lib/librte_vhost/vhost_user.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2016 Intel Corporation
   3  */
   4
   5 /* Security model
   6  * --------------
   7  * The vhost-user protocol connection is an external interface, so it must be
   8  * robust against invalid inputs.
   9  *
  10  * This is important because the vhost-user master is only one step removed
  11  * from the guest.  Malicious guests that have escaped will then launch further
  12  * attacks from the vhost-user master.
  13  *
  14  * Even in deployments where guests are trusted, a bug in the vhost-user master
  15  * can still cause invalid messages to be sent.  Such messages must not
  16  * compromise the stability of the DPDK application by causing crashes, memory
  17  * corruption, or other problematic behavior.
  18  *
  19  * Do not assume received VhostUserMsg fields contain sensible values!
  20  */
  21
  22 #include <stdint.h>
  23 #include <stdio.h>
  24 #include <stdlib.h>
  25 #include <string.h>
  26 #include <unistd.h>
  27 #include <sys/mman.h>
  28 #include <sys/types.h>
  29 #include <sys/stat.h>
  30 #include <assert.h>
  31 #ifdef RTE_LIBRTE_VHOST_NUMA
  32 #include <numaif.h>
  33 #endif
  34
  35 #include <rte_common.h>
  36 #include <rte_malloc.h>
  37 #include <rte_log.h>
  38
  39 #include "iotlb.h"
  40 #include "vhost.h"
  41 #include "vhost_user.h"
  42
  43 #define VIRTIO_MIN_MTU 68
  44 #define VIRTIO_MAX_MTU 65535
  45
  46 static const char *vhost_message_str[VHOST_USER_MAX] = {
  47         [VHOST_USER_NONE] = "VHOST_USER_NONE",
  48         [VHOST_USER_GET_FEATURES] = "VHOST_USER_GET_FEATURES",
  49         [VHOST_USER_SET_FEATURES] = "VHOST_USER_SET_FEATURES",
  50         [VHOST_USER_SET_OWNER] = "VHOST_USER_SET_OWNER",
  51         [VHOST_USER_RESET_OWNER] = "VHOST_USER_RESET_OWNER",
  52         [VHOST_USER_SET_MEM_TABLE] = "VHOST_USER_SET_MEM_TABLE",
  53         [VHOST_USER_SET_LOG_BASE] = "VHOST_USER_SET_LOG_BASE",
  54         [VHOST_USER_SET_LOG_FD] = "VHOST_USER_SET_LOG_FD",
  55         [VHOST_USER_SET_VRING_NUM] = "VHOST_USER_SET_VRING_NUM",
  56         [VHOST_USER_SET_VRING_ADDR] = "VHOST_USER_SET_VRING_ADDR",
  57         [VHOST_USER_SET_VRING_BASE] = "VHOST_USER_SET_VRING_BASE",
  58         [VHOST_USER_GET_VRING_BASE] = "VHOST_USER_GET_VRING_BASE",
  59         [VHOST_USER_SET_VRING_KICK] = "VHOST_USER_SET_VRING_KICK",
  60         [VHOST_USER_SET_VRING_CALL] = "VHOST_USER_SET_VRING_CALL",
  61         [VHOST_USER_SET_VRING_ERR]  = "VHOST_USER_SET_VRING_ERR",
  62         [VHOST_USER_GET_PROTOCOL_FEATURES]  = "VHOST_USER_GET_PROTOCOL_FEATURES",
  63         [VHOST_USER_SET_PROTOCOL_FEATURES]  = "VHOST_USER_SET_PROTOCOL_FEATURES",
  64         [VHOST_USER_GET_QUEUE_NUM]  = "VHOST_USER_GET_QUEUE_NUM",
  65         [VHOST_USER_SET_VRING_ENABLE]  = "VHOST_USER_SET_VRING_ENABLE",
  66         [VHOST_USER_SEND_RARP]  = "VHOST_USER_SEND_RARP",
  67         [VHOST_USER_NET_SET_MTU]  = "VHOST_USER_NET_SET_MTU",
  68         [VHOST_USER_SET_SLAVE_REQ_FD]  = "VHOST_USER_SET_SLAVE_REQ_FD",
  69         [VHOST_USER_IOTLB_MSG]  = "VHOST_USER_IOTLB_MSG",
  70 };
  71
  72 static uint64_t
  73 get_blk_size(int fd)
  74 {
  75         struct stat stat;
  76         int ret;
  77
  78         ret = fstat(fd, &stat);
  79         return ret == -1 ? (uint64_t)-1 : (uint64_t)stat.st_blksize;
  80 }
  81
  82 static void
  83 free_mem_region(struct virtio_net *dev)
  84 {
  85         uint32_t i;
  86         struct rte_vhost_mem_region *reg;
  87
  88         if (!dev || !dev->mem)
  89                 return;
  90
  91         for (i = 0; i < dev->mem->nregions; i++) {
  92                 reg = &dev->mem->regions[i];
  93                 if (reg->host_user_addr) {
  94                         munmap(reg->mmap_addr, reg->mmap_size);
  95                         close(reg->fd);
  96                 }
  97         }
  98 }
  99
 100 void
 101 vhost_backend_cleanup(struct virtio_net *dev)
 102 {
 103         if (dev->mem) {
 104                 free_mem_region(dev);
 105                 rte_free(dev->mem);
 106                 dev->mem = NULL;
 107         }
 108
 109         free(dev->guest_pages);
 110         dev->guest_pages = NULL;
 111
 112         if (dev->log_addr) {
 113                 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
 114                 dev->log_addr = 0;
 115         }
 116
 117         if (dev->slave_req_fd >= 0) {
 118                 close(dev->slave_req_fd);
 119                 dev->slave_req_fd = -1;
 120         }
 121 }
 122
 123 /*
 124  * This function just returns success at the moment unless
 125  * the device hasn't been initialised.
 126  */
 127 static int
 128 vhost_user_set_owner(void)
 129 {
 130         return 0;
 131 }
 132
 133 static int
 134 vhost_user_reset_owner(struct virtio_net *dev)
 135 {
 136         if (dev->flags & VIRTIO_DEV_RUNNING) {
 137                 dev->flags &= ~VIRTIO_DEV_RUNNING;
 138                 dev->notify_ops->destroy_device(dev->vid);
 139         }
 140
 141         cleanup_device(dev, 0);
 142         reset_device(dev);
 143         return 0;
 144 }
 145
 146 /*
 147  * The features that we support are requested.
 148  */
 149 static uint64_t
 150 vhost_user_get_features(struct virtio_net *dev)
 151 {
 152         uint64_t features = 0;
 153
 154         rte_vhost_driver_get_features(dev->ifname, &features);
 155         return features;
 156 }
 157
 158 /*
 159  * We receive the negotiated features supported by us and the virtio device.
 160  */
 161 static int
 162 vhost_user_set_features(struct virtio_net *dev, uint64_t features)
 163 {
 164         uint64_t vhost_features = 0;
 165
 166         rte_vhost_driver_get_features(dev->ifname, &vhost_features);
 167         if (features & ~vhost_features) {
 168                 RTE_LOG(ERR, VHOST_CONFIG,
 169                         "(%d) received invalid negotiated features.\n",
 170                         dev->vid);
 171                 return -1;
 172         }
 173
 174         if (dev->flags & VIRTIO_DEV_RUNNING) {
 175                 if (dev->features == features)
 176                         return 0;
 177
 178                 /*
 179                  * Error out if master tries to change features while device is
 180                  * in running state. The exception being VHOST_F_LOG_ALL, which
 181                  * is enabled when the live-migration starts.
