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         if (unlikely(desc->len < vq->vhost_hlen))
 155                 return -1;
 156
 157         desc_addr = gpa_to_vva(dev, desc->addr);
 158         rte_prefetch0((void *)(uintptr_t)desc_addr);
 159
 160         virtio_enqueue_offload(m, &virtio_hdr.hdr);
 161         copy_virtio_net_hdr(vq, desc_addr, virtio_hdr);
 162         vhost_log_write(dev, desc->addr, vq->vhost_hlen);
 163         PRINT_PACKET(dev, (uintptr_t)desc_addr, vq->vhost_hlen, 0);
 164
 165         desc_offset = vq->vhost_hlen;
 166         desc_avail  = desc->len - vq->vhost_hlen;
 167
 168         *copied = rte_pktmbuf_pkt_len(m);
 169         mbuf_avail  = rte_pktmbuf_data_len(m);
 170         mbuf_offset = 0;
 171         while (mbuf_avail != 0 || m->next != NULL) {
 172                 /* done with current mbuf, fetch next */
 173                 if (mbuf_avail == 0) {
 174                         m = m->next;
 175
 176                         mbuf_offset = 0;
 177                         mbuf_avail  = rte_pktmbuf_data_len(m);
 178                 }
 179
 180                 /* done with current desc buf, fetch next */
 181                 if (desc_avail == 0) {
 182                         if ((desc->flags & VRING_DESC_F_NEXT) == 0) {
 183                                 /* Room in vring buffer is not enough */
 184                                 return -1;
 185                         }
 186                         if (unlikely(desc->next >= vq->size))
 187                                 return -1;
 188
 189                         desc = &vq->desc[desc->next];
 190                         desc_addr   = gpa_to_vva(dev, desc->addr);
 191                         desc_offset = 0;
 192                         desc_avail  = desc->len;
 193                 }
 194
 195                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 196                 rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)),
 197                         rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 198                         cpy_len);
 199                 vhost_log_write(dev, desc->addr + desc_offset, cpy_len);
 200                 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
 201                              cpy_len, 0);
 202
 203                 mbuf_avail  -= cpy_len;
 204                 mbuf_offset += cpy_len;
 205                 desc_avail  -= cpy_len;
 206                 desc_offset += cpy_len;
 207         }
 208
 209         return 0;
 210 }
 211
 212 /*
 213  * As many data cores may want to access available buffers
 214  * they need to be reserved.
 215  */
 216 static inline uint32_t
 217 reserve_avail_buf(struct vhost_virtqueue *vq, uint32_t count,
 218                   uint16_t *start, uint16_t *end)
 219 {
 220         uint16_t res_start_idx;
 221         uint16_t res_end_idx;
 222         uint16_t avail_idx;
 223         uint16_t free_entries;
 224         int success;
 225
 226         count = RTE_MIN(count, (uint32_t)MAX_PKT_BURST);
 227
 228 again:
 229         res_start_idx = vq->last_used_idx_res;
 230         avail_idx = *((volatile uint16_t *)&vq->avail->idx);
 231
 232         free_entries = avail_idx - res_start_idx;
 233         count = RTE_MIN(count, free_entries);
 234         if (count == 0)
 235                 return 0;
 236
 237         res_end_idx = res_start_idx + count;
 238
 239         /*
 240          * update vq->last_used_idx_res atomically; try again if failed.
 241          *
 242          * TODO: Allow to disable cmpset if no concurrency in application.
 243          */
 244         success = rte_atomic16_cmpset(&vq->last_used_idx_res,
 245                                       res_start_idx, res_end_idx);
 246         if (unlikely(!success))
 247                 goto again;
 248
 249         *start = res_start_idx;
 250         *end   = res_end_idx;
 251
 252         return count;
 253 }
 254
 255 /**
 256  * This function adds buffers to the virtio devices RX virtqueue. Buffers can
 257  * be received from the physical port or from another virtio device. A packet
 258  * count is returned to indicate the number of packets that are succesfully
 259  * added to the RX queue. This function works when the mbuf is scattered, but
 260  * it doesn't support the mergeable feature.
