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 #include "virtio-net.h"
  50
  51 #define MAX_PKT_BURST 32
  52 #define VHOST_LOG_PAGE  4096
  53
  54 static inline void __attribute__((always_inline))
  55 vhost_log_page(uint8_t *log_base, uint64_t page)
  56 {
  57         log_base[page / 8] |= 1 << (page % 8);
  58 }
  59
  60 static inline void __attribute__((always_inline))
  61 vhost_log_write(struct virtio_net *dev, uint64_t addr, uint64_t len)
  62 {
  63         uint64_t page;
  64
  65         if (likely(((dev->features & (1ULL << VHOST_F_LOG_ALL)) == 0) ||
  66                    !dev->log_base || !len))
  67                 return;
  68
  69         if (unlikely(dev->log_size <= ((addr + len - 1) / VHOST_LOG_PAGE / 8)))
  70                 return;
  71
  72         /* To make sure guest memory updates are committed before logging */
  73         rte_smp_wmb();
  74
  75         page = addr / VHOST_LOG_PAGE;
  76         while (page * VHOST_LOG_PAGE < addr + len) {
  77                 vhost_log_page((uint8_t *)(uintptr_t)dev->log_base, page);
  78                 page += 1;
  79         }
  80 }
  81
  82 static inline void __attribute__((always_inline))
  83 vhost_log_used_vring(struct virtio_net *dev, struct vhost_virtqueue *vq,
  84                      uint64_t offset, uint64_t len)
  85 {
  86         vhost_log_write(dev, vq->log_guest_addr + offset, len);
  87 }
  88
  89 static bool
  90 is_valid_virt_queue_idx(uint32_t idx, int is_tx, uint32_t qp_nb)
  91 {
  92         return (is_tx ^ (idx & 1)) == 0 && idx < qp_nb * VIRTIO_QNUM;
  93 }
  94
  95 static void
  96 virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
  97 {
  98         if (m_buf->ol_flags & PKT_TX_L4_MASK) {
  99                 net_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 100                 net_hdr->csum_start = m_buf->l2_len + m_buf->l3_len;
 101
 102                 switch (m_buf->ol_flags & PKT_TX_L4_MASK) {
 103                 case PKT_TX_TCP_CKSUM:
 104                         net_hdr->csum_offset = (offsetof(struct tcp_hdr,
 105                                                 cksum));
 106                         break;
 107                 case PKT_TX_UDP_CKSUM:
 108                         net_hdr->csum_offset = (offsetof(struct udp_hdr,
 109                                                 dgram_cksum));
 110                         break;
 111                 case PKT_TX_SCTP_CKSUM:
 112                         net_hdr->csum_offset = (offsetof(struct sctp_hdr,
 113                                                 cksum));
 114                         break;
 115                 }
 116         }
 117
 118         if (m_buf->ol_flags & PKT_TX_TCP_SEG) {
 119                 if (m_buf->ol_flags & PKT_TX_IPV4)
 120                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 121                 else
 122                         net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
 123                 net_hdr->gso_size = m_buf->tso_segsz;
 124                 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len
 125                                         + m_buf->l4_len;
 126         }
 127 }
 128
 129 static inline void
 130 copy_virtio_net_hdr(struct vhost_virtqueue *vq, uint64_t desc_addr,
 131                     struct virtio_net_hdr_mrg_rxbuf hdr)
 132 {
 133         if (vq->vhost_hlen == sizeof(struct virtio_net_hdr_mrg_rxbuf))
 134                 *(struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)desc_addr = hdr;
 135         else
 136                 *(struct virtio_net_hdr *)(uintptr_t)desc_addr = hdr.hdr;
