING_DESC_F_NEXT) == 0)
 358                         break;
 359
 360                 idx = vq->desc[idx].next;
 361         }
 362
 363         *allocated = len;
 364         *vec_idx   = vec_id;
 365
 366         return 0;
 367 }
 368
 369 /*
 370  * As many data cores may want to access available buffers concurrently,
 371  * they need to be reserved.
 372  *
 373  * Returns -1 on fail, 0 on success
 374  */
 375 static inline int
 376 reserve_avail_buf_mergeable(struct vhost_virtqueue *vq, uint32_t size,
 377                             uint16_t *start, uint16_t *end,
 378                             struct buf_vector *buf_vec)
 379 {
 380         uint16_t res_start_idx;
 381         uint16_t res_cur_idx;
 382         uint16_t avail_idx;
 383         uint32_t allocated;
 384         uint32_t vec_idx;
 385         uint16_t tries;
 386
 387 again:
 388         res_start_idx = vq->last_used_idx_res;
 389         res_cur_idx  = res_start_idx;
 390
 391         allocated = 0;
 392         vec_idx   = 0;
 393         tries     = 0;
 394         while (1) {
 395                 avail_idx = *((volatile uint16_t *)&vq->avail->idx);
 396                 if (unlikely(res_cur_idx == avail_idx))
 397                         return -1;
 398
 399                 if (unlikely(fill_vec_buf(vq, res_cur_idx, &allocated,
 400                                           &vec_idx, buf_vec) < 0))
 401                         return -1;
 402
 403                 res_cur_idx++;
 404                 tries++;
 405
 406                 if (allocated >= size)
 407                         break;
 408
 409                 /*
 410                  * if we tried all available ring items, and still
 411                  * can't get enough buf, it means something abnormal
 412                  * happened.
 413                  */
 414                 if (unlikely(tries >= vq->size))
 415                         return -1;
 416         }
 417
 418         /*
 419          * update vq->last_used_idx_res atomically.
 420          * retry again if failed.
 421          */
 422         if (rte_atomic16_cmpset(&vq->last_used_idx_res,
 423                                 res_start_idx, res_cur_idx) == 0)
 424                 goto again;
 425
 426         *start = res_start_idx;
 427         *end   = res_cur_idx;
 428         return 0;
 429 }
 430
 431 static inline uint32_t __attribute__((always_inline))
 432 copy_mbuf_to_desc_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
 433                             uint16_t res_start_idx, uint16_t res_end_idx,
 434                             struct rte_mbuf *m, struct buf_vector *buf_vec)
 435 {
 436         struct virtio_net_hdr_mrg_rxbuf virtio_hdr = {{0, 0, 0, 0, 0, 0}, 0};
 437         uint32_t vec_idx = 0;
 438         uint16_t cur_idx = res_start_idx;
 439         uint64_t desc_addr;
 440         uint32_t mbuf_offset, mbuf_avail;
 441         uint32_t desc_offset, desc_avail;
 442         uint32_t cpy_len;
 443         uint16_t desc_idx, used_idx;
 444
 445         if (unlikely(m == NULL))
 446                 return 0;
 447
 448         LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
 449                 dev->vid, cur_idx, res_end_idx);
 450
 451         if (buf_vec[vec_idx].buf_len < dev->vhost_hlen)
 452                 return -1;
 453
 454         desc_addr = gpa_to_vva(dev, buf_vec[vec_idx].buf_addr);
 455         rte_prefetch0((void *)(uintptr_t)desc_addr);
 456
 457         virtio_hdr.num_buffers = res_end_idx - res_start_idx;
 458         LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
 459                 dev->vid, virtio_hdr.num_buffers);
 460
 461         virtio_enqueue_offload(m, &virtio_hdr.hdr);
 462         copy_virtio_net_hdr(dev, desc_addr, virtio_hdr);
 463         vhost_log_write(dev, buf_vec[vec_idx].buf_addr, dev->vhost_hlen);
 464         PRINT_PACKET(dev, (uintptr_t)desc_addr, dev->vhost_hlen, 0);
 465
 466         desc_avail  = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
 467         desc_offset = dev->vhost_hlen;
 468
 469         mbuf_avail  = rte_pktmbuf_data_len(m);
 470         mbuf_offset = 0;
 471         while (mbuf_avail != 0 || m->next != NULL) {
 472                 /* done with current desc buf, get the next one */
 473                 if (desc_avail == 0) {
 474                         desc_idx = buf_vec[vec_idx].desc_idx;
 475
