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         return;
 128 }
 129
 130 static inline void
 131 copy_virtio_net_hdr(struct vhost_virtqueue *vq, uint64_t desc_addr,
 132                     struct virtio_net_hdr_mrg_rxbuf hdr)
 133 {
 134         if (vq->vhost_hlen == sizeof(struct virtio_net_hdr_mrg_rxbuf))
 135                 *(struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)desc_addr = hdr;
 136         else
 137                 *(struct virtio_net_hdr *)(uintptr_t)desc_addr = hdr.hdr;
 138 }
 139
 140 static inline int __attribute__((always_inline))
 141 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 142                   struct rte_mbuf *m, uint16_t desc_idx, uint32_t *copied)
 143 {
 144         uint32_t desc_avail, desc_offset;
 145         uint32_t mbuf_avail, mbuf_offset;
 146         uint32_t cpy_len;
 147         struct vring_desc *desc;
 148         uint64_t desc_addr;
 149         struct virtio_net_hdr_mrg_rxbuf virtio_hdr = {{0, 0, 0, 0, 0, 0}, 0};
 150
 151         desc = &vq->desc[desc_idx];
 152         if (unlikely(desc->len < vq->vhost_hlen))
 153                 return -1;
 154
 155         desc_addr = gpa_to_vva(dev, desc->addr);
 156         rte_prefetch0((void *)(uintptr_t)desc_addr);
 157
 158         virtio_enqueue_offload(m, &virtio_hdr.hdr);
 159         copy_virtio_net_hdr(vq, desc_addr, virtio_hdr);
 160         vhost_log_write(dev, desc->addr, vq->vhost_hlen);
 161         PRINT_PACKET(dev, (uintptr_t)desc_addr, vq->vhost_hlen, 0);
 162
 163         desc_offset = vq->vhost_hlen;
 164         desc_avail  = desc->len - vq->vhost_hlen;
 165
 166         *copied = rte_pktmbuf_pkt_len(m);
 167         mbuf_avail  = rte_pktmbuf_data_len(m);
 168         mbuf_offset = 0;
 169         while (mbuf_avail != 0 || m->next != NULL) {
 170                 /* done with current mbuf, fetch next */
 171                 if (mbuf_avail == 0) {
 172                         m = m->next;
 173
 174                         mbuf_offset = 0;
 175                         mbuf_avail  = rte_pktmbuf_data_len(m);
 176                 }
 177
 178                 /* done with current desc buf, fetch next */
 179                 if (desc_avail == 0) {
 180                         if ((desc->flags & VRING_DESC_F_NEXT) == 0) {
 181                                 /* Room in vring buffer is not enough */
 182                                 return -1;
 183                         }
 184                         if (unlikely(desc->next >= vq->size))
 185                                 return -1;
 186
 187                         desc = &vq->desc[desc->next];
 188                         desc_addr   = gpa_to_vva(dev, desc->addr);
 189                         desc_offset = 0;
 190                         desc_avail  = desc->len;
 191                 }
 192
 193                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 194                 rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)),
 195                         rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 196                         cpy_len);
 197                 vhost_log_write(dev, desc->addr + desc_offset, cpy_len);
 198                 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
 199                              cpy_len, 0);
 200
 201                 mbuf_avail  -= cpy_len;
 202                 mbuf_offset += cpy_len;
 203                 desc_avail  -= cpy_len;
 204                 desc_offset += cpy_len;
 205         }
 206
 207         return 0;
 208 }
 209
 210 /*
 211  * As many data cores may want to access available buffers
 212  * they need to be reserved.
 213  */
 214 static inline uint32_t
 215 reserve_avail_buf(struct vhost_virtqueue *vq, uint32_t count,
 216                   uint16_t *start, uint16_t *end)
 217 {
 218         uint16_t res_start_idx;
 219         uint16_t res_end_idx;
 220         uint16_t avail_idx;
 221         uint16_t free_entries;
 222         int success;
 223
 224         count = RTE_MIN(count, (uint32_t)MAX_PKT_BURST);
 225
 226 again:
 227         res_start_idx = vq->last_used_idx_res;
 228         avail_idx = *((volatile uint16_t *)&vq->avail->idx);
 229
 230         free_entries = avail_idx - res_start_idx;
 231         count = RTE_MIN(count, free_entries);
 232         if (count == 0)
 233                 return 0;
 234
 235         res_end_idx = res_start_idx + count;
 236
 237         /*
 238          * update vq->last_used_idx_res atomically; try again if failed.
