lib/librte_vhost/vhost_user.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2018 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         [VHOST_USER_CRYPTO_CREATE_SESS] = "VHOST_USER_CRYPTO_CREATE_SESS",
  71         [VHOST_USER_CRYPTO_CLOSE_SESS] = "VHOST_USER_CRYPTO_CLOSE_SESS",
  72 };
  73
  74 static uint64_t
  75 get_blk_size(int fd)
  76 {
  77         struct stat stat;
  78         int ret;
  79
  80         ret = fstat(fd, &stat);
  81         return ret == -1 ? (uint64_t)-1 : (uint64_t)stat.st_blksize;
  82 }
  83
  84 static void
  85 free_mem_region(struct virtio_net *dev)
  86 {
  87         uint32_t i;
  88         struct rte_vhost_mem_region *reg;
  89
  90         if (!dev || !dev->mem)
  91                 return;
  92
  93         for (i = 0; i < dev->mem->nregions; i++) {
  94                 reg = &dev->mem->regions[i];
  95                 if (reg->host_user_addr) {
  96                         munmap(reg->mmap_addr, reg->mmap_size);
  97                         close(reg->fd);
  98                 }
  99         }
 100 }
 101
 102 void
 103 vhost_backend_cleanup(struct virtio_net *dev)
 104 {
 105         if (dev->mem) {
 106                 free_mem_region(dev);
 107                 rte_free(dev->mem);
 108                 dev->mem = NULL;
 109         }
 110
 111         free(dev->guest_pages);
 112         dev->guest_pages = NULL;
 113
 114         if (dev->log_addr) {
 115                 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
 116                 dev->log_addr = 0;
 117         }
 118
 119         if (dev->slave_req_fd >= 0) {
 120                 close(dev->slave_req_fd);
 121                 dev->slave_req_fd = -1;
 122         }
 123 }
 124
 125 /*
 126  * This function just returns success at the moment unless
 127  * the device hasn't been initialised.
 128  */
 129 static int
 130 vhost_user_set_owner(void)
 131 {
 132         return 0;
 133 }
 134
 135 static int
 136 vhost_user_reset_owner(struct virtio_net *dev)
 137 {
 138         vhost_destroy_device_notify(dev);
 139
 140         cleanup_device(dev, 0);
 141         reset_device(dev);
 142         return 0;
 143 }
 144
 145 /*
 146  * The features that we support are requested.
 147  */
 148 static uint64_t
 149 vhost_user_get_features(struct virtio_net *dev, struct VhostUserMsg *msg)
 150 {
 151         uint64_t features = 0;
 152
 153         rte_vhost_driver_get_features(dev->ifname, &features);
 154
 155         msg->payload.u64 = features;
 156         msg->size = sizeof(msg->payload.u64);
 157
 158         return features;
 159 }
 160
 161 /*
 162  * The queue number that we support are requested.
 163  */
 164 static uint32_t
 165 vhost_user_get_queue_num(struct virtio_net *dev, struct VhostUserMsg *msg)
 166 {
 167         uint32_t queue_num = 0;
 168
 169         rte_vhost_driver_get_queue_num(dev->ifname, &queue_num);
 170
 171         msg->payload.u64 = (uint64_t)queue_num;
 172         msg->size = sizeof(msg->payload.u64);
 173
 174         return queue_num;
 175 }
 176
 177 /*
 178  * We receive the negotiated features supported by us and the virtio device.
 179  */
 180 static int
 181 vhost_user_set_features(struct virtio_net *dev, uint64_t features)
 182 {
 183         uint64_t vhost_features = 0;
 184         struct rte_vdpa_device *vdpa_dev;
 185         int did = -1;
 186
 187         rte_vhost_driver_get_features(dev->ifname, &vhost_features);
 188         if (features & ~vhost_features) {
 189                 RTE_LOG(ERR, VHOST_CONFIG,
 190                         "(%d) received invalid negotiated features.\n",
 191                         dev->vid);
 192                 return -1;
 193         }
 194
 195         if (dev->flags & VIRTIO_DEV_RUNNING) {
 196                 if (dev->features == features)
 197                         return 0;
 198
 199                 /*
 200                  * Error out if master tries to change features while device is
 201                  * in running state. The exception being VHOST_F_LOG_ALL, which
 202                  * is enabled when the live-migration starts.
 203                  */
 204                 if ((dev->features ^ features) & ~(1ULL << VHOST_F_LOG_ALL)) {
 205                         RTE_LOG(ERR, VHOST_CONFIG,
 206                                 "(%d) features changed while device is running.\n",
 207                                 dev->vid);
 208                         return -1;
 209                 }
 210
 211                 if (dev->notify_ops->features_changed)
 212                         dev->notify_ops->features_changed(dev->vid, features);
 213         }
 214
 215         dev->features = features;
 216         if (dev->features &
 217                 ((1 << VIRTIO_NET_F_MRG_RXBUF) | (1ULL << VIRTIO_F_VERSION_1))) {
 218                 dev->vhost_hlen = sizeof(struct virtio_net_hdr_mrg_rxbuf);
 219         } else {
 220                 dev->vhost_hlen = sizeof(struct virtio_net_hdr);
 221         }
 222         VHOST_LOG_DEBUG(VHOST_CONFIG,
 223                 "(%d) mergeable RX buffers %s, virtio 1 %s\n",
 224                 dev->vid,
 225                 (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ? "on" : "off",
 226                 (dev->features & (1ULL << VIRTIO_F_VERSION_1)) ? "on" : "off");
 227
 228         if ((dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET) &&
 229             !(dev->features & (1ULL << VIRTIO_NET_F_MQ))) {
 230                 /*
 231                  * Remove all but first queue pair if MQ hasn't been
 232                  * negotiated. This is safe because the device is not
 233                  * running at this stage.
 234                  */
 235                 while (dev->nr_vring > 2) {
 236                         struct vhost_virtqueue *vq;
 237
 238                         vq = dev->virtqueue[--dev->nr_vring];
 239                         if (!vq)
 240                                 continue;
 241
 242                         dev->virtqueue[dev->nr_vring] = NULL;
 243                         cleanup_vq(vq, 1);
 244                         free_vq(dev, vq);
 245                 }
 246         }
 247
 248         did = dev->vdpa_dev_id;
 249         vdpa_dev = rte_vdpa_get_device(did);
 250         if (vdpa_dev && vdpa_dev->ops->set_features)
 251                 vdpa_dev->ops->set_features(dev->vid);
 252
 253         return 0;
 254 }
 255
 256 /*
 257  * The virtio device sends us the size of the descriptor ring.
 258  */
 259 static int
 260 vhost_user_set_vring_num(struct virtio_net *dev,
 261                          struct VhostUserMsg *msg)
 262 {
 263         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
 264
 265         vq->size = msg->payload.state.num;
 266
 267         /* VIRTIO 1.0, 2.4 Virtqueues says:
 268          *
 269          *   Queue Size value is always a power of 2. The maximum Queue Size
 270          *   value is 32768.