 182                  */
 183                 if ((dev->features ^ features) & ~(1ULL << VHOST_F_LOG_ALL)) {
 184                         RTE_LOG(ERR, VHOST_CONFIG,
 185                                 "(%d) features changed while device is running.\n",
 186                                 dev->vid);
 187                         return -1;
 188                 }
 189
 190                 if (dev->notify_ops->features_changed)
 191                         dev->notify_ops->features_changed(dev->vid, features);
 192         }
 193
 194         dev->features = features;
 195         if (dev->features &
 196                 ((1 << VIRTIO_NET_F_MRG_RXBUF) | (1ULL << VIRTIO_F_VERSION_1))) {
 197                 dev->vhost_hlen = sizeof(struct virtio_net_hdr_mrg_rxbuf);
 198         } else {
 199                 dev->vhost_hlen = sizeof(struct virtio_net_hdr);
 200         }
 201         LOG_DEBUG(VHOST_CONFIG,
 202                 "(%d) mergeable RX buffers %s, virtio 1 %s\n",
 203                 dev->vid,
 204                 (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ? "on" : "off",
 205                 (dev->features & (1ULL << VIRTIO_F_VERSION_1)) ? "on" : "off");
 206
 207         if ((dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET) &&
 208             !(dev->features & (1ULL << VIRTIO_NET_F_MQ))) {
 209                 /*
 210                  * Remove all but first queue pair if MQ hasn't been
 211                  * negotiated. This is safe because the device is not
 212                  * running at this stage.
 213                  */
 214                 while (dev->nr_vring > 2) {
 215                         struct vhost_virtqueue *vq;
 216
 217                         vq = dev->virtqueue[--dev->nr_vring];
 218                         if (!vq)
 219                                 continue;
 220
 221                         dev->virtqueue[dev->nr_vring] = NULL;
 222                         cleanup_vq(vq, 1);
 223                         free_vq(vq);
 224                 }
 225         }
 226
 227         return 0;
 228 }
 229
 230 /*
 231  * The virtio device sends us the size of the descriptor ring.
 232  */
 233 static int
 234 vhost_user_set_vring_num(struct virtio_net *dev,
 235                          VhostUserMsg *msg)
 236 {
 237         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
 238
 239         vq->size = msg->payload.state.num;
 240
 241         /* VIRTIO 1.0, 2.4 Virtqueues says:
 242          *
 243          *   Queue Size value is always a power of 2. The maximum Queue Size
 244          *   value is 32768.
 245          */
 246         if ((vq->size & (vq->size - 1)) || vq->size > 32768) {
 247                 RTE_LOG(ERR, VHOST_CONFIG,
 248                         "invalid virtqueue size %u\n", vq->size);
 249                 return -1;
 250         }
 251
 252         if (dev->dequeue_zero_copy) {
 253                 vq->nr_zmbuf = 0;
 254                 vq->last_zmbuf_idx = 0;
 255                 vq->zmbuf_size = vq->size;
 256                 vq->zmbufs = rte_zmalloc(NULL, vq->zmbuf_size *
 257                                          sizeof(struct zcopy_mbuf), 0);
 258                 if (vq->zmbufs == NULL) {
 259                         RTE_LOG(WARNING, VHOST_CONFIG,
 260                                 "failed to allocate mem for zero copy; "
 261                                 "zero copy is force disabled\n");
 262                         dev->dequeue_zero_copy = 0;
 263                 }
 264                 TAILQ_INIT(&vq->zmbuf_list);
 265         }
 266
 267         vq->shadow_used_ring = rte_malloc(NULL,
 268                                 vq->size * sizeof(struct vring_used_elem),
 269                                 RTE_CACHE_LINE_SIZE);
 270         if (!vq->shadow_used_ring) {
 271                 RTE_LOG(ERR, VHOST_CONFIG,
 272                         "failed to allocate memory for shadow used ring.\n");
 273                 return -1;
 274         }
 275
 276         vq->batch_copy_elems = rte_malloc(NULL,
 277                                 vq->size * sizeof(struct batch_copy_elem),
 278                                 RTE_CACHE_LINE_SIZE);
 279         if (!vq->batch_copy_elems) {
 280                 RTE_LOG(ERR, VHOST_CONFIG,
 281                         "failed to allocate memory for batching copy.\n");
 282                 return -1;
 283         }
 284
 285         return 0;
 286 }
 287
 288 /*
 289  * Reallocate virtio_dev and vhost_virtqueue data structure to make them on the
 290  * same numa node as the memory of vring descriptor.
 291  */
 292 #ifdef RTE_LIBRTE_VHOST_NUMA
 293 static struct virtio_net*
 294 numa_realloc(struct virtio_net *dev, int index)
 295 {
 296         int oldnode, newnode;
 297         struct virtio_net *old_dev;
 298         struct vhost_virtqueue *old_vq, *vq;
 299         struct zcopy_mbuf *new_zmbuf;
 300         struct vring_used_elem *new_shadow_used_ring;
 301         struct batch_copy_elem *new_batch_copy_elems;
 302         int ret;
 303
 304         old_dev = dev;
 305         vq = old_vq = dev->virtqueue[index];
 306
 307         ret = get_mempolicy(&newnode, NULL, 0, old_vq->desc,
 308                             MPOL_F_NODE | MPOL_F_ADDR);
 309
 310         /* check if we need to reallocate vq */
 311         ret |= get_mempolicy(&oldnode, NULL, 0, old_vq,
 312                              MPOL_F_NODE | MPOL_F_ADDR);
 313         if (ret) {
 314                 RTE_LOG(ERR, VHOST_CONFIG,
 315                         "Unable to get vq numa information.\n");
 316                 return dev;
 317         }
 318         if (oldnode != newnode) {
 319                 RTE_LOG(INFO, VHOST_CONFIG,
 320                         "reallocate vq from %d to %d node\n", oldnode, newnode);
 321                 vq = rte_malloc_socket(NULL, sizeof(*vq), 0, newnode);
 322                 if (!vq)
 323                         return dev;
 324
 325                 memcpy(vq, old_vq, sizeof(*vq));
 326                 TAILQ_INIT(&vq->zmbuf_list);
 327
 328                 new_zmbuf = rte_malloc_socket(NULL, vq->zmbuf_size *
 329                         sizeof(struct zcopy_mbuf), 0, newnode);
 330                 if (new_zmbuf) {
 331                         rte_free(vq->zmbufs);
 332                         vq->zmbufs = new_zmbuf;
 333                 }
 334
 335                 new_shadow_used_ring = rte_malloc_socket(NULL,
 336                         vq->size * sizeof(struct vring_used_elem),
 337                         RTE_CACHE_LINE_SIZE,
 338                         newnode);
 339                 if (new_shadow_used_ring) {
 340                         rte_free(vq->shadow_used_ring);
 341                         vq->shadow_used_ring = new_shadow_used_ring;
 342                 }
 343
 344                 new_batch_copy_elems = rte_malloc_socket(NULL,
 345                         vq->size * sizeof(struct batch_copy_elem),
 346                         RTE_CACHE_LINE_SIZE,
 347                         newnode);
 348                 if (new_batch_copy_elems) {
 349                         rte_free(vq->batch_copy_elems);
 350                         vq->batch_copy_elems = new_batch_copy_elems;
 351                 }
 352
 353                 rte_free(old_vq);
 354         }
 355
 356         /* check if we need to reallocate dev */
 357         ret = get_mempolicy(&oldnode, NULL, 0, old_dev,
 358                             MPOL_F_NODE | MPOL_F_ADDR);
 359         if (ret) {
 360                 RTE_LOG(ERR, VHOST_CONFIG,
 361                         "Unable to get dev numa information.\n");
 362                 goto out;
 363         }
 364         if (oldnode != newnode) {
 365                 RTE_LOG(INFO, VHOST_CONFIG,
 366                         "reallocate dev from %d to %d node\n",
 367                         oldnode, newnode);
 368                 dev = rte_malloc_socket(NULL, sizeof(*dev), 0, newnode);
 369                 if (!dev) {
 370                         dev = old_dev;
 371                         goto out;
 372                 }
 373
 374                 memcpy(dev, old_dev, sizeof(*dev));
 375                 rte_free(old_dev);
 376         }
 377
 378 out:
 379         dev->virtqueue[index] = vq;
 380         vhost_devices[dev->vid] = dev;
 381
 382         if (old_vq != vq)
 383                 vhost_user_iotlb_init(dev, index);
 384
 385         return dev;
 386 }
 387 #else
 388 static struct virtio_net*
 389 numa_realloc(struct virtio_net *dev, int index __rte_unused)
 390 {
 391         return dev;
 392 }
 393 #endif
 394
 395 /* Converts QEMU virtual address to Vhost virtual address. */
 396 static uint64_t
 397 qva_to_vva(struct virtio_net *dev, uint64_t qva)
 398 {
 399         struct rte_vhost_mem_region *reg;
 400         uint32_t i;
 401
 402         /* Find the region where the address lives. */
 403         for (i = 0; i < dev->mem->nregions; i++) {
 404                 reg = &dev->mem->regions[i];
 405
 406                 if (qva >= reg->guest_user_addr &&
 407                     qva <  reg->guest_user_addr + reg->size) {
 408                         return qva - reg->guest_user_addr +
 409                                reg->host_user_addr;
 410                 }
 411         }
 412
 413         return 0;
 414 }
 415
 416
 417 /*
 418  * Converts ring address to Vhost virtual address.
