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, uint32_t *copied)
 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         *copied = rte_pktmbuf_pkt_len(m);
 165         mbuf_avail  = rte_pktmbuf_data_len(m);
 166         mbuf_offset = 0;
 167         while (mbuf_avail != 0 || m->next != NULL) {
 168                 /* done with current mbuf, fetch next */
 169                 if (mbuf_avail == 0) {
 170                         m = m->next;
 171
 172                         mbuf_offset = 0;
 173                         mbuf_avail  = rte_pktmbuf_data_len(m);
 174                 }
 175
 176                 /* done with current desc buf, fetch next */
 177                 if (desc_avail == 0) {
 178                         if ((desc->flags & VRING_DESC_F_NEXT) == 0) {
 179                                 /* Room in vring buffer is not enough */
 180                                 return -1;
 181                         }
 182                         if (unlikely(desc->next >= vq->size))
 183                                 return -1;
 184
 185                         desc = &vq->desc[desc->next];
 186                         desc_addr   = gpa_to_vva(dev, desc->addr);
 187                         desc_offset = 0;
 188                         desc_avail  = desc->len;
 189                 }
 190
 191                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 192                 rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)),
 193                         rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 194                         cpy_len);
 195                 vhost_log_write(dev, desc->addr + desc_offset, cpy_len);
 196                 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
 197                              cpy_len, 0);
 198
 199                 mbuf_avail  -= cpy_len;
 200                 mbuf_offset += cpy_len;
 201                 desc_avail  -= cpy_len;
 202                 desc_offset += cpy_len;
 203         }
 204
 205         return 0;
 206 }
 207
 208 /*
 209  * As many data cores may want to access available buffers
 210  * they need to be reserved.
 211  */
 212 static inline uint32_t
 213 reserve_avail_buf(struct vhost_virtqueue *vq, uint32_t count,
 214                   uint16_t *start, uint16_t *end)
 215 {
 216         uint16_t res_start_idx;
 217         uint16_t res_end_idx;
 218         uint16_t avail_idx;
 219         uint16_t free_entries;
 220         int success;
 221
 222         count = RTE_MIN(count, (uint32_t)MAX_PKT_BURST);
 223
 224 again:
 225         res_start_idx = vq->last_used_idx_res;
 226         avail_idx = *((volatile uint16_t *)&vq->avail->idx);
 227
 228         free_entries = avail_idx - res_start_idx;
 229         count = RTE_MIN(count, free_entries);
 230         if (count == 0)
 231                 return 0;
 232
 233         res_end_idx = res_start_idx + count;
 234
 235         /*
 236          * update vq->last_used_idx_res atomically; try again if failed.
 237          *
 238          * TODO: Allow to disable cmpset if no concurrency in application.
 239          */
 240         success = rte_atomic16_cmpset(&vq->last_used_idx_res,
 241                                       res_start_idx, res_end_idx);
 242         if (unlikely(!success))
 243                 goto again;
 244
 245         *start = res_start_idx;
 246         *end   = res_end_idx;
 247
 248         return count;
 249 }
 250
 251 /**
 252  * This function adds buffers to the virtio devices RX virtqueue. Buffers can
 253  * be received from the physical port or from another virtio device. A packet
 254  * count is returned to indicate the number of packets that are succesfully
 255  * added to the RX queue. This function works when the mbuf is scattered, but
 256  * it doesn't support the mergeable feature.