 261  */
 262 static inline uint32_t __attribute__((always_inline))
 263 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
 264               struct rte_mbuf **pkts, uint32_t count)
 265 {
 266         struct vhost_virtqueue *vq;
 267         uint16_t res_start_idx, res_end_idx;
 268         uint16_t desc_indexes[MAX_PKT_BURST];
 269         uint32_t i;
 270
 271         LOG_DEBUG(VHOST_DATA, "(%"PRIu64") virtio_dev_rx()\n", dev->device_fh);
 272         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->virt_qp_nb))) {
 273                 RTE_LOG(ERR, VHOST_DATA,
 274                         "%s (%"PRIu64"): virtqueue idx:%d invalid.\n",
 275                         __func__, dev->device_fh, queue_id);
 276                 return 0;
 277         }
 278
 279         vq = dev->virtqueue[queue_id];
 280         if (unlikely(vq->enabled == 0))
 281                 return 0;
 282
 283         count = reserve_avail_buf(vq, count, &res_start_idx, &res_end_idx);
 284         if (count == 0)
 285                 return 0;
 286
 287         LOG_DEBUG(VHOST_DATA,
 288                 "(%"PRIu64") res_start_idx %d| res_end_idx Index %d\n",
 289                 dev->device_fh, res_start_idx, res_end_idx);
 290
 291         /* Retrieve all of the desc indexes first to avoid caching issues. */
 292         rte_prefetch0(&vq->avail->ring[res_start_idx & (vq->size - 1)]);
 293         for (i = 0; i < count; i++) {
 294                 desc_indexes[i] = vq->avail->ring[(res_start_idx + i) &
 295                                                   (vq->size - 1)];
 296         }
 297
 298         rte_prefetch0(&vq->desc[desc_indexes[0]]);
 299         for (i = 0; i < count; i++) {
 300                 uint16_t desc_idx = desc_indexes[i];
 301                 uint16_t used_idx = (res_start_idx + i) & (vq->size - 1);
 302                 uint32_t copied;
 303                 int err;
 304
 305                 err = copy_mbuf_to_desc(dev, vq, pkts[i], desc_idx, &copied);
 306
 307                 vq->used->ring[used_idx].id = desc_idx;
 308                 if (unlikely(err))
 309                         vq->used->ring[used_idx].len = vq->vhost_hlen;
 310                 else
 311                         vq->used->ring[used_idx].len = copied + vq->vhost_hlen;
 312                 vhost_log_used_vring(dev, vq,
 313                         offsetof(struct vring_used, ring[used_idx]),
 314                         sizeof(vq->used->ring[used_idx]));
 315
 316                 if (i + 1 < count)
 317                         rte_prefetch0(&vq->desc[desc_indexes[i+1]]);
 318         }
 319
 320         rte_compiler_barrier();
 321
 322         /* Wait until it's our turn to add our buffer to the used ring. */
 323         while (unlikely(vq->last_used_idx != res_start_idx))
 324                 rte_pause();
 325
 326         *(volatile uint16_t *)&vq->used->idx += count;
 327         vq->last_used_idx = res_end_idx;
 328         vhost_log_used_vring(dev, vq,
 329                 offsetof(struct vring_used, idx),
 330                 sizeof(vq->used->idx));
 331
 332         /* flush used->idx update before we read avail->flags. */
 333         rte_mb();
 334
 335         /* Kick the guest if necessary. */
 336         if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
 337                         && (vq->callfd >= 0))
 338                 eventfd_write(vq->callfd, (eventfd_t)1);
 339         return count;
 340 }
 341
 342 static inline int
 343 fill_vec_buf(struct vhost_virtqueue *vq, uint32_t avail_idx,
 344              uint32_t *allocated, uint32_t *vec_idx)
 345 {
 346         uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
 347         uint32_t vec_id = *vec_idx;
 348         uint32_t len    = *allocated;
 349
 350         while (1) {
 351                 if (unlikely(vec_id >= BUF_VECTOR_MAX || idx >= vq->size))
 352                         return -1;
 353
 354                 len += vq->desc[idx].len;
 355                 vq->buf_vec[vec_id].buf_addr = vq->desc[idx].addr;
 356                 vq->buf_vec[vec_id].buf_len  = vq->desc[idx].len;
 357                 vq->buf_vec[vec_id].desc_idx = idx;
 358                 vec_id++;
 359
 360                 if ((vq->desc[idx].flags & VRING_DESC_F_NEXT) == 0)
 361                         break;
 362
 363                 idx = vq->desc[idx].next;
 364         }
 365
 366         *allocated = len;
 367         *vec_idx   = vec_id;
 368
 369         return 0;
 370 }
 371
 372 /*
 373  * As many data cores may want to access available buffers concurrently,
 374  * they need to be reserved.
