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         memset(net_hdr, 0, sizeof(struct virtio_net_hdr));
  98
  99         if (m_buf->ol_flags & PKT_TX_L4_MASK) {
 100                 net_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 101                 net_hdr->csum_start = m_buf->l2_len + m_buf->l3_len;
 102
 103                 switch (m_buf->ol_flags & PKT_TX_L4_MASK) {
 104                 case PKT_TX_TCP_CKSUM:
 105                         net_hdr->csum_offset = (offsetof(struct tcp_hdr,
 106                                                 cksum));
 107                         break;
 108                 case PKT_TX_UDP_CKSUM:
 109                         net_hdr->csum_offset = (offsetof(struct udp_hdr,
 110                                                 dgram_cksum));
 111                         break;
 112                 case PKT_TX_SCTP_CKSUM:
 113                         net_hdr->csum_offset = (offsetof(struct sctp_hdr,
 114                                                 cksum));
 115                         break;
 116                 }
 117         }
 118
 119         if (m_buf->ol_flags & PKT_TX_TCP_SEG) {
 120                 if (m_buf->ol_flags & PKT_TX_IPV4)
 121                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 122                 else
 123                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 124                 net_hdr->gso_size = m_buf->tso_segsz;
 125                 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len
 126                                         + m_buf->l4_len;
 127         }
 128
 129         return;
 130 }
 131
 132 static inline void
 133 copy_virtio_net_hdr(struct vhost_virtqueue *vq, uint64_t desc_addr,
 134                     struct virtio_net_hdr_mrg_rxbuf hdr)
 135 {
 136         if (vq->vhost_hlen == sizeof(struct virtio_net_hdr_mrg_rxbuf))
 137                 *(struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)desc_addr = hdr;
 138         else
 139                 *(struct virtio_net_hdr *)(uintptr_t)desc_addr = hdr.hdr;
 140 }
 141
 142 static inline int __attribute__((always_inline))
 143 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 144                   struct rte_mbuf *m, uint16_t desc_idx, uint32_t *copied)
 145 {
 146         uint32_t desc_avail, desc_offset;
 147         uint32_t mbuf_avail, mbuf_offset;
 148         uint32_t cpy_len;
 149         struct vring_desc *desc;
 150         uint64_t desc_addr;
 151         struct virtio_net_hdr_mrg_rxbuf virtio_hdr = {{0, 0, 0, 0, 0, 0}, 0};
 152
 153         desc = &vq->desc[desc_idx];
 154         desc_addr = gpa_to_vva(dev, desc->addr);
 155         rte_prefetch0((void *)(uintptr_t)desc_addr);
 156
 157         virtio_enqueue_offload(m, &virtio_hdr.hdr);
 158         copy_virtio_net_hdr(vq, desc_addr, virtio_hdr);
 159         vhost_log_write(dev, desc->addr, vq->vhost_hlen);
 160         PRINT_PACKET(dev, (uintptr_t)desc_addr, vq->vhost_hlen, 0);
 161
 162         desc_offset = vq->vhost_hlen;
 163         desc_avail  = desc->len - vq->vhost_hlen;
 164
 165         *copied = rte_pktmbuf_pkt_len(m);
 166         mbuf_avail  = rte_pktmbuf_data_len(m);
 167         mbuf_offset = 0;
 168         while (mbuf_avail != 0 || m->next != NULL) {
 169                 /* done with current mbuf, fetch next */
 170                 if (mbuf_avail == 0) {
 171                         m = m->next;
 172
 173                         mbuf_offset = 0;
 174                         mbuf_avail  = rte_pktmbuf_data_len(m);
 175                 }
 176
 177                 /* done with current desc buf, fetch next */
 178                 if (desc_avail == 0) {
 179                         if ((desc->flags & VRING_DESC_F_NEXT) == 0) {
 180                                 /* Room in vring buffer is not enough */
 181                                 return -1;
 182                         }
 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         uint32_t i;
 265
 266         LOG_DEBUG(VHOST_DATA, "(%"PRIu64") virtio_dev_rx()\n", dev->device_fh);
 267         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->virt_qp_nb))) {
 268                 RTE_LOG(ERR, VHOST_DATA,
 269                         "%s (%"PRIu64"): virtqueue idx:%d invalid.\n",
 270                         __func__, dev->device_fh, 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,