 476                         if (!(vq->desc[desc_idx].flags & VRING_DESC_F_NEXT)) {
 477                                 /* Update used ring with desc information */
 478                                 used_idx = cur_idx++ & (vq->size - 1);
 479                                 vq->used->ring[used_idx].id  = desc_idx;
 480                                 vq->used->ring[used_idx].len = desc_offset;
 481                                 vhost_log_used_vring(dev, vq,
 482                                         offsetof(struct vring_used,
 483                                                  ring[used_idx]),
 484                                         sizeof(vq->used->ring[used_idx]));
 485                         }
 486
 487                         vec_idx++;
 488                         desc_addr = gpa_to_vva(dev, buf_vec[vec_idx].buf_addr);
 489
 490                         /* Prefetch buffer address. */
 491                         rte_prefetch0((void *)(uintptr_t)desc_addr);
 492                         desc_offset = 0;
 493                         desc_avail  = buf_vec[vec_idx].buf_len;
 494                 }
 495
 496                 /* done with current mbuf, get the next one */
 497                 if (mbuf_avail == 0) {
 498                         m = m->next;
 499
 500                         mbuf_offset = 0;
 501                         mbuf_avail  = rte_pktmbuf_data_len(m);
 502                 }
 503
 504                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 505                 rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)),
 506                         rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 507                         cpy_len);
 508                 vhost_log_write(dev, buf_vec[vec_idx].buf_addr + desc_offset,
 509                         cpy_len);
 510                 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
 511                         cpy_len, 0);
 512
 513                 mbuf_avail  -= cpy_len;
 514                 mbuf_offset += cpy_len;
 515                 desc_avail  -= cpy_len;
 516                 desc_offset += cpy_len;
 517         }
 518
 519         used_idx = cur_idx & (vq->size - 1);
 520         vq->used->ring[used_idx].id = buf_vec[vec_idx].desc_idx;
 521         vq->used->ring[used_idx].len = desc_offset;
 522         vhost_log_used_vring(dev, vq,
 523                 offsetof(struct vring_used, ring[used_idx]),
 524                 sizeof(vq->used->ring[used_idx]));
 525
 526         return res_end_idx - res_start_idx;
 527 }
 528
 529 static inline uint32_t __attribute__((always_inline))
 530 virtio_dev_merge_rx(struct virtio_net *dev, uint16_t queue_id,
 531         struct rte_mbuf **pkts, uint32_t count)
 532 {
 533         struct vhost_virtqueue *vq;
 534         uint32_t pkt_idx = 0, nr_used = 0;
 535         uint16_t start, end;
 536         struct buf_vector buf_vec[BUF_VECTOR_MAX];
 537
 538         LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
 539         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->virt_qp_nb))) {
 540                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
 541                         dev->vid, __func__, queue_id);
 542                 return 0;
 543         }
 544
 545         vq = dev->virtqueue[queue_id];
 546         if (unlikely(vq->enabled == 0))
 547                 return 0;
 548
 549         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
 550         if (count == 0)
 551                 return 0;
 552
 553         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
 554                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
 555
 556                 if (unlikely(reserve_avail_buf_mergeable(vq, pkt_len, &start,
 557                                                          &end, buf_vec) < 0)) {
 558                         LOG_DEBUG(VHOST_DATA,
 559                                 "(%d) failed to get enough desc from vring\n",
 560                                 dev->vid);
 561                         break;
 562                 }
 563
 564                 nr_used = copy_mbuf_to_desc_mergeable(dev, vq, start, end,
 565                                                       pkts[pkt_idx], buf_vec);
 566                 rte_smp_wmb();
 567
 568                 /*
 569                  * Wait until it's our turn to add our buffer
 570                  * to the used ring.