 375  *
 376  * Returns -1 on fail, 0 on success
 377  */
 378 static inline int
 379 reserve_avail_buf_mergeable(struct vhost_virtqueue *vq, uint32_t size,
 380                             uint16_t *start, uint16_t *end)
 381 {
 382         uint16_t res_start_idx;
 383         uint16_t res_cur_idx;
 384         uint16_t avail_idx;
 385         uint32_t allocated;
 386         uint32_t vec_idx;
 387         uint16_t tries;
 388
 389 again:
 390         res_start_idx = vq->last_used_idx_res;
 391         res_cur_idx  = res_start_idx;
 392
 393         allocated = 0;
 394         vec_idx   = 0;
 395         tries     = 0;
 396         while (1) {
 397                 avail_idx = *((volatile uint16_t *)&vq->avail->idx);
 398                 if (unlikely(res_cur_idx == avail_idx))
 399                         return -1;
 400
 401                 if (unlikely(fill_vec_buf(vq, res_cur_idx, &allocated,
 402                                           &vec_idx) < 0))
 403                         return -1;
 404
 405                 res_cur_idx++;
 406                 tries++;
 407
 408                 if (allocated >= size)
 409                         break;
 410
 411                 /*
 412                  * if we tried all available ring items, and still
 413                  * can't get enough buf, it means something abnormal
 414                  * happened.
 415                  */
 416                 if (unlikely(tries >= vq->size))
 417                         return -1;
 418         }
 419
 420         /*
 421          * update vq->last_used_idx_res atomically.
 422          * retry again if failed.
 423          */
 424         if (rte_atomic16_cmpset(&vq->last_used_idx_res,
 425                                 res_start_idx, res_cur_idx) == 0)
 426                 goto again;
 427
 428         *start = res_start_idx;
 429         *end   = res_cur_idx;
 430         return 0;
 431 }
 432
 433 static inline uint32_t __attribute__((always_inline))
 434 copy_mbuf_to_desc_mergeable(struct virtio_net *dev, struct vhost_virtqueue *vq,
 435                             uint16_t res_start_idx, uint16_t res_end_idx,
 436                             struct rte_mbuf *m)
 437 {
 438         struct virtio_net_hdr_mrg_rxbuf virtio_hdr = {{0, 0, 0, 0, 0, 0}, 0};
 439         uint32_t vec_idx = 0;
 440         uint16_t cur_idx = res_start_idx;
 441         uint64_t desc_addr;
 442         uint32_t mbuf_offset, mbuf_avail;
 443         uint32_t desc_offset, desc_avail;
 444         uint32_t cpy_len;
 445         uint16_t desc_idx, used_idx;
 446
 447         if (unlikely(m == NULL))
 448                 return 0;
 449
 450         LOG_DEBUG(VHOST_DATA,
 451                 "(%"PRIu64") Current Index %d| End Index %d\n",
 452                 dev->device_fh, cur_idx, res_end_idx);
 453
 454         if (vq->buf_vec[vec_idx].buf_len < vq->vhost_hlen)
 455                 return -1;
 456
 457         desc_addr = gpa_to_vva(dev, vq->buf_vec[vec_idx].buf_addr);
 458         rte_prefetch0((void *)(uintptr_t)desc_addr);
 459
 460         virtio_hdr.num_buffers = res_end_idx - res_start_idx;
 461         LOG_DEBUG(VHOST_DATA, "(%"PRIu64") RX: Num merge buffers %d\n",
 462                 dev->device_fh, virtio_hdr.num_buffers);
 463
 464         virtio_enqueue_offload(m, &virtio_hdr.hdr);
 465         copy_virtio_net_hdr(vq, desc_addr, virtio_hdr);
 466         vhost_log_write(dev, vq->buf_vec[vec_idx].buf_addr, vq->vhost_hlen);
 467         PRINT_PACKET(dev, (uintptr_t)desc_addr, vq->vhost_hlen, 0);
 468
 469         desc_avail  = vq->buf_vec[vec_idx].buf_len - vq->vhost_hlen;
 470         desc_offset = vq->vhost_hlen;
 471
 472         mbuf_avail  = rte_pktmbuf_data_len(m);
 473         mbuf_offset = 0;
 474         while (mbuf_avail != 0 || m->next != NULL) {
 475                 /* done with current desc buf, get the next one */
 476                 if (desc_avail == 0) {
 477                         desc_idx = vq->buf_vec[vec_idx].desc_idx;
 478
 479                         if (!(vq->desc[desc_idx].flags & VRING_DESC_F_NEXT)) {
 480                                 /* Update used ring with desc information */
 481                                 used_idx = cur_idx++ & (vq->size - 1);