 283                 "(%"PRIu64") res_start_idx %d| res_end_idx Index %d\n",
 284                 dev->device_fh, res_start_idx, res_end_idx);
 285
 286         /* Retrieve all of the desc indexes first to avoid caching issues. */
 287         rte_prefetch0(&vq->avail->ring[res_start_idx & (vq->size - 1)]);
 288         for (i = 0; i < count; i++) {
 289                 desc_indexes[i] = vq->avail->ring[(res_start_idx + i) &
 290                                                   (vq->size - 1)];
 291         }
 292
 293         rte_prefetch0(&vq->desc[desc_indexes[0]]);
 294         for (i = 0; i < count; i++) {
 295                 uint16_t desc_idx = desc_indexes[i];
 296                 uint16_t used_idx = (res_start_idx + i) & (vq->size - 1);
 297                 uint32_t copied;
 298                 int err;
 299
 300                 err = copy_mbuf_to_desc(dev, vq, pkts[i], desc_idx, &copied);
 301
 302                 vq->used->ring[used_idx].id = desc_idx;
 303                 if (unlikely(err))
 304                         vq->used->ring[used_idx].len = vq->vhost_hlen;
 305                 else
 306                         vq->used->ring[used_idx].len = copied + vq->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                 if (i + 1 < count)
 312                         rte_prefetch0(&vq->desc[desc_indexes[i+1]]);
 313         }
 314
 315         rte_compiler_barrier();
 316
 317         /* Wait until it's our turn to add our buffer to the used ring. */
 318         while (unlikely(vq->last_used_idx != res_start_idx))
 319                 rte_pause();
 320
 321         *(volatile uint16_t *)&vq->used->idx += count;
 322         vq->last_used_idx = res_end_idx;
 323         vhost_log_used_vring(dev, vq,
 324                 offsetof(struct vring_used, idx),
 325                 sizeof(vq->used->idx));
 326
 327         /* flush used->idx update before we read avail->flags. */
 328         rte_mb();
 329
 330         /* Kick the guest if necessary. */
 331         if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT))
 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 (vec_id >= BUF_VECTOR_MAX)
 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,
 445                 "(%"PRIu64") Current Index %d| End Index %d\n",
 446                 dev->device_fh, cur_idx, res_end_idx);
 447
 448         desc_addr = gpa_to_vva(dev, vq->buf_vec[vec_idx].buf_addr);
 449         rte_prefetch0((void *)(uintptr_t)desc_addr);
 450
 451         virtio_hdr.num_buffers = res_end_idx - res_start_idx;
 452         LOG_DEBUG(VHOST_DATA, "(%"PRIu64") RX: Num merge buffers %d\n",
 453                 dev->device_fh, virtio_hdr.num_buffers);
 454
 455         virtio_enqueue_offload(m, &virtio_hdr.hdr);
 456         copy_virtio_net_hdr(vq, desc_addr, virtio_hdr);
 457         vhost_log_write(dev, vq->buf_vec[vec_idx].buf_addr, vq->vhost_hlen);
 458         PRINT_PACKET(dev, (uintptr_t)desc_addr, vq->vhost_hlen, 0);
 459
 460         desc_avail  = vq->buf_vec[vec_idx].buf_len - vq->vhost_hlen;
 461         desc_offset = vq->vhost_hlen;
 462
 463         mbuf_avail  = rte_pktmbuf_data_len(m);
 464         mbuf_offset = 0;
 465         while (mbuf_avail != 0 || m->next != NULL) {
 466                 /* done with current desc buf, get the next one */
 467                 if (desc_avail == 0) {
 468                         desc_idx = vq->buf_vec[vec_idx].desc_idx;
 469
 470                         if (!(vq->desc[desc_idx].flags & VRING_DESC_F_NEXT)) {
 471                                 /* Update used ring with desc information */
 472                                 used_idx = cur_idx++ & (vq->size - 1);
 473                                 vq->used->ring[used_idx].id  = desc_idx;
 474                                 vq->used->ring[used_idx].len = desc_offset;
 475                                 vhost_log_used_vring(dev, vq,
 476                                         offsetof(struct vring_used,
 477                                                  ring[used_idx]),
 478                                         sizeof(vq->used->ring[used_idx]));
 479                         }
 480
 481                         vec_idx++;
 482                         desc_addr = gpa_to_vva(dev, vq->buf_vec[vec_idx].buf_addr);
 483
 484                         /* Prefetch buffer address. */
 485                         rte_prefetch0((void *)(uintptr_t)desc_addr);