 137 }
 138
 139 static inline int __attribute__((always_inline))
 140 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 141                   struct rte_mbuf *m, uint16_t desc_idx, uint32_t *copied)
 142 {
 143         uint32_t desc_avail, desc_offset;
 144         uint32_t mbuf_avail, mbuf_offset;
 145         uint32_t cpy_len;
 146         struct vring_desc *desc;
 147         uint64_t desc_addr;
 148         struct virtio_net_hdr_mrg_rxbuf virtio_hdr = {{0, 0, 0, 0, 0, 0}, 0};
 149
 150         desc = &vq->desc[desc_idx];
 151         if (unlikely(desc->len < vq->vhost_hlen))
 152                 return -1;
 153
 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                         if (unlikely(desc->next >= vq->size))
 184                                 return -1;
 185
 186                         desc = &vq->desc[desc->next];
 187                         desc_addr   = gpa_to_vva(dev, desc->addr);
 188                         desc_offset = 0;
 189                         desc_avail  = desc->len;
 190                 }
 191
 192                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 193                 rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)),
 194                         rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 195                         cpy_len);
 196                 vhost_log_write(dev, desc->addr + desc_offset, cpy_len);
 197                 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
 198                              cpy_len, 0);
 199
 200                 mbuf_avail  -= cpy_len;
 201                 mbuf_offset += cpy_len;
 202                 desc_avail  -= cpy_len;
 203                 desc_offset += cpy_len;
 204         }
 205
 206         return 0;
 207 }
 208
 209 /*
 210  * As many data cores may want to access available buffers
 211  * they need to be reserved.
 212  */
 213 static inline uint32_t
 214 reserve_avail_buf(struct vhost_virtqueue *vq, uint32_t count,
 215                   uint16_t *start, uint16_t *end)
 216 {
 217         uint16_t res_start_idx;
 218         uint16_t res_end_idx;
 219         uint16_t avail_idx;
 220         uint16_t free_entries;
 221         int success;
 222
 223         count = RTE_MIN(count, (uint32_t)MAX_PKT_BURST);
 224
 225 again:
 226         res_start_idx = vq->last_used_idx_res;
 227         avail_idx = *((volatile uint16_t *)&vq->avail->idx);
 228
 229         free_entries = avail_idx - res_start_idx;
 230         count = RTE_MIN(count, free_entries);
 231         if (count == 0)
 232                 return 0;
 233
 234         res_end_idx = res_start_idx + count;
 235
 236         /*
 237          * update vq->last_used_idx_res atomically; try again if failed.
 238          *
 239          * TODO: Allow to disable cmpset if no concurrency in application.
 240          */
 241         success = rte_atomic16_cmpset(&vq->last_used_idx_res,
 242                                       res_start_idx, res_end_idx);
 243         if (unlikely(!success))
 244                 goto again;
 245
 246         *start = res_start_idx;
 247         *end   = res_end_idx;
 248
 249         return count;
 250 }
 251
 252 /**
 253  * This function adds buffers to the virtio devices RX virtqueue. Buffers can
 254  * be received from the physical port or from another virtio device. A packet
 255  * count is returned to indicate the number of packets that are succesfully
 256  * added to the RX queue. This function works when the mbuf is scattered, but
 257  * it doesn't support the mergeable feature.