 482                                 vq->used->ring[used_idx].id  = desc_idx;
 483                                 vq->used->ring[used_idx].len = desc_offset;
 484                                 vhost_log_used_vring(dev, vq,
 485                                         offsetof(struct vring_used,
 486                                                  ring[used_idx]),
 487                                         sizeof(vq->used->ring[used_idx]));
 488                         }
 489
 490                         vec_idx++;
 491                         desc_addr = gpa_to_vva(dev, vq->buf_vec[vec_idx].buf_addr);
 492
 493                         /* Prefetch buffer address. */
 494                         rte_prefetch0((void *)(uintptr_t)desc_addr);
 495                         desc_offset = 0;
 496                         desc_avail  = vq->buf_vec[vec_idx].buf_len;
 497                 }
 498
 499                 /* done with current mbuf, get the next one */
 500                 if (mbuf_avail == 0) {
 501                         m = m->next;
 502
 503                         mbuf_offset = 0;
 504                         mbuf_avail  = rte_pktmbuf_data_len(m);
 505                 }
 506
 507                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 508                 rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)),
 509                         rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 510                         cpy_len);
 511                 vhost_log_write(dev, vq->buf_vec[vec_idx].buf_addr + desc_offset,
 512                         cpy_len);
 513                 PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
 514                         cpy_len, 0);
 515
 516                 mbuf_avail  -= cpy_len;
 517                 mbuf_offset += cpy_len;
 518                 desc_avail  -= cpy_len;
 519                 desc_offset += cpy_len;
 520         }
 521
 522         used_idx = cur_idx & (vq->size - 1);
 523         vq->used->ring[used_idx].id = vq->buf_vec[vec_idx].desc_idx;
 524         vq->used->ring[used_idx].len = desc_offset;
 525         vhost_log_used_vring(dev, vq,
 526                 offsetof(struct vring_used, ring[used_idx]),
 527                 sizeof(vq->used->ring[used_idx]));
 528
 529         return res_end_idx - res_start_idx;
 530 }
 531
 532 static inline uint32_t __attribute__((always_inline))
 533 virtio_dev_merge_rx(struct virtio_net *dev, uint16_t queue_id,
 534         struct rte_mbuf **pkts, uint32_t count)
 535 {
 536         struct vhost_virtqueue *vq;
 537         uint32_t pkt_idx = 0, nr_used = 0;
 538         uint16_t start, end;
 539
 540         LOG_DEBUG(VHOST_DATA, "(%"PRIu64") virtio_dev_merge_rx()\n",
 541                 dev->device_fh);
 542         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->virt_qp_nb))) {
 543                 RTE_LOG(ERR, VHOST_DATA,
 544                         "%s (%"PRIu64"): virtqueue idx:%d invalid.\n",
 545                         __func__, dev->device_fh, queue_id);
 546                 return 0;
 547         }
 548
 549         vq = dev->virtqueue[queue_id];
 550         if (unlikely(vq->enabled == 0))
 551                 return 0;
 552
 553         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
 554         if (count == 0)
 555                 return 0;
 556
 557         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
 558                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + vq->vhost_hlen;
 559
 560                 if (unlikely(reserve_avail_buf_mergeable(vq, pkt_len,
 561                                                          &start, &end) < 0)) {
 562                         LOG_DEBUG(VHOST_DATA,
 563                                 "(%" PRIu64 ") Failed to get enough desc from vring\n",
 564                                 dev->device_fh);
 565                         break;
 566                 }
 567
 568                 nr_used = copy_mbuf_to_desc_mergeable(dev, vq, start, end,
 569                                                       pkts[pkt_idx]);
 570                 rte_compiler_barrier();
 571
 572                 /*
 573                  * Wait until it's our turn to add our buffer
 574                  * to the used ring.