 486                         desc_offset = 0;
 487                         desc_avail  = vq->buf_vec[vec_idx].buf_len;
 488                 }
 489
 490                 /* done with current mbuf, get the next one */
 491                 if (mbuf_avail == 0) {
 492                         m = m->next;
 493
 494                         mbuf_offset = 0;
 495                         mbuf_avail  = rte_pktmbuf_data_len(m);
 496                 }
 497
 498                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 499                 rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)),
 500                         rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 501                         cpy_len);
 502                 vhost_log_write(dev, vq->buf_vec[vec_idx].buf_addr + desc_offset,
 503                         cpy_len);
 504                 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
 505                         cpy_len, 0);
 506
 507                 mbuf_avail  -= cpy_len;
 508                 mbuf_offset += cpy_len;
 509                 desc_avail  -= cpy_len;
 510                 desc_offset += cpy_len;
 511         }
 512
 513         used_idx = cur_idx & (vq->size - 1);
 514         vq->used->ring[used_idx].id = vq->buf_vec[vec_idx].desc_idx;
 515         vq->used->ring[used_idx].len = desc_offset;
 516         vhost_log_used_vring(dev, vq,
 517                 offsetof(struct vring_used, ring[used_idx]),
 518                 sizeof(vq->used->ring[used_idx]));
 519
 520         return res_end_idx - res_start_idx;
 521 }
 522
 523 static inline uint32_t __attribute__((always_inline))
 524 virtio_dev_merge_rx(struct virtio_net *dev, uint16_t queue_id,
 525         struct rte_mbuf **pkts, uint32_t count)
 526 {
 527         struct vhost_virtqueue *vq;
 528         uint32_t pkt_idx = 0, nr_used = 0;
 529         uint16_t start, end;
 530
 531         LOG_DEBUG(VHOST_DATA, "(%"PRIu64") virtio_dev_merge_rx()\n",
 532                 dev->device_fh);
 533         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->virt_qp_nb))) {
 534                 RTE_LOG(ERR, VHOST_DATA,
 535                         "%s (%"PRIu64"): virtqueue idx:%d invalid.\n",
 536                         __func__, dev->device_fh, 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                                 "(%" PRIu64 ") Failed to get enough desc from vring\n",
 555                                 dev->device_fh);
 556                         break;
 557                 }
 558
 559                 nr_used = copy_mbuf_to_desc_mergeable(dev, vq, start, end,
 560                                                       pkts[pkt_idx]);
 561                 rte_compiler_barrier();
 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                         eventfd_write(vq->callfd, (eventfd_t)1);
 583         }
 584
 585         return pkt_idx;
 586 }
 587
 588 uint16_t
 589 rte_vhost_enqueue_burst(struct virtio_net *dev, uint16_t queue_id,
 590         struct rte_mbuf **pkts, uint16_t count)
 591 {
 592         if (unlikely(dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF)))
 593                 return virtio_dev_merge_rx(dev, queue_id, pkts, count);
 594         else
 595                 return virtio_dev_rx(dev, queue_id, pkts, count);
 596 }
 597
 598 static void
 599 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
 600 {
 601         struct ipv4_hdr *ipv4_hdr;
 602         struct ipv6_hdr *ipv6_hdr;
 603         void *l3_hdr = NULL;
 604         struct ether_hdr *eth_hdr;
 605         uint16_t ethertype;
 606
 607         eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
 608
 609         m->l2_len = sizeof(struct ether_hdr);
 610         ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
 611
 612         if (ethertype == ETHER_TYPE_VLAN) {
 613                 struct vlan_hdr *vlan_hdr = (struct vlan_hdr *)(eth_hdr + 1);
 614
 615                 m->l2_len += sizeof(struct vlan_hdr);
 616                 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
 617         }
 618
 619         l3_hdr = (char *)eth_hdr + m->l2_len;
 620
 621         switch (ethertype) {
 622         case ETHER_TYPE_IPv4:
 623                 ipv4_hdr = (struct ipv4_hdr *)l3_hdr;
 624                 *l4_proto = ipv4_hdr->next_proto_id;
 625                 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