 239          *
 240          * TODO: Allow to disable cmpset if no concurrency in application.
 241          */
 242         success = rte_atomic16_cmpset(&vq->last_used_idx_res,
 243                                       res_start_idx, res_end_idx);
 244         if (unlikely(!success))
 245                 goto again;
 246
 247         *start = res_start_idx;
 248         *end   = res_end_idx;
 249
 250         return count;
 251 }
 252
 253 /**
 254  * This function adds buffers to the virtio devices RX virtqueue. Buffers can
 255  * be received from the physical port or from another virtio device. A packet
 256  * count is returned to indicate the number of packets that are succesfully
 257  * added to the RX queue. This function works when the mbuf is scattered, but
 258  * it doesn't support the mergeable feature.
 259  */
 260 static inline uint32_t __attribute__((always_inline))
 261 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
 262               struct rte_mbuf **pkts, uint32_t count)
 263 {
 264         struct vhost_virtqueue *vq;
 265         uint16_t res_start_idx, res_end_idx;
 266         uint16_t desc_indexes[MAX_PKT_BURST];
 267         uint32_t i;
 268
 269         LOG_DEBUG(VHOST_DATA, "(%"PRIu64") virtio_dev_rx()\n", dev->device_fh);
 270         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->virt_qp_nb))) {
 271                 RTE_LOG(ERR, VHOST_DATA,
 272                         "%s (%"PRIu64"): virtqueue idx:%d invalid.\n",
 273                         __func__, dev->device_fh, queue_id);
 274                 return 0;
 275         }
 276
 277         vq = dev->virtqueue[queue_id];
 278         if (unlikely(vq->enabled == 0))
 279                 return 0;
 280
 281         count = reserve_avail_buf(vq, count, &res_start_idx, &res_end_idx);
 282         if (count == 0)
 283                 return 0;
 284
 285         LOG_DEBUG(VHOST_DATA,
 286                 "(%"PRIu64") res_start_idx %d| res_end_idx Index %d\n",
 287                 dev->device_fh, res_start_idx, res_end_idx);
 288
 289         /* Retrieve all of the desc indexes first to avoid caching issues. */
 290         rte_prefetch0(&vq->avail->ring[res_start_idx & (vq->size - 1)]);
 291         for (i = 0; i < count; i++) {
 292                 desc_indexes[i] = vq->avail->ring[(res_start_idx + i) &
 293                                                   (vq->size - 1)];
 294         }
 295
 296         rte_prefetch0(&vq->desc[desc_indexes[0]]);
 297         for (i = 0; i < count; i++) {
 298                 uint16_t desc_idx = desc_indexes[i];
 299                 uint16_t used_idx = (res_start_idx + i) & (vq->size - 1);
 300                 uint32_t copied;
 301                 int err;
 302
 303                 err = copy_mbuf_to_desc(dev, vq, pkts[i], desc_idx, &copied);
 304
 305                 vq->used->ring[used_idx].id = desc_idx;
 306                 if (unlikely(err))
 307                         vq->used->ring[used_idx].len = vq->vhost_hlen;
 308                 else
 309                         vq->used->ring[used_idx].len = copied + vq->vhost_hlen;
 310                 vhost_log_used_vring(dev, vq,
 311                         offsetof(struct vring_used, ring[used_idx]),
 312                         sizeof(vq->used->ring[used_idx]));
 313
 314                 if (i + 1 < count)
 315                         rte_prefetch0(&vq->desc[desc_indexes[i+1]]);
 316         }
 317
 318         rte_compiler_barrier();
 319
 320         /* Wait until it's our turn to add our buffer to the used ring. */