 258  */
 259 static inline uint32_t __attribute__((always_inline))
 260 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
 261               struct rte_mbuf **pkts, uint32_t count)
 262 {
 263         struct vhost_virtqueue *vq;
 264         uint16_t res_start_idx, res_end_idx;
 265         uint16_t desc_indexes[MAX_PKT_BURST];
 266         uint32_t i;
 267
 268         LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
 269         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->virt_qp_nb))) {
 270                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
 271                         dev->vid, __func__, queue_id);
 272                 return 0;
 273         }
 274
 275         vq = dev->virtqueue[queue_id];
 276         if (unlikely(vq->enabled == 0))
 277                 return 0;
 278
 279         count = reserve_avail_buf(vq, count, &res_start_idx, &res_end_idx);
 280         if (count == 0)
 281                 return 0;
 282
 283         LOG_DEBUG(VHOST_DATA, "(%d) res_start_idx %d | res_end_idx Index %d\n",
 284                 dev->vid, 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_smp_wmb();
 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                         && (vq->callfd >= 0))
 333                 eventfd_write(vq->callfd, (eventfd_t)1);
 334         return count;
 335 }
 336
 337 static inline int
 338 fill_vec_buf(struct vhost_virtqueue *vq, uint32_t avail_idx,
 339              uint32_t *allocated, uint32_t *vec_idx)
 340 {
 341         uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
 342         uint32_t vec_id = *vec_idx;
 343         uint32_t len    = *allocated;
 344
 345         while (1) {
 346                 if (unlikely(vec_id >= BUF_VECTOR_MAX || idx >= vq->size))
 347                         return -1;
 348
 349                 len += vq->desc[idx].len;
 350                 vq->buf_vec[vec_id].buf_addr = vq->desc[idx].addr;
 351                 vq->buf_vec[vec_id].buf_len  = vq->desc[idx].len;
 352                 vq->buf_vec[vec_id].desc_idx = idx;
 353                 vec_id++;
 354
 355                 if ((vq->desc[idx].flags & VRING_DESC_F_NEXT) == 0)
 356                         break;
 357
 358                 idx = vq->desc[idx].next;
 359         }
 360
 361         *allocated = len;
 362         *vec_idx   = vec_id;
 363
 364         return 0;
 365 }
 366
 367 /*
 368  * As many data cores may want to access available buffers concurrently,
 369  * they need to be reserved.
 370  *
 371  * Returns -1 on fail, 0 on success
 372  */
 373 static inline int
 374 reserve_avail_buf_mergeable(struct vhost_virtqueue *vq, uint32_t size,
 375                             uint16_t *start, uint16_t *end)
 376 {
 377         uint16_t res_start_idx;
 378         uint16_t res_cur_idx;
 379         uint16_t avail_idx;
 380         uint32_t allocated;
 381         uint32_t vec_idx;
 382         uint16_t tries;
 383
 384 again:
 385         res_start_idx = vq->last_used_idx_res;
 386         res_cur_idx  = res_start_idx;
 387
 388         allocated = 0;
 389         vec_idx   = 0;
 390         tries     = 0;
 391         while (1) {
 392                 avail_idx = *((volatile uint16_t *)&vq->avail->idx);
 393                 if (unlikely(res_cur_idx == avail_idx))
 394                         return -1;
 395
 396                 if (unlikely(fill_vec_buf(vq, res_cur_idx, &allocated,
 397                                           &vec_idx) < 0))
 398                         return -1;
 399
 400                 res_cur_idx++;
 401                 tries++;
 402
 403                 if (allocated >= size)
 404                         break;
 405
 406                 /*
 407                  * if we tried all available ring items, and still
 408                  * can't get enough buf, it means something abnormal
 409                  * happened.
 410                  */
 411                 if (unlikely(tries >= vq->size))
 412                         return -1;
 413         }
 414
 415         /*
 416          * update vq->last_used_idx_res atomically.
 417          * retry again if failed.