 271          */
 272         if ((vq->size & (vq->size - 1)) || vq->size > 32768) {
 273                 RTE_LOG(ERR, VHOST_CONFIG,
 274                         "invalid virtqueue size %u\n", vq->size);
 275                 return -1;
 276         }
 277
 278         if (dev->dequeue_zero_copy) {
 279                 vq->nr_zmbuf = 0;
 280                 vq->last_zmbuf_idx = 0;
 281                 vq->zmbuf_size = vq->size;
 282                 vq->zmbufs = rte_zmalloc(NULL, vq->zmbuf_size *
 283                                          sizeof(struct zcopy_mbuf), 0);
 284                 if (vq->zmbufs == NULL) {
 285                         RTE_LOG(WARNING, VHOST_CONFIG,
 286                                 "failed to allocate mem for zero copy; "
 287                                 "zero copy is force disabled\n");
 288                         dev->dequeue_zero_copy = 0;
 289                 }
 290                 TAILQ_INIT(&vq->zmbuf_list);
 291         }
 292
 293         if (vq_is_packed(dev)) {
 294                 vq->shadow_used_packed = rte_malloc(NULL,
 295                                 vq->size *
 296                                 sizeof(struct vring_used_elem_packed),
 297                                 RTE_CACHE_LINE_SIZE);
 298                 if (!vq->shadow_used_packed) {
 299                         RTE_LOG(ERR, VHOST_CONFIG,
 300                                         "failed to allocate memory for shadow used ring.\n");
 301                         return -1;
 302                 }
 303
 304         } else {
 305                 vq->shadow_used_split = rte_malloc(NULL,
 306                                 vq->size * sizeof(struct vring_used_elem),
 307                                 RTE_CACHE_LINE_SIZE);
 308                 if (!vq->shadow_used_split) {
 309                         RTE_LOG(ERR, VHOST_CONFIG,
 310                                         "failed to allocate memory for shadow used ring.\n");
 311                         return -1;
 312                 }
 313         }
 314
 315         vq->batch_copy_elems = rte_malloc(NULL,
 316                                 vq->size * sizeof(struct batch_copy_elem),
 317                                 RTE_CACHE_LINE_SIZE);
 318         if (!vq->batch_copy_elems) {
 319                 RTE_LOG(ERR, VHOST_CONFIG,
 320                         "failed to allocate memory for batching copy.\n");
 321                 return -1;
 322         }
 323
 324         return 0;
 325 }
 326
 327 /*
 328  * Reallocate virtio_dev and vhost_virtqueue data structure to make them on the
 329  * same numa node as the memory of vring descriptor.
 330  */
 331 #ifdef RTE_LIBRTE_VHOST_NUMA
 332 static struct virtio_net*
 333 numa_realloc(struct virtio_net *dev, int index)
 334 {
 335         int oldnode, newnode;
 336         struct virtio_net *old_dev;
 337         struct vhost_virtqueue *old_vq, *vq;
 338         struct zcopy_mbuf *new_zmbuf;
 339         struct vring_used_elem *new_shadow_used_split;
 340         struct vring_used_elem_packed *new_shadow_used_packed;
 341         struct batch_copy_elem *new_batch_copy_elems;
 342         int ret;
 343
 344         old_dev = dev;
 345         vq = old_vq = dev->virtqueue[index];
 346
 347         ret = get_mempolicy(&newnode, NULL, 0, old_vq->desc,
 348                             MPOL_F_NODE | MPOL_F_ADDR);
 349
 350         /* check if we need to reallocate vq */
 351         ret |= get_mempolicy(&oldnode, NULL, 0, old_vq,
 352                              MPOL_F_NODE | MPOL_F_ADDR);
 353         if (ret) {
 354                 RTE_LOG(ERR, VHOST_CONFIG,
 355                         "Unable to get vq numa information.\n");
 356                 return dev;
 357         }
 358         if (oldnode != newnode) {
 359                 RTE_LOG(INFO, VHOST_CONFIG,
 360                         "reallocate vq from %d to %d node\n", oldnode, newnode);
 361                 vq = rte_malloc_socket(NULL, sizeof(*vq), 0, newnode);
 362                 if (!vq)
 363                         return dev;
 364
 365                 memcpy(vq, old_vq, sizeof(*vq));
 366                 TAILQ_INIT(&vq->zmbuf_list);
 367
 368                 if (dev->dequeue_zero_copy) {
 369                         new_zmbuf = rte_malloc_socket(NULL, vq->zmbuf_size *
 370                                         sizeof(struct zcopy_mbuf), 0, newnode);
 371                         if (new_zmbuf) {
 372                                 rte_free(vq->zmbufs);
 373                                 vq->zmbufs = new_zmbuf;
 374                         }
 375                 }
 376
 377                 if (vq_is_packed(dev)) {
 378                         new_shadow_used_packed = rte_malloc_socket(NULL,
 379                                         vq->size *
 380                                         sizeof(struct vring_used_elem_packed),
 381                                         RTE_CACHE_LINE_SIZE,
 382                                         newnode);
 383                         if (new_shadow_used_packed) {
 384                                 rte_free(vq->shadow_used_packed);
 385                                 vq->shadow_used_packed = new_shadow_used_packed;
 386                         }
 387                 } else {
 388                         new_shadow_used_split = rte_malloc_socket(NULL,
 389                                         vq->size *
 390                                         sizeof(struct vring_used_elem),
 391                                         RTE_CACHE_LINE_SIZE,
 392                                         newnode);
 393                         if (new_shadow_used_split) {
 394                                 rte_free(vq->shadow_used_split);
 395                                 vq->shadow_used_split = new_shadow_used_split;
 396                         }
 397                 }
 398
 399                 new_batch_copy_elems = rte_malloc_socket(NULL,
 400                         vq->size * sizeof(struct batch_copy_elem),
 401                         RTE_CACHE_LINE_SIZE,
 402                         newnode);
 403                 if (new_batch_copy_elems) {
 404                         rte_free(vq->batch_copy_elems);
 405                         vq->batch_copy_elems = new_batch_copy_elems;
 406                 }
 407
 408                 rte_free(old_vq);
 409         }
 410
 411         /* check if we need to reallocate dev */
 412         ret = get_mempolicy(&oldnode, NULL, 0, old_dev,
 413                             MPOL_F_NODE | MPOL_F_ADDR);
 414         if (ret) {
 415                 RTE_LOG(ERR, VHOST_CONFIG,
 416                         "Unable to get dev numa information.\n");
 417                 goto out;
 418         }
 419         if (oldnode != newnode) {
 420                 RTE_LOG(INFO, VHOST_CONFIG,
 421                         "reallocate dev from %d to %d node\n",
 422                         oldnode, newnode);
 423                 dev = rte_malloc_socket(NULL, sizeof(*dev), 0, newnode);
 424                 if (!dev) {
 425                         dev = old_dev;
 426                         goto out;
 427                 }
 428
 429                 memcpy(dev, old_dev, sizeof(*dev));
 430                 rte_free(old_dev);
 431         }
 432
 433 out:
 434         dev->virtqueue[index] = vq;
 435         vhost_devices[dev->vid] = dev;
 436
 437         if (old_vq != vq)
 438                 vhost_user_iotlb_init(dev, index);
 439
 440         return dev;
 441 }
 442 #else
 443 static struct virtio_net*
 444 numa_realloc(struct virtio_net *dev, int index __rte_unused)
 445 {
 446         return dev;
 447 }
 448 #endif
 449
 450 /* Converts QEMU virtual address to Vhost virtual address. */
 451 static uint64_t
 452 qva_to_vva(struct virtio_net *dev, uint64_t qva, uint64_t *len)
 453 {
 454         struct rte_vhost_mem_region *r;
 455         uint32_t i;
 456
 457         /* Find the region where the address lives. */
 458         for (i = 0; i < dev->mem->nregions; i++) {
 459                 r = &dev->mem->regions[i];
 460
 461                 if (qva >= r->guest_user_addr &&
 462                     qva <  r->guest_user_addr + r->size) {
 463
 464                         if (unlikely(*len > r->guest_user_addr + r->size - qva))
 465                                 *len = r->guest_user_addr + r->size - qva;
 466
 467                         return qva - r->guest_user_addr +
 468                                r->host_user_addr;
 469                 }
 470         }
 471         *len = 0;
 472
 473         return 0;
 474 }
 475
 476
 477 /*
 478  * Converts ring address to Vhost virtual address.
 479  * If IOMMU is enabled, the ring address is a guest IO virtual address,
 480  * else it is a QEMU virtual address.