 571                  */
 572                 while (unlikely(vq->last_used_idx != start))
 573                         rte_pause();
 574
 575                 *(volatile uint16_t *)&vq->used->idx += nr_used;
 576                 vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
 577                         sizeof(vq->used->idx));
 578                 vq->last_used_idx = end;
 579         }
 580
 581         if (likely(pkt_idx)) {
 582                 /* flush used->idx update before we read avail->flags. */
 583                 rte_mb();
 584
 585                 /* Kick the guest if necessary. */
 586                 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
 587                                 && (vq->callfd >= 0))
 588                         eventfd_write(vq->callfd, (eventfd_t)1);
 589         }
 590
 591         return pkt_idx;
 592 }
 593
 594 uint16_t
 595 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
 596         struct rte_mbuf **pkts, uint16_t count)
 597 {
 598         struct virtio_net *dev = get_device(vid);
 599
 600         if (!dev)
 601                 return 0;
 602
 603         if (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF))
 604                 return virtio_dev_merge_rx(dev, queue_id, pkts, count);
 605         else
 606                 return virtio_dev_rx(dev, queue_id, pkts, count);
 607 }
 608
 609 static void
 610 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
 611 {
 612         struct ipv4_hdr *ipv4_hdr;
 613         struct ipv6_hdr *ipv6_hdr;
 614         void *l3_hdr = NULL;
 615         struct ether_hdr *eth_hdr;
 616         uint16_t ethertype;
 617
 618         eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
 619
 620         m->l2_len = sizeof(struct ether_hdr);
 621         ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
 622
 623         if (ethertype == ETHER_TYPE_VLAN) {
 624                 struct vlan_hdr *vlan_hdr = (struct vlan_hdr *)(eth_hdr + 1);
 625
 626                 m->l2_len += sizeof(struct vlan_hdr);
 627                 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
 628         }
 629
 630         l3_hdr = (char *)eth_hdr + m->l2_len;
 631
 632         switch (ethertype) {
 633         case ETHER_TYPE_IPv4:
 634                 ipv4_hdr = (struct ipv4_hdr *)l3_hdr;
 635                 *l4_proto = ipv4_hdr->next_proto_id;
 636                 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
 637                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 638                 m->ol_flags |= PKT_TX_IPV4;
 639                 break;
 640         case ETHER_TYPE_IPv6:
 641                 ipv6_hdr = (struct ipv6_hdr *)l3_hdr;
 642                 *l4_proto = ipv6_hdr->proto;
 643                 m->l3_len = sizeof(struct ipv6_hdr);
 644                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 645                 m->ol_flags |= PKT_TX_IPV6;
 646                 break;
 647         default:
 648                 m->l3_len = 0;
 649                 *l4_proto = 0;
 650                 break;
 651         }
 652 }
 653
 654 static inline void __attribute__((always_inline))
 655 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
 656 {
 657         uint16_t l4_proto = 0;
 658         void *l4_hdr = NULL;
 659         struct tcp_hdr *tcp_hdr = NULL;
 660
 661         parse_ethernet(m, &l4_proto, &l4_hdr);
 662         if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 663                 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
 664                         switch (hdr->csum_offset) {
 665                         case (offsetof(struct tcp_hdr, cksum)):
 666                                 if (l4_proto == IPPROTO_TCP)
 667                                         m->ol_flags |= PKT_TX_TCP_CKSUM;
 668                                 break;