 418          */
 419         if (rte_atomic16_cmpset(&vq->last_used_idx_res,
 420                                 res_start_idx, res_cur_idx) == 0)
 421                 goto again;
 422
 423         *start = res_start_idx;
 424         *end   = res_cur_idx;
 425         return 0;
 426 }
 427
 428 static inline uint32_t __attribute__((always_inline))
 429 copy_mbuf_to_desc_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
 430                             uint16_t res_start_idx, uint16_t res_end_idx,
 431                             struct rte_mbuf *m)
 432 {
 433         struct virtio_net_hdr_mrg_rxbuf virtio_hdr = {{0, 0, 0, 0, 0, 0}, 0};
 434         uint32_t vec_idx = 0;
 435         uint16_t cur_idx = res_start_idx;
 436         uint64_t desc_addr;
 437         uint32_t mbuf_offset, mbuf_avail;
 438         uint32_t desc_offset, desc_avail;
 439         uint32_t cpy_len;
 440         uint16_t desc_idx, used_idx;
 441
 442         if (unlikely(m == NULL))
 443                 return 0;
 444
 445         LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
 446                 dev->vid, cur_idx, res_end_idx);
 447
 448         if (vq->buf_vec[vec_idx].buf_len < vq->vhost_hlen)
 449                 return -1;
 450
 451         desc_addr = gpa_to_vva(dev, vq->buf_vec[vec_idx].buf_addr);
 452         rte_prefetch0((void *)(uintptr_t)desc_addr);
 453
 454         virtio_hdr.num_buffers = res_end_idx - res_start_idx;
 455         LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
 456                 dev->vid, virtio_hdr.num_buffers);
 457
 458         virtio_enqueue_offload(m, &virtio_hdr.hdr);
 459         copy_virtio_net_hdr(vq, desc_addr, virtio_hdr);
 460         vhost_log_write(dev, vq->buf_vec[vec_idx].buf_addr, vq->vhost_hlen);
 461         PRINT_PACKET(dev, (uintptr_t)desc_addr, vq->vhost_hlen, 0);
 462
 463         desc_avail  = vq->buf_vec[vec_idx].buf_len - vq->vhost_hlen;
 464         desc_offset = vq->vhost_hlen;
 465
 466         mbuf_avail  = rte_pktmbuf_data_len(m);
 467         mbuf_offset = 0;
 468         while (mbuf_avail != 0 || m->next != NULL) {
 469                 /* done with current desc buf, get the next one */
 470                 if (desc_avail == 0) {
 471                         desc_idx = vq->buf_vec[vec_idx].desc_idx;
 472
 473                         if (!(vq->desc[desc_idx].flags & VRING_DESC_F_NEXT)) {
 474                                 /* Update used ring with desc information */
 475                                 used_idx = cur_idx++ & (vq->size - 1);
 476                                 vq->used->ring[used_idx].id  = desc_idx;
 477                                 vq->used->ring[used_idx].len = desc_offset;
 478                                 vhost_log_used_vring(dev, vq,
 479                                         offsetof(struct vring_used,
 480                                                  ring[used_idx]),
 481                                         sizeof(vq->used->ring[used_idx]));
 482                         }
 483
 484                         vec_idx++;
 485                         desc_addr = gpa_to_vva(dev, vq->buf_vec[vec_idx].buf_addr);
 486
 487                         /* Prefetch buffer address. */
 488                         rte_prefetch0((void *)(uintptr_t)desc_addr);
 489                         desc_offset = 0;
 490                         desc_avail  = vq->buf_vec[vec_idx].buf_len;
 491                 }
 492
 493                 /* done with current mbuf, get the next one */
 494                 if (mbuf_avail == 0) {
 495                         m = m->next;
 496
 497                         mbuf_offset = 0;
 498                         mbuf_avail  = rte_pktmbuf_data_len(m);
 499                 }
 500
 501                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 502                 rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)),
 503                         rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 504                         cpy_len);
 505                 vhost_log_write(dev, vq->buf_vec[vec_idx].buf_addr + desc_offset,
 506                         cpy_len);
 507                 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
 508                         cpy_len, 0);
 509
 510                 mbuf_avail  -= cpy_len;
 511                 mbuf_offset += cpy_len;
 512                 desc_avail  -= cpy_len;
 513                 desc_offset += cpy_len;
 514         }
 515
 516         used_idx = cur_idx & (vq->size - 1);
 517         vq->used->ring[used_idx].id = vq->buf_vec[vec_idx].desc_idx;
 518         vq->used->ring[used_idx].len = desc_offset;
 519         vhost_log_used_vring(dev, vq,
 520                 offsetof(struct vring_used, ring[used_idx]),
 521                 sizeof(vq->used->ring[used_idx]));
 522
 523         return res_end_idx - res_start_idx;
 524 }
 525
 526 static inline uint32_t __attribute__((always_inline))
 527 virtio_dev_merge_rx(struct virtio_net *dev, uint16_t queue_id,
 528         struct rte_mbuf **pkts, uint32_t count)
 529 {
 530         struct vhost_virtqueue *vq;
 531         uint32_t pkt_idx = 0, nr_used = 0;
 532         uint16_t start, end;
 533
 534         LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
 535         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->virt_qp_nb))) {
 536                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
 537                         dev->vid, __func__, queue_id);
 538                 return 0;
 539         }
 540
 541         vq = dev->virtqueue[queue_id];
 542         if (unlikely(vq->enabled == 0))
 543                 return 0;
 544
 545         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
 546         if (count == 0)
 547                 return 0;
 548
 549         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
 550                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + vq->vhost_hlen;
 551
 552                 if (unlikely(reserve_avail_buf_mergeable(vq, pkt_len,
 553                                                          &start, &end) < 0)) {
 554                         LOG_DEBUG(VHOST_DATA,
 555                                 "(%d) failed to get enough desc from vring\n",
 556                                 dev->vid);
 557                         break;
 558                 }
 559
 560                 nr_used = copy_mbuf_to_desc_mergeable(dev, vq, start, end,
 561                                                       pkts[pkt_idx]);
 562                 rte_smp_wmb();
 563
 564                 /*
 565                  * Wait until it's our turn to add our buffer
 566                  * to the used ring.