 481  */
 482 static uint64_t
 483 ring_addr_to_vva(struct virtio_net *dev, struct vhost_virtqueue *vq,
 484                 uint64_t ra, uint64_t *size)
 485 {
 486         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)) {
 487                 uint64_t vva;
 488
 489                 vva = vhost_user_iotlb_cache_find(vq, ra,
 490                                         size, VHOST_ACCESS_RW);
 491                 if (!vva)
 492                         vhost_user_iotlb_miss(dev, ra, VHOST_ACCESS_RW);
 493
 494                 return vva;
 495         }
 496
 497         return qva_to_vva(dev, ra, size);
 498 }
 499
 500 static struct virtio_net *
 501 translate_ring_addresses(struct virtio_net *dev, int vq_index)
 502 {
 503         struct vhost_virtqueue *vq = dev->virtqueue[vq_index];
 504         struct vhost_vring_addr *addr = &vq->ring_addrs;
 505         uint64_t len;
 506
 507         if (vq_is_packed(dev)) {
 508                 len = sizeof(struct vring_packed_desc) * vq->size;
 509                 vq->desc_packed = (struct vring_packed_desc *)(uintptr_t)
 510                         ring_addr_to_vva(dev, vq, addr->desc_user_addr, &len);
 511                 vq->log_guest_addr = 0;
 512                 if (vq->desc_packed == NULL ||
 513                                 len != sizeof(struct vring_packed_desc) *
 514                                 vq->size) {
 515                         RTE_LOG(DEBUG, VHOST_CONFIG,
 516                                 "(%d) failed to map desc_packed ring.\n",
 517                                 dev->vid);
 518                         return dev;
 519                 }
 520
 521                 dev = numa_realloc(dev, vq_index);
 522                 vq = dev->virtqueue[vq_index];
 523                 addr = &vq->ring_addrs;
 524
 525                 len = sizeof(struct vring_packed_desc_event);
 526                 vq->driver_event = (struct vring_packed_desc_event *)
 527                                         (uintptr_t)ring_addr_to_vva(dev,
 528                                         vq, addr->avail_user_addr, &len);
 529                 if (vq->driver_event == NULL ||
 530                                 len != sizeof(struct vring_packed_desc_event)) {
 531                         RTE_LOG(DEBUG, VHOST_CONFIG,
 532                                 "(%d) failed to find driver area address.\n",
 533                                 dev->vid);
 534                         return dev;
 535                 }
 536
 537                 len = sizeof(struct vring_packed_desc_event);
 538                 vq->device_event = (struct vring_packed_desc_event *)
 539                                         (uintptr_t)ring_addr_to_vva(dev,
 540                                         vq, addr->used_user_addr, &len);
 541                 if (vq->device_event == NULL ||
 542                                 len != sizeof(struct vring_packed_desc_event)) {
 543                         RTE_LOG(DEBUG, VHOST_CONFIG,
 544                                 "(%d) failed to find device area address.\n",
 545                                 dev->vid);
 546                         return dev;
 547                 }
 548
 549                 return dev;
 550         }
 551
 552         /* The addresses are converted from QEMU virtual to Vhost virtual. */
 553         if (vq->desc && vq->avail && vq->used)
 554                 return dev;
 555
 556         len = sizeof(struct vring_desc) * vq->size;
 557         vq->desc = (struct vring_desc *)(uintptr_t)ring_addr_to_vva(dev,
 558                         vq, addr->desc_user_addr, &len);
 559         if (vq->desc == 0 || len != sizeof(struct vring_desc) * vq->size) {
 560                 RTE_LOG(DEBUG, VHOST_CONFIG,
 561                         "(%d) failed to map desc ring.\n",
 562                         dev->vid);
 563                 return dev;
 564         }
 565
 566         dev = numa_realloc(dev, vq_index);
 567         vq = dev->virtqueue[vq_index];
 568         addr = &vq->ring_addrs;
 569
 570         len = sizeof(struct vring_avail) + sizeof(uint16_t) * vq->size;
 571         vq->avail = (struct vring_avail *)(uintptr_t)ring_addr_to_vva(dev,
 572                         vq, addr->avail_user_addr, &len);
 573         if (vq->avail == 0 ||
 574                         len != sizeof(struct vring_avail) +
 575                         sizeof(uint16_t) * vq->size) {
 576                 RTE_LOG(DEBUG, VHOST_CONFIG,
 577                         "(%d) failed to map avail ring.\n",
 578                         dev->vid);
 579                 return dev;
 580         }
 581
 582         len = sizeof(struct vring_used) +
 583                 sizeof(struct vring_used_elem) * vq->size;
 584         vq->used = (struct vring_used *)(uintptr_t)ring_addr_to_vva(dev,
 585                         vq, addr->used_user_addr, &len);
 586         if (vq->used == 0 || len != sizeof(struct vring_used) +
 587                         sizeof(struct vring_used_elem) * vq->size) {
 588                 RTE_LOG(DEBUG, VHOST_CONFIG,
 589                         "(%d) failed to map used ring.\n",
 590                         dev->vid);
 591                 return dev;
 592         }
 593
 594         if (vq->last_used_idx != vq->used->idx) {
 595                 RTE_LOG(WARNING, VHOST_CONFIG,
 596                         "last_used_idx (%u) and vq->used->idx (%u) mismatches; "
 597                         "some packets maybe resent for Tx and dropped for Rx\n",
 598                         vq->last_used_idx, vq->used->idx);
 599                 vq->last_used_idx  = vq->used->idx;
 600                 vq->last_avail_idx = vq->used->idx;
 601         }
 602
 603         vq->log_guest_addr = addr->log_guest_addr;
 604
 605         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address desc: %p\n",
 606                         dev->vid, vq->desc);
 607         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address avail: %p\n",
 608                         dev->vid, vq->avail);
 609         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address used: %p\n",
 610                         dev->vid, vq->used);
 611         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) log_guest_addr: %" PRIx64 "\n",
 612                         dev->vid, vq->log_guest_addr);
 613
 614         return dev;
 615 }
 616
 617 /*
 618  * The virtio device sends us the desc, used and avail ring addresses.
 619  * This function then converts these to our address space.
 620  */
 621 static int
 622 vhost_user_set_vring_addr(struct virtio_net **pdev, struct VhostUserMsg *msg)
 623 {
 624         struct vhost_virtqueue *vq;
 625         struct vhost_vring_addr *addr = &msg->payload.addr;
 626         struct virtio_net *dev = *pdev;
 627
 628         if (dev->mem == NULL)
 629                 return -1;
 630
 631         /* addr->index refers to the queue index. The txq 1, rxq is 0. */
 632         vq = dev->virtqueue[msg->payload.addr.index];
 633
 634         /*
 635          * Rings addresses should not be interpreted as long as the ring is not
 636          * started and enabled
 637          */
 638         memcpy(&vq->ring_addrs, addr, sizeof(*addr));
 639
 640         vring_invalidate(dev, vq);
 641
 642         if (vq->enabled && (dev->features &
 643                                 (1ULL << VHOST_USER_F_PROTOCOL_FEATURES))) {
 644                 dev = translate_ring_addresses(dev, msg->payload.addr.index);
 645                 if (!dev)
 646                         return -1;
 647
 648                 *pdev = dev;
 649         }
 650
 651         return 0;
 652 }
 653
 654 /*
 655  * The virtio device sends us the available ring last used index.
