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 vhost_virtqueue *vq, uint64_t desc_addr,
 130                     struct virtio_net_hdr_mrg_rxbuf hdr)
 131 {
 132         if (vq->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 < vq->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(vq, desc_addr, virtio_hdr);
 158         vhost_log_write(dev, desc->addr, vq->vhost_hlen);
 159         PRINT_PACKET(dev, (uintptr_t)desc_addr, vq->vhost_hlen, 0);
 160
 161         desc_offset = vq->vhost_hlen;
 162         desc_avail  = desc->len - vq->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 = vq->vhost_hlen;
 304                 else
 305                         vq->used->ring[used_idx].len = copied + vq->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 {
 340         uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
 341         uint32_t vec_id = *vec_idx;
 342         uint32_t len    = *allocated;
 343
 344         while (1) {
 345                 if (unlikely(vec_id >= BUF_VECTOR_MAX || idx >= vq->size))
 346                         return -1;
 347
 348                 len += vq->desc[idx].len;
 349                 vq->buf_vec[vec_id].buf_addr = vq->desc[idx].addr;
 350                 vq->buf_vec[vec_id].buf_len  = vq->desc[idx].len;
 351                 vq->buf_vec[vec_id].desc_idx = idx;
 352                 vec_id++;
 353
 354                 if ((vq->desc[idx].flags & VRING_DESC_F_NEXT) == 0)
 355                         break;
 356
 357                 idx = vq->desc[idx].next;
 358         }
 359
 360         *allocated = len;
 361         *vec_idx   = vec_id;
 362
 363         return 0;
 364 }
 365
 366 /*
 367  * As many data cores may want to access available buffers concurrently,
 368  * they need to be reserved.
 369  *
 370  * Returns -1 on fail, 0 on success
 371  */
 372 static inline int
 373 reserve_avail_buf_mergeable(struct vhost_virtqueue *vq, uint32_t size,
 374                             uint16_t *start, uint16_t *end)
 375 {
 376         uint16_t res_start_idx;
 377         uint16_t res_cur_idx;
 378         uint16_t avail_idx;
 379         uint32_t allocated;
 380         uint32_t vec_idx;
 381         uint16_t tries;
 382
 383 again:
 384         res_start_idx = vq->last_used_idx_res;
 385         res_cur_idx  = res_start_idx;
 386
 387         allocated = 0;
 388         vec_idx   = 0;
 389         tries     = 0;
 390         while (1) {
 391                 avail_idx = *((volatile uint16_t *)&vq->avail->idx);
 392                 if (unlikely(res_cur_idx == avail_idx))
 393                         return -1;
 394
 395                 if (unlikely(fill_vec_buf(vq, res_cur_idx, &allocated,
 396                                           &vec_idx) < 0))
 397                         return -1;
 398
 399                 res_cur_idx++;
 400                 tries++;
 401
 402                 if (allocated >= size)
 403                         break;
 404
 405                 /*
 406                  * if we tried all available ring items, and still
 407                  * can't get enough buf, it means something abnormal
 408                  * happened.
 409                  */
 410                 if (unlikely(tries >= vq->size))
 411                         return -1;
 412         }
 413
 414         /*
 415          * update vq->last_used_idx_res atomically.
 416          * retry again if failed.