 567                  */
 568                 while (unlikely(vq->last_used_idx != start))
 569                         rte_pause();
 570
 571                 *(volatile uint16_t *)&vq->used->idx += nr_used;
 572                 vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
 573                         sizeof(vq->used->idx));
 574                 vq->last_used_idx = end;
 575         }
 576
 577         if (likely(pkt_idx)) {
 578                 /* flush used->idx update before we read avail->flags. */
 579                 rte_mb();
 580
 581                 /* Kick the guest if necessary. */
 582                 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
 583                                 && (vq->callfd >= 0))
 584                         eventfd_write(vq->callfd, (eventfd_t)1);
 585         }
 586
 587         return pkt_idx;
 588 }
 589
 590 uint16_t
 591 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
 592         struct rte_mbuf **pkts, uint16_t count)
 593 {
 594         struct virtio_net *dev = get_device(vid);
 595
 596         if (!dev)
 597                 return 0;
 598
 599         if (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF))
 600                 return virtio_dev_merge_rx(dev, queue_id, pkts, count);
 601         else
 602                 return virtio_dev_rx(dev, queue_id, pkts, count);
 603 }
 604
 605 static void
 606 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
 607 {
 608         struct ipv4_hdr *ipv4_hdr;
 609         struct ipv6_hdr *ipv6_hdr;
 610         void *l3_hdr = NULL;
 611         struct ether_hdr *eth_hdr;
 612         uint16_t ethertype;
 613
 614         eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
 615
 616         m->l2_len = sizeof(struct ether_hdr);
 617         ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
 618
 619         if (ethertype == ETHER_TYPE_VLAN) {
 620                 struct vlan_hdr *vlan_hdr = (struct vlan_hdr *)(eth_hdr + 1);
 621
 622                 m->l2_len += sizeof(struct vlan_hdr);
 623                 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
 624         }
 625
 626         l3_hdr = (char *)eth_hdr + m->l2_len;
 627
 628         switch (ethertype) {
 629         case ETHER_TYPE_IPv4:
 630                 ipv4_hdr = (struct ipv4_hdr *)l3_hdr;
 631                 *l4_proto = ipv4_hdr->next_proto_id;
 632                 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
 633                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 634                 m->ol_flags |= PKT_TX_IPV4;
 635                 break;
 636         case ETHER_TYPE_IPv6:
 637                 ipv6_hdr = (struct ipv6_hdr *)l3_hdr;
 638                 *l4_proto = ipv6_hdr->proto;
 639                 m->l3_len = sizeof(struct ipv6_hdr);
 640                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 641                 m->ol_flags |= PKT_TX_IPV6;
 642                 break;
 643         default:
 644                 m->l3_len = 0;
 645                 *l4_proto = 0;
 646                 break;
 647         }
 648 }
 649
 650 static inline void __attribute__((always_inline))
 651 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
 652 {
 653         uint16_t l4_proto = 0;
 654         void *l4_hdr = NULL;
 655         struct tcp_hdr *tcp_hdr = NULL;
 656
 657         parse_ethernet(m, &l4_proto, &l4_hdr);
 658         if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 659                 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
 660                         switch (hdr->csum_offset) {
 661                         case (offsetof(struct tcp_hdr, cksum)):
 662                                 if (l4_proto == IPPROTO_TCP)
 663                                         m->ol_flags |= PKT_TX_TCP_CKSUM;
 664                                 break;