 626                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 627                 m->ol_flags |= PKT_TX_IPV4;
 628                 break;
 629         case ETHER_TYPE_IPv6:
 630                 ipv6_hdr = (struct ipv6_hdr *)l3_hdr;
 631                 *l4_proto = ipv6_hdr->proto;
 632                 m->l3_len = sizeof(struct ipv6_hdr);
 633                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 634                 m->ol_flags |= PKT_TX_IPV6;
 635                 break;
 636         default:
 637                 m->l3_len = 0;
 638                 *l4_proto = 0;
 639                 break;
 640         }
 641 }
 642
 643 static inline void __attribute__((always_inline))
 644 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
 645 {
 646         uint16_t l4_proto = 0;
 647         void *l4_hdr = NULL;
 648         struct tcp_hdr *tcp_hdr = NULL;
 649
 650         parse_ethernet(m, &l4_proto, &l4_hdr);
 651         if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 652                 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
 653                         switch (hdr->csum_offset) {
 654                         case (offsetof(struct tcp_hdr, cksum)):
 655                                 if (l4_proto == IPPROTO_TCP)
 656                                         m->ol_flags |= PKT_TX_TCP_CKSUM;
 657                                 break;
 658                         case (offsetof(struct udp_hdr, dgram_cksum)):
 659                                 if (l4_proto == IPPROTO_UDP)
 660                                         m->ol_flags |= PKT_TX_UDP_CKSUM;
 661                                 break;
 662                         case (offsetof(struct sctp_hdr, cksum)):
 663                                 if (l4_proto == IPPROTO_SCTP)
 664                                         m->ol_flags |= PKT_TX_SCTP_CKSUM;
 665                                 break;
 666                         default:
 667                                 break;
 668                         }
 669                 }
 670         }
 671
 672         if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 673                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 674                 case VIRTIO_NET_HDR_GSO_TCPV4:
 675                 case VIRTIO_NET_HDR_GSO_TCPV6:
 676                         tcp_hdr = (struct tcp_hdr *)l4_hdr;
 677                         m->ol_flags |= PKT_TX_TCP_SEG;
 678                         m->tso_segsz = hdr->gso_size;
 679                         m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
 680                         break;
 681                 default:
 682                         RTE_LOG(WARNING, VHOST_DATA,
 683                                 "unsupported gso type %u.\n", hdr->gso_type);
 684                         break;
 685                 }
 686         }
 687 }
 688
 689 #define RARP_PKT_SIZE   64
 690
 691 static int
 692 make_rarp_packet(struct rte_mbuf *rarp_mbuf, const struct ether_addr *mac)
 693 {
 694         struct ether_hdr *eth_hdr;
 695         struct arp_hdr  *rarp;
 696
 697         if (rarp_mbuf->buf_len < 64) {
 698                 RTE_LOG(WARNING, VHOST_DATA,
 699                         "failed to make RARP; mbuf size too small %u (< %d)\n",
 700                         rarp_mbuf->buf_len, RARP_PKT_SIZE);
 701                 return -1;
 702         }
 703
 704         /* Ethernet header. */
 705         eth_hdr = rte_pktmbuf_mtod_offset(rarp_mbuf, struct ether_hdr *, 0);
 706         memset(eth_hdr->d_addr.addr_bytes, 0xff, ETHER_ADDR_LEN);
 707         ether_addr_copy(mac, &eth_hdr->s_addr);
 708         eth_hdr->ether_type = htons(ETHER_TYPE_RARP);
 709
 710         /* RARP header. */
 711         rarp = (struct arp_hdr *)(eth_hdr + 1);
 712         rarp->arp_hrd = htons(ARP_HRD_ETHER);
 713         rarp->arp_pro = htons(ETHER_TYPE_IPv4);
 714         rarp->arp_hln = ETHER_ADDR_LEN;
 715         rarp->arp_pln = 4;
 716         rarp->arp_op  = htons(ARP_OP_REVREQUEST);
 717
 718         ether_addr_copy(mac, &rarp->arp_data.arp_sha);
 719         ether_addr_copy(mac, &rarp->arp_data.arp_tha);
 720         memset(&rarp->arp_data.arp_sip, 0x00, 4);
 721         memset(&rarp->arp_data.arp_tip, 0x00, 4);
 722
 723         rarp_mbuf->pkt_len  = rarp_mbuf->data_len = RARP_PKT_SIZE;
 724
 725         return 0;
 726 }
 727