 417          */
 418         if (rte_atomic16_cmpset(&vq->last_used_idx_res,
 419                                 res_start_idx, res_cur_idx) == 0)
 420                 goto again;
 421
 422         *start = res_start_idx;
 423         *end   = res_cur_idx;
 424         return 0;
 425 }
 426
 427 static inline uint32_t __attribute__((always_inline))
 428 copy_mbuf_to_desc_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
 429                             uint16_t res_start_idx, uint16_t res_end_idx,
 430                             struct rte_mbuf *m)
 431 {
 432         struct virtio_net_hdr_mrg_rxbuf virtio_hdr = {{0, 0, 0, 0, 0, 0}, 0};
 433         uint32_t vec_idx = 0;
 434         uint16_t cur_idx = res_start_idx;
 435         uint64_t desc_addr;
 436         uint32_t mbuf_offset, mbuf_avail;
 437         uint32_t desc_offset, desc_avail;
 438         uint32_t cpy_len;
 439         uint16_t desc_idx, used_idx;
 440
 441         if (unlikely(m == NULL))
 442                 return 0;
 443
 444         LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
 445                 dev->vid, cur_idx, res_end_idx);
 446
 447         if (vq->buf_vec[vec_idx].buf_len < vq->vhost_hlen)
 448                 return -1;
 449
 450         desc_addr = gpa_to_vva(dev, vq->buf_vec[vec_idx].buf_addr);
 451         rte_prefetch0((void *)(uintptr_t)desc_addr);
 452
 453         virtio_hdr.num_buffers = res_end_idx - res_start_idx;
 454         LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
 455                 dev->vid, virtio_hdr.num_buffers);
 456
 457         virtio_enqueue_offload(m, &virtio_hdr.hdr);
 458         copy_virtio_net_hdr(vq, desc_addr, virtio_hdr);
 459         vhost_log_write(dev, vq->buf_vec[vec_idx].buf_addr, vq->vhost_hlen);
 460         PRINT_PACKET(dev, (uintptr_t)desc_addr, vq->vhost_hlen, 0);
 461
 462         desc_avail  = vq->buf_vec[vec_idx].buf_len - vq->vhost_hlen;
 463         desc_offset = vq->vhost_hlen;
 464
 465         mbuf_avail  = rte_pktmbuf_data_len(m);
 466         mbuf_offset = 0;
 467         while (mbuf_avail != 0 || m->next != NULL) {
 468                 /* done with current desc buf, get the next one */
 469                 if (desc_avail == 0) {
 470                         desc_idx = vq->buf_vec[vec_idx].desc_idx;
 471
 472                         if (!(vq->desc[desc_idx].flags & VRING_DESC_F_NEXT)) {
 473                                 /* Update used ring with desc information */
 474                                 used_idx = cur_idx++ & (vq->size - 1);
 475                                 vq->used->ring[used_idx].id  = desc_idx;
 476                                 vq->used->ring[used_idx].len = desc_offset;
 477                                 vhost_log_used_vring(dev, vq,
 478                                         offsetof(struct vring_used,
 479                                                  ring[used_idx]),
 480                                         sizeof(vq->used->ring[used_idx]));
 481                         }
 482
 483                         vec_idx++;
 484                         desc_addr = gpa_to_vva(dev, vq->buf_vec[vec_idx].buf_addr);
 485
 486                         /* Prefetch buffer address. */
 487                         rte_prefetch0((void *)(uintptr_t)desc_addr);
 488                         desc_offset = 0;
 489                         desc_avail  = vq->buf_vec[vec_idx].buf_len;
 490                 }
 491
 492                 /* done with current mbuf, get the next one */
 493                 if (mbuf_avail == 0) {
 494                         m = m->next;
 495
 496                         mbuf_offset = 0;
 497                         mbuf_avail  = rte_pktmbuf_data_len(m);
 498                 }
 499
 500                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 501                 rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)),
 502                         rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 503                         cpy_len);
 504                 vhost_log_write(dev, vq->buf_vec[vec_idx].buf_addr + desc_offset,
 505                         cpy_len);
 506                 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
 507                         cpy_len, 0);
 508
 509                 mbuf_avail  -= cpy_len;
 510                 mbuf_offset += cpy_len;
 511                 desc_avail  -= cpy_len;
 512                 desc_offset += cpy_len;
 513         }
 514
 515         used_idx = cur_idx & (vq->size - 1);