 665                         case (offsetof(struct udp_hdr, dgram_cksum)):
 666                                 if (l4_proto == IPPROTO_UDP)
 667                                         m->ol_flags |= PKT_TX_UDP_CKSUM;
 668                                 break;
 669                         case (offsetof(struct sctp_hdr, cksum)):
 670                                 if (l4_proto == IPPROTO_SCTP)
 671                                         m->ol_flags |= PKT_TX_SCTP_CKSUM;
 672                                 break;
 673                         default:
 674                                 break;
 675                         }
 676                 }
 677         }
 678
 679         if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 680                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 681                 case VIRTIO_NET_HDR_GSO_TCPV4:
 682                 case VIRTIO_NET_HDR_GSO_TCPV6:
 683                         tcp_hdr = (struct tcp_hdr *)l4_hdr;
 684                         m->ol_flags |= PKT_TX_TCP_SEG;
 685                         m->tso_segsz = hdr->gso_size;
 686                         m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
 687                         break;
 688                 default:
 689                         RTE_LOG(WARNING, VHOST_DATA,
 690                                 "unsupported gso type %u.\n", hdr->gso_type);
 691                         break;
 692                 }
 693         }
 694 }
 695
 696 #define RARP_PKT_SIZE   64
 697
 698 static int
 699 make_rarp_packet(struct rte_mbuf *rarp_mbuf, const struct ether_addr *mac)
 700 {
 701         struct ether_hdr *eth_hdr;
 702         struct arp_hdr  *rarp;
 703
 704         if (rarp_mbuf->buf_len < 64) {
 705                 RTE_LOG(WARNING, VHOST_DATA,
 706                         "failed to make RARP; mbuf size too small %u (< %d)\n",
 707                         rarp_mbuf->buf_len, RARP_PKT_SIZE);
 708                 return -1;
 709         }
 710
 711         /* Ethernet header. */
 712         eth_hdr = rte_pktmbuf_mtod_offset(rarp_mbuf, struct ether_hdr *, 0);
 713         memset(eth_hdr->d_addr.addr_bytes, 0xff, ETHER_ADDR_LEN);
 714         ether_addr_copy(mac, &eth_hdr->s_addr);
 715         eth_hdr->ether_type = htons(ETHER_TYPE_RARP);
 716
 717         /* RARP header. */
 718         rarp = (struct arp_hdr *)(eth_hdr + 1);
 719         rarp->arp_hrd = htons(ARP_HRD_ETHER);
 720         rarp->arp_pro = htons(ETHER_TYPE_IPv4);
 721         rarp->arp_hln = ETHER_ADDR_LEN;
 722         rarp->arp_pln = 4;
 723         rarp->arp_op  = htons(ARP_OP_REVREQUEST);
 724
 725         ether_addr_copy(mac, &rarp->arp_data.arp_sha);
 726         ether_addr_copy(mac, &rarp->arp_data.arp_tha);
 727         memset(&rarp->arp_data.arp_sip, 0x00, 4);
 728         memset(&rarp->arp_data.arp_tip, 0x00, 4);
 729
 730         rarp_mbuf->pkt_len  = rarp_mbuf->data_len = RARP_PKT_SIZE;
 731
 732         return 0;
 733 }
 734
 735 static inline int __attribute__((always_inline))
 736 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 737                   struct rte_mbuf *m, uint16_t desc_idx,
 738                   struct rte_mempool *mbuf_pool)
 739 {
 740         struct vring_desc *desc;
 741         uint64_t desc_addr;
 742         uint32_t desc_avail, desc_offset;
 743         uint32_t mbuf_avail, mbuf_offset;
 744         uint32_t cpy_len;
 745         struct rte_mbuf *cur = m, *prev = m;
 746         struct virtio_net_hdr *hdr;
 747         /* A counter to avoid desc dead loop chain */
 748         uint32_t nr_desc = 1;
 749
 750         desc = &vq->desc[desc_idx];
 751         if (unlikely(desc->len < vq->vhost_hlen))
 752                 return -1;
 753
 754         desc_addr = gpa_to_vva(dev, desc->addr);
 755         rte_prefetch0((void *)(uintptr_t)desc_addr);