 656  */
 657 static int
 658 vhost_user_set_vring_base(struct virtio_net *dev,
 659                           struct VhostUserMsg *msg)
 660 {
 661         dev->virtqueue[msg->payload.state.index]->last_used_idx  =
 662                         msg->payload.state.num;
 663         dev->virtqueue[msg->payload.state.index]->last_avail_idx =
 664                         msg->payload.state.num;
 665
 666         return 0;
 667 }
 668
 669 static int
 670 add_one_guest_page(struct virtio_net *dev, uint64_t guest_phys_addr,
 671                    uint64_t host_phys_addr, uint64_t size)
 672 {
 673         struct guest_page *page, *last_page;
 674
 675         if (dev->nr_guest_pages == dev->max_guest_pages) {
 676                 dev->max_guest_pages *= 2;
 677                 dev->guest_pages = realloc(dev->guest_pages,
 678                                         dev->max_guest_pages * sizeof(*page));
 679                 if (!dev->guest_pages) {
 680                         RTE_LOG(ERR, VHOST_CONFIG, "cannot realloc guest_pages\n");
 681                         return -1;
 682                 }
 683         }
 684
 685         if (dev->nr_guest_pages > 0) {
 686                 last_page = &dev->guest_pages[dev->nr_guest_pages - 1];
 687                 /* merge if the two pages are continuous */
 688                 if (host_phys_addr == last_page->host_phys_addr +
 689                                       last_page->size) {
 690                         last_page->size += size;
 691                         return 0;
 692                 }
 693         }
 694
 695         page = &dev->guest_pages[dev->nr_guest_pages++];
 696         page->guest_phys_addr = guest_phys_addr;
 697         page->host_phys_addr  = host_phys_addr;
 698         page->size = size;
 699
 700         return 0;
 701 }
 702
 703 static int
 704 add_guest_pages(struct virtio_net *dev, struct rte_vhost_mem_region *reg,
 705                 uint64_t page_size)
 706 {
 707         uint64_t reg_size = reg->size;
 708         uint64_t host_user_addr  = reg->host_user_addr;
 709         uint64_t guest_phys_addr = reg->guest_phys_addr;
 710         uint64_t host_phys_addr;
 711         uint64_t size;
 712
 713         host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)host_user_addr);
 714         size = page_size - (guest_phys_addr & (page_size - 1));
 715         size = RTE_MIN(size, reg_size);
 716
 717         if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr, size) < 0)
 718                 return -1;
 719
 720         host_user_addr  += size;
 721         guest_phys_addr += size;
 722         reg_size -= size;
 723
 724         while (reg_size > 0) {
 725                 size = RTE_MIN(reg_size, page_size);
 726                 host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)
 727                                                   host_user_addr);
 728                 if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr,
 729                                 size) < 0)
 730                         return -1;
 731
 732                 host_user_addr  += size;
 733                 guest_phys_addr += size;
 734                 reg_size -= size;
 735         }
 736
 737         return 0;
 738 }
 739
 740 #ifdef RTE_LIBRTE_VHOST_DEBUG
 741 /* TODO: enable it only in debug mode? */
 742 static void
 743 dump_guest_pages(struct virtio_net *dev)
 744 {
 745         uint32_t i;
 746         struct guest_page *page;
 747
 748         for (i = 0; i < dev->nr_guest_pages; i++) {
 749                 page = &dev->guest_pages[i];
 750
 751                 RTE_LOG(INFO, VHOST_CONFIG,
 752                         "guest physical page region %u\n"
 753                         "\t guest_phys_addr: %" PRIx64 "\n"
 754                         "\t host_phys_addr : %" PRIx64 "\n"
 755                         "\t size           : %" PRIx64 "\n",
 756                         i,
 757                         page->guest_phys_addr,
 758                         page->host_phys_addr,
 759                         page->size);
 760         }
 761 }
 762 #else
 763 #define dump_guest_pages(dev)
 764 #endif
 765
 766 static bool
 767 vhost_memory_changed(struct VhostUserMemory *new,
 768                      struct rte_vhost_memory *old)
 769 {
 770         uint32_t i;
 771
 772         if (new->nregions != old->nregions)
 773                 return true;
 774
 775         for (i = 0; i < new->nregions; ++i) {
 776                 VhostUserMemoryRegion *new_r = &new->regions[i];
 777                 struct rte_vhost_mem_region *old_r = &old->regions[i];
 778
 779                 if (new_r->guest_phys_addr != old_r->guest_phys_addr)
 780                         return true;
 781                 if (new_r->memory_size != old_r->size)
 782                         return true;
 783                 if (new_r->userspace_addr != old_r->guest_user_addr)
 784                         return true;
 785         }
 786
 787         return false;
 788 }
 789
 790 static int
 791 vhost_user_set_mem_table(struct virtio_net **pdev, struct VhostUserMsg *msg)
 792 {
 793         struct virtio_net *dev = *pdev;
 794         struct VhostUserMemory memory = msg->payload.memory;
 795         struct rte_vhost_mem_region *reg;
 796         void *mmap_addr;
 797         uint64_t mmap_size;
 798         uint64_t mmap_offset;
 799         uint64_t alignment;
 800         uint32_t i;
 801         int populate;
 802         int fd;
 803
 804         if (memory.nregions > VHOST_MEMORY_MAX_NREGIONS) {
 805                 RTE_LOG(ERR, VHOST_CONFIG,
 806                         "too many memory regions (%u)\n", memory.nregions);
 807                 return -1;
 808         }
 809
 810         if (dev->mem && !vhost_memory_changed(&memory, dev->mem)) {
 811                 RTE_LOG(INFO, VHOST_CONFIG,
 812                         "(%d) memory regions not changed\n", dev->vid);
 813
 814                 for (i = 0; i < memory.nregions; i++)
 815                         close(msg->fds[i]);
 816
 817                 return 0;
 818         }
 819
 820         if (dev->mem) {
 821                 free_mem_region(dev);
 822                 rte_free(dev->mem);
 823                 dev->mem = NULL;
 824         }
 825
 826         /* Flush IOTLB cache as previous HVAs are now invalid */
 827         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 828                 for (i = 0; i < dev->nr_vring; i++)
 829                         vhost_user_iotlb_flush_all(dev->virtqueue[i]);
 830
 831         dev->nr_guest_pages = 0;
 832         if (!dev->guest_pages) {
 833                 dev->max_guest_pages = 8;
 834                 dev->guest_pages = malloc(dev->max_guest_pages *
 835                                                 sizeof(struct guest_page));
 836                 if (dev->guest_pages == NULL) {
 837                         RTE_LOG(ERR, VHOST_CONFIG,
 838                                 "(%d) failed to allocate memory "
 839                                 "for dev->guest_pages\n",
 840                                 dev->vid);
 841                         return -1;
 842                 }
 843         }
 844
 845         dev->mem = rte_zmalloc("vhost-mem-table", sizeof(struct rte_vhost_memory) +
 846                 sizeof(struct rte_vhost_mem_region) * memory.nregions, 0);
 847         if (dev->mem == NULL) {
 848                 RTE_LOG(ERR, VHOST_CONFIG,
 849                         "(%d) failed to allocate memory for dev->mem\n",
 850                         dev->vid);
 851                 return -1;
 852         }
 853         dev->mem->nregions = memory.nregions;
 854
 855         for (i = 0; i < memory.nregions; i++) {
 856                 fd  = msg->fds[i];
 857                 reg = &dev->mem->regions[i];
 858
 859                 reg->guest_phys_addr = memory.regions[i].guest_phys_addr;
 860                 reg->guest_user_addr = memory.regions[i].userspace_addr;
 861                 reg->size            = memory.regions[i].memory_size;
 862                 reg->fd              = fd;
 863
 864                 mmap_offset = memory.regions[i].mmap_offset;
 865
 866                 /* Check for memory_size + mmap_offset overflow */
 867                 if (mmap_offset >= -reg->size) {
 868                         RTE_LOG(ERR, VHOST_CONFIG,
 869                                 "mmap_offset (%#"PRIx64") and memory_size "
 870                                 "(%#"PRIx64") overflow\n",
 871                                 mmap_offset, reg->size);
 872                         goto err_mmap;
 873                 }
 874
 875                 mmap_size = reg->size + mmap_offset;
 876
 877                 /* mmap() without flag of MAP_ANONYMOUS, should be called
 878                  * with length argument aligned with hugepagesz at older
 879                  * longterm version Linux, like 2.6.32 and 3.2.72, or
 880                  * mmap() will fail with EINVAL.
 881                  *
 882                  * to avoid failure, make sure in caller to keep length
 883                  * aligned.