 419  * If IOMMU is enabled, the ring address is a guest IO virtual address,
 420  * else it is a QEMU virtual address.
 421  */
 422 static uint64_t
 423 ring_addr_to_vva(struct virtio_net *dev, struct vhost_virtqueue *vq,
 424                 uint64_t ra, uint64_t size)
 425 {
 426         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)) {
 427                 uint64_t vva;
 428
 429                 vva = vhost_user_iotlb_cache_find(vq, ra,
 430                                         &size, VHOST_ACCESS_RW);
 431                 if (!vva)
 432                         vhost_user_iotlb_miss(dev, ra, VHOST_ACCESS_RW);
 433
 434                 return vva;
 435         }
 436
 437         return qva_to_vva(dev, ra);
 438 }
 439
 440 static struct virtio_net *
 441 translate_ring_addresses(struct virtio_net *dev, int vq_index)
 442 {
 443         struct vhost_virtqueue *vq = dev->virtqueue[vq_index];
 444         struct vhost_vring_addr *addr = &vq->ring_addrs;
 445
 446         /* The addresses are converted from QEMU virtual to Vhost virtual. */
 447         if (vq->desc && vq->avail && vq->used)
 448                 return dev;
 449
 450         vq->desc = (struct vring_desc *)(uintptr_t)ring_addr_to_vva(dev,
 451                         vq, addr->desc_user_addr, sizeof(struct vring_desc));
 452         if (vq->desc == 0) {
 453                 RTE_LOG(DEBUG, VHOST_CONFIG,
 454                         "(%d) failed to find desc ring address.\n",
 455                         dev->vid);
 456                 return dev;
 457         }
 458
 459         dev = numa_realloc(dev, vq_index);
 460         vq = dev->virtqueue[vq_index];
 461         addr = &vq->ring_addrs;
 462
 463         vq->avail = (struct vring_avail *)(uintptr_t)ring_addr_to_vva(dev,
 464                         vq, addr->avail_user_addr, sizeof(struct vring_avail));
 465         if (vq->avail == 0) {
 466                 RTE_LOG(DEBUG, VHOST_CONFIG,
 467                         "(%d) failed to find avail ring address.\n",
 468                         dev->vid);
 469                 return dev;
 470         }
 471
 472         vq->used = (struct vring_used *)(uintptr_t)ring_addr_to_vva(dev,
 473                         vq, addr->used_user_addr, sizeof(struct vring_used));
 474         if (vq->used == 0) {
 475                 RTE_LOG(DEBUG, VHOST_CONFIG,
 476                         "(%d) failed to find used ring address.\n",
 477                         dev->vid);
 478                 return dev;
 479         }
 480
 481         if (vq->last_used_idx != vq->used->idx) {
 482                 RTE_LOG(WARNING, VHOST_CONFIG,
 483                         "last_used_idx (%u) and vq->used->idx (%u) mismatches; "
 484                         "some packets maybe resent for Tx and dropped for Rx\n",
 485                         vq->last_used_idx, vq->used->idx);
 486                 vq->last_used_idx  = vq->used->idx;
 487                 vq->last_avail_idx = vq->used->idx;
 488         }
 489
 490         vq->log_guest_addr = addr->log_guest_addr;
 491
 492         LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address desc: %p\n",
 493                         dev->vid, vq->desc);
 494         LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address avail: %p\n",
 495                         dev->vid, vq->avail);
 496         LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address used: %p\n",
 497                         dev->vid, vq->used);
 498         LOG_DEBUG(VHOST_CONFIG, "(%d) log_guest_addr: %" PRIx64 "\n",
 499                         dev->vid, vq->log_guest_addr);
 500
 501         return dev;
 502 }
 503
 504 /*
 505  * The virtio device sends us the desc, used and avail ring addresses.
 506  * This function then converts these to our address space.
 507  */
 508 static int
 509 vhost_user_set_vring_addr(struct virtio_net **pdev, VhostUserMsg *msg)
 510 {
 511         struct vhost_virtqueue *vq;
 512         struct vhost_vring_addr *addr = &msg->payload.addr;
 513         struct virtio_net *dev = *pdev;
 514
 515         if (dev->mem == NULL)
 516                 return -1;
 517
 518         /* addr->index refers to the queue index. The txq 1, rxq is 0. */
 519         vq = dev->virtqueue[msg->payload.addr.index];
 520
 521         /*
 522          * Rings addresses should not be interpreted as long as the ring is not
 523          * started and enabled
 524          */
 525         memcpy(&vq->ring_addrs, addr, sizeof(*addr));
 526
 527         vring_invalidate(dev, vq);
 528
 529         if (vq->enabled && (dev->features &
 530                                 (1ULL << VHOST_USER_F_PROTOCOL_FEATURES))) {
 531                 dev = translate_ring_addresses(dev, msg->payload.addr.index);
 532                 if (!dev)
 533                         return -1;
 534
 535                 *pdev = dev;
 536         }
 537
 538         return 0;
 539 }
 540
 541 /*
 542  * The virtio device sends us the available ring last used index.