 321         while (unlikely(vq->last_used_idx != res_start_idx))
 322                 rte_pause();
 323
 324         *(volatile uint16_t *)&vq->used->idx += count;
 325         vq->last_used_idx = res_end_idx;
 326         vhost_log_used_vring(dev, vq,
 327                 offsetof(struct vring_used, idx),
 328                 sizeof(vq->used->idx));
 329
 330         /* flush used->idx update before we read avail->flags. */
 331         rte_mb();
 332
 333         /* Kick the guest if necessary. */
 334         if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
 335                         && (vq->callfd >= 0))
 336                 eventfd_write(vq->callfd, (eventfd_t)1);
 337         return count;
 338 }
 339
 340 static inline int
 341 fill_vec_buf(struct vhost_virtqueue *vq, uint32_t avail_idx,
 342              uint32_t *allocated, uint32_t *vec_idx)
 343 {
 344         uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
 345         uint32_t vec_id = *vec_idx;
 346         uint32_t len    = *allocated;
 347
 348         while (1) {
 349                 if (unlikely(vec_id >= BUF_VECTOR_MAX || idx >= vq->size))
 350                         return -1;
 351
 352                 len += vq->desc[idx].len;
 353                 vq->buf_vec[vec_id].buf_addr = vq->desc[idx].addr;
 354                 vq->buf_vec[vec_id].buf_len  = vq->desc[idx].len;
 355                 vq->buf_vec[vec_id].desc_idx = idx;
 356                 vec_id++;
 357
 358                 if ((vq->desc[idx].flags & VRING_DESC_F_NEXT) == 0)
 359                         break;
 360
 361                 idx = vq->desc[idx].next;
 362         }
 363
 364         *allocated = len;
 365         *vec_idx   = vec_id;
 366
 367         return 0;
 368 }
 369
 370 /*
 371  * As many data cores may want to access available buffers concurrently,
 372  * they need to be reserved.
 373  *
 374  * Returns -1 on fail, 0 on success
 375  */
 376 static inline int
 377 reserve_avail_buf_mergeable(struct vhost_virtqueue *vq, uint32_t size,
 378                             uint16_t *start, uint16_t *end)
 379 {
 380         uint16_t res_start_idx;
 381         uint16_t res_cur_idx;
 382         uint16_t avail_idx;
 383         uint32_t allocated;
 384         uint32_t vec_idx;
 385         uint16_t tries;
 386
 387 again:
 388         res_start_idx = vq->last_used_idx_res;
 389         res_cur_idx  = res_start_idx;
 390
 391         allocated = 0;
 392         vec_idx   = 0;
 393         tries     = 0;
 394         while (1) {
 395                 avail_idx = *((volatile uint16_t *)&vq->avail->idx);
 396                 if (unlikely(res_cur_idx == avail_idx))
 397                         return -1;
 398
 399                 if (unlikely(fill_vec_buf(vq, res_cur_idx, &allocated,
 400                                           &vec_idx) < 0))
 401                         return -1;
 402
 403                 res_cur_idx++;
 404                 tries++;
 405
 406                 if (allocated >= size)
 407                         break;
 408
 409                 /*
 410                  * if we tried all available ring items, and still
 411                  * can't get enough buf, it means something abnormal
 412                  * happened.
 413                  */
 414                 if (unlikely(tries >= vq->size))
 415                         return -1;
 416         }
 417
 418         /*
 419          * update vq->last_used_idx_res atomically.
 420          * retry again if failed.