 728 static inline int __attribute__((always_inline))
 729 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 730                   struct rte_mbuf *m, uint16_t desc_idx,
 731                   struct rte_mempool *mbuf_pool)
 732 {
 733         struct vring_desc *desc;
 734         uint64_t desc_addr;
 735         uint32_t desc_avail, desc_offset;
 736         uint32_t mbuf_avail, mbuf_offset;
 737         uint32_t cpy_len;
 738         struct rte_mbuf *cur = m, *prev = m;
 739         struct virtio_net_hdr *hdr;
 740
 741         desc = &vq->desc[desc_idx];
 742         desc_addr = gpa_to_vva(dev, desc->addr);
 743         rte_prefetch0((void *)(uintptr_t)desc_addr);
 744
 745         /* Retrieve virtio net header */
 746         hdr = (struct virtio_net_hdr *)((uintptr_t)desc_addr);
 747         desc_avail  = desc->len - vq->vhost_hlen;
 748         desc_offset = vq->vhost_hlen;
 749
 750         mbuf_offset = 0;
 751         mbuf_avail  = m->buf_len - RTE_PKTMBUF_HEADROOM;
 752         while (desc_avail != 0 || (desc->flags & VRING_DESC_F_NEXT) != 0) {
 753                 /* This desc reaches to its end, get the next one */
 754                 if (desc_avail == 0) {
 755                         desc = &vq->desc[desc->next];
 756
 757                         desc_addr = gpa_to_vva(dev, desc->addr);
 758                         rte_prefetch0((void *)(uintptr_t)desc_addr);
 759
 760                         desc_offset = 0;
 761                         desc_avail  = desc->len;
 762
 763                         PRINT_PACKET(dev, (uintptr_t)desc_addr, desc->len, 0);
 764                 }
 765
 766                 /*
 767                  * This mbuf reaches to its end, get a new one
 768                  * to hold more data.
 769                  */
 770                 if (mbuf_avail == 0) {
 771                         cur = rte_pktmbuf_alloc(mbuf_pool);
 772                         if (unlikely(cur == NULL)) {
 773                                 RTE_LOG(ERR, VHOST_DATA, "Failed to "
 774                                         "allocate memory for mbuf.\n");
 775                                 return -1;
 776                         }
 777
 778                         prev->next = cur;
 779                         prev->data_len = mbuf_offset;
 780                         m->nb_segs += 1;
 781                         m->pkt_len += mbuf_offset;
 782                         prev = cur;
 783
 784                         mbuf_offset = 0;
 785                         mbuf_avail  = cur->buf_len - RTE_PKTMBUF_HEADROOM;
 786                 }
 787
 788                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 789                 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *, mbuf_offset),
 790                         (void *)((uintptr_t)(desc_addr + desc_offset)),
 791                         cpy_len);
 792
 793                 mbuf_avail  -= cpy_len;
 794                 mbuf_offset += cpy_len;
 795                 desc_avail  -= cpy_len;
 796                 desc_offset += cpy_len;
 797         }
 798
 799         prev->data_len = mbuf_offset;
 800         m->pkt_len    += mbuf_offset;
 801
 802         if (hdr->flags != 0 || hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE)
 803                 vhost_dequeue_offload(hdr, m);
 804
 805         return 0;
 806 }
 807
 808 uint16_t
 809 rte_vhost_dequeue_burst(struct virtio_net *dev, uint16_t queue_id,
 810         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
 811 {
 812         struct rte_mbuf *rarp_mbuf = NULL;
 813         struct vhost_virtqueue *vq;
 814         uint32_t desc_indexes[MAX_PKT_BURST];
 815         uint32_t used_idx;
 816         uint32_t i = 0;
 817         uint16_t free_entries;
 818         uint16_t avail_idx;
 819
 820         if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->virt_qp_nb))) {
 821                 RTE_LOG(ERR, VHOST_DATA,
 822                         "%s (%"PRIu64"): virtqueue idx:%d invalid.\n",
 823                         __func__, dev->device_fh, queue_id);
 824                 return 0;
 825         }
 826
 827         vq = dev->virtqueue[queue_id];
 828         if (unlikely(vq->enabled == 0))
 829                 return 0;
 830
 831         /*
 832          * Construct a RARP broadcast packet, and inject it to the "pkts"
 833          * array, to looks like that guest actually send such packet.