 543  */
 544 static int
 545 vhost_user_set_vring_base(struct virtio_net *dev,
 546                           VhostUserMsg *msg)
 547 {
 548         dev->virtqueue[msg->payload.state.index]->last_used_idx  =
 549                         msg->payload.state.num;
 550         dev->virtqueue[msg->payload.state.index]->last_avail_idx =
 551                         msg->payload.state.num;
 552
 553         return 0;
 554 }
 555
 556 static void
 557 add_one_guest_page(struct virtio_net *dev, uint64_t guest_phys_addr,
 558                    uint64_t host_phys_addr, uint64_t size)
 559 {
 560         struct guest_page *page, *last_page;
 561
 562         if (dev->nr_guest_pages == dev->max_guest_pages) {
 563                 dev->max_guest_pages *= 2;
 564                 dev->guest_pages = realloc(dev->guest_pages,
 565                                         dev->max_guest_pages * sizeof(*page));
 566         }
 567
 568         if (dev->nr_guest_pages > 0) {
 569                 last_page = &dev->guest_pages[dev->nr_guest_pages - 1];
 570                 /* merge if the two pages are continuous */
 571                 if (host_phys_addr == last_page->host_phys_addr +
 572                                       last_page->size) {
 573                         last_page->size += size;
 574                         return;
 575                 }
 576         }
 577
 578         page = &dev->guest_pages[dev->nr_guest_pages++];
 579         page->guest_phys_addr = guest_phys_addr;
 580         page->host_phys_addr  = host_phys_addr;
 581         page->size = size;
 582 }
 583
 584 static void
 585 add_guest_pages(struct virtio_net *dev, struct rte_vhost_mem_region *reg,
 586                 uint64_t page_size)
 587 {
 588         uint64_t reg_size = reg->size;
 589         uint64_t host_user_addr  = reg->host_user_addr;
 590         uint64_t guest_phys_addr = reg->guest_phys_addr;
 591         uint64_t host_phys_addr;
 592         uint64_t size;
 593
 594         host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)host_user_addr);
 595         size = page_size - (guest_phys_addr & (page_size - 1));
 596         size = RTE_MIN(size, reg_size);
 597
 598         add_one_guest_page(dev, guest_phys_addr, host_phys_addr, size);
 599         host_user_addr  += size;
 600         guest_phys_addr += size;
 601         reg_size -= size;
 602
 603         while (reg_size > 0) {
 604                 size = RTE_MIN(reg_size, page_size);
 605                 host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)
 606                                                   host_user_addr);
 607                 add_one_guest_page(dev, guest_phys_addr, host_phys_addr, size);
 608
 609                 host_user_addr  += size;
 610                 guest_phys_addr += size;
 611                 reg_size -= size;
 612         }
 613 }
 614
 615 #ifdef RTE_LIBRTE_VHOST_DEBUG
 616 /* TODO: enable it only in debug mode? */
 617 static void
 618 dump_guest_pages(struct virtio_net *dev)
 619 {
 620         uint32_t i;
 621         struct guest_page *page;
 622
 623         for (i = 0; i < dev->nr_guest_pages; i++) {
 624                 page = &dev->guest_pages[i];
 625
 626                 RTE_LOG(INFO, VHOST_CONFIG,
 627                         "guest physical page region %u\n"
 628                         "\t guest_phys_addr: %" PRIx64 "\n"
 629                         "\t host_phys_addr : %" PRIx64 "\n"
 630                         "\t size           : %" PRIx64 "\n",
 631                         i,
 632                         page->guest_phys_addr,
 633                         page->host_phys_addr,
 634                         page->size);
 635         }
 636 }
 637 #else
 638 #define dump_guest_pages(dev)
 639 #endif
 640
 641 static bool
 642 vhost_memory_changed(struct VhostUserMemory *new,
 643                      struct rte_vhost_memory *old)
 644 {
 645         uint32_t i;
 646
 647         if (new->nregions != old->nregions)
 648                 return true;
 649
 650         for (i = 0; i < new->nregions; ++i) {
 651                 VhostUserMemoryRegion *new_r = &new->regions[i];
 652                 struct rte_vhost_mem_region *old_r = &old->regions[i];
 653
 654                 if (new_r->guest_phys_addr != old_r->guest_phys_addr)
 655                         return true;
 656                 if (new_r->memory_size != old_r->size)
 657                         return true;
 658                 if (new_r->userspace_addr != old_r->guest_user_addr)
 659                         return true;
 660         }
 661
 662         return false;
 663 }
 664
 665 static int
 666 vhost_user_set_mem_table(struct virtio_net *dev, struct VhostUserMsg *pmsg)
 667 {
 668         struct VhostUserMemory memory = pmsg->payload.memory;
 669         struct rte_vhost_mem_region *reg;
 670         void *mmap_addr;
 671         uint64_t mmap_size;
 672         uint64_t mmap_offset;
 673         uint64_t alignment;
 674         uint32_t i;
 675         int fd;
 676
 677         if (memory.nregions > VHOST_MEMORY_MAX_NREGIONS) {
 678                 RTE_LOG(ERR, VHOST_CONFIG,
 679                         "too many memory regions (%u)\n", memory.nregions);
 680                 return -1;
 681         }
 682
 683         if (dev->mem && !vhost_memory_changed(&memory, dev->mem)) {
 684                 RTE_LOG(INFO, VHOST_CONFIG,
 685                         "(%d) memory regions not changed\n", dev->vid);
 686
 687                 for (i = 0; i < memory.nregions; i++)
 688                         close(pmsg->fds[i]);
 689
 690                 return 0;
 691         }
 692
 693         if (dev->mem) {
 694                 free_mem_region(dev);
 695                 rte_free(dev->mem);
 696                 dev->mem = NULL;
 697         }
 698
 699         dev->nr_guest_pages = 0;
 700         if (!dev->guest_pages) {
 701                 dev->max_guest_pages = 8;
 702                 dev->guest_pages = malloc(dev->max_guest_pages *
 703                                                 sizeof(struct guest_page));
 704                 if (dev->guest_pages == NULL) {
 705                         RTE_LOG(ERR, VHOST_CONFIG,
 706                                 "(%d) failed to allocate memory "
 707                                 "for dev->guest_pages\n",
 708                                 dev->vid);
 709                         return -1;
 710                 }
 711         }
 712
 713         dev->mem = rte_zmalloc("vhost-mem-table", sizeof(struct rte_vhost_memory) +
 714                 sizeof(struct rte_vhost_mem_region) * memory.nregions, 0);
 715         if (dev->mem == NULL) {
 716                 RTE_LOG(ERR, VHOST_CONFIG,
 717                         "(%d) failed to allocate memory for dev->mem\n",
 718                         dev->vid);
 719                 return -1;
 720         }
 721         dev->mem->nregions = memory.nregions;
 722
 723         for (i = 0; i < memory.nregions; i++) {
 724                 fd  = pmsg->fds[i];
 725                 reg = &dev->mem->regions[i];
 726
 727                 reg->guest_phys_addr = memory.regions[i].guest_phys_addr;
 728                 reg->guest_user_addr = memory.regions[i].userspace_addr;
 729                 reg->size            = memory.regions[i].memory_size;
 730                 reg->fd              = fd;
 731
 732                 mmap_offset = memory.regions[i].mmap_offset;
 733                 mmap_size   = reg->size + mmap_offset;
 734
 735                 /* mmap() without flag of MAP_ANONYMOUS, should be called
 736                  * with length argument aligned with hugepagesz at older
 737                  * longterm version Linux, like 2.6.32 and 3.2.72, or
 738                  * mmap() will fail with EINVAL.
 739                  *
 740                  * to avoid failure, make sure in caller to keep length
 741                  * aligned.