 257  */
 258 static inline uint32_t __attribute__((always_inline))
 259 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
 260               struct rte_mbuf **pkts, uint32_t count)
 261 {
 262         struct vhost_virtqueue *vq;
 263         uint16_t res_start_idx, res_end_idx;
 264         uint16_t desc_indexes[MAX_PKT_BURST];
 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                 desc_indexes[i] = vq->avail->ring[(res_start_idx + i) &
 289                                                   (vq->size - 1)];
 290         }
 291
 292         rte_prefetch0(&vq->desc[desc_indexes[0]]);
 293         for (i = 0; i < count; i++) {
 294                 uint16_t desc_idx = desc_indexes[i];
 295                 uint16_t used_idx = (res_start_idx + i) & (vq->size - 1);
 296                 uint32_t copied;
 297                 int err;
 298
 299                 err = copy_mbuf_to_desc(dev, vq, pkts[i], desc_idx, &copied);
 300
 301                 vq->used->ring[used_idx].id = desc_idx;
 302                 if (unlikely(err))
 303                         vq->used->ring[used_idx].len = dev->vhost_hlen;
 304                 else
 305                         vq->used->ring[used_idx].len = copied + dev->vhost_hlen;
 306                 vhost_log_used_vring(dev, vq,
 307                         offsetof(struct vring_used, ring[used_idx]),
 308                         sizeof(vq->used->ring[used_idx]));
 309
 310                 if (i + 1 < count)
 311                         rte_prefetch0(&vq->desc[desc_indexes[i+1]]);
 312         }
 313
 314         rte_smp_wmb();
 315
 316         /* Wait until it's our turn to add our buffer to the used ring. */
 317         while (unlikely(vq->last_used_idx != res_start_idx))
 318                 rte_pause();
 319
 320         *(volatile uint16_t *)&vq->used->idx += count;
 321         vq->last_used_idx = res_end_idx;
 322         vhost_log_used_vring(dev, vq,
 323                 offsetof(struct vring_used, idx),
 324                 sizeof(vq->used->idx));
 325
 326         /* flush used->idx update before we read avail->flags. */
 327         rte_mb();
 328
 329         /* Kick the guest if necessary. */
 330         if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
 331                         && (vq->callfd >= 0))
 332                 eventfd_write(vq->callfd, (eventfd_t)1);
 333         return count;
 334 }
 335
 336 static inline int
 337 fill_vec_buf(struct vhost_virtqueue *vq, uint32_t avail_idx,
 338              uint32_t *allocated, uint32_t *vec_idx,
 339              struct buf_vector *buf_vec)
 340 {
 341         uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
 342         uint32_t vec_id = *vec_idx;
 343         uint32_t len    = *allocated;
 344
 345         while (1) {
 346                 if (unlikely(vec_id >= BUF_VECTOR_MAX || idx >= vq->size))
 347                         return -1;
 348
 349                 len += vq->desc[idx].len;
 350                 buf_vec[vec_id].buf_addr = vq->desc[idx].addr;
 351                 buf_vec[vec_id].buf_len  = vq->desc[idx].len;
 352                 buf_vec[vec_id].desc_idx = idx;
 353                 vec_id++;
 354
 355                 if ((vq->desc[idx].flags & VRING_DESC_F_NEXT) == 0)
 356                         break;
 357
 358                 idx = vq->desc[idx].next;
 359         }
 360
 361         *allocated = len;
 362         *vec_idx   = vec_id;
 363
 364         return 0;
 365 }
 366
 367 /*
 368  * As many data cores may want to access available buffers concurrently,
 369  * they need to be reserved.
 370  *
 371  * Returns -1 on fail, 0 on success
 372  */
 373 static inline int
 374 reserve_avail_buf_mergeable(struct vhost_virtqueue *vq, uint32_t size,
 375                             uint16_t *start, uint16_t *end,
 376                             struct buf_vector *buf_vec)
 377 {
 378         uint16_t res_start_idx;
 379         uint16_t res_cur_idx;
 380         uint16_t avail_idx;
 381         uint32_t allocated;
 382         uint32_t vec_idx;
 383         uint16_t tries;
 384
 385 again:
 386         res_start_idx = vq->last_used_idx_res;
 387         res_cur_idx  = res_start_idx;
 388
 389         allocated = 0;
 390         vec_idx   = 0;
 391         tries     = 0;
 392         while (1) {
 393                 avail_idx = *((volatile uint16_t *)&vq->avail->idx);
 394                 if (unlikely(res_cur_idx == avail_idx))
 395                         return -1;
 396
 397                 if (unlikely(fill_vec_buf(vq, res_cur_idx, &allocated,
 398                                           &vec_idx, buf_vec) < 0))
 399                         return -1;
 400
 401                 res_cur_idx++;
 402                 tries++;
 403
 404                 if (allocated >= size)
 405                         break;
 406
 407                 /*
 408                  * if we tried all available ring items, and still
 409                  * can't get enough buf, it means something abnormal
 410                  * happened.
 411                  */
 412                 if (unlikely(tries >= vq->size))
 413                         return -1;
 414         }
 415
 416         /*
 417          * update vq->last_used_idx_res atomically.
 418          * retry again if failed.