 421          */
 422         if (rte_atomic16_cmpset(&vq->last_used_idx_res,
 423                                 res_start_idx, res_cur_idx) == 0)
 424                 goto again;
 425
 426         *start = res_start_idx;
 427         *end   = res_cur_idx;
 428         return 0;
 429 }
 430
 431 static inline uint32_t __attribute__((always_inline))
 432 copy_mbuf_to_desc_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
 433                             uint16_t res_start_idx, uint16_t res_end_idx,
 434                             struct rte_mbuf *m)
 435 {
 436         struct virtio_net_hdr_mrg_rxbuf virtio_hdr = {{0, 0, 0, 0, 0, 0}, 0};
 437         uint32_t vec_idx = 0;
 438         uint16_t cur_idx = res_start_idx;
 439         uint64_t desc_addr;
 440         uint32_t mbuf_offset, mbuf_avail;
 441         uint32_t desc_offset, desc_avail;
 442         uint32_t cpy_len;
 443         uint16_t desc_idx, used_idx;
 444
 445         if (unlikely(m == NULL))
 446                 return 0;
 447
 448         LOG_DEBUG(VHOST_DATA,
 449                 "(%"PRIu64") Current Index %d| End Index %d\n",
 450                 dev->device_fh, cur_idx, res_end_idx);
 451
 452         if (vq->buf_vec[vec_idx].buf_len < vq->vhost_hlen)
 453                 return -1;
 454
 455         desc_addr = gpa_to_vva(dev, vq->buf_vec[vec_idx].buf_addr);
 456         rte_prefetch0((void *)(uintptr_t)desc_addr);
 457
 458         virtio_hdr.num_buffers = res_end_idx - res_start_idx;
 459         LOG_DEBUG(VHOST_DATA, "(%"PRIu64") RX: Num merge buffers %d\n",
 460                 dev->device_fh, virtio_hdr.num_buffers);
 461
 462         virtio_enqueue_offload(m, &virtio_hdr.hdr);
 463         copy_virtio_net_hdr(vq, desc_addr, virtio_hdr);
 464         vhost_log_write(dev, vq->buf_vec[vec_idx].buf_addr, vq->vhost_hlen);
 465         PRINT_PACKET(dev, (uintptr_t)desc_addr, vq->vhost_hlen, 0);
 466
 467         desc_avail  = vq->buf_vec[vec_idx].buf_len - vq->vhost_hlen;
 468         desc_offset = vq->vhost_hlen;
 469
 470         mbuf_avail  = rte_pktmbuf_data_len(m);
 471         mbuf_offset = 0;
 472         while (mbuf_avail != 0 || m->next != NULL) {
 473                 /* done with current desc buf, get the next one */
 474                 if (desc_avail == 0) {
 475                         desc_idx = vq->buf_vec[vec_idx].desc_idx;
 476
 477                         if (!(vq->desc[desc_idx].flags & VRING_DESC_F_NEXT)) {
 478                                 /* Update used ring with desc information */
 479                                 used_idx = cur_idx++ & (vq->size - 1);
 480                                 vq->used->ring[used_idx].id  = desc_idx;
 481                                 vq->used->ring[used_idx].len = desc_offset;
 482                                 vhost_log_used_vring(dev, vq,
 483                                         offsetof(struct vring_used,
 484                                                  ring[used_idx]),
 485                                         sizeof(vq->used->ring[used_idx]));
 486                         }
 487
 488                         vec_idx++;
 489                         desc_addr = gpa_to_vva(dev, vq->buf_vec[vec_idx].buf_addr);
 490
 491                         /* Prefetch buffer address. */
 492                         rte_prefetch0((void *)(uintptr_t)desc_addr);
 493                         desc_offset = 0;
 494                         desc_avail  = vq->buf_vec[vec_idx].buf_len;
 495                 }
 496
 497                 /* done with current mbuf, get the next one */
 498                 if (mbuf_avail == 0) {
 499                         m = m->next;
 500
 501                         mbuf_offset = 0;
 502                         mbuf_avail  = rte_pktmbuf_data_len(m);
 503                 }
 504
 505                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 506                 rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)),
 507                         rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 508                         cpy_len);
 509                 vhost_log_write(dev, vq->buf_vec[vec_idx].buf_addr + desc_offset,
 510                         cpy_len);
 511                 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
 512                         cpy_len, 0);
 513
 514                 mbuf_avail  -= cpy_len;
 515                 mbuf_offset += cpy_len;
 516                 desc_avail  -= cpy_len;
 517                 desc_offset += cpy_len;
 518         }
 519
 520         used_idx = cur_idx & (vq->size - 1);