 516         vq->used->ring[used_idx].id = vq->buf_vec[vec_idx].desc_idx;
 517         vq->used->ring[used_idx].len = desc_offset;
 518         vhost_log_used_vring(dev, vq,
 519                 offsetof(struct vring_used, ring[used_idx]),
 520                 sizeof(vq->used->ring[used_idx]));
 521
 522         return res_end_idx - res_start_idx;
 523 }
 524
 525 static inline uint32_t __attribute__((always_inline))
 526 virtio_dev_merge_rx(struct virtio_net *dev, uint16_t queue_id,
 527         struct rte_mbuf **pkts, uint32_t count)
 528 {
 529         struct vhost_virtqueue *vq;
 530         uint32_t pkt_idx = 0, nr_used = 0;
 531         uint16_t start, end;
 532
 533         LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
 534         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->virt_qp_nb))) {
 535                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
 536                         dev->vid, __func__, queue_id);
 537                 return 0;
 538         }
 539
 540         vq = dev->virtqueue[queue_id];
 541         if (unlikely(vq->enabled == 0))
 542                 return 0;
 543
 544         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
 545         if (count == 0)
 546                 return 0;
 547
 548         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
 549                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + vq->vhost_hlen;
 550
 551                 if (unlikely(reserve_avail_buf_mergeable(vq, pkt_len,
 552                                                          &start, &end) < 0)) {
 553                         LOG_DEBUG(VHOST_DATA,
 554                                 "(%d) failed to get enough desc from vring\n",
 555                                 dev->vid);
 556                         break;
 557                 }
 558
 559                 nr_used = copy_mbuf_to_desc_mergeable(dev, vq, start, end,
 560                                                       pkts[pkt_idx]);
 561                 rte_smp_wmb();
 562
 563                 /*
 564                  * Wait until it's our turn to add our buffer
 565                  * to the used ring.
 566                  */
 567                 while (unlikely(vq->last_used_idx != start))
 568                         rte_pause();
 569
 570                 *(volatile uint16_t *)&vq->used->idx += nr_used;
 571                 vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
 572                         sizeof(vq->used->idx));
 573                 vq->last_used_idx = end;
 574         }
 575
 576         if (likely(pkt_idx)) {
 577                 /* flush used->idx update before we read avail->flags. */
 578                 rte_mb();
 579
 580                 /* Kick the guest if necessary. */
 581                 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
 582                                 && (vq->callfd >= 0))
 583                         eventfd_write(vq->callfd, (eventfd_t)1);
 584         }
 585
 586         return pkt_idx;
 587 }
 588
 589 uint16_t
 590 rte_vhost_enqueue_burst(struct virtio_net *dev, uint16_t queue_id,
 591         struct rte_mbuf **pkts, uint16_t count)
 592 {
 593         if (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF))
 594                 return virtio_dev_merge_rx(dev, queue_id, pkts, count);
 595         else
 596                 return virtio_dev_rx(dev, queue_id, pkts, count);
 597 }
 598
 599 static void
 600 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
 601 {
 602         struct ipv4_hdr *ipv4_hdr;
 603         struct ipv6_hdr *ipv6_hdr;
 604         void *l3_hdr = NULL;
 605         struct ether_hdr *eth_hdr;
 606         uint16_t ethertype;
 607
 608         eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
 609
 610         m->l2_len = sizeof(struct ether_hdr);
 611         ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
 612
 613         if (ethertype == ETHER_TYPE_VLAN) {
 614                 struct vlan_hdr *vlan_hdr = (struct vlan_hdr *)(eth_hdr + 1);
 615
 616                 m->l2_len += sizeof(struct vlan_hdr);
 617                 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
 618         }
 619
 620         l3_hdr = (char *)eth_hdr + m->l2_len;
 621
 622         switch (ethertype) {
 623         case ETHER_TYPE_IPv4:
 624                 ipv4_hdr = (struct ipv4_hdr *)l3_hdr;
 625                 *l4_proto = ipv4_hdr->next_proto_id;
 626                 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