 419          */
 420         if (rte_atomic16_cmpset(&vq->last_used_idx_res,
 421                                 res_start_idx, res_cur_idx) == 0)
 422                 goto again;
 423
 424         *start = res_start_idx;
 425         *end   = res_cur_idx;
 426         return 0;
 427 }
 428
 429 static inline uint32_t __attribute__((always_inline))
 430 copy_mbuf_to_desc_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
 431                             uint16_t res_start_idx, uint16_t res_end_idx,
 432                             struct rte_mbuf *m, struct buf_vector *buf_vec)
 433 {
 434         struct virtio_net_hdr_mrg_rxbuf virtio_hdr = {{0, 0, 0, 0, 0, 0}, 0};
 435         uint32_t vec_idx = 0;
 436         uint16_t cur_idx = res_start_idx;
 437         uint64_t desc_addr;
 438         uint32_t mbuf_offset, mbuf_avail;
 439         uint32_t desc_offset, desc_avail;
 440         uint32_t cpy_len;
 441         uint16_t desc_idx, used_idx;
 442
 443         if (unlikely(m == NULL))
 444                 return 0;
 445
 446         LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
 447                 dev->vid, cur_idx, res_end_idx);
 448
 449         if (buf_vec[vec_idx].buf_len < dev->vhost_hlen)
 450                 return -1;
 451
 452         desc_addr = gpa_to_vva(dev, buf_vec[vec_idx].buf_addr);
 453         rte_prefetch0((void *)(uintptr_t)desc_addr);
 454
 455         virtio_hdr.num_buffers = res_end_idx - res_start_idx;
 456         LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
 457                 dev->vid, virtio_hdr.num_buffers);
 458
 459         virtio_enqueue_offload(m, &virtio_hdr.hdr);
 460         copy_virtio_net_hdr(dev, desc_addr, virtio_hdr);
 461         vhost_log_write(dev, buf_vec[vec_idx].buf_addr, dev->vhost_hlen);
 462         PRINT_PACKET(dev, (uintptr_t)desc_addr, dev->vhost_hlen, 0);
 463
 464         desc_avail  = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
 465         desc_offset = dev->vhost_hlen;
 466
 467         mbuf_avail  = rte_pktmbuf_data_len(m);
 468         mbuf_offset = 0;
 469         while (mbuf_avail != 0 || m->next != NULL) {
 470                 /* done with current desc buf, get the next one */
 471                 if (desc_avail == 0) {
 472                         desc_idx = buf_vec[vec_idx].desc_idx;
 473
 474                         if (!(vq->desc[desc_idx].flags & VRING_DESC_F_NEXT)) {
 475                                 /* Update used ring with desc information */
 476                                 used_idx = cur_idx++ & (vq->size - 1);
 477                                 vq->used->ring[used_idx].id  = desc_idx;
 478                                 vq->used->ring[used_idx].len = desc_offset;
 479                                 vhost_log_used_vring(dev, vq,
 480                                         offsetof(struct vring_used,
 481                                                  ring[used_idx]),
 482                                         sizeof(vq->used->ring[used_idx]));
 483                         }
 484
 485                         vec_idx++;
 486                         desc_addr = gpa_to_vva(dev, buf_vec[vec_idx].buf_addr);
 487
 488                         /* Prefetch buffer address. */
 489                         rte_prefetch0((void *)(uintptr_t)desc_addr);
 490                         desc_offset = 0;
 491                         desc_avail  = buf_vec[vec_idx].buf_len;
 492                 }
 493
 494                 /* done with current mbuf, get the next one */
 495                 if (mbuf_avail == 0) {
 496                         m = m->next;
 497
 498                         mbuf_offset = 0;
 499                         mbuf_avail  = rte_pktmbuf_data_len(m);
 500                 }
 501
 502                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 503                 rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)),
 504                         rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 505                         cpy_len);
 506                 vhost_log_write(dev, buf_vec[vec_idx].buf_addr + desc_offset,
 507                         cpy_len);
 508                 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
 509                         cpy_len, 0);
 510
 511                 mbuf_avail  -= cpy_len;
 512                 mbuf_offset += cpy_len;
 513                 desc_avail  -= cpy_len;
 514                 desc_offset += cpy_len;
 515         }
 516
 517         used_idx = cur_idx & (vq->size - 1);
 518         vq->used->ring[used_idx].id = buf_vec[vec_idx].desc_idx;