 575                  */
 576                 while (unlikely(vq->last_used_idx != start))
 577                         rte_pause();
 578
 579                 *(volatile uint16_t *)&vq->used->idx += nr_used;
 580                 vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
 581                         sizeof(vq->used->idx));
 582                 vq->last_used_idx = end;
 583         }
 584
 585         if (likely(pkt_idx)) {
 586                 /* flush used->idx update before we read avail->flags. */
 587                 rte_mb();
 588
 589                 /* Kick the guest if necessary. */
 590                 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
 591                                 && (vq->callfd >= 0))
 592                         eventfd_write(vq->callfd, (eventfd_t)1);
 593         }
 594
 595         return pkt_idx;
 596 }
 597
 598 uint16_t
 599 rte_vhost_enqueue_burst(struct virtio_net *dev, uint16_t queue_id,
 600         struct rte_mbuf **pkts, uint16_t count)
 601 {
 602         if (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF))
 603                 return virtio_dev_merge_rx(dev, queue_id, pkts, count);
 604         else
 605                 return virtio_dev_rx(dev, queue_id, pkts, count);
 606 }
 607
 608 static void
 609 parse_ethernet(struct rte_mbuf *m, uint16_t *l4_proto, void **l4_hdr)
 610 {
 611         struct ipv4_hdr *ipv4_hdr;
 612         struct ipv6_hdr *ipv6_hdr;
 613         void *l3_hdr = NULL;
 614         struct ether_hdr *eth_hdr;
 615         uint16_t ethertype;
 616
 617         eth_hdr = rte_pktmbuf_mtod(m, struct ether_hdr *);
 618
 619         m->l2_len = sizeof(struct ether_hdr);
 620         ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
 621
 622         if (ethertype == ETHER_TYPE_VLAN) {
 623                 struct vlan_hdr *vlan_hdr = (struct vlan_hdr *)(eth_hdr + 1);
 624
 625                 m->l2_len += sizeof(struct vlan_hdr);
 626                 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
 627         }
 628
 629         l3_hdr = (char *)eth_hdr + m->l2_len;
 630
 631         switch (ethertype) {
 632         case ETHER_TYPE_IPv4:
 633                 ipv4_hdr = (struct ipv4_hdr *)l3_hdr;
 634                 *l4_proto = ipv4_hdr->next_proto_id;
 635                 m->l3_len = (ipv4_hdr->version_ihl & 0x0f) * 4;
 636                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 637                 m->ol_flags |= PKT_TX_IPV4;
 638                 break;
 639         case ETHER_TYPE_IPv6:
 640                 ipv6_hdr = (struct ipv6_hdr *)l3_hdr;
 641                 *l4_proto = ipv6_hdr->proto;
 642                 m->l3_len = sizeof(struct ipv6_hdr);
 643                 *l4_hdr = (char *)l3_hdr + m->l3_len;
 644                 m->ol_flags |= PKT_TX_IPV6;
 645                 break;
 646         default:
 647                 m->l3_len = 0;
 648                 *l4_proto = 0;
 649                 break;
 650         }
 651 }
 652
 653 static inline void __attribute__((always_inline))
 654 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
 655 {
 656         uint16_t l4_proto = 0;
 657         void *l4_hdr = NULL;
 658         struct tcp_hdr *tcp_hdr = NULL;
 659
 660         parse_ethernet(m, &l4_proto, &l4_hdr);
 661         if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 662                 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
 663                         switch (hdr->csum_offset) {
 664                         case (offsetof(struct tcp_hdr, cksum)):
 665                                 if (l4_proto == IPPROTO_TCP)
 666                                         m->ol_flags |= PKT_TX_TCP_CKSUM;
 667                                 break;
 668                         case (offsetof(struct udp_hdr, dgram_cksum)):
 669                                 if (l4_proto == IPPROTO_UDP)