 627                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 628                 m->ol_flags |= PKT_TX_IPV4;
 629                 break;
 630         case ETHER_TYPE_IPv6:
 631                 ipv6_hdr = (struct ipv6_hdr *)l3_hdr;
 632                 *l4_proto = ipv6_hdr->proto;
 633                 m->l3_len = sizeof(struct ipv6_hdr);
 634                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 635                 m->ol_flags |= PKT_TX_IPV6;
 636                 break;
 637         default:
 638                 m->l3_len = 0;
 639                 *l4_proto = 0;
 640                 break;
 641         }
 642 }
 643
 644 static inline void __attribute__((always_inline))
 645 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
 646 {
 647         uint16_t l4_proto = 0;
 648         void *l4_hdr = NULL;
 649         struct tcp_hdr *tcp_hdr = NULL;
 650
 651         parse_ethernet(m, &l4_proto, &l4_hdr);
 652         if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 653                 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
 654                         switch (hdr->csum_offset) {
 655                         case (offsetof(struct tcp_hdr, cksum)):
 656                                 if (l4_proto == IPPROTO_TCP)
 657                                         m->ol_flags |= PKT_TX_TCP_CKSUM;
 658                                 break;
 659                         case (offsetof(struct udp_hdr, dgram_cksum)):
 660                                 if (l4_proto == IPPROTO_UDP)
 661                                         m->ol_flags |= PKT_TX_UDP_CKSUM;
 662                                 break;
 663                         case (offsetof(struct sctp_hdr, cksum)):
 664                                 if (l4_proto == IPPROTO_SCTP)
 665                                         m->ol_flags |= PKT_TX_SCTP_CKSUM;
 666                                 break;
 667                         default:
 668                                 break;
 669                         }
 670                 }
 671         }
 672
 673         if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 674                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 675                 case VIRTIO_NET_HDR_GSO_TCPV4:
 676                 case VIRTIO_NET_HDR_GSO_TCPV6:
 677                         tcp_hdr = (struct tcp_hdr *)l4_hdr;
 678                         m->ol_flags |= PKT_TX_TCP_SEG;
 679                         m->tso_segsz = hdr->gso_size;
 680                         m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
 681                         break;
 682                 default:
 683                         RTE_LOG(WARNING, VHOST_DATA,
 684                                 "unsupported gso type %u.\n", hdr->gso_type);
 685                         break;
 686                 }
 687         }
 688 }
 689
 690 #define RARP_PKT_SIZE   64
 691
 692 static int
 693 make_rarp_packet(struct rte_mbuf *rarp_mbuf, const struct ether_addr *mac)
 694 {
 695         struct ether_hdr *eth_hdr;
 696         struct arp_hdr  *rarp;
 697
 698         if (rarp_mbuf->buf_len < 64) {
 699                 RTE_LOG(WARNING, VHOST_DATA,
 700                         "failed to make RARP; mbuf size too small %u (< %d)\n",
 701                         rarp_mbuf->buf_len, RARP_PKT_SIZE);
 702                 return -1;
 703         }
 704
 705         /* Ethernet header. */
 706         eth_hdr = rte_pktmbuf_mtod_offset(rarp_mbuf, struct ether_hdr *, 0);
 707         memset(eth_hdr->d_addr.addr_bytes, 0xff, ETHER_ADDR_LEN);
 708         ether_addr_copy(mac, &eth_hdr->s_addr);
 709         eth_hdr->ether_type = htons(ETHER_TYPE_RARP);
 710
 711         /* RARP header. */
 712         rarp = (struct arp_hdr *)(eth_hdr + 1);
 713         rarp->arp_hrd = htons(ARP_HRD_ETHER);
 714         rarp->arp_pro = htons(ETHER_TYPE_IPv4);
 715         rarp->arp_hln = ETHER_ADDR_LEN;
 716         rarp->arp_pln = 4;
 717         rarp->arp_op  = htons(ARP_OP_REVREQUEST);
 718
 719         ether_addr_copy(mac, &rarp->arp_data.arp_sha);
 720         ether_addr_copy(mac, &rarp->arp_data.arp_tha);
 721         memset(&rarp->arp_data.arp_sip, 0x00, 4);
 722         memset(&rarp->arp_data.arp_tip, 0x00, 4);
 723
 724         rarp_mbuf->pkt_len  = rarp_mbuf->data_len = RARP_PKT_SIZE;
 725
 726         return 0;
 727 }
 728
 729 static inline int __attribute__((always_inline))
 730 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 731                   struct rte_mbuf *m, uint16_t desc_idx,