 519         vq->used->ring[used_idx].len = desc_offset;
 520         vhost_log_used_vring(dev, vq,
 521                 offsetof(struct vring_used, ring[used_idx]),
 522                 sizeof(vq->used->ring[used_idx]));
 523
 524         return res_end_idx - res_start_idx;
 525 }
 526
 527 static inline uint32_t __attribute__((always_inline))
 528 virtio_dev_merge_rx(struct virtio_net *dev, uint16_t queue_id,
 529         struct rte_mbuf **pkts, uint32_t count)
 530 {
 531         struct vhost_virtqueue *vq;
 532         uint32_t pkt_idx = 0, nr_used = 0;
 533         uint16_t start, end;
 534         struct buf_vector buf_vec[BUF_VECTOR_MAX];
 535
 536         LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
 537         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->virt_qp_nb))) {
 538                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
 539                         dev->vid, __func__, queue_id);
 540                 return 0;
 541         }
 542
 543         vq = dev->virtqueue[queue_id];
 544         if (unlikely(vq->enabled == 0))
 545                 return 0;
 546
 547         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
 548         if (count == 0)
 549                 return 0;
 550
 551         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
 552                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
 553
 554                 if (unlikely(reserve_avail_buf_mergeable(vq, pkt_len, &start,
 555                                                          &end, buf_vec) < 0)) {
 556                         LOG_DEBUG(VHOST_DATA,
 557                                 "(%d) failed to get enough desc from vring\n",
 558                                 dev->vid);
 559                         break;
 560                 }
 561
 562                 nr_used = copy_mbuf_to_desc_mergeable(dev, vq, start, end,
 563                                                       pkts[pkt_idx], buf_vec);
 564                 rte_smp_wmb();
 565
 566                 /*
 567                  * Wait until it's our turn to add our buffer
 568                  * to the used ring.
 569                  */
 570                 while (unlikely(vq->last_used_idx != start))
 571                         rte_pause();
 572
 573                 *(volatile uint16_t *)&vq->used->idx += nr_used;
 574                 vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
 575                         sizeof(vq->used->idx));
 576                 vq->last_used_idx = end;
 577         }
 578
 579         if (likely(pkt_idx)) {
 580                 /* flush used->idx update before we read avail->flags. */
 581                 rte_mb();
 582
 583                 /* Kick the guest if necessary. */
 584                 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
 585                                 && (vq->callfd >= 0))
 586                         eventfd_write(vq->callfd, (eventfd_t)1);
 587         }
 588
 589         return pkt_idx;
 590 }
 591
 592 uint16_t
 593 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
 594         struct rte_mbuf **pkts, uint16_t count)
 595 {
 596         struct virtio_net *dev = get_device(vid);
 597
 598         if (!dev)
 599                 return 0;
 600
 601         if (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF))
 602                 return virtio_dev_merge_rx(dev, queue_id, pkts, count);
 603         else
 604                 return virtio_dev_rx(dev, queue_id, pkts, count);
 605 }
 606
 607 static void
 608 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
 609 {
 610         struct ipv4_hdr *ipv4_hdr;
 611         struct ipv6_hdr *ipv6_hdr;
 612         void *l3_hdr = NULL;
 613         struct ether_hdr *eth_hdr;
 614         uint16_t ethertype;
 615
 616         eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
 617
 618         m->l2_len = sizeof(struct ether_hdr);
 619         ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
 620
 621         if (ethertype == ETHER_TYPE_VLAN) {
 622                 struct vlan_hdr *vlan_hdr = (struct vlan_hdr *)(eth_hdr + 1);
 623
 624                 m->l2_len += sizeof(struct vlan_hdr);
 625                 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
 626         }
 627
 628         l3_hdr = (char *)eth_hdr + m->l2_len;
 629
 630         switch (ethertype) {
 631         case ETHER_TYPE_IPv4:
 632                 ipv4_hdr = (struct ipv4_hdr *)l3_hdr;
 633                 *l4_proto = ipv4_hdr->next_proto_id;
 634                 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