 669                         case (offsetof(struct udp_hdr, dgram_cksum)):
 670                                 if (l4_proto == IPPROTO_UDP)
 671                                         m->ol_flags |= PKT_TX_UDP_CKSUM;
 672                                 break;
 673                         case (offsetof(struct sctp_hdr, cksum)):
 674                                 if (l4_proto == IPPROTO_SCTP)
 675                                         m->ol_flags |= PKT_TX_SCTP_CKSUM;
 676                                 break;
 677                         default:
 678                                 break;
 679                         }
 680                 }
 681         }
 682
 683         if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 684                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 685                 case VIRTIO_NET_HDR_GSO_TCPV4:
 686                 case VIRTIO_NET_HDR_GSO_TCPV6:
 687                         tcp_hdr = (struct tcp_hdr *)l4_hdr;
 688                         m->ol_flags |= PKT_TX_TCP_SEG;
 689                         m->tso_segsz = hdr->gso_size;
 690                         m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
 691                         break;
 692                 default:
 693                         RTE_LOG(WARNING, VHOST_DATA,
 694                                 "unsupported gso type %u.\n", hdr->gso_type);
 695                         break;
 696                 }
 697         }
 698 }
 699
 700 #define RARP_PKT_SIZE   64
 701
 702 static int
 703 make_rarp_packet(struct rte_mbuf *rarp_mbuf, const struct ether_addr *mac)
 704 {
 705         struct ether_hdr *eth_hdr;
 706         struct arp_hdr  *rarp;
 707
 708         if (rarp_mbuf->buf_len < 64) {
 709                 RTE_LOG(WARNING, VHOST_DATA,
 710                         "failed to make RARP; mbuf size too small %u (< %d)\n",
 711                         rarp_mbuf->buf_len, RARP_PKT_SIZE);
 712                 return -1;
 713         }
 714
 715         /* Ethernet header. */
 716         eth_hdr = rte_pktmbuf_mtod_offset(rarp_mbuf, struct ether_hdr *, 0);
 717         memset(eth_hdr->d_addr.addr_bytes, 0xff, ETHER_ADDR_LEN);
 718         ether_addr_copy(mac, &eth_hdr->s_addr);
 719         eth_hdr->ether_type = htons(ETHER_TYPE_RARP);
 720
 721         /* RARP header. */
 722         rarp = (struct arp_hdr *)(eth_hdr + 1);
 723         rarp->arp_hrd = htons(ARP_HRD_ETHER);
 724         rarp->arp_pro = htons(ETHER_TYPE_IPv4);
 725         rarp->arp_hln = ETHER_ADDR_LEN;
 726         rarp->arp_pln = 4;
 727         rarp->arp_op  = htons(ARP_OP_REVREQUEST);
 728
 729         ether_addr_copy(mac, &rarp->arp_data.arp_sha);
 730         ether_addr_copy(mac, &rarp->arp_data.arp_tha);
 731         memset(&rarp->arp_data.arp_sip, 0x00, 4);
 732         memset(&rarp->arp_data.arp_tip, 0x00, 4);
 733
 734         rarp_mbuf->pkt_len  = rarp_mbuf->data_len = RARP_PKT_SIZE;
 735
 736         return 0;
 737 }
 738
 739 static inline int __attribute__((always_inline))
 740 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 741                   struct rte_mbuf *m, uint16_t desc_idx,
 742                   struct rte_mempool *mbuf_pool)
 743 {
 744         struct vring_desc *desc;
 745         uint64_t desc_addr;
 746         uint32_t desc_avail, desc_offset;
 747         uint32_t mbuf_avail, mbuf_offset;
 748         uint32_t cpy_len;
 749         struct rte_mbuf *cur = m, *prev = m;
 750         struct virtio_net_hdr *hdr;
 751         /* A counter to avoid desc dead loop chain */
 752         uint32_t nr_desc = 1;
 753
 754         desc = &vq->desc[desc_idx];
 755         if (unlikely(desc->len < dev->vhost_hlen))
 756                 return -1;
 757
 758         desc_addr = gpa_to_vva(dev, desc->addr);
 759         rte_prefetch0((void *)(uintptr_t)desc_addr);
 760
 761         /* Retrieve virtio net header */
