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