 742                  */
 743                 alignment = get_blk_size(fd);
 744                 if (alignment == (uint64_t)-1) {
 745                         RTE_LOG(ERR, VHOST_CONFIG,
 746                                 "couldn't get hugepage size through fstat\n");
 747                         goto err_mmap;
 748                 }
 749                 mmap_size = RTE_ALIGN_CEIL(mmap_size, alignment);
 750
 751                 mmap_addr = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE,
 752                                  MAP_SHARED | MAP_POPULATE, fd, 0);
 753
 754                 if (mmap_addr == MAP_FAILED) {
 755                         RTE_LOG(ERR, VHOST_CONFIG,
 756                                 "mmap region %u failed.\n", i);
 757                         goto err_mmap;
 758                 }
 759
 760                 reg->mmap_addr = mmap_addr;
 761                 reg->mmap_size = mmap_size;
 762                 reg->host_user_addr = (uint64_t)(uintptr_t)mmap_addr +
 763                                       mmap_offset;
 764
 765                 if (dev->dequeue_zero_copy)
 766                         add_guest_pages(dev, reg, alignment);
 767
 768                 RTE_LOG(INFO, VHOST_CONFIG,
 769                         "guest memory region %u, size: 0x%" PRIx64 "\n"
 770                         "\t guest physical addr: 0x%" PRIx64 "\n"
 771                         "\t guest virtual  addr: 0x%" PRIx64 "\n"
 772                         "\t host  virtual  addr: 0x%" PRIx64 "\n"
 773                         "\t mmap addr : 0x%" PRIx64 "\n"
 774                         "\t mmap size : 0x%" PRIx64 "\n"
 775                         "\t mmap align: 0x%" PRIx64 "\n"
 776                         "\t mmap off  : 0x%" PRIx64 "\n",
 777                         i, reg->size,
 778                         reg->guest_phys_addr,
 779                         reg->guest_user_addr,
 780                         reg->host_user_addr,
 781                         (uint64_t)(uintptr_t)mmap_addr,
 782                         mmap_size,
 783                         alignment,
 784                         mmap_offset);
 785         }
 786
 787         dump_guest_pages(dev);
 788
 789         return 0;
 790
 791 err_mmap:
 792         free_mem_region(dev);
 793         rte_free(dev->mem);
 794         dev->mem = NULL;
 795         return -1;
 796 }
 797
 798 static int
 799 vq_is_ready(struct vhost_virtqueue *vq)
 800 {
 801         return vq && vq->desc && vq->avail && vq->used &&
 802                vq->kickfd != VIRTIO_UNINITIALIZED_EVENTFD &&
 803                vq->callfd != VIRTIO_UNINITIALIZED_EVENTFD;
 804 }
 805
 806 static int
 807 virtio_is_ready(struct virtio_net *dev)
 808 {
 809         struct vhost_virtqueue *vq;
 810         uint32_t i;
 811
 812         if (dev->nr_vring == 0)
 813                 return 0;
 814
 815         for (i = 0; i < dev->nr_vring; i++) {
 816                 vq = dev->virtqueue[i];
 817
 818                 if (!vq_is_ready(vq))
 819                         return 0;
 820         }
 821
 822         RTE_LOG(INFO, VHOST_CONFIG,
 823                 "virtio is now ready for processing.\n");
 824         return 1;
 825 }
 826
 827 static void
 828 vhost_user_set_vring_call(struct virtio_net *dev, struct VhostUserMsg *pmsg)
 829 {
 830         struct vhost_vring_file file;
 831         struct vhost_virtqueue *vq;
 832
 833         file.index = pmsg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
 834         if (pmsg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
 835                 file.fd = VIRTIO_INVALID_EVENTFD;
 836         else
 837                 file.fd = pmsg->fds[0];
 838         RTE_LOG(INFO, VHOST_CONFIG,
 839                 "vring call idx:%d file:%d\n", file.index, file.fd);
 840
 841         vq = dev->virtqueue[file.index];
 842         if (vq->callfd >= 0)
 843                 close(vq->callfd);
 844
 845         vq->callfd = file.fd;
 846 }
 847
 848 static void
 849 vhost_user_set_vring_kick(struct virtio_net **pdev, struct VhostUserMsg *pmsg)
 850 {
 851         struct vhost_vring_file file;
 852         struct vhost_virtqueue *vq;
 853         struct virtio_net *dev = *pdev;
 854
 855         file.index = pmsg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
 856         if (pmsg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
 857                 file.fd = VIRTIO_INVALID_EVENTFD;
 858         else
 859                 file.fd = pmsg->fds[0];
 860         RTE_LOG(INFO, VHOST_CONFIG,
 861                 "vring kick idx:%d file:%d\n", file.index, file.fd);
 862
 863         /* Interpret ring addresses only when ring is started. */
 864         dev = translate_ring_addresses(dev, file.index);
 865         if (!dev)
 866                 return;
 867
 868         *pdev = dev;
 869
 870         vq = dev->virtqueue[file.index];
 871
 872         /*
 873          * When VHOST_USER_F_PROTOCOL_FEATURES is not negotiated,
 874          * the ring starts already enabled. Otherwise, it is enabled via
 875          * the SET_VRING_ENABLE message.
 876          */
 877         if (!(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES)))
 878                 vq->enabled = 1;
 879
 880         if (vq->kickfd >= 0)
 881                 close(vq->kickfd);
 882         vq->kickfd = file.fd;
 883 }
 884
 885 static void
 886 free_zmbufs(struct vhost_virtqueue *vq)
 887 {
 888         struct zcopy_mbuf *zmbuf, *next;
 889
 890         for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
 891              zmbuf != NULL; zmbuf = next) {
 892                 next = TAILQ_NEXT(zmbuf, next);
 893
 894                 rte_pktmbuf_free(zmbuf->mbuf);
 895                 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
 896         }
 897
 898         rte_free(vq->zmbufs);
 899 }
 900
 901 /*
 902  * when virtio is stopped, qemu will send us the GET_VRING_BASE message.
 903  */
 904 static int
 905 vhost_user_get_vring_base(struct virtio_net *dev,
 906                           VhostUserMsg *msg)
 907 {
 908         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
 909
 910         /* We have to stop the queue (virtio) if it is running. */
 911         if (dev->flags & VIRTIO_DEV_RUNNING) {
 912                 dev->flags &= ~VIRTIO_DEV_RUNNING;
 913                 dev->notify_ops->destroy_device(dev->vid);
 914         }
 915
 916         dev->flags &= ~VIRTIO_DEV_READY;
 917
 918         /* Here we are safe to get the last used index */
 919         msg->payload.state.num = vq->last_used_idx;
 920
 921         RTE_LOG(INFO, VHOST_CONFIG,
 922                 "vring base idx:%d file:%d\n", msg->payload.state.index,
 923                 msg->payload.state.num);
 924         /*
 925          * Based on current qemu vhost-user implementation, this message is
 926          * sent and only sent in vhost_vring_stop.
 927          * TODO: cleanup the vring, it isn't usable since here.
 928          */
 929         if (vq->kickfd >= 0)
 930                 close(vq->kickfd);
 931
 932         vq->kickfd = VIRTIO_UNINITIALIZED_EVENTFD;
 933
 934         if (dev->dequeue_zero_copy)
 935                 free_zmbufs(vq);
 936         rte_free(vq->shadow_used_ring);
 937         vq->shadow_used_ring = NULL;
 938
 939         rte_free(vq->batch_copy_elems);
 940         vq->batch_copy_elems = NULL;
 941
 942         return 0;
 943 }
 944
 945 /*
 946  * when virtio queues are ready to work, qemu will send us to
 947  * enable the virtio queue pair.
 948  */
 949 static int
 950 vhost_user_set_vring_enable(struct virtio_net *dev,
 951                             VhostUserMsg *msg)
 952 {
 953         int enable = (int)msg->payload.state.num;
 954
 955         RTE_LOG(INFO, VHOST_CONFIG,
 956                 "set queue enable: %d to qp idx: %d\n",
 957                 enable, msg->payload.state.index);
 958
 959         if (dev->notify_ops->vring_state_changed)
 960                 dev->notify_ops->vring_state_changed(dev->vid,
 961                                 msg->payload.state.index, enable);
 962
 963         dev->virtqueue[msg->payload.state.index]->enabled = enable;
 964
 965         return 0;
 966 }
 967
 968 static void
 969 vhost_user_get_protocol_features(struct virtio_net *dev,
 970                                  struct VhostUserMsg *msg)
 971 {
 972         uint64_t features, protocol_features = VHOST_USER_PROTOCOL_FEATURES;
 973
 974         rte_vhost_driver_get_features(dev->ifname, &features);
 975
 976         /*
 977          * REPLY_ACK protocol feature is only mandatory for now
 978          * for IOMMU feature. If IOMMU is explicitly disabled by the
 979          * application, disable also REPLY_ACK feature for older buggy
 980          * Qemu versions (from v2.7.0 to v2.9.0).