 521         vq->used->ring[used_idx].id = vq->buf_vec[vec_idx].desc_idx;
 522         vq->used->ring[used_idx].len = desc_offset;
 523         vhost_log_used_vring(dev, vq,
 524                 offsetof(struct vring_used, ring[used_idx]),
 525                 sizeof(vq->used->ring[used_idx]));
 526
 527         return res_end_idx - res_start_idx;
 528 }
 529
 530 static inline uint32_t __attribute__((always_inline))
 531 virtio_dev_merge_rx(struct virtio_net *dev, uint16_t queue_id,
 532         struct rte_mbuf **pkts, uint32_t count)
 533 {
 534         struct vhost_virtqueue *vq;
 535         uint32_t pkt_idx = 0, nr_used = 0;
 536         uint16_t start, end;
 537
 538         LOG_DEBUG(VHOST_DATA, "(%"PRIu64") virtio_dev_merge_rx()\n",
 539                 dev->device_fh);
 540         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->virt_qp_nb))) {
 541                 RTE_LOG(ERR, VHOST_DATA,
 542                         "%s (%"PRIu64"): virtqueue idx:%d invalid.\n",
 543                         __func__, dev->device_fh, queue_id);
 544                 return 0;
 545         }
 546
 547         vq = dev->virtqueue[queue_id];
 548         if (unlikely(vq->enabled == 0))
 549                 return 0;
 550
 551         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
 552         if (count == 0)
 553                 return 0;
 554
 555         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
 556                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + vq->vhost_hlen;
 557
 558                 if (unlikely(reserve_avail_buf_mergeable(vq, pkt_len,
 559                                                          &start, &end) < 0)) {
 560                         LOG_DEBUG(VHOST_DATA,
 561                                 "(%" PRIu64 ") Failed to get enough desc from vring\n",
 562                                 dev->device_fh);
 563                         break;
 564                 }
 565
 566                 nr_used = copy_mbuf_to_desc_mergeable(dev, vq, start, end,
 567                                                       pkts[pkt_idx]);
 568                 rte_compiler_barrier();
 569
 570                 /*
 571                  * Wait until it's our turn to add our buffer
 572                  * to the used ring.
 573                  */
 574                 while (unlikely(vq->last_used_idx != start))
 575                         rte_pause();
 576
 577                 *(volatile uint16_t *)&vq->used->idx += nr_used;
 578                 vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
 579                         sizeof(vq->used->idx));
 580                 vq->last_used_idx = end;
 581         }
 582
 583         if (likely(pkt_idx)) {
 584                 /* flush used->idx update before we read avail->flags. */
 585                 rte_mb();
 586
 587                 /* Kick the guest if necessary. */
 588                 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
 589                                 && (vq->callfd >= 0))
 590                         eventfd_write(vq->callfd, (eventfd_t)1);
 591         }
 592
 593         return pkt_idx;
 594 }
 595
 596 uint16_t
 597 rte_vhost_enqueue_burst(struct virtio_net *dev, uint16_t queue_id,
 598         struct rte_mbuf **pkts, uint16_t count)
 599 {
 600         if (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF))
 601                 return virtio_dev_merge_rx(dev, queue_id, pkts, count);
 602         else
 603                 return virtio_dev_rx(dev, queue_id, pkts, count);
 604 }
 605
 606 static void
 607 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
 608 {
 609         struct ipv4_hdr *ipv4_hdr;
 610         struct ipv6_hdr *ipv6_hdr;
 611         void *l3_hdr = NULL;
 612         struct ether_hdr *eth_hdr;
 613         uint16_t ethertype;
 614
 615         eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
 616
 617         m->l2_len = sizeof(struct ether_hdr);
 618         ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
 619
 620         if (ethertype == ETHER_TYPE_VLAN) {
 621                 struct vlan_hdr *vlan_hdr = (struct vlan_hdr *)(eth_hdr + 1);
 622
 623                 m->l2_len += sizeof(struct vlan_hdr);
 624                 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
 625         }
 626
 627         l3_hdr = (char *)eth_hdr + m->l2_len;
 628
 629         switch (ethertype) {
 630         case ETHER_TYPE_IPv4:
 631                 ipv4_hdr = (struct ipv4_hdr *)l3_hdr;
 632                 *l4_proto = ipv4_hdr->next_proto_id;
 633                 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