 670                                         m->ol_flags |= PKT_TX_UDP_CKSUM;
 671                                 break;
 672                         case (offsetof(struct sctp_hdr, cksum)):
 673                                 if (l4_proto == IPPROTO_SCTP)
 674                                         m->ol_flags |= PKT_TX_SCTP_CKSUM;
 675                                 break;
 676                         default:
 677                                 break;
 678                         }
 679                 }
 680         }
 681
 682         if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 683                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 684                 case VIRTIO_NET_HDR_GSO_TCPV4:
 685                 case VIRTIO_NET_HDR_GSO_TCPV6:
 686                         tcp_hdr = (struct tcp_hdr *)l4_hdr;
 687                         m->ol_flags |= PKT_TX_TCP_SEG;
 688                         m->tso_segsz = hdr->gso_size;
 689                         m->l4_len = (tcp_hdr->data_off & 0xf0) >> 2;
 690                         break;
 691                 default:
 692                         RTE_LOG(WARNING, VHOST_DATA,
 693                                 "unsupported gso type %u.\n", hdr->gso_type);
 694                         break;
 695                 }
 696         }
 697 }
 698
 699 #define RARP_PKT_SIZE   64
 700
 701 static int
 702 make_rarp_packet(struct rte_mbuf *rarp_mbuf, const struct ether_addr *mac)
 703 {
 704         struct ether_hdr *eth_hdr;
 705         struct arp_hdr  *rarp;
 706
 707         if (rarp_mbuf->buf_len < 64) {
 708                 RTE_LOG(WARNING, VHOST_DATA,
 709                         "failed to make RARP; mbuf size too small %u (< %d)\n",
 710                         rarp_mbuf->buf_len, RARP_PKT_SIZE);
 711                 return -1;
 712         }
 713
 714         /* Ethernet header. */
 715         eth_hdr = rte_pktmbuf_mtod_offset(rarp_mbuf, struct ether_hdr *, 0);
 716         memset(eth_hdr->d_addr.addr_bytes, 0xff, ETHER_ADDR_LEN);
 717         ether_addr_copy(mac, &eth_hdr->s_addr);
 718         eth_hdr->ether_type = htons(ETHER_TYPE_RARP);
 719
 720         /* RARP header. */
 721         rarp = (struct arp_hdr *)(eth_hdr + 1);
 722         rarp->arp_hrd = htons(ARP_HRD_ETHER);
 723         rarp->arp_pro = htons(ETHER_TYPE_IPv4);
 724         rarp->arp_hln = ETHER_ADDR_LEN;
 725         rarp->arp_pln = 4;
 726         rarp->arp_op  = htons(ARP_OP_REVREQUEST);
 727
 728         ether_addr_copy(mac, &rarp->arp_data.arp_sha);
 729         ether_addr_copy(mac, &rarp->arp_data.arp_tha);
 730         memset(&rarp->arp_data.arp_sip, 0x00, 4);
 731         memset(&rarp->arp_data.arp_tip, 0x00, 4);
 732
 733         rarp_mbuf->pkt_len  = rarp_mbuf->data_len = RARP_PKT_SIZE;
 734
 735         return 0;
 736 }
 737
 738 static inline int __attribute__((always_inline))
 739 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 740                   struct rte_mbuf *m, uint16_t desc_idx,
 741                   struct rte_mempool *mbuf_pool)
 742 {
 743         struct vring_desc *desc;
 744         uint64_t desc_addr;
 745         uint32_t desc_avail, desc_offset;
 746         uint32_t mbuf_avail, mbuf_offset;
 747         uint32_t cpy_len;
 748         struct rte_mbuf *cur = m, *prev = m;
 749         struct virtio_net_hdr *hdr;
 750         /* A counter to avoid desc dead loop chain */
 751         uint32_t nr_desc = 1;
 752
 753         desc = &vq->desc[desc_idx];
 754         if (unlikely(desc->len < vq->vhost_hlen))
 755                 return -1;
 756
 757         desc_addr = gpa_to_vva(dev, desc->addr);
 758         rte_prefetch0((void *)(uintptr_t)desc_addr);
 759
 760         /* Retrieve virtio net header */
 761         hdr = (struct virtio_net_hdr *)((uintptr_t)desc_addr);