 884                  */
 885                 alignment = get_blk_size(fd);
 886                 if (alignment == (uint64_t)-1) {
 887                         RTE_LOG(ERR, VHOST_CONFIG,
 888                                 "couldn't get hugepage size through fstat\n");
 889                         goto err_mmap;
 890                 }
 891                 mmap_size = RTE_ALIGN_CEIL(mmap_size, alignment);
 892
 893                 populate = (dev->dequeue_zero_copy) ? MAP_POPULATE : 0;
 894                 mmap_addr = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE,
 895                                  MAP_SHARED | populate, fd, 0);
 896
 897                 if (mmap_addr == MAP_FAILED) {
 898                         RTE_LOG(ERR, VHOST_CONFIG,
 899                                 "mmap region %u failed.\n", i);
 900                         goto err_mmap;
 901                 }
 902
 903                 reg->mmap_addr = mmap_addr;
 904                 reg->mmap_size = mmap_size;
 905                 reg->host_user_addr = (uint64_t)(uintptr_t)mmap_addr +
 906                                       mmap_offset;
 907
 908                 if (dev->dequeue_zero_copy)
 909                         if (add_guest_pages(dev, reg, alignment) < 0) {
 910                                 RTE_LOG(ERR, VHOST_CONFIG,
 911                                         "adding guest pages to region %u failed.\n",
 912                                         i);
 913                                 goto err_mmap;
 914                         }
 915
 916                 RTE_LOG(INFO, VHOST_CONFIG,
 917                         "guest memory region %u, size: 0x%" PRIx64 "\n"
 918                         "\t guest physical addr: 0x%" PRIx64 "\n"
 919                         "\t guest virtual  addr: 0x%" PRIx64 "\n"
 920                         "\t host  virtual  addr: 0x%" PRIx64 "\n"
 921                         "\t mmap addr : 0x%" PRIx64 "\n"
 922                         "\t mmap size : 0x%" PRIx64 "\n"
 923                         "\t mmap align: 0x%" PRIx64 "\n"
 924                         "\t mmap off  : 0x%" PRIx64 "\n",
 925                         i, reg->size,
 926                         reg->guest_phys_addr,
 927                         reg->guest_user_addr,
 928                         reg->host_user_addr,
 929                         (uint64_t)(uintptr_t)mmap_addr,
 930                         mmap_size,
 931                         alignment,
 932                         mmap_offset);
 933         }
 934
 935         for (i = 0; i < dev->nr_vring; i++) {
 936                 struct vhost_virtqueue *vq = dev->virtqueue[i];
 937
 938                 if (vq->desc || vq->avail || vq->used) {
 939                         /*
 940                          * If the memory table got updated, the ring addresses
 941                          * need to be translated again as virtual addresses have
 942                          * changed.
 943                          */
 944                         vring_invalidate(dev, vq);
 945
 946                         dev = translate_ring_addresses(dev, i);
 947                         if (!dev)
 948                                 return -1;
 949
 950                         *pdev = dev;
 951                 }
 952         }
 953
 954         dump_guest_pages(dev);
 955
 956         return 0;
 957
 958 err_mmap:
 959         free_mem_region(dev);
 960         rte_free(dev->mem);
 961         dev->mem = NULL;
 962         return -1;
 963 }
 964
 965 static bool
 966 vq_is_ready(struct virtio_net *dev, struct vhost_virtqueue *vq)
 967 {
 968         bool rings_ok;
 969
 970         if (!vq)
 971                 return false;
 972
 973         if (vq_is_packed(dev))
 974                 rings_ok = !!vq->desc_packed;
 975         else
 976                 rings_ok = vq->desc && vq->avail && vq->used;
 977
 978         return rings_ok &&
 979                vq->kickfd != VIRTIO_UNINITIALIZED_EVENTFD &&
 980                vq->callfd != VIRTIO_UNINITIALIZED_EVENTFD;
 981 }
 982
 983 static int
 984 virtio_is_ready(struct virtio_net *dev)
 985 {
 986         struct vhost_virtqueue *vq;
 987         uint32_t i;
 988
 989         if (dev->nr_vring == 0)
 990                 return 0;
 991
 992         for (i = 0; i < dev->nr_vring; i++) {
 993                 vq = dev->virtqueue[i];
 994
 995                 if (!vq_is_ready(dev, vq))
 996                         return 0;
 997         }
 998
 999         RTE_LOG(INFO, VHOST_CONFIG,
1000                 "virtio is now ready for processing.\n");
1001         return 1;
1002 }
1003
1004 static void
1005 vhost_user_set_vring_call(struct virtio_net *dev, struct VhostUserMsg *msg)
1006 {
1007         struct vhost_vring_file file;
1008         struct vhost_virtqueue *vq;
1009
1010         file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1011         if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1012                 file.fd = VIRTIO_INVALID_EVENTFD;
1013         else
1014                 file.fd = msg->fds[0];
1015         RTE_LOG(INFO, VHOST_CONFIG,
1016                 "vring call idx:%d file:%d\n", file.index, file.fd);
1017
1018         vq = dev->virtqueue[file.index];
1019         if (vq->callfd >= 0)
1020                 close(vq->callfd);
1021
1022         vq->callfd = file.fd;
1023 }
1024
1025 static void vhost_user_set_vring_err(struct virtio_net **pdev __rte_unused,
1026                         struct VhostUserMsg *msg)
1027 {
1028         if (!(msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK))
1029                 close(msg->fds[0]);
1030         RTE_LOG(INFO, VHOST_CONFIG, "not implemented\n");
1031 }
1032
1033 static int
1034 vhost_user_set_vring_kick(struct virtio_net **pdev, struct VhostUserMsg *msg)
1035 {
1036         struct vhost_vring_file file;
1037         struct vhost_virtqueue *vq;
1038         struct virtio_net *dev = *pdev;
1039
1040         file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1041         if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1042                 file.fd = VIRTIO_INVALID_EVENTFD;
1043         else
1044                 file.fd = msg->fds[0];
1045         RTE_LOG(INFO, VHOST_CONFIG,
1046                 "vring kick idx:%d file:%d\n", file.index, file.fd);
1047
1048         /* Interpret ring addresses only when ring is started. */
1049         dev = translate_ring_addresses(dev, file.index);
1050         if (!dev)
1051                 return -1;
1052
1053         *pdev = dev;
1054
1055         vq = dev->virtqueue[file.index];
1056
1057         /*
1058          * When VHOST_USER_F_PROTOCOL_FEATURES is not negotiated,
1059          * the ring starts already enabled. Otherwise, it is enabled via
1060          * the SET_VRING_ENABLE message.
1061          */
1062         if (!(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES)))
1063                 vq->enabled = 1;
1064
1065         if (vq->kickfd >= 0)
1066                 close(vq->kickfd);
1067         vq->kickfd = file.fd;
1068         return 0;
1069 }
1070
1071 static void
1072 free_zmbufs(struct vhost_virtqueue *vq)
1073 {
1074         struct zcopy_mbuf *zmbuf, *next;
1075
1076         for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1077              zmbuf != NULL; zmbuf = next) {
1078                 next = TAILQ_NEXT(zmbuf, next);
1079
1080                 rte_pktmbuf_free(zmbuf->mbuf);
1081                 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1082         }
1083
1084         rte_free(vq->zmbufs);
1085 }
1086
1087 /*
1088  * when virtio is stopped, qemu will send us the GET_VRING_BASE message.
1089  */
1090 static int
1091 vhost_user_get_vring_base(struct virtio_net *dev,
1092                           struct VhostUserMsg *msg)
1093 {
1094         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
1095
1096         /* We have to stop the queue (virtio) if it is running. */
1097         vhost_destroy_device_notify(dev);
1098
1099         dev->flags &= ~VIRTIO_DEV_READY;
1100         dev->flags &= ~VIRTIO_DEV_VDPA_CONFIGURED;
1101
1102         /* Here we are safe to get the last avail index */
1103         msg->payload.state.num = vq->last_avail_idx;
1104
1105         RTE_LOG(INFO, VHOST_CONFIG,
1106                 "vring base idx:%d file:%d\n", msg->payload.state.index,
1107                 msg->payload.state.num);
1108         /*
1109          * Based on current qemu vhost-user implementation, this message is
1110          * sent and only sent in vhost_vring_stop.