 762         hdr = (struct virtio_net_hdr *)((uintptr_t)desc_addr);
 763         desc_avail  = desc->len - dev->vhost_hlen;
 764         desc_offset = dev->vhost_hlen;
 765
 766         mbuf_offset = 0;
 767         mbuf_avail  = m->buf_len - RTE_PKTMBUF_HEADROOM;
 768         while (desc_avail != 0 || (desc->flags & VRING_DESC_F_NEXT) != 0) {
 769                 /* This desc reaches to its end, get the next one */
 770                 if (desc_avail == 0) {
 771                         if (unlikely(desc->next >= vq->size ||
 772                                      ++nr_desc >= vq->size))
 773                                 return -1;
 774                         desc = &vq->desc[desc->next];
 775
 776                         desc_addr = gpa_to_vva(dev, desc->addr);
 777                         rte_prefetch0((void *)(uintptr_t)desc_addr);
 778
 779                         desc_offset = 0;
 780                         desc_avail  = desc->len;
 781
 782                         PRINT_PACKET(dev, (uintptr_t)desc_addr, desc->len, 0);
 783                 }
 784
 785                 /*
 786                  * This mbuf reaches to its end, get a new one
 787                  * to hold more data.
 788                  */
 789                 if (mbuf_avail == 0) {
 790                         cur = rte_pktmbuf_alloc(mbuf_pool);
 791                         if (unlikely(cur == NULL)) {
 792                                 RTE_LOG(ERR, VHOST_DATA, "Failed to "
 793                                         "allocate memory for mbuf.\n");
 794                                 return -1;
 795                         }
 796
 797                         prev->next = cur;
 798                         prev->data_len = mbuf_offset;
 799                         m->nb_segs += 1;
 800                         m->pkt_len += mbuf_offset;
 801                         prev = cur;
 802
 803                         mbuf_offset = 0;
 804                         mbuf_avail  = cur->buf_len - RTE_PKTMBUF_HEADROOM;
 805                 }
 806
 807                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 808                 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *, mbuf_offset),
 809                         (void *)((uintptr_t)(desc_addr + desc_offset)),
 810                         cpy_len);
 811
 812                 mbuf_avail  -= cpy_len;
 813                 mbuf_offset += cpy_len;
 814                 desc_avail  -= cpy_len;
 815                 desc_offset += cpy_len;
 816         }
 817
 818         prev->data_len = mbuf_offset;
 819         m->pkt_len    += mbuf_offset;
 820
 821         if (hdr->flags != 0 || hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE)
 822                 vhost_dequeue_offload(hdr, m);
 823
 824         return 0;
 825 }
 826
 827 uint16_t
 828 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
 829         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
 830 {
 831         struct virtio_net *dev;
 832         struct rte_mbuf *rarp_mbuf = NULL;
 833         struct vhost_virtqueue *vq;
 834         uint32_t desc_indexes[MAX_PKT_BURST];
 835         uint32_t used_idx;
 836         uint32_t i = 0;
 837         uint16_t free_entries;
 838         uint16_t avail_idx;
 839
 840         dev = get_device(vid);
 841         if (!dev)
 842                 return 0;
 843
 844         if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->virt_qp_nb))) {
 845                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
 846                         dev->vid, __func__, queue_id);
 847                 return 0;
 848         }
 849
 850         vq = dev->virtqueue[queue_id];
 851         if (unlikely(vq->enabled == 0))
 852                 return 0;
 853
 854         /*
 855          * Construct a RARP broadcast packet, and inject it to the "pkts"
 856          * array, to looks like that guest actually send such packet.