 634                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 635                 m->ol_flags |= PKT_TX_IPV4;
 636                 break;
 637         case ETHER_TYPE_IPv6:
 638                 ipv6_hdr = (struct ipv6_hdr *)l3_hdr;
 639                 *l4_proto = ipv6_hdr->proto;
 640                 m->l3_len = sizeof(struct ipv6_hdr);
 641                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 642                 m->ol_flags |= PKT_TX_IPV6;
 643                 break;
 644         default:
 645                 m->l3_len = 0;
 646                 *l4_proto = 0;
 647                 break;
 648         }
 649 }
 650
 651 static inline void __attribute__((always_inline))
 652 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
 653 {
 654         uint16_t l4_proto = 0;
 655         void *l4_hdr = NULL;
 656         struct tcp_hdr *tcp_hdr = NULL;
 657
 658         parse_ethernet(m, &l4_proto, &l4_hdr);
 659         if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 660                 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
 661                         switch (hdr->csum_offset) {
 662                         case (offsetof(struct tcp_hdr, cksum)):
 663                                 if (l4_proto == IPPROTO_TCP)
 664                                         m->ol_flags |= PKT_TX_TCP_CKSUM;
 665                                 break;
 666                         case (offsetof(struct udp_hdr, dgram_cksum)):
 667                                 if (l4_proto == IPPROTO_UDP)
 668                                         m->ol_flags |= PKT_TX_UDP_CKSUM;
 669                                 break;
 670                         case (offsetof(struct sctp_hdr, cksum)):
 671                                 if (l4_proto == IPPROTO_SCTP)
 672                                         m->ol_flags |= PKT_TX_SCTP_CKSUM;
 673                                 break;
 674                         default:
 675                                 break;
 676                         }
 677                 }
 678         }
 679
 680         if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 681                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 682                 case VIRTIO_NET_HDR_GSO_TCPV4:
 683                 case VIRTIO_NET_HDR_GSO_TCPV6:
 684                         tcp_hdr = (struct tcp_hdr *)l4_hdr;
 685                         m->ol_flags |= PKT_TX_TCP_SEG;
 686                         m->tso_segsz = hdr->gso_size;
 687                         m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
 688                         break;
 689                 default:
 690                         RTE_LOG(WARNING, VHOST_DATA,
 691                                 "unsupported gso type %u.\n", hdr->gso_type);
 692                         break;
 693                 }
 694         }
 695 }
 696
 697 #define RARP_PKT_SIZE   64
 698
 699 static int
 700 make_rarp_packet(struct rte_mbuf *rarp_mbuf, const struct ether_addr *mac)
 701 {
 702         struct ether_hdr *eth_hdr;
 703         struct arp_hdr  *rarp;
 704
 705         if (rarp_mbuf->buf_len < 64) {
 706                 RTE_LOG(WARNING, VHOST_DATA,
 707                         "failed to make RARP; mbuf size too small %u (< %d)\n",
 708                         rarp_mbuf->buf_len, RARP_PKT_SIZE);
 709                 return -1;
 710         }
 711
 712         /* Ethernet header. */
 713         eth_hdr = rte_pktmbuf_mtod_offset(rarp_mbuf, struct ether_hdr *, 0);
 714         memset(eth_hdr->d_addr.addr_bytes, 0xff, ETHER_ADDR_LEN);
 715         ether_addr_copy(mac, &eth_hdr->s_addr);
 716         eth_hdr->ether_type = htons(ETHER_TYPE_RARP);
 717
 718         /* RARP header. */
 719         rarp = (struct arp_hdr *)(eth_hdr + 1);
 720         rarp->arp_hrd = htons(ARP_HRD_ETHER);
 721         rarp->arp_pro = htons(ETHER_TYPE_IPv4);
 722         rarp->arp_hln = ETHER_ADDR_LEN;
 723         rarp->arp_pln = 4;
 724         rarp->arp_op  = htons(ARP_OP_REVREQUEST);
 725
 726         ether_addr_copy(mac, &rarp->arp_data.arp_sha);
 727         ether_addr_copy(mac, &rarp->arp_data.arp_tha);
 728         memset(&rarp->arp_data.arp_sip, 0x00, 4);
 729         memset(&rarp->arp_data.arp_tip, 0x00, 4);
 730
 731         rarp_mbuf->pkt_len  = rarp_mbuf->data_len = RARP_PKT_SIZE;
 732
 733         return 0;
 734 }
 735
 736 static inline int __attribute__((always_inline))
 737 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 738                   struct rte_mbuf *m, uint16_t desc_idx,