1111          * TODO: cleanup the vring, it isn't usable since here.
1112          */
1113         if (vq->kickfd >= 0)
1114                 close(vq->kickfd);
1115
1116         vq->kickfd = VIRTIO_UNINITIALIZED_EVENTFD;
1117
1118         if (vq->callfd >= 0)
1119                 close(vq->callfd);
1120
1121         vq->callfd = VIRTIO_UNINITIALIZED_EVENTFD;
1122
1123         if (dev->dequeue_zero_copy)
1124                 free_zmbufs(vq);
1125         if (vq_is_packed(dev)) {
1126                 rte_free(vq->shadow_used_packed);
1127                 vq->shadow_used_packed = NULL;
1128         } else {
1129                 rte_free(vq->shadow_used_split);
1130                 vq->shadow_used_split = NULL;
1131         }
1132
1133         rte_free(vq->batch_copy_elems);
1134         vq->batch_copy_elems = NULL;
1135
1136         msg->size = sizeof(msg->payload.state);
1137
1138         return 0;
1139 }
1140
1141 /*
1142  * when virtio queues are ready to work, qemu will send us to
1143  * enable the virtio queue pair.
1144  */
1145 static int
1146 vhost_user_set_vring_enable(struct virtio_net *dev,
1147                             struct VhostUserMsg *msg)
1148 {
1149         int enable = (int)msg->payload.state.num;
1150         int index = (int)msg->payload.state.index;
1151         struct rte_vdpa_device *vdpa_dev;
1152         int did = -1;
1153
1154         RTE_LOG(INFO, VHOST_CONFIG,
1155                 "set queue enable: %d to qp idx: %d\n",
1156                 enable, index);
1157
1158         did = dev->vdpa_dev_id;
1159         vdpa_dev = rte_vdpa_get_device(did);
1160         if (vdpa_dev && vdpa_dev->ops->set_vring_state)
1161                 vdpa_dev->ops->set_vring_state(dev->vid, index, enable);
1162
1163         if (dev->notify_ops->vring_state_changed)
1164                 dev->notify_ops->vring_state_changed(dev->vid,
1165                                 index, enable);
1166
1167         dev->virtqueue[index]->enabled = enable;
1168
1169         return 0;
1170 }
1171
1172 static void
1173 vhost_user_get_protocol_features(struct virtio_net *dev,
1174                                  struct VhostUserMsg *msg)
1175 {
1176         uint64_t features, protocol_features;
1177
1178         rte_vhost_driver_get_features(dev->ifname, &features);
1179         rte_vhost_driver_get_protocol_features(dev->ifname, &protocol_features);
1180
1181         /*
1182          * REPLY_ACK protocol feature is only mandatory for now
1183          * for IOMMU feature. If IOMMU is explicitly disabled by the
1184          * application, disable also REPLY_ACK feature for older buggy
1185          * Qemu versions (from v2.7.0 to v2.9.0).
1186          */
1187         if (!(features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)))
1188                 protocol_features &= ~(1ULL << VHOST_USER_PROTOCOL_F_REPLY_ACK);
1189
1190         msg->payload.u64 = protocol_features;
1191         msg->size = sizeof(msg->payload.u64);
1192 }
1193
1194 static int
1195 vhost_user_set_protocol_features(struct virtio_net *dev,
1196                                  uint64_t protocol_features)
1197 {
1198         if (protocol_features & ~VHOST_USER_PROTOCOL_FEATURES) {
1199                 RTE_LOG(ERR, VHOST_CONFIG,
1200                         "(%d) received invalid protocol features.\n",
1201                         dev->vid);
1202                 return -1;
1203         }
1204
1205         dev->protocol_features = protocol_features;
1206         return 0;
1207 }
1208
1209 static int
1210 vhost_user_set_log_base(struct virtio_net *dev, struct VhostUserMsg *msg)
1211 {
1212         int fd = msg->fds[0];
1213         uint64_t size, off;
1214         void *addr;
1215
1216         if (fd < 0) {
1217                 RTE_LOG(ERR, VHOST_CONFIG, "invalid log fd: %d\n", fd);
1218                 return -1;
1219         }
1220
1221         if (msg->size != sizeof(VhostUserLog)) {
1222                 RTE_LOG(ERR, VHOST_CONFIG,
1223                         "invalid log base msg size: %"PRId32" != %d\n",
1224                         msg->size, (int)sizeof(VhostUserLog));
1225                 return -1;
1226         }
1227
1228         size = msg->payload.log.mmap_size;
1229         off  = msg->payload.log.mmap_offset;
1230
1231         /* Don't allow mmap_offset to point outside the mmap region */
1232         if (off > size) {
1233                 RTE_LOG(ERR, VHOST_CONFIG,
1234                         "log offset %#"PRIx64" exceeds log size %#"PRIx64"\n",
1235                         off, size);
1236                 return -1;
1237         }
1238
1239         RTE_LOG(INFO, VHOST_CONFIG,
1240                 "log mmap size: %"PRId64", offset: %"PRId64"\n",
1241                 size, off);
1242
1243         /*
1244          * mmap from 0 to workaround a hugepage mmap bug: mmap will
1245          * fail when offset is not page size aligned.
1246          */
1247         addr = mmap(0, size + off, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
1248         close(fd);
1249         if (addr == MAP_FAILED) {
1250                 RTE_LOG(ERR, VHOST_CONFIG, "mmap log base failed!\n");
1251                 return -1;
1252         }
1253
1254         /*
1255          * Free previously mapped log memory on occasionally
1256          * multiple VHOST_USER_SET_LOG_BASE.
1257          */
1258         if (dev->log_addr) {
1259                 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
1260         }
1261         dev->log_addr = (uint64_t)(uintptr_t)addr;
1262         dev->log_base = dev->log_addr + off;
1263         dev->log_size = size;
1264
1265         msg->size = sizeof(msg->payload.u64);
1266
1267         return 0;
1268 }
1269
1270 static void vhost_user_set_log_fd(struct virtio_net **pdev __rte_unused,
1271                         struct VhostUserMsg *msg)
1272 {
1273         close(msg->fds[0]);
1274         RTE_LOG(INFO, VHOST_CONFIG, "not implemented.\n");
1275 }
1276
1277 /*
1278  * An rarp packet is constructed and broadcasted to notify switches about
1279  * the new location of the migrated VM, so that packets from outside will
1280  * not be lost after migration.
1281  *
1282  * However, we don't actually "send" a rarp packet here, instead, we set
1283  * a flag 'broadcast_rarp' to let rte_vhost_dequeue_burst() inject it.
1284  */
1285 static int
1286 vhost_user_send_rarp(struct virtio_net *dev, struct VhostUserMsg *msg)
1287 {
1288         uint8_t *mac = (uint8_t *)&msg->payload.u64;
1289         struct rte_vdpa_device *vdpa_dev;
1290         int did = -1;
1291
1292         RTE_LOG(DEBUG, VHOST_CONFIG,
1293                 ":: mac: %02x:%02x:%02x:%02x:%02x:%02x\n",
1294                 mac[0], mac[1], mac[2], mac[3], mac[4], mac[5]);
1295         memcpy(dev->mac.addr_bytes, mac, 6);
1296
1297         /*
1298          * Set the flag to inject a RARP broadcast packet at
1299          * rte_vhost_dequeue_burst().
1300          *
1301          * rte_smp_wmb() is for making sure the mac is copied
1302          * before the flag is set.