 762         desc_avail  = desc->len - vq->vhost_hlen;
 763         desc_offset = vq->vhost_hlen;
 764
 765         mbuf_offset = 0;
 766         mbuf_avail  = m->buf_len - RTE_PKTMBUF_HEADROOM;
 767         while (desc_avail != 0 || (desc->flags & VRING_DESC_F_NEXT) != 0) {
 768                 /* This desc reaches to its end, get the next one */
 769                 if (desc_avail == 0) {
 770                         if (unlikely(desc->next >= vq->size ||
 771                                      ++nr_desc >= vq->size))
 772                                 return -1;
 773                         desc = &vq->desc[desc->next];
 774
 775                         desc_addr = gpa_to_vva(dev, desc->addr);
 776                         rte_prefetch0((void *)(uintptr_t)desc_addr);
 777
 778                         desc_offset = 0;
 779                         desc_avail  = desc->len;
 780
 781                         PRINT_PACKET(dev, (uintptr_t)desc_addr, desc->len, 0);
 782                 }
 783
 784                 /*
 785                  * This mbuf reaches to its end, get a new one
 786                  * to hold more data.
 787                  */
 788                 if (mbuf_avail == 0) {
 789                         cur = rte_pktmbuf_alloc(mbuf_pool);
 790                         if (unlikely(cur == NULL)) {
 791                                 RTE_LOG(ERR, VHOST_DATA, "Failed to "
 792                                         "allocate memory for mbuf.\n");
 793                                 return -1;
 794                         }
 795
 796                         prev->next = cur;
 797                         prev->data_len = mbuf_offset;
 798                         m->nb_segs += 1;
 799                         m->pkt_len += mbuf_offset;
 800                         prev = cur;
 801
 802                         mbuf_offset = 0;
 803                         mbuf_avail  = cur->buf_len - RTE_PKTMBUF_HEADROOM;
 804                 }
 805
 806                 cpy_len = RTE_MIN(desc_avail, mbuf_avail);
 807                 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *, mbuf_offset),
 808                         (void *)((uintptr_t)(desc_addr + desc_offset)),
 809                         cpy_len);
 810
 811                 mbuf_avail  -= cpy_len;
 812                 mbuf_offset += cpy_len;
 813                 desc_avail  -= cpy_len;
 814                 desc_offset += cpy_len;
 815         }
 816
 817         prev->data_len = mbuf_offset;
 818         m->pkt_len    += mbuf_offset;
 819
 820         if (hdr->flags != 0 || hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE)
 821                 vhost_dequeue_offload(hdr, m);
 822
 823         return 0;
 824 }
 825
 826 uint16_t
 827 rte_vhost_dequeue_burst(struct virtio_net *dev, uint16_t queue_id,
 828         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
 829 {
 830         struct rte_mbuf *rarp_mbuf = NULL;
 831         struct vhost_virtqueue *vq;
 832         uint32_t desc_indexes[MAX_PKT_BURST];
 833         uint32_t used_idx;
 834         uint32_t i = 0;
 835         uint16_t free_entries;
 836         uint16_t avail_idx;
 837
 838         if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->virt_qp_nb))) {
 839                 RTE_LOG(ERR, VHOST_DATA,
 840                         "%s (%"PRIu64"): virtqueue idx:%d invalid.\n",
 841                         __func__, dev->device_fh, queue_id);
 842                 return 0;
 843         }
 844
 845         vq = dev->virtqueue[queue_id];
 846         if (unlikely(vq->enabled == 0))
 847                 return 0;
 848
 849         /*
 850          * Construct a RARP broadcast packet, and inject it to the "pkts"
 851          * array, to looks like that guest actually send such packet.
 852          *
 853          * Check user_send_rarp() for more information.