 739                   struct rte_mempool *mbuf_pool)
 740 {
 741         struct vring_desc *desc;
 742         uint64_t desc_addr;
 743         uint32_t desc_avail, desc_offset;
 744         uint32_t mbuf_avail, mbuf_offset;
 745         uint32_t cpy_len;
 746         struct rte_mbuf *cur = m, *prev = m;
 747         struct virtio_net_hdr *hdr;
 748         /* A counter to avoid desc dead loop chain */
 749         uint32_t nr_desc = 1;
 750
 751         desc = &vq->desc[desc_idx];
 752         if (unlikely(desc->len < vq->vhost_hlen))
 753                 return -1;
 754
 755         desc_addr = gpa_to_vva(dev, desc->addr);
 756         rte_prefetch0((void *)(uintptr_t)desc_addr);
 757
 758         /* Retrieve virtio net header */
 759         hdr = (struct virtio_net_hdr *)((uintptr_t)desc_addr);
 760         desc_avail  = desc->len - vq->vhost_hlen;
 761         desc_offset = vq->vhost_hlen;
 762
 763         mbuf_offset = 0;
 764         mbuf_avail  = m->buf_len - RTE_PKTMBUF_HEADROOM;
 765         while (desc_avail != 0 || (desc->flags & VRING_DESC_F_NEXT) != 0) {
 766                 /* This desc reaches to its end, get the next one */
 767                 if (desc_avail == 0) {
 768                         if (unlikely(desc->next >= vq->size ||
 769                                      ++nr_desc >= vq->size))
 770                                 return -1;
 771                         desc = &vq->desc[desc->next];
 772
 773                         desc_addr = gpa_to_vva(dev, desc->addr);
 774                         rte_prefetch0((void *)(uintptr_t)desc_addr);
 775
 776                         desc_offset = 0;
 777                         desc_avail  = desc->len;
 778
 779                         PRINT_PACKET(dev, (uintptr_t)desc_addr, desc->len, 0);
 780                 }
 781
 782                 /*
 783                  * This mbuf reaches to its end, get a new one
 784                  * to hold more data.
 785                  */
 786                 if (mbuf_avail == 0) {
 787                         cur = rte_pktmbuf_alloc(mbuf_pool);
 788                         if (unlikely(cur == NULL)) {
 789                                 RTE_LOG(ERR, VHOST_DATA, "Failed to "
 790                                         "allocate memory for mbuf.\n");
 791                                 return -1;
 792                         }
 793
 794                         prev->next = cur;
 795                         prev->data_len = mbuf_offset;
 796                         m->nb_segs += 1;
 797                         m->pkt_len += mbuf_offset;
 798                         prev = cur;
 799
 800                         mbuf_offset = 0;
 801                         mbuf_avail  = cur->buf_len - RTE_PKTMBUF_HEADROOM;
 802                 }
 803
 804                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 805                 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *, mbuf_offset),
 806                         (void *)((uintptr_t)(desc_addr + desc_offset)),
 807                         cpy_len);
 808
 809                 mbuf_avail  -= cpy_len;
 810                 mbuf_offset += cpy_len;
 811                 desc_avail  -= cpy_len;
 812                 desc_offset += cpy_len;
 813         }
 814
 815         prev->data_len = mbuf_offset;
 816         m->pkt_len    += mbuf_offset;
 817
 818         if (hdr->flags != 0 || hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE)
 819                 vhost_dequeue_offload(hdr, m);
 820
 821         return 0;
 822 }
 823
 824 uint16_t
 825 rte_vhost_dequeue_burst(struct virtio_net *dev, uint16_t queue_id,
 826         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
 827 {
 828         struct rte_mbuf *rarp_mbuf = NULL;
 829         struct vhost_virtqueue *vq;
 830         uint32_t desc_indexes[MAX_PKT_BURST];
 831         uint32_t used_idx;
 832         uint32_t i = 0;
 833         uint16_t free_entries;
 834         uint16_t avail_idx;
 835
 836         if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->virt_qp_nb))) {
 837                 RTE_LOG(ERR, VHOST_DATA,
 838                         "%s (%"PRIu64"): virtqueue idx:%d invalid.\n",
 839                         __func__, dev->device_fh, queue_id);
 840                 return 0;
 841         }
 842
 843         vq = dev->virtqueue[queue_id];
 844         if (unlikely(vq->enabled == 0))
 845                 return 0;
 846
 847         /*
 848          * Construct a RARP broadcast packet, and inject it to the "pkts"
 849          * array, to looks like that guest actually send such packet.