 981          */
 982         if (!(features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)))
 983                 protocol_features &= ~(1ULL << VHOST_USER_PROTOCOL_F_REPLY_ACK);
 984
 985         msg->payload.u64 = protocol_features;
 986         msg->size = sizeof(msg->payload.u64);
 987 }
 988
 989 static void
 990 vhost_user_set_protocol_features(struct virtio_net *dev,
 991                                  uint64_t protocol_features)
 992 {
 993         if (protocol_features & ~VHOST_USER_PROTOCOL_FEATURES)
 994                 return;
 995
 996         dev->protocol_features = protocol_features;
 997 }
 998
 999 static int
1000 vhost_user_set_log_base(struct virtio_net *dev, struct VhostUserMsg *msg)
1001 {
1002         int fd = msg->fds[0];
1003         uint64_t size, off;
1004         void *addr;
1005
1006         if (fd < 0) {
1007                 RTE_LOG(ERR, VHOST_CONFIG, "invalid log fd: %d\n", fd);
1008                 return -1;
1009         }
1010
1011         if (msg->size != sizeof(VhostUserLog)) {
1012                 RTE_LOG(ERR, VHOST_CONFIG,
1013                         "invalid log base msg size: %"PRId32" != %d\n",
1014                         msg->size, (int)sizeof(VhostUserLog));
1015                 return -1;
1016         }
1017
1018         size = msg->payload.log.mmap_size;
1019         off  = msg->payload.log.mmap_offset;
1020
1021         /* Don't allow mmap_offset to point outside the mmap region */
1022         if (off > size) {
1023                 RTE_LOG(ERR, VHOST_CONFIG,
1024                         "log offset %#"PRIx64" exceeds log size %#"PRIx64"\n",
1025                         off, size);
1026                 return -1;
1027         }
1028
1029         RTE_LOG(INFO, VHOST_CONFIG,
1030                 "log mmap size: %"PRId64", offset: %"PRId64"\n",
1031                 size, off);
1032
1033         /*
1034          * mmap from 0 to workaround a hugepage mmap bug: mmap will
1035          * fail when offset is not page size aligned.
1036          */
1037         addr = mmap(0, size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
1038         close(fd);
1039         if (addr == MAP_FAILED) {
1040                 RTE_LOG(ERR, VHOST_CONFIG, "mmap log base failed!\n");
1041                 return -1;
1042         }
1043
1044         /*
1045          * Free previously mapped log memory on occasionally
1046          * multiple VHOST_USER_SET_LOG_BASE.
1047          */
1048         if (dev->log_addr) {
1049                 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
1050         }
1051         dev->log_addr = (uint64_t)(uintptr_t)addr;
1052         dev->log_base = dev->log_addr + off;
1053         dev->log_size = size;
1054
1055         return 0;
1056 }
1057
1058 /*
1059  * An rarp packet is constructed and broadcasted to notify switches about
1060  * the new location of the migrated VM, so that packets from outside will
1061  * not be lost after migration.
1062  *
1063  * However, we don't actually "send" a rarp packet here, instead, we set
1064  * a flag 'broadcast_rarp' to let rte_vhost_dequeue_burst() inject it.
1065  */
1066 static int
1067 vhost_user_send_rarp(struct virtio_net *dev, struct VhostUserMsg *msg)
1068 {
1069         uint8_t *mac = (uint8_t *)&msg->payload.u64;
1070
1071         RTE_LOG(DEBUG, VHOST_CONFIG,
1072                 ":: mac: %02x:%02x:%02x:%02x:%02x:%02x\n",
1073                 mac[0], mac[1], mac[2], mac[3], mac[4], mac[5]);
1074         memcpy(dev->mac.addr_bytes, mac, 6);
1075
1076         /*
1077          * Set the flag to inject a RARP broadcast packet at
1078          * rte_vhost_dequeue_burst().
1079          *
1080          * rte_smp_wmb() is for making sure the mac is copied
1081          * before the flag is set.
1082          */
1083         rte_smp_wmb();
1084         rte_atomic16_set(&dev->broadcast_rarp, 1);
1085
1086         return 0;
1087 }
1088
1089 static int
1090 vhost_user_net_set_mtu(struct virtio_net *dev, struct VhostUserMsg *msg)
1091 {
1092         if (msg->payload.u64 < VIRTIO_MIN_MTU ||
1093                         msg->payload.u64 > VIRTIO_MAX_MTU) {
1094                 RTE_LOG(ERR, VHOST_CONFIG, "Invalid MTU size (%"PRIu64")\n",
1095                                 msg->payload.u64);
1096
1097                 return -1;
1098         }
1099
1100         dev->mtu = msg->payload.u64;
1101
1102         return 0;
1103 }
1104
1105 static int
1106 vhost_user_set_req_fd(struct virtio_net *dev, struct VhostUserMsg *msg)
1107 {
1108         int fd = msg->fds[0];
1109
1110         if (fd < 0) {
1111                 RTE_LOG(ERR, VHOST_CONFIG,
1112                                 "Invalid file descriptor for slave channel (%d)\n",
1113                                 fd);
1114                 return -1;
1115         }
1116
1117         dev->slave_req_fd = fd;
1118
1119         return 0;
1120 }
1121
1122 static int
1123 is_vring_iotlb_update(struct vhost_virtqueue *vq, struct vhost_iotlb_msg *imsg)
1124 {
1125         struct vhost_vring_addr *ra;
1126         uint64_t start, end;
1127
1128         start = imsg->iova;
1129         end = start + imsg->size;
1130
1131         ra = &vq->ring_addrs;
1132         if (ra->desc_user_addr >= start && ra->desc_user_addr < end)
1133                 return 1;
1134         if (ra->avail_user_addr >= start && ra->avail_user_addr < end)
1135                 return 1;
1136         if (ra->used_user_addr >= start && ra->used_user_addr < end)
1137                 return 1;
1138
1139         return 0;
1140 }
1141
1142 static int
1143 is_vring_iotlb_invalidate(struct vhost_virtqueue *vq,
1144                                 struct vhost_iotlb_msg *imsg)
1145 {
1146         uint64_t istart, iend, vstart, vend;
1147
1148         istart = imsg->iova;
1149         iend = istart + imsg->size - 1;
1150
1151         vstart = (uintptr_t)vq->desc;
1152         vend = vstart + sizeof(struct vring_desc) * vq->size - 1;
1153         if (vstart <= iend && istart <= vend)
1154                 return 1;
1155
1156         vstart = (uintptr_t)vq->avail;
1157         vend = vstart + sizeof(struct vring_avail);
1158         vend += sizeof(uint16_t) * vq->size - 1;
1159         if (vstart <= iend && istart <= vend)
1160                 return 1;
1161
1162         vstart = (uintptr_t)vq->used;
1163         vend = vstart + sizeof(struct vring_used);
1164         vend += sizeof(struct vring_used_elem) * vq->size - 1;
1165         if (vstart <= iend && istart <= vend)
1166                 return 1;
1167
1168         return 0;
1169 }
1170
1171 static int
1172 vhost_user_iotlb_msg(struct virtio_net **pdev, struct VhostUserMsg *msg)
1173 {
1174         struct virtio_net *dev = *pdev;
1175         struct vhost_iotlb_msg *imsg = &msg->payload.iotlb;
1176         uint16_t i;
1177         uint64_t vva;
1178
1179         switch (imsg->type) {
1180         case VHOST_IOTLB_UPDATE:
1181                 vva = qva_to_vva(dev, imsg->uaddr);
1182                 if (!vva)
1183                         return -1;
1184
1185                 for (i = 0; i < dev->nr_vring; i++) {
1186                         struct vhost_virtqueue *vq = dev->virtqueue[i];
1187
1188                         vhost_user_iotlb_cache_insert(vq, imsg->iova, vva,
1189                                         imsg->size, imsg->perm);
1190
1191                         if (is_vring_iotlb_update(vq, imsg))