 857          *
 858          * Check user_send_rarp() for more information.
 859          */
 860         if (unlikely(rte_atomic16_cmpset((volatile uint16_t *)
 861                                          &dev->broadcast_rarp.cnt, 1, 0))) {
 862                 rarp_mbuf = rte_pktmbuf_alloc(mbuf_pool);
 863                 if (rarp_mbuf == NULL) {
 864                         RTE_LOG(ERR, VHOST_DATA,
 865                                 "Failed to allocate memory for mbuf.\n");
 866                         return 0;
 867                 }
 868
 869                 if (make_rarp_packet(rarp_mbuf, &dev->mac)) {
 870                         rte_pktmbuf_free(rarp_mbuf);
 871                         rarp_mbuf = NULL;
 872                 } else {
 873                         count -= 1;
 874                 }
 875         }
 876
 877         avail_idx =  *((volatile uint16_t *)&vq->avail->idx);
 878         free_entries = avail_idx - vq->last_used_idx;
 879         if (free_entries == 0)
 880                 goto out;
 881
 882         LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
 883
 884         /* Prefetch available ring to retrieve head indexes. */
 885         used_idx = vq->last_used_idx & (vq->size - 1);
 886         rte_prefetch0(&vq->avail->ring[used_idx]);
 887         rte_prefetch0(&vq->used->ring[used_idx]);
 888
 889         count = RTE_MIN(count, MAX_PKT_BURST);
 890         count = RTE_MIN(count, free_entries);
 891         LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
 892                         dev->vid, count);
 893
 894         /* Retrieve all of the head indexes first to avoid caching issues. */
 895         for (i = 0; i < count; i++) {
 896                 used_idx = (vq->last_used_idx + i) & (vq->size - 1);
 897                 desc_indexes[i] = vq->avail->ring[used_idx];
 898
 899                 vq->used->ring[used_idx].id  = desc_indexes[i];
 900                 vq->used->ring[used_idx].len = 0;
 901                 vhost_log_used_vring(dev, vq,
 902                                 offsetof(struct vring_used, ring[used_idx]),
 903                                 sizeof(vq->used->ring[used_idx]));
 904         }
 905
 906         /* Prefetch descriptor index. */
 907         rte_prefetch0(&vq->desc[desc_indexes[0]]);
 908         for (i = 0; i < count; i++) {
 909                 int err;
 910
 911                 if (likely(i + 1 < count))
 912                         rte_prefetch0(&vq->desc[desc_indexes[i + 1]]);
 913
 914                 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
 915                 if (unlikely(pkts[i] == NULL)) {
 916                         RTE_LOG(ERR, VHOST_DATA,
 917                                 "Failed to allocate memory for mbuf.\n");
 918                         break;
 919                 }
 920                 err = copy_desc_to_mbuf(dev, vq, pkts[i], desc_indexes[i],
 921                                         mbuf_pool);
 922                 if (unlikely(err)) {
 923                         rte_pktmbuf_free(pkts[i]);
 924                         break;
 925                 }
 926         }
 927
 928         rte_smp_wmb();
 929         rte_smp_rmb();
 930         vq->used->idx += i;
 931         vq->last_used_idx += i;
 932         vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
 933                         sizeof(vq->used->idx));
 934
 935         /* Kick guest if required. */
 936         if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
 937                         && (vq->callfd >= 0))
 938                 eventfd_write(vq->callfd, (eventfd_t)1);
 939
 940 out:
 941         if (unlikely(rarp_mbuf != NULL)) {
 942                 /*
 943                  * Inject it to the head of "pkts" array, so that switch's mac
 944                  * learning table will get updated first.
 945                  */
 946                 memmove(&pkts[1], pkts, i * sizeof(struct rte_mbuf *));
 947                 pkts[0] = rarp_mbuf;
 948                 i += 1;
 949         }
 950
 951         return i;
 952 }