 756
 757         /* Retrieve virtio net header */
 758         hdr = (struct virtio_net_hdr *)((uintptr_t)desc_addr);
 759         desc_avail  = desc->len - vq->vhost_hlen;
 760         desc_offset = vq->vhost_hlen;
 761
 762         mbuf_offset = 0;
 763         mbuf_avail  = m->buf_len - RTE_PKTMBUF_HEADROOM;
 764         while (desc_avail != 0 || (desc->flags & VRING_DESC_F_NEXT) != 0) {
 765                 /* This desc reaches to its end, get the next one */
 766                 if (desc_avail == 0) {
 767                         if (unlikely(desc->next >= vq->size ||
 768                                      ++nr_desc >= vq->size))
 769                                 return -1;
 770                         desc = &vq->desc[desc->next];
 771
 772                         desc_addr = gpa_to_vva(dev, desc->addr);
 773                         rte_prefetch0((void *)(uintptr_t)desc_addr);
 774
 775                         desc_offset = 0;
 776                         desc_avail  = desc->len;
 777
 778                         PRINT_PACKET(dev, (uintptr_t)desc_addr, desc->len, 0);
 779                 }
 780
 781                 /*
 782                  * This mbuf reaches to its end, get a new one
 783                  * to hold more data.
 784                  */
 785                 if (mbuf_avail == 0) {
 786                         cur = rte_pktmbuf_alloc(mbuf_pool);
 787                         if (unlikely(cur == NULL)) {
 788                                 RTE_LOG(ERR, VHOST_DATA, "Failed to "
 789                                         "allocate memory for mbuf.\n");
 790                                 return -1;
 791                         }
 792
 793                         prev->next = cur;
 794                         prev->data_len = mbuf_offset;
 795                         m->nb_segs += 1;
 796                         m->pkt_len += mbuf_offset;
 797                         prev = cur;
 798
 799                         mbuf_offset = 0;
 800                         mbuf_avail  = cur->buf_len - RTE_PKTMBUF_HEADROOM;
 801                 }
 802
 803                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 804                 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *, mbuf_offset),
 805                         (void *)((uintptr_t)(desc_addr + desc_offset)),
 806                         cpy_len);
 807
 808                 mbuf_avail  -= cpy_len;
 809                 mbuf_offset += cpy_len;
 810                 desc_avail  -= cpy_len;
 811                 desc_offset += cpy_len;
 812         }
 813
 814         prev->data_len = mbuf_offset;
 815         m->pkt_len    += mbuf_offset;
 816
 817         if (hdr->flags != 0 || hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE)
 818                 vhost_dequeue_offload(hdr, m);
 819
 820         return 0;
 821 }
 822
 823 uint16_t
 824 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
 825         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
 826 {
 827         struct virtio_net *dev;
 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         dev = get_device(vid);
 837         if (!dev)
 838                 return 0;
 839
 840         if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->virt_qp_nb))) {
 841                 RTE_LOG(ERR, VHOST_DATA, "(%d) %s: invalid virtqueue idx %d.\n",
 842                         dev->vid, __func__, queue_id);
 843                 return 0;
 844         }
 845
 846         vq = dev->virtqueue[queue_id];
 847         if (unlikely(vq->enabled == 0))
 848                 return 0;
 849
 850         /*
 851          * Construct a RARP broadcast packet, and inject it to the "pkts"
 852          * array, to looks like that guest actually send such packet.
 853          *
 854          * Check user_send_rarp() for more information.