 635                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 636                 m->ol_flags |= PKT_TX_IPV4;
 637                 break;
 638         case ETHER_TYPE_IPv6:
 639                 ipv6_hdr = (struct ipv6_hdr *)l3_hdr;
 640                 *l4_proto = ipv6_hdr->proto;
 641                 m->l3_len = sizeof(struct ipv6_hdr);
 642                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 643                 m->ol_flags |= PKT_TX_IPV6;
 644                 break;
 645         default:
 646                 m->l3_len = 0;
 647                 *l4_proto = 0;
 648                 break;
 649         }
 650 }
 651
 652 static inline void __attribute__((always_inline))
 653 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
 654 {
 655         uint16_t l4_proto = 0;
 656         void *l4_hdr = NULL;
 657         struct tcp_hdr *tcp_hdr = NULL;
 658
 659         parse_ethernet(m, &l4_proto, &l4_hdr);
 660         if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 661                 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
 662                         switch (hdr->csum_offset) {
 663                         case (offsetof(struct tcp_hdr, cksum)):
 664                                 if (l4_proto == IPPROTO_TCP)
 665                                         m->ol_flags |= PKT_TX_TCP_CKSUM;
 666                                 break;
 667                         case (offsetof(struct udp_hdr, dgram_cksum)):
 668                                 if (l4_proto == IPPROTO_UDP)
 669                                         m->ol_flags |= PKT_TX_UDP_CKSUM;
 670                                 break;
 671                         case (offsetof(struct sctp_hdr, cksum)):
 672                                 if (l4_proto == IPPROTO_SCTP)
 673                                         m->ol_flags |= PKT_TX_SCTP_CKSUM;
 674                                 break;
 675                         default:
 676                                 break;
 677                         }
 678                 }
 679         }
 680
 681         if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 682                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 683                 case VIRTIO_NET_HDR_GSO_TCPV4:
 684                 case VIRTIO_NET_HDR_GSO_TCPV6:
 685                         tcp_hdr = (struct tcp_hdr *)l4_hdr;
 686                         m->ol_flags |= PKT_TX_TCP_SEG;
 687                         m->tso_segsz = hdr->gso_size;
 688                         m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
 689                         break;
 690                 default:
 691                         RTE_LOG(WARNING, VHOST_DATA,
 692                                 "unsupported gso type %u.\n", hdr->gso_type);
 693                         break;
 694                 }
 695         }
 696 }
 697
 698 #define RARP_PKT_SIZE   64
 699
 700 static int
 701 make_rarp_packet(struct rte_mbuf *rarp_mbuf, const struct ether_addr *mac)
 702 {
 703         struct ether_hdr *eth_hdr;
 704         struct arp_hdr  *rarp;
 705
 706         if (rarp_mbuf->buf_len < 64) {
 707                 RTE_LOG(WARNING, VHOST_DATA,
 708                         "failed to make RARP; mbuf size too small %u (< %d)\n",
 709                         rarp_mbuf->buf_len, RARP_PKT_SIZE);
 710                 return -1;
 711         }
 712
 713         /* Ethernet header. */
 714         eth_hdr = rte_pktmbuf_mtod_offset(rarp_mbuf, struct ether_hdr *, 0);
 715         memset(eth_hdr->d_addr.addr_bytes, 0xff, ETHER_ADDR_LEN);
 716         ether_addr_copy(mac, &eth_hdr->s_addr);
 717         eth_hdr->ether_type = htons(ETHER_TYPE_RARP);
 718
 719         /* RARP header. */
 720         rarp = (struct arp_hdr *)(eth_hdr + 1);
 721         rarp->arp_hrd = htons(ARP_HRD_ETHER);
 722         rarp->arp_pro = htons(ETHER_TYPE_IPv4);
 723         rarp->arp_hln = ETHER_ADDR_LEN;
 724         rarp->arp_pln = 4;
 725         rarp->arp_op  = htons(ARP_OP_REVREQUEST);
 726
 727         ether_addr_copy(mac, &rarp->arp_data.arp_sha);
 728         ether_addr_copy(mac, &rarp->arp_data.arp_tha);
 729         memset(&rarp->arp_data.arp_sip, 0x00, 4);
 730         memset(&rarp->arp_data.arp_tip, 0x00, 4);
 731
 732         rarp_mbuf->pkt_len  = rarp_mbuf->data_len = RARP_PKT_SIZE;
 733
 734         return 0;
 735 }
 736
 737 static inline int __attribute__((always_inline))
 738 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 739                   struct rte_mbuf *m, uint16_t desc_idx,
 740                   struct rte_mempool *mbuf_pool)
 741 {