 834          *
 835          * Check user_send_rarp() for more information.
 836          */
 837         if (unlikely(rte_atomic16_cmpset((volatile uint16_t *)
 838                                          &dev->broadcast_rarp.cnt, 1, 0))) {
 839                 rarp_mbuf = rte_pktmbuf_alloc(mbuf_pool);
 840                 if (rarp_mbuf == NULL) {
 841                         RTE_LOG(ERR, VHOST_DATA,
 842                                 "Failed to allocate memory for mbuf.\n");
 843                         return 0;
 844                 }
 845
 846                 if (make_rarp_packet(rarp_mbuf, &dev->mac)) {
 847                         rte_pktmbuf_free(rarp_mbuf);
 848                         rarp_mbuf = NULL;
 849                 } else {
 850                         count -= 1;
 851                 }
 852         }
 853
 854         avail_idx =  *((volatile uint16_t *)&vq->avail->idx);
 855         free_entries = avail_idx - vq->last_used_idx;
 856         if (free_entries == 0)
 857                 goto out;
 858
 859         LOG_DEBUG(VHOST_DATA, "%s (%"PRIu64")\n", __func__, dev->device_fh);
 860
 861         /* Prefetch available ring to retrieve head indexes. */
 862         used_idx = vq->last_used_idx & (vq->size - 1);
 863         rte_prefetch0(&vq->avail->ring[used_idx]);
 864
 865         count = RTE_MIN(count, MAX_PKT_BURST);
 866         count = RTE_MIN(count, free_entries);
 867         LOG_DEBUG(VHOST_DATA, "(%"PRIu64") about to dequeue %u buffers\n",
 868                         dev->device_fh, count);
 869
 870         /* Retrieve all of the head indexes first to avoid caching issues. */
 871         for (i = 0; i < count; i++) {
 872                 desc_indexes[i] = vq->avail->ring[(vq->last_used_idx + i) &
 873                                         (vq->size - 1)];
 874         }
 875
 876         /* Prefetch descriptor index. */
 877         rte_prefetch0(&vq->desc[desc_indexes[0]]);
 878         rte_prefetch0(&vq->used->ring[vq->last_used_idx & (vq->size - 1)]);
 879
 880         for (i = 0; i < count; i++) {
 881                 int err;
 882
 883                 if (likely(i + 1 < count)) {
 884                         rte_prefetch0(&vq->desc[desc_indexes[i + 1]]);
 885                         rte_prefetch0(&vq->used->ring[(used_idx + 1) &
 886                                                       (vq->size - 1)]);
 887                 }
 888
 889                 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
 890                 if (unlikely(pkts[i] == NULL)) {
 891                         RTE_LOG(ERR, VHOST_DATA,
 892                                 "Failed to allocate memory for mbuf.\n");
 893                         break;
 894                 }
 895                 err = copy_desc_to_mbuf(dev, vq, pkts[i], desc_indexes[i],
 896                                         mbuf_pool);
 897                 if (unlikely(err)) {
 898                         rte_pktmbuf_free(pkts[i]);
 899                         break;
 900                 }
 901
 902                 used_idx = vq->last_used_idx++ & (vq->size - 1);
 903                 vq->used->ring[used_idx].id  = desc_indexes[i];
 904                 vq->used->ring[used_idx].len = 0;
 905                 vhost_log_used_vring(dev, vq,
 906                                 offsetof(struct vring_used, ring[used_idx]),
 907                                 sizeof(vq->used->ring[used_idx]));
 908         }
 909
 910         rte_compiler_barrier();
 911         vq->used->idx += i;
 912         vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
 913                         sizeof(vq->used->idx));
 914
 915         /* Kick guest if required. */
 916         if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT))
 917                 eventfd_write(vq->callfd, (eventfd_t)1);
 918
 919 out:
 920         if (unlikely(rarp_mbuf != NULL)) {
 921                 /*
 922                  * Inject it to the head of "pkts" array, so that switch's mac
 923                  * learning table will get updated first.
 924                  */
 925                 memmove(&pkts[1], pkts, i * sizeof(struct rte_mbuf *));
 926                 pkts[0] = rarp_mbuf;
 927                 i += 1;
 928         }
 929
 930         return i;
 931 }