 855          */
 856         if (unlikely(rte_atomic16_cmpset((volatile uint16_t *)
 857                                          &dev->broadcast_rarp.cnt, 1, 0))) {
 858                 rarp_mbuf = rte_pktmbuf_alloc(mbuf_pool);
 859                 if (rarp_mbuf == NULL) {
 860                         RTE_LOG(ERR, VHOST_DATA,
 861                                 "Failed to allocate memory for mbuf.\n");
 862                         return 0;
 863                 }
 864
 865                 if (make_rarp_packet(rarp_mbuf, &dev->mac)) {
 866                         rte_pktmbuf_free(rarp_mbuf);
 867                         rarp_mbuf = NULL;
 868                 } else {
 869                         count -= 1;
 870                 }
 871         }
 872
 873         avail_idx =  *((volatile uint16_t *)&vq->avail->idx);
 874         free_entries = avail_idx - vq->last_used_idx;
 875         if (free_entries == 0)
 876                 goto out;
 877
 878         LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
 879
 880         /* Prefetch available ring to retrieve head indexes. */
 881         used_idx = vq->last_used_idx & (vq->size - 1);
 882         rte_prefetch0(&vq->avail->ring[used_idx]);
 883
 884         count = RTE_MIN(count, MAX_PKT_BURST);
 885         count = RTE_MIN(count, free_entries);
 886         LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
 887                         dev->vid, count);
 888
 889         /* Retrieve all of the head indexes first to avoid caching issues. */
 890         for (i = 0; i < count; i++) {
 891                 desc_indexes[i] = vq->avail->ring[(vq->last_used_idx + i) &
 892                                         (vq->size - 1)];
 893         }
 894
 895         /* Prefetch descriptor index. */
 896         rte_prefetch0(&vq->desc[desc_indexes[0]]);
 897         rte_prefetch0(&vq->used->ring[vq->last_used_idx & (vq->size - 1)]);
 898
 899         for (i = 0; i < count; i++) {
 900                 int err;
 901
 902                 if (likely(i + 1 < count)) {
 903                         rte_prefetch0(&vq->desc[desc_indexes[i + 1]]);
 904                         rte_prefetch0(&vq->used->ring[(used_idx + 1) &
 905                                                       (vq->size - 1)]);
 906                 }
 907
 908                 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
 909                 if (unlikely(pkts[i] == NULL)) {
 910                         RTE_LOG(ERR, VHOST_DATA,
 911                                 "Failed to allocate memory for mbuf.\n");
 912                         break;
 913                 }
 914                 err = copy_desc_to_mbuf(dev, vq, pkts[i], desc_indexes[i],
 915                                         mbuf_pool);
 916                 if (unlikely(err)) {
 917                         rte_pktmbuf_free(pkts[i]);
 918                         break;
 919                 }
 920
 921                 used_idx = vq->last_used_idx++ & (vq->size - 1);
 922                 vq->used->ring[used_idx].id  = desc_indexes[i];
 923                 vq->used->ring[used_idx].len = 0;
 924                 vhost_log_used_vring(dev, vq,
 925                                 offsetof(struct vring_used, ring[used_idx]),
 926                                 sizeof(vq->used->ring[used_idx]));
 927         }
 928
 929         rte_smp_wmb();
 930         rte_smp_rmb();
 931         vq->used->idx += i;
 932         vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
 933                         sizeof(vq->used->idx));
 934
 935         /* Kick guest if required. */
 936         if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
 937                         && (vq->callfd >= 0))
 938                 eventfd_write(vq->callfd, (eventfd_t)1);
 939
 940 out:
 941         if (unlikely(rarp_mbuf != NULL)) {
 942                 /*
 943                  * Inject it to the head of "pkts" array, so that switch's mac
 944                  * learning table will get updated first.
 945                  */
 946                 memmove(&pkts[1], pkts, i * sizeof(struct rte_mbuf *));
 947                 pkts[0] = rarp_mbuf;
 948                 i += 1;
 949         }
 950
 951         return i;
 952 }