 850          *
 851          * Check user_send_rarp() for more information.
 852          */
 853         if (unlikely(rte_atomic16_cmpset((volatile uint16_t *)
 854                                          &dev->broadcast_rarp.cnt, 1, 0))) {
 855                 rarp_mbuf = rte_pktmbuf_alloc(mbuf_pool);
 856                 if (rarp_mbuf == NULL) {
 857                         RTE_LOG(ERR, VHOST_DATA,
 858                                 "Failed to allocate memory for mbuf.\n");
 859                         return 0;
 860                 }
 861
 862                 if (make_rarp_packet(rarp_mbuf, &dev->mac)) {
 863                         rte_pktmbuf_free(rarp_mbuf);
 864                         rarp_mbuf = NULL;
 865                 } else {
 866                         count -= 1;
 867                 }
 868         }
 869
 870         avail_idx =  *((volatile uint16_t *)&vq->avail->idx);
 871         free_entries = avail_idx - vq->last_used_idx;
 872         if (free_entries == 0)
 873                 goto out;
 874
 875         LOG_DEBUG(VHOST_DATA, "%s (%"PRIu64")\n", __func__, dev->device_fh);
 876
 877         /* Prefetch available ring to retrieve head indexes. */
 878         used_idx = vq->last_used_idx & (vq->size - 1);
 879         rte_prefetch0(&vq->avail->ring[used_idx]);
 880
 881         count = RTE_MIN(count, MAX_PKT_BURST);
 882         count = RTE_MIN(count, free_entries);
 883         LOG_DEBUG(VHOST_DATA, "(%"PRIu64") about to dequeue %u buffers\n",
 884                         dev->device_fh, count);
 885
 886         /* Retrieve all of the head indexes first to avoid caching issues. */
 887         for (i = 0; i < count; i++) {
 888                 desc_indexes[i] = vq->avail->ring[(vq->last_used_idx + i) &
 889                                         (vq->size - 1)];
 890         }
 891
 892         /* Prefetch descriptor index. */
 893         rte_prefetch0(&vq->desc[desc_indexes[0]]);
 894         rte_prefetch0(&vq->used->ring[vq->last_used_idx & (vq->size - 1)]);
 895
 896         for (i = 0; i < count; i++) {
 897                 int err;
 898
 899                 if (likely(i + 1 < count)) {
 900                         rte_prefetch0(&vq->desc[desc_indexes[i + 1]]);
 901                         rte_prefetch0(&vq->used->ring[(used_idx + 1) &
 902                                                       (vq->size - 1)]);
 903                 }
 904
 905                 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
 906                 if (unlikely(pkts[i] == NULL)) {
 907                         RTE_LOG(ERR, VHOST_DATA,
 908                                 "Failed to allocate memory for mbuf.\n");
 909                         break;
 910                 }
 911                 err = copy_desc_to_mbuf(dev, vq, pkts[i], desc_indexes[i],
 912                                         mbuf_pool);
 913                 if (unlikely(err)) {
 914                         rte_pktmbuf_free(pkts[i]);
 915                         break;
 916                 }
 917
 918                 used_idx = vq->last_used_idx++ & (vq->size - 1);
 919                 vq->used->ring[used_idx].id  = desc_indexes[i];
 920                 vq->used->ring[used_idx].len = 0;
 921                 vhost_log_used_vring(dev, vq,
 922                                 offsetof(struct vring_used, ring[used_idx]),
 923                                 sizeof(vq->used->ring[used_idx]));
 924         }
 925
 926         rte_compiler_barrier();
 927         vq->used->idx += i;
 928         vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
 929                         sizeof(vq->used->idx));
 930
 931         /* Kick guest if required. */
 932         if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
 933                         && (vq->callfd >= 0))
 934                 eventfd_write(vq->callfd, (eventfd_t)1);
 935
 936 out:
 937         if (unlikely(rarp_mbuf != NULL)) {
 938                 /*
 939                  * Inject it to the head of "pkts" array, so that switch's mac
 940                  * learning table will get updated first.
 941                  */
 942                 memmove(&pkts[1], pkts, i * sizeof(struct rte_mbuf *));
 943                 pkts[0] = rarp_mbuf;
 944                 i += 1;
 945         }
 946
 947         return i;
 948 }