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