 732                   struct rte_mempool *mbuf_pool)
 733 {
 734         struct vring_desc *desc;
 735         uint64_t desc_addr;
 736         uint32_t desc_avail, desc_offset;
 737         uint32_t mbuf_avail, mbuf_offset;
 738         uint32_t cpy_len;
 739         struct rte_mbuf *cur = m, *prev = m;
 740         struct virtio_net_hdr *hdr;
 741         /* A counter to avoid desc dead loop chain */
 742         uint32_t nr_desc = 1;
 743
 744         desc = &vq->desc[desc_idx];
 745         if (unlikely(desc->len < vq->vhost_hlen))
 746                 return -1;
 747
 748         desc_addr = gpa_to_vva(dev, desc->addr);
 749         rte_prefetch0((void *)(uintptr_t)desc_addr);
 750
 751         /* Retrieve virtio net header */
 752         hdr = (struct virtio_net_hdr *)((uintptr_t)desc_addr);
 753         desc_avail  = desc->len - vq->vhost_hlen;
 754         desc_offset = vq->vhost_hlen;
 755
 756         mbuf_offset = 0;
 757         mbuf_avail  = m->buf_len - RTE_PKTMBUF_HEADROOM;
 758         while (desc_avail != 0 || (desc->flags & VRING_DESC_F_NEXT) != 0) {
 759                 /* This desc reaches to its end, get the next one */
 760                 if (desc_avail == 0) {
 761                         if (unlikely(desc->next >= vq->size ||
 762                                      ++nr_desc >= vq->size))
 763                                 return -1;
 764                         desc = &vq->desc[desc->next];
 765
 766                         desc_addr = gpa_to_vva(dev, desc->addr);
 767                         rte_prefetch0((void *)(uintptr_t)desc_addr);
 768
 769                         desc_offset = 0;
 770                         desc_avail  = desc->len;
 771
 772                         PRINT_PACKET(dev, (uintptr_t)desc_addr, desc->len, 0);
 773                 }
 774
 775                 /*
 776                  * This mbuf reaches to its end, get a new one
 777                  * to hold more data.
 778                  */
 779                 if (mbuf_avail == 0) {
 780                         cur = rte_pktmbuf_alloc(mbuf_pool);
 781                         if (unlikely(cur == NULL)) {
 782                                 RTE_LOG(ERR, VHOST_DATA, "Failed to "
 783                                         "allocate memory for mbuf.\n");
 784                                 return -1;
 785                         }
 786
 787                         prev->next = cur;
 788                         prev->data_len = mbuf_offset;
 789                         m->nb_segs += 1;
 790                         m->pkt_len += mbuf_offset;
 791                         prev = cur;
 792
 793                         mbuf_offset = 0;
 794                         mbuf_avail  = cur->buf_len - RTE_PKTMBUF_HEADROOM;
 795                 }
 796
 797                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 798                 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *, mbuf_offset),
 799                         (void *)((uintptr_t)(desc_addr + desc_offset)),
 800                         cpy_len);
 801
 802                 mbuf_avail  -= cpy_len;
 803                 mbuf_offset += cpy_len;
 804                 desc_avail  -= cpy_len;
 805                 desc_offset += cpy_len;
 806         }
 807
 808         prev->data_len = mbuf_offset;
 809         m->pkt_len    += mbuf_offset;
 810
 811         if (hdr->flags != 0 || hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE)
 812                 vhost_dequeue_offload(hdr, m);
 813
 814         return 0;
 815 }
 816
 817 uint16_t
 818 rte_vhost_dequeue_burst(struct virtio_net *dev, uint16_t queue_id,
 819         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
 820 {
 821         struct rte_mbuf *rarp_mbuf = NULL;
 822         struct vhost_virtqueue *vq;
 823         uint32_t desc_indexes[MAX_PKT_BURST];
 824         uint32_t used_idx;
 825         uint32_t i = 0;
 826         uint16_t free_entries;
 827         uint16_t avail_idx;
 828
 829         if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->virt_qp_nb))) {
 830                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
 831                         dev->vid, __func__, queue_id);
 832                 return 0;
 833         }
 834
 835         vq = dev->virtqueue[queue_id];
 836         if (unlikely(vq->enabled == 0))
 837                 return 0;
 838
 839         /*
 840          * Construct a RARP broadcast packet, and inject it to the "pkts"
 841          * array, to looks like that guest actually send such packet.