 854          */
 855         if (unlikely(rte_atomic16_cmpset((volatile uint16_t *)
 856                                          &dev->broadcast_rarp.cnt, 1, 0))) {
 857                 rarp_mbuf = rte_pktmbuf_alloc(mbuf_pool);
 858                 if (rarp_mbuf == NULL) {
 859                         RTE_LOG(ERR, VHOST_DATA,
 860                                 "Failed to allocate memory for mbuf.\n");
 861                         return 0;
 862                 }
 863
 864                 if (make_rarp_packet(rarp_mbuf, &dev->mac)) {
 865                         rte_pktmbuf_free(rarp_mbuf);
 866                         rarp_mbuf = NULL;
 867                 } else {
 868                         count -= 1;
 869                 }
 870         }
 871
 872         avail_idx =  *((volatile uint16_t *)&vq->avail->idx);
 873         free_entries = avail_idx - vq->last_used_idx;
 874         if (free_entries == 0)
 875                 goto out;
 876
 877         LOG_DEBUG(VHOST_DATA, "%s (%"PRIu64")\n", __func__, dev->device_fh);
 878
 879         /* Prefetch available ring to retrieve head indexes. */
 880         used_idx = vq->last_used_idx & (vq->size - 1);
 881         rte_prefetch0(&vq->avail->ring[used_idx]);
 882
 883         count = RTE_MIN(count, MAX_PKT_BURST);
 884         count = RTE_MIN(count, free_entries);
 885         LOG_DEBUG(VHOST_DATA, "(%"PRIu64") about to dequeue %u buffers\n",
 886                         dev->device_fh, count);
 887
 888         /* Retrieve all of the head indexes first to avoid caching issues. */
 889         for (i = 0; i < count; i++) {
 890                 desc_indexes[i] = vq->avail->ring[(vq->last_used_idx + i) &
 891                                         (vq->size - 1)];
 892         }
 893
 894         /* Prefetch descriptor index. */
 895         rte_prefetch0(&vq->desc[desc_indexes[0]]);
 896         rte_prefetch0(&vq->used->ring[vq->last_used_idx & (vq->size - 1)]);
 897
 898         for (i = 0; i < count; i++) {
 899                 int err;
 900
 901                 if (likely(i + 1 < count)) {
 902                         rte_prefetch0(&vq->desc[desc_indexes[i + 1]]);
 903                         rte_prefetch0(&vq->used->ring[(used_idx + 1) &
 904                                                       (vq->size - 1)]);
 905                 }
 906
 907                 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
 908                 if (unlikely(pkts[i] == NULL)) {
 909                         RTE_LOG(ERR, VHOST_DATA,
 910                                 "Failed to allocate memory for mbuf.\n");
 911                         break;
 912                 }
 913                 err = copy_desc_to_mbuf(dev, vq, pkts[i], desc_indexes[i],
 914                                         mbuf_pool);
 915                 if (unlikely(err)) {
 916                         rte_pktmbuf_free(pkts[i]);
 917                         break;
 918                 }
 919
 920                 used_idx = vq->last_used_idx++ & (vq->size - 1);
 921                 vq->used->ring[used_idx].id  = desc_indexes[i];
 922                 vq->used->ring[used_idx].len = 0;
 923                 vhost_log_used_vring(dev, vq,
 924                                 offsetof(struct vring_used, ring[used_idx]),
 925                                 sizeof(vq->used->ring[used_idx]));
 926         }
 927
 928         rte_compiler_barrier();
 929         vq->used->idx += i;
 930         vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
 931                         sizeof(vq->used->idx));
 932
 933         /* Kick guest if required. */
 934         if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT)
 935                         && (vq->callfd >= 0))
 936                 eventfd_write(vq->callfd, (eventfd_t)1);
 937
 938 out:
 939         if (unlikely(rarp_mbuf != NULL)) {
 940                 /*
 941                  * Inject it to the head of "pkts" array, so that switch's mac
 942                  * learning table will get updated first.
 943                  */
 944                 memmove(&pkts[1], pkts, i * sizeof(struct rte_mbuf *));
 945                 pkts[0] = rarp_mbuf;
 946                 i += 1;
 947         }
 948
 949         return i;
 950 }