1192                                 *pdev = dev = translate_ring_addresses(dev, i);
1193                 }
1194                 break;
1195         case VHOST_IOTLB_INVALIDATE:
1196                 for (i = 0; i < dev->nr_vring; i++) {
1197                         struct vhost_virtqueue *vq = dev->virtqueue[i];
1198
1199                         vhost_user_iotlb_cache_remove(vq, imsg->iova,
1200                                         imsg->size);
1201
1202                         if (is_vring_iotlb_invalidate(vq, imsg))
1203                                 vring_invalidate(dev, vq);
1204                 }
1205                 break;
1206         default:
1207                 RTE_LOG(ERR, VHOST_CONFIG, "Invalid IOTLB message type (%d)\n",
1208                                 imsg->type);
1209                 return -1;
1210         }
1211
1212         return 0;
1213 }
1214
1215 /* return bytes# of read on success or negative val on failure. */
1216 static int
1217 read_vhost_message(int sockfd, struct VhostUserMsg *msg)
1218 {
1219         int ret;
1220
1221         ret = read_fd_message(sockfd, (char *)msg, VHOST_USER_HDR_SIZE,
1222                 msg->fds, VHOST_MEMORY_MAX_NREGIONS);
1223         if (ret <= 0)
1224                 return ret;
1225
1226         if (msg && msg->size) {
1227                 if (msg->size > sizeof(msg->payload)) {
1228                         RTE_LOG(ERR, VHOST_CONFIG,
1229                                 "invalid msg size: %d\n", msg->size);
1230                         return -1;
1231                 }
1232                 ret = read(sockfd, &msg->payload, msg->size);
1233                 if (ret <= 0)
1234                         return ret;
1235                 if (ret != (int)msg->size) {
1236                         RTE_LOG(ERR, VHOST_CONFIG,
1237                                 "read control message failed\n");
1238                         return -1;
1239                 }
1240         }
1241
1242         return ret;
1243 }
1244
1245 static int
1246 send_vhost_message(int sockfd, struct VhostUserMsg *msg)
1247 {
1248         if (!msg)
1249                 return 0;
1250
1251         return send_fd_message(sockfd, (char *)msg,
1252                 VHOST_USER_HDR_SIZE + msg->size, NULL, 0);
1253 }
1254
1255 static int
1256 send_vhost_reply(int sockfd, struct VhostUserMsg *msg)
1257 {
1258         if (!msg)
1259                 return 0;
1260
1261         msg->flags &= ~VHOST_USER_VERSION_MASK;
1262         msg->flags &= ~VHOST_USER_NEED_REPLY;
1263         msg->flags |= VHOST_USER_VERSION;
1264         msg->flags |= VHOST_USER_REPLY_MASK;
1265
1266         return send_vhost_message(sockfd, msg);
1267 }
1268
1269 /*
1270  * Allocate a queue pair if it hasn't been allocated yet
1271  */
1272 static int
1273 vhost_user_check_and_alloc_queue_pair(struct virtio_net *dev, VhostUserMsg *msg)
1274 {
1275         uint16_t vring_idx;
1276
1277         switch (msg->request.master) {
1278         case VHOST_USER_SET_VRING_KICK:
1279         case VHOST_USER_SET_VRING_CALL:
1280         case VHOST_USER_SET_VRING_ERR:
1281                 vring_idx = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1282                 break;
1283         case VHOST_USER_SET_VRING_NUM:
1284         case VHOST_USER_SET_VRING_BASE:
1285         case VHOST_USER_SET_VRING_ENABLE:
1286                 vring_idx = msg->payload.state.index;
1287                 break;
1288         case VHOST_USER_SET_VRING_ADDR:
1289                 vring_idx = msg->payload.addr.index;
1290                 break;
1291         default:
1292                 return 0;
1293         }
1294
1295         if (vring_idx >= VHOST_MAX_VRING) {
1296                 RTE_LOG(ERR, VHOST_CONFIG,
1297                         "invalid vring index: %u\n", vring_idx);
1298                 return -1;
1299         }
1300
1301         if (dev->virtqueue[vring_idx])
1302                 return 0;
1303
1304         return alloc_vring_queue(dev, vring_idx);
1305 }
1306
1307 static void
1308 vhost_user_lock_all_queue_pairs(struct virtio_net *dev)
1309 {
1310         unsigned int i = 0;
1311         unsigned int vq_num = 0;
1312
1313         while (vq_num < dev->nr_vring) {
1314                 struct vhost_virtqueue *vq = dev->virtqueue[i];
1315
1316                 if (vq) {
1317                         rte_spinlock_lock(&vq->access_lock);
1318                         vq_num++;
1319                 }
1320                 i++;
1321         }
1322 }
1323
1324 static void
1325 vhost_user_unlock_all_queue_pairs(struct virtio_net *dev)
1326 {
1327         unsigned int i = 0;
1328         unsigned int vq_num = 0;
1329
1330         while (vq_num < dev->nr_vring) {
1331                 struct vhost_virtqueue *vq = dev->virtqueue[i];
1332
1333                 if (vq) {
1334                         rte_spinlock_unlock(&vq->access_lock);
1335                         vq_num++;
1336                 }
1337                 i++;
1338         }
1339 }
1340
1341 int
1342 vhost_user_msg_handler(int vid, int fd)
1343 {
1344         struct virtio_net *dev;
1345         struct VhostUserMsg msg;
1346         int ret;
1347         int unlock_required = 0;
1348
1349         dev = get_device(vid);
1350         if (dev == NULL)
1351                 return -1;
1352
1353         if (!dev->notify_ops) {
1354                 dev->notify_ops = vhost_driver_callback_get(dev->ifname);
1355                 if (!dev->notify_ops) {
1356                         RTE_LOG(ERR, VHOST_CONFIG,
1357                                 "failed to get callback ops for driver %s\n",
1358                                 dev->ifname);
1359                         return -1;
1360                 }
1361         }
1362
1363         ret = read_vhost_message(fd, &msg);
1364         if (ret <= 0 || msg.request.master >= VHOST_USER_MAX) {
1365                 if (ret < 0)
1366                         RTE_LOG(ERR, VHOST_CONFIG,
1367                                 "vhost read message failed\n");
1368                 else if (ret == 0)
1369                         RTE_LOG(INFO, VHOST_CONFIG,
1370                                 "vhost peer closed\n");
1371                 else
1372                         RTE_LOG(ERR, VHOST_CONFIG,
1373                                 "vhost read incorrect message\n");
1374
1375                 return -1;
1376         }
1377
1378         ret = 0;
1379         if (msg.request.master != VHOST_USER_IOTLB_MSG)
1380                 RTE_LOG(INFO, VHOST_CONFIG, "read message %s\n",
1381                         vhost_message_str[msg.request.master]);
1382         else
1383                 RTE_LOG(DEBUG, VHOST_CONFIG, "read message %s\n",
1384                         vhost_message_str[msg.request.master]);
1385
1386         ret = vhost_user_check_and_alloc_queue_pair(dev, &msg);
1387         if (ret < 0) {
1388                 RTE_LOG(ERR, VHOST_CONFIG,
1389                         "failed to alloc queue\n");
1390                 return -1;
1391         }
1392
1393         /*
1394          * Note: we don't lock all queues on VHOST_USER_GET_VRING_BASE
1395          * and VHOST_USER_RESET_OWNER, since it is sent when virtio stops
1396          * and device is destroyed. destroy_device waits for queues to be
1397          * inactive, so it is safe. Otherwise taking the access_lock
1398          * would cause a dead lock.