 842          *
 843          * Check user_send_rarp() for more information.
 844          */
 845         if (unlikely(rte_atomic16_cmpset((volatile uint16_t *)
 846                                          &dev->broadcast_rarp.cnt, 1, 0))) {
 847                 rarp_mbuf = rte_pktmbuf_alloc(mbuf_pool);
 848                 if (rarp_mbuf == NULL) {
 849                         RTE_LOG(ERR, VHOST_DATA,
 850                                 "Failed to allocate memory for mbuf.\n");
 851                         return 0;
 852                 }
 853
 854                 if (make_rarp_packet(rarp_mbuf, &dev->mac)) {
 855                         rte_pktmbuf_free(rarp_mbuf);
 856                         rarp_mbuf = NULL;
 857                 } else {
 858                         count -= 1;
 859                 }
 860         }
 861
 862         avail_idx =  *((volatile uint16_t *)&vq->avail->idx);
 863         free_entries = avail_idx - vq->last_used_idx;
 864         if (free_entries == 0)
 865                 goto out;
 866
 867         LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
 868
 869         /* Prefetch available ring to retrieve head indexes. */
 870         used_idx = vq->last_used_idx & (vq->size - 1);
 871         rte_prefetch0(&vq->avail->ring[used_idx]);
 872
 873         count = RTE_MIN(count, MAX_PKT_BURST);
 874         count = RTE_MIN(count, free_entries);
 875         LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
 876                         dev->vid, count);
 877
 878         /* Retrieve all of the head indexes first to avoid caching issues. */
 879         for (i = 0; i < count; i++) {
 880                 desc_indexes[i] = vq->avail->ring[(vq->last_used_idx + i) &
 881                                         (vq->size - 1)];
 882         }
 883
 884         /* Prefetch descriptor index. */
 885         rte_prefetch0(&vq->desc[desc_indexes[0]]);
 886         rte_prefetch0(&vq->used->ring[vq->last_used_idx & (vq->size - 1)]);
 887
 888         for (i = 0; i < count; i++) {
 889                 int err;
 890
 891                 if (likely(i + 1 < count)) {
 892                         rte_prefetch0(&vq->desc[desc_indexes[i + 1]]);
 893                         rte_prefetch0(&vq->used->ring[(used_idx + 1) &
 894                                                       (vq->size - 1)]);
 895                 }
 896
 897                 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
 898                 if (unlikely(pkts[i] == NULL)) {
 899                         RTE_LOG(ERR, VHOST_DATA,
 900                                 "Failed to allocate memory for mbuf.\n");
 901                         break;
 902                 }
 903                 err = copy_desc_to_mbuf(dev, vq, pkts[i], desc_indexes[i],
 904                                         mbuf_pool);
 905                 if (unlikely(err)) {
 906                         rte_pktmbuf_free(pkts[i]);
 907                         break;
 908                 }
 909
 910                 used_idx = vq->last_used_idx++ & (vq->size - 1);
 911                 vq->used->ring[used_idx].id  = desc_indexes[i];
 912                 vq->used->ring[used_idx].len = 0;
 913                 vhost_log_used_vring(dev, vq,
 914                                 offsetof(struct vring_used, ring[used_idx]),
 915                                 sizeof(vq->used->ring[used_idx]));
 916         }
 917
 918         rte_smp_wmb();
 919         rte_smp_rmb();
 920         vq->used->idx += i;
 921         vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
 922                         sizeof(vq->used->idx));
 923
 924         /* Kick guest if required. */
 925         if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
 926                         && (vq->callfd >= 0))
 927                 eventfd_write(vq->callfd, (eventfd_t)1);
 928
 929 out:
 930         if (unlikely(rarp_mbuf != NULL)) {
 931                 /*
 932                  * Inject it to the head of "pkts" array, so that switch's mac
 933                  * learning table will get updated first.
 934                  */
 935                 memmove(&pkts[1], pkts, i * sizeof(struct rte_mbuf *));
 936                 pkts[0] = rarp_mbuf;
 937                 i += 1;
 938         }
 939
 940         return i;
 941 }