 742         struct vring_desc *desc;
 743         uint64_t desc_addr;
 744         uint32_t desc_avail, desc_offset;
 745         uint32_t mbuf_avail, mbuf_offset;
 746         uint32_t cpy_len;
 747         struct rte_mbuf *cur = m, *prev = m;
 748         struct virtio_net_hdr *hdr;
 749         /* A counter to avoid desc dead loop chain */
 750         uint32_t nr_desc = 1;
 751
 752         desc = &vq->desc[desc_idx];
 753         if (unlikely(desc->len < dev->vhost_hlen))
 754                 return -1;
 755
 756         desc_addr = gpa_to_vva(dev, desc->addr);
 757         rte_prefetch0((void *)(uintptr_t)desc_addr);
 758
 759         /* Retrieve virtio net header */
 760         hdr = (struct virtio_net_hdr *)((uintptr_t)desc_addr);
 761         desc_avail  = desc->len - dev->vhost_hlen;
 762         desc_offset = dev->vhost_hlen;
 763
 764         mbuf_offset = 0;
 765         mbuf_avail  = m->buf_len - RTE_PKTMBUF_HEADROOM;
 766         while (desc_avail != 0 || (desc->flags & VRING_DESC_F_NEXT) != 0) {
 767                 /* This desc reaches to its end, get the next one */
 768                 if (desc_avail == 0) {
 769                         if (unlikely(desc->next >= vq->size ||
 770                                      ++nr_desc >= vq->size))
 771                                 return -1;
 772                         desc = &vq->desc[desc->next];
 773
 774                         desc_addr = gpa_to_vva(dev, desc->addr);
 775                         rte_prefetch0((void *)(uintptr_t)desc_addr);
 776
 777                         desc_offset = 0;
 778                         desc_avail  = desc->len;
 779
 780                         PRINT_PACKET(dev, (uintptr_t)desc_addr, desc->len, 0);
 781                 }
 782
 783                 /*
 784                  * This mbuf reaches to its end, get a new one
 785                  * to hold more data.
 786                  */
 787                 if (mbuf_avail == 0) {
 788                         cur = rte_pktmbuf_alloc(mbuf_pool);
 789                         if (unlikely(cur == NULL)) {
 790                                 RTE_LOG(ERR, VHOST_DATA, "Failed to "
 791                                         "allocate memory for mbuf.\n");
 792                                 return -1;
 793                         }
 794
 795                         prev->next = cur;
 796                         prev->data_len = mbuf_offset;
 797                         m->nb_segs += 1;
 798                         m->pkt_len += mbuf_offset;
 799                         prev = cur;
 800
 801                         mbuf_offset = 0;
 802                         mbuf_avail  = cur->buf_len - RTE_PKTMBUF_HEADROOM;
 803                 }
 804
 805                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 806                 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *, mbuf_offset),
 807                         (void *)((uintptr_t)(desc_addr + desc_offset)),
 808                         cpy_len);
 809
 810                 mbuf_avail  -= cpy_len;
 811                 mbuf_offset += cpy_len;
 812                 desc_avail  -= cpy_len;
 813                 desc_offset += cpy_len;
 814         }
 815
 816         prev->data_len = mbuf_offset;
 817         m->pkt_len    += mbuf_offset;
 818
 819         if (hdr->flags != 0 || hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE)
 820                 vhost_dequeue_offload(hdr, m);
 821
 822         return 0;
 823 }
 824
 825 uint16_t
 826 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
 827         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
 828 {
 829         struct virtio_net *dev;
 830         struct rte_mbuf *rarp_mbuf = NULL;
 831         struct vhost_virtqueue *vq;
 832         uint32_t desc_indexes[MAX_PKT_BURST];
 833         uint32_t used_idx;
 834         uint32_t i = 0;
 835         uint16_t free_entries;
 836         uint16_t avail_idx;
 837
 838         dev = get_device(vid);
 839         if (!dev)
 840                 return 0;
 841
 842         if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->virt_qp_nb))) {
 843                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
 844                         dev->vid, __func__, queue_id);
 845                 return 0;
 846         }
 847
 848         vq = dev->virtqueue[queue_id];
 849         if (unlikely(vq->enabled == 0))
 850                 return 0;
 851
 852         /*
 853          * Construct a RARP broadcast packet, and inject it to the "pkts"
 854          * array, to looks like that guest actually send such packet.