1399          */
1400         switch (msg.request.master) {
1401         case VHOST_USER_SET_FEATURES:
1402         case VHOST_USER_SET_PROTOCOL_FEATURES:
1403         case VHOST_USER_SET_OWNER:
1404         case VHOST_USER_SET_MEM_TABLE:
1405         case VHOST_USER_SET_LOG_BASE:
1406         case VHOST_USER_SET_LOG_FD:
1407         case VHOST_USER_SET_VRING_NUM:
1408         case VHOST_USER_SET_VRING_ADDR:
1409         case VHOST_USER_SET_VRING_BASE:
1410         case VHOST_USER_SET_VRING_KICK:
1411         case VHOST_USER_SET_VRING_CALL:
1412         case VHOST_USER_SET_VRING_ERR:
1413         case VHOST_USER_SET_VRING_ENABLE:
1414         case VHOST_USER_SEND_RARP:
1415         case VHOST_USER_NET_SET_MTU:
1416         case VHOST_USER_SET_SLAVE_REQ_FD:
1417                 vhost_user_lock_all_queue_pairs(dev);
1418                 unlock_required = 1;
1419                 break;
1420         default:
1421                 break;
1422
1423         }
1424
1425         switch (msg.request.master) {
1426         case VHOST_USER_GET_FEATURES:
1427                 msg.payload.u64 = vhost_user_get_features(dev);
1428                 msg.size = sizeof(msg.payload.u64);
1429                 send_vhost_reply(fd, &msg);
1430                 break;
1431         case VHOST_USER_SET_FEATURES:
1432                 ret = vhost_user_set_features(dev, msg.payload.u64);
1433                 if (ret)
1434                         return -1;
1435                 break;
1436
1437         case VHOST_USER_GET_PROTOCOL_FEATURES:
1438                 vhost_user_get_protocol_features(dev, &msg);
1439                 send_vhost_reply(fd, &msg);
1440                 break;
1441         case VHOST_USER_SET_PROTOCOL_FEATURES:
1442                 vhost_user_set_protocol_features(dev, msg.payload.u64);
1443                 break;
1444
1445         case VHOST_USER_SET_OWNER:
1446                 vhost_user_set_owner();
1447                 break;
1448         case VHOST_USER_RESET_OWNER:
1449                 vhost_user_reset_owner(dev);
1450                 break;
1451
1452         case VHOST_USER_SET_MEM_TABLE:
1453                 ret = vhost_user_set_mem_table(dev, &msg);
1454                 break;
1455
1456         case VHOST_USER_SET_LOG_BASE:
1457                 vhost_user_set_log_base(dev, &msg);
1458
1459                 /* it needs a reply */
1460                 msg.size = sizeof(msg.payload.u64);
1461                 send_vhost_reply(fd, &msg);
1462                 break;
1463         case VHOST_USER_SET_LOG_FD:
1464                 close(msg.fds[0]);
1465                 RTE_LOG(INFO, VHOST_CONFIG, "not implemented.\n");
1466                 break;
1467
1468         case VHOST_USER_SET_VRING_NUM:
1469                 vhost_user_set_vring_num(dev, &msg);
1470                 break;
1471         case VHOST_USER_SET_VRING_ADDR:
1472                 vhost_user_set_vring_addr(&dev, &msg);
1473                 break;
1474         case VHOST_USER_SET_VRING_BASE:
1475                 vhost_user_set_vring_base(dev, &msg);
1476                 break;
1477
1478         case VHOST_USER_GET_VRING_BASE:
1479                 vhost_user_get_vring_base(dev, &msg);
1480                 msg.size = sizeof(msg.payload.state);
1481                 send_vhost_reply(fd, &msg);
1482                 break;
1483
1484         case VHOST_USER_SET_VRING_KICK:
1485                 vhost_user_set_vring_kick(&dev, &msg);
1486                 break;
1487         case VHOST_USER_SET_VRING_CALL:
1488                 vhost_user_set_vring_call(dev, &msg);
1489                 break;
1490
1491         case VHOST_USER_SET_VRING_ERR:
1492                 if (!(msg.payload.u64 & VHOST_USER_VRING_NOFD_MASK))
1493                         close(msg.fds[0]);
1494                 RTE_LOG(INFO, VHOST_CONFIG, "not implemented\n");
1495                 break;
1496
1497         case VHOST_USER_GET_QUEUE_NUM:
1498                 msg.payload.u64 = VHOST_MAX_QUEUE_PAIRS;
1499                 msg.size = sizeof(msg.payload.u64);
1500                 send_vhost_reply(fd, &msg);
1501                 break;
1502
1503         case VHOST_USER_SET_VRING_ENABLE:
1504                 vhost_user_set_vring_enable(dev, &msg);
1505                 break;
1506         case VHOST_USER_SEND_RARP:
1507                 vhost_user_send_rarp(dev, &msg);
1508                 break;
1509
1510         case VHOST_USER_NET_SET_MTU:
1511                 ret = vhost_user_net_set_mtu(dev, &msg);
1512                 break;
1513
1514         case VHOST_USER_SET_SLAVE_REQ_FD:
1515                 ret = vhost_user_set_req_fd(dev, &msg);
1516                 break;
1517
1518         case VHOST_USER_IOTLB_MSG:
1519                 ret = vhost_user_iotlb_msg(&dev, &msg);
1520                 break;
1521
1522         default:
1523                 ret = -1;
1524                 break;
1525
1526         }
1527
1528         if (unlock_required)
1529                 vhost_user_unlock_all_queue_pairs(dev);
1530
1531         if (msg.flags & VHOST_USER_NEED_REPLY) {
1532                 msg.payload.u64 = !!ret;
1533                 msg.size = sizeof(msg.payload.u64);
1534                 send_vhost_reply(fd, &msg);
1535         }
1536
1537         if (!(dev->flags & VIRTIO_DEV_RUNNING) && virtio_is_ready(dev)) {
1538                 dev->flags |= VIRTIO_DEV_READY;
1539
1540                 if (!(dev->flags & VIRTIO_DEV_RUNNING)) {
1541                         if (dev->dequeue_zero_copy) {
1542                                 RTE_LOG(INFO, VHOST_CONFIG,
1543                                                 "dequeue zero copy is enabled\n");
1544                         }
1545
1546                         if (dev->notify_ops->new_device(dev->vid) == 0)
1547                                 dev->flags |= VIRTIO_DEV_RUNNING;
1548                 }
1549         }
1550
1551         return 0;
1552 }
1553
1554 int
1555 vhost_user_iotlb_miss(struct virtio_net *dev, uint64_t iova, uint8_t perm)
1556 {
1557         int ret;
1558         struct VhostUserMsg msg = {
1559                 .request.slave = VHOST_USER_SLAVE_IOTLB_MSG,
1560                 .flags = VHOST_USER_VERSION,
1561                 .size = sizeof(msg.payload.iotlb),
1562                 .payload.iotlb = {
1563                         .iova = iova,
1564                         .perm = perm,
1565                         .type = VHOST_IOTLB_MISS,
1566                 },
1567         };
1568
1569         ret = send_vhost_message(dev->slave_req_fd, &msg);
1570         if (ret < 0) {
1571                 RTE_LOG(ERR, VHOST_CONFIG,
1572                                 "Failed to send IOTLB miss message (%d)\n",
1573                                 ret);
1574                 return ret;
1575         }
1576
1577         return 0;
1578 }