 855          *
 856          * Check user_send_rarp() for more information.
 857          */
 858         if (unlikely(rte_atomic16_cmpset((volatile uint16_t *)
 859                                          &dev->broadcast_rarp.cnt, 1, 0))) {
 860                 rarp_mbuf = rte_pktmbuf_alloc(mbuf_pool);
 861                 if (rarp_mbuf == NULL) {
 862                         RTE_LOG(ERR, VHOST_DATA,
 863                                 "Failed to allocate memory for mbuf.\n");
 864                         return 0;
 865                 }
 866
 867                 if (make_rarp_packet(rarp_mbuf, &dev->mac)) {
 868                         rte_pktmbuf_free(rarp_mbuf);
 869                         rarp_mbuf = NULL;
 870                 } else {
 871                         count -= 1;
 872                 }
 873         }
 874
 875         avail_idx =  *((volatile uint16_t *)&vq->avail->idx);
 876         free_entries = avail_idx - vq->last_used_idx;
 877         if (free_entries == 0)
 878                 goto out;
 879
 880         LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
 881
 882         /* Prefetch available ring to retrieve head indexes. */
 883         used_idx = vq->last_used_idx & (vq->size - 1);
 884         rte_prefetch0(&vq->avail->ring[used_idx]);
 885
 886         count = RTE_MIN(count, MAX_PKT_BURST);
 887         count = RTE_MIN(count, free_entries);
 888         LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
 889                         dev->vid, count);
 890
 891         /* Retrieve all of the head indexes first to avoid caching issues. */
 892         for (i = 0; i < count; i++) {
 893                 desc_indexes[i] = vq->avail->ring[(vq->last_used_idx + i) &
 894                                         (vq->size - 1)];
 895         }
 896
 897         /* Prefetch descriptor index. */
 898         rte_prefetch0(&vq->desc[desc_indexes[0]]);
 899         rte_prefetch0(&vq->used->ring[vq->last_used_idx & (vq->size - 1)]);
 900
 901         for (i = 0; i < count; i++) {
 902                 int err;
 903
 904                 if (likely(i + 1 < count)) {
 905                         rte_prefetch0(&vq->desc[desc_indexes[i + 1]]);
 906                         rte_prefetch0(&vq->used->ring[(used_idx + 1) &
 907                                                       (vq->size - 1)]);
 908                 }
 909
 910                 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
 911                 if (unlikely(pkts[i] == NULL)) {
 912                         RTE_LOG(ERR, VHOST_DATA,
 913                                 "Failed to allocate memory for mbuf.\n");
 914                         break;
 915                 }
 916                 err = copy_desc_to_mbuf(dev, vq, pkts[i], desc_indexes[i],
 917                                         mbuf_pool);
 918                 if (unlikely(err)) {
 919                         rte_pktmbuf_free(pkts[i]);
 920                         break;
 921                 }
 922
 923                 used_idx = vq->last_used_idx++ & (vq->size - 1);
 924                 vq->used->ring[used_idx].id  = desc_indexes[i];
 925                 vq->used->ring[used_idx].len = 0;
 926                 vhost_log_used_vring(dev, vq,
 927                                 offsetof(struct vring_used, ring[used_idx]),
 928                                 sizeof(vq->used->ring[used_idx]));
 929         }
 930
 931         rte_smp_wmb();
 932         rte_smp_rmb();
 933         vq->used->idx += i;
 934         vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
 935                         sizeof(vq->used->idx));
 936
 937         /* Kick guest if required. */
 938         if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
 939                         && (vq->callfd >= 0))
 940                 eventfd_write(vq->callfd, (eventfd_t)1);
 941
 942 out:
 943         if (unlikely(rarp_mbuf != NULL)) {
 944                 /*
 945                  * Inject it to the head of "pkts" array, so that switch's mac
 946                  * learning table will get updated first.
 947                  */
 948                 memmove(&pkts[1], pkts, i * sizeof(struct rte_mbuf *));
 949                 pkts[0] = rarp_mbuf;
 950                 i += 1;
 951         }
 952
 953         return i;
 954 }