1303          */
1304         rte_smp_wmb();
1305         rte_atomic16_set(&dev->broadcast_rarp, 1);
1306         did = dev->vdpa_dev_id;
1307         vdpa_dev = rte_vdpa_get_device(did);
1308         if (vdpa_dev && vdpa_dev->ops->migration_done)
1309                 vdpa_dev->ops->migration_done(dev->vid);
1310
1311         return 0;
1312 }
1313
1314 static int
1315 vhost_user_net_set_mtu(struct virtio_net *dev, struct VhostUserMsg *msg)
1316 {
1317         if (msg->payload.u64 < VIRTIO_MIN_MTU ||
1318                         msg->payload.u64 > VIRTIO_MAX_MTU) {
1319                 RTE_LOG(ERR, VHOST_CONFIG, "Invalid MTU size (%"PRIu64")\n",
1320                                 msg->payload.u64);
1321
1322                 return -1;
1323         }
1324
1325         dev->mtu = msg->payload.u64;
1326
1327         return 0;
1328 }
1329
1330 static int
1331 vhost_user_set_req_fd(struct virtio_net *dev, struct VhostUserMsg *msg)
1332 {
1333         int fd = msg->fds[0];
1334
1335         if (fd < 0) {
1336                 RTE_LOG(ERR, VHOST_CONFIG,
1337                                 "Invalid file descriptor for slave channel (%d)\n",
1338                                 fd);
1339                 return -1;
1340         }
1341
1342         dev->slave_req_fd = fd;
1343
1344         return 0;
1345 }
1346
1347 static int
1348 is_vring_iotlb_update(struct vhost_virtqueue *vq, struct vhost_iotlb_msg *imsg)
1349 {
1350         struct vhost_vring_addr *ra;
1351         uint64_t start, end;
1352
1353         start = imsg->iova;
1354         end = start + imsg->size;
1355
1356         ra = &vq->ring_addrs;
1357         if (ra->desc_user_addr >= start && ra->desc_user_addr < end)
1358                 return 1;
1359         if (ra->avail_user_addr >= start && ra->avail_user_addr < end)
1360                 return 1;
1361         if (ra->used_user_addr >= start && ra->used_user_addr < end)
1362                 return 1;
1363
1364         return 0;
1365 }
1366
1367 static int
1368 is_vring_iotlb_invalidate(struct vhost_virtqueue *vq,
1369                                 struct vhost_iotlb_msg *imsg)
1370 {
1371         uint64_t istart, iend, vstart, vend;
1372
1373         istart = imsg->iova;
1374         iend = istart + imsg->size - 1;
1375
1376         vstart = (uintptr_t)vq->desc;
1377         vend = vstart + sizeof(struct vring_desc) * vq->size - 1;
1378         if (vstart <= iend && istart <= vend)
1379                 return 1;
1380
1381         vstart = (uintptr_t)vq->avail;
1382         vend = vstart + sizeof(struct vring_avail);
1383         vend += sizeof(uint16_t) * vq->size - 1;
1384         if (vstart <= iend && istart <= vend)
1385                 return 1;
1386
1387         vstart = (uintptr_t)vq->used;
1388         vend = vstart + sizeof(struct vring_used);
1389         vend += sizeof(struct vring_used_elem) * vq->size - 1;
1390         if (vstart <= iend && istart <= vend)
1391                 return 1;
1392
1393         return 0;
1394 }
1395
1396 static int
1397 vhost_user_iotlb_msg(struct virtio_net **pdev, struct VhostUserMsg *msg)
1398 {
1399         struct virtio_net *dev = *pdev;
1400         struct vhost_iotlb_msg *imsg = &msg->payload.iotlb;
1401         uint16_t i;
1402         uint64_t vva, len;
1403
1404         switch (imsg->type) {
1405         case VHOST_IOTLB_UPDATE:
1406                 len = imsg->size;
1407                 vva = qva_to_vva(dev, imsg->uaddr, &len);
1408                 if (!vva)
1409                         return -1;
1410
1411                 for (i = 0; i < dev->nr_vring; i++) {
1412                         struct vhost_virtqueue *vq = dev->virtqueue[i];
1413
1414                         vhost_user_iotlb_cache_insert(vq, imsg->iova, vva,
1415                                         len, imsg->perm);
1416
1417                         if (is_vring_iotlb_update(vq, imsg))
1418                                 *pdev = dev = translate_ring_addresses(dev, i);
1419                 }
1420                 break;
1421         case VHOST_IOTLB_INVALIDATE:
1422                 for (i = 0; i < dev->nr_vring; i++) {
1423                         struct vhost_virtqueue *vq = dev->virtqueue[i];
1424
1425                         vhost_user_iotlb_cache_remove(vq, imsg->iova,
1426                                         imsg->size);
1427
1428                         if (is_vring_iotlb_invalidate(vq, imsg))
1429                                 vring_invalidate(dev, vq);
1430                 }
1431                 break;
1432         default:
1433                 RTE_LOG(ERR, VHOST_CONFIG, "Invalid IOTLB message type (%d)\n",
1434                                 imsg->type);
1435                 return -1;
1436         }
1437
1438         return 0;
1439 }
1440
1441 /* return bytes# of read on success or negative val on failure. */
1442 static int
1443 read_vhost_message(int sockfd, struct VhostUserMsg *msg)
1444 {
1445         int ret;
1446
1447         ret = read_fd_message(sockfd, (char *)msg, VHOST_USER_HDR_SIZE,
1448                 msg->fds, VHOST_MEMORY_MAX_NREGIONS);
1449         if (ret <= 0)
1450                 return ret;
1451
1452         if (msg && msg->size) {
1453                 if (msg->size > sizeof(msg->payload)) {
1454                         RTE_LOG(ERR, VHOST_CONFIG,
1455                                 "invalid msg size: %d\n", msg->size);
1456                         return -1;
1457                 }
1458                 ret = read(sockfd, &msg->payload, msg->size);
1459                 if (ret <= 0)
1460                         return ret;
1461                 if (ret != (int)msg->size) {
1462                         RTE_LOG(ERR, VHOST_CONFIG,
1463                                 "read control message failed\n");
1464                         return -1;
1465                 }
1466         }
1467
1468         return ret;
1469 }
1470
1471 static int
1472 send_vhost_message(int sockfd, struct VhostUserMsg *msg, int *fds, int fd_num)
1473 {
1474         if (!msg)
1475                 return 0;
1476
1477         return send_fd_message(sockfd, (char *)msg,
1478                 VHOST_USER_HDR_SIZE + msg->size, fds, fd_num);
1479 }
1480
1481 static int
1482 send_vhost_reply(int sockfd, struct VhostUserMsg *msg)
1483 {
1484         if (!msg)
1485                 return 0;
1486
1487         msg->flags &= ~VHOST_USER_VERSION_MASK;
1488         msg->flags &= ~VHOST_USER_NEED_REPLY;
1489         msg->flags |= VHOST_USER_VERSION;
1490         msg->flags |= VHOST_USER_REPLY_MASK;
1491
1492         return send_vhost_message(sockfd, msg, NULL, 0);
1493 }
1494
1495 static int
1496 send_vhost_slave_message(struct virtio_net *dev, struct VhostUserMsg *msg,
1497                          int *fds, int fd_num)
1498 {
1499         int ret;
1500
1501         if (msg->flags & VHOST_USER_NEED_REPLY)
1502                 rte_spinlock_lock(&dev->slave_req_lock);
1503
1504         ret = send_vhost_message(dev->slave_req_fd, msg, fds, fd_num);
1505         if (ret < 0 && (msg->flags & VHOST_USER_NEED_REPLY))
1506                 rte_spinlock_unlock(&dev->slave_req_lock);
1507
1508         return ret;
1509 }
1510
1511 /*
1512  * Allocate a queue pair if it hasn't been allocated yet
1513  */
1514 static int
1515 vhost_user_check_and_alloc_queue_pair(struct virtio_net *dev,
1516                         struct VhostUserMsg *msg)
1517 {
1518         uint16_t vring_idx;
1519
1520         switch (msg->request.master) {
1521         case VHOST_USER_SET_VRING_KICK:
1522         case VHOST_USER_SET_VRING_CALL:
1523         case VHOST_USER_SET_VRING_ERR:
1524                 vring_idx = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1525                 break;
1526         case VHOST_USER_SET_VRING_NUM:
1527         case VHOST_USER_SET_VRING_BASE:
1528         case VHOST_USER_SET_VRING_ENABLE:
1529                 vring_idx = msg->payload.state.index;
1530                 break;
1531         case VHOST_USER_SET_VRING_ADDR:
1532                 vring_idx = msg->payload.addr.index;
1533                 break;
1534         default:
1535                 return 0;
1536         }
1537
1538         if (vring_idx >= VHOST_MAX_VRING) {
1539                 RTE_LOG(ERR, VHOST_CONFIG,
1540                         "invalid vring index: %u\n", vring_idx);
1541                 return -1;
1542         }
1543
1544         if (dev->virtqueue[vring_idx])
1545                 return 0;
1546
1547         return alloc_vring_queue(dev, vring_idx);
1548 }
1549
1550 static void
1551 vhost_user_lock_all_queue_pairs(struct virtio_net *dev)
1552 {
1553         unsigned int i = 0;
1554         unsigned int vq_num = 0;
1555
1556         while (vq_num < dev->nr_vring) {
1557                 struct vhost_virtqueue *vq = dev->virtqueue[i];
1558
1559                 if (vq) {
1560                         rte_spinlock_lock(&vq->access_lock);
1561                         vq_num++;
1562                 }
1563                 i++;
1564         }
1565 }
1566
1567 static void
1568 vhost_user_unlock_all_queue_pairs(struct virtio_net *dev)
1569 {
1570         unsigned int i = 0;
1571         unsigned int vq_num = 0;
1572
1573         while (vq_num < dev->nr_vring) {
1574                 struct vhost_virtqueue *vq = dev->virtqueue[i];
1575
1576                 if (vq) {
1577                         rte_spinlock_unlock(&vq->access_lock);
1578                         vq_num++;
1579                 }
1580                 i++;
1581         }
1582 }
1583
1584 int
1585 vhost_user_msg_handler(int vid, int fd)
1586 {
1587         struct virtio_net *dev;
1588         struct VhostUserMsg msg;
1589         struct rte_vdpa_device *vdpa_dev;
1590         int did = -1;
1591         int ret;
1592         int unlock_required = 0;
1593         uint32_t skip_master = 0;
1594
1595         dev = get_device(vid);
1596         if (dev == NULL)
1597                 return -1;
1598
1599         if (!dev->notify_ops) {
1600                 dev->notify_ops = vhost_driver_callback_get(dev->ifname);
1601                 if (!dev->notify_ops) {
1602                         RTE_LOG(ERR, VHOST_CONFIG,
1603                                 "failed to get callback ops for driver %s\n",
1604                                 dev->ifname);
1605                         return -1;
1606                 }
1607         }
1608
1609         ret = read_vhost_message(fd, &msg);
1610         if (ret <= 0 || msg.request.master >= VHOST_USER_MAX) {
1611                 if (ret < 0)
1612                         RTE_LOG(ERR, VHOST_CONFIG,
1613                                 "vhost read message failed\n");
1614                 else if (ret == 0)
1615                         RTE_LOG(INFO, VHOST_CONFIG,
1616                                 "vhost peer closed\n");
1617                 else
1618                         RTE_LOG(ERR, VHOST_CONFIG,
1619                                 "vhost read incorrect message\n");
1620
1621                 return -1;
1622         }
1623
1624         ret = 0;
1625         if (msg.request.master != VHOST_USER_IOTLB_MSG)
1626                 RTE_LOG(INFO, VHOST_CONFIG, "read message %s\n",
1627                         vhost_message_str[msg.request.master]);
1628         else
1629                 RTE_LOG(DEBUG, VHOST_CONFIG, "read message %s\n",
1630                         vhost_message_str[msg.request.master]);
1631
1632         ret = vhost_user_check_and_alloc_queue_pair(dev, &msg);
1633         if (ret < 0) {
1634                 RTE_LOG(ERR, VHOST_CONFIG,
1635                         "failed to alloc queue\n");
1636                 return -1;
1637         }
1638
1639         /*
1640          * Note: we don't lock all queues on VHOST_USER_GET_VRING_BASE
1641          * and VHOST_USER_RESET_OWNER, since it is sent when virtio stops
1642          * and device is destroyed. destroy_device waits for queues to be
1643          * inactive, so it is safe. Otherwise taking the access_lock
1644          * would cause a dead lock.
1645          */
1646         switch (msg.request.master) {
1647         case VHOST_USER_SET_FEATURES:
1648         case VHOST_USER_SET_PROTOCOL_FEATURES:
1649         case VHOST_USER_SET_OWNER:
1650         case VHOST_USER_SET_MEM_TABLE:
1651         case VHOST_USER_SET_LOG_BASE:
1652         case VHOST_USER_SET_LOG_FD:
1653         case VHOST_USER_SET_VRING_NUM:
1654         case VHOST_USER_SET_VRING_ADDR:
1655         case VHOST_USER_SET_VRING_BASE:
1656         case VHOST_USER_SET_VRING_KICK:
1657         case VHOST_USER_SET_VRING_CALL:
1658         case VHOST_USER_SET_VRING_ERR:
1659         case VHOST_USER_SET_VRING_ENABLE:
1660         case VHOST_USER_SEND_RARP:
1661         case VHOST_USER_NET_SET_MTU:
1662         case VHOST_USER_SET_SLAVE_REQ_FD:
1663                 vhost_user_lock_all_queue_pairs(dev);
1664                 unlock_required = 1;
1665                 break;
1666         default:
1667                 break;
1668
1669         }
1670
1671         if (dev->extern_ops.pre_msg_handle) {
1672                 uint32_t need_reply;
1673
1674                 ret = (*dev->extern_ops.pre_msg_handle)(dev->vid,
1675                                 (void *)&msg, &need_reply, &skip_master);
1676                 if (ret < 0)
1677                         goto skip_to_reply;
1678
1679                 if (need_reply)
1680                         send_vhost_reply(fd, &msg);
1681
1682                 if (skip_master)
1683                         goto skip_to_post_handle;
1684         }
1685
1686         switch (msg.request.master) {
1687         case VHOST_USER_GET_FEATURES:
1688                 vhost_user_get_features(dev, &msg);
1689                 send_vhost_reply(fd, &msg);
1690                 break;
1691         case VHOST_USER_SET_FEATURES:
1692                 ret = vhost_user_set_features(dev, msg.payload.u64);
1693                 break;
1694
1695         case VHOST_USER_GET_PROTOCOL_FEATURES:
1696                 vhost_user_get_protocol_features(dev, &msg);
1697                 send_vhost_reply(fd, &msg);
1698                 break;
1699         case VHOST_USER_SET_PROTOCOL_FEATURES:
1700                 ret = vhost_user_set_protocol_features(dev, msg.payload.u64);
1701                 break;
1702
1703         case VHOST_USER_SET_OWNER:
1704                 ret = vhost_user_set_owner();
1705                 break;
1706         case VHOST_USER_RESET_OWNER:
1707                 ret = vhost_user_reset_owner(dev);
1708                 break;
1709
1710         case VHOST_USER_SET_MEM_TABLE:
1711                 ret = vhost_user_set_mem_table(&dev, &msg);
1712                 break;
1713
1714         case VHOST_USER_SET_LOG_BASE:
1715                 ret = vhost_user_set_log_base(dev, &msg);
1716                 if (ret)
1717                         goto skip_to_reply;
1718                 /* it needs a reply */
1719                 send_vhost_reply(fd, &msg);
1720                 break;
1721         case VHOST_USER_SET_LOG_FD:
1722                 vhost_user_set_log_fd(&dev, &msg);
1723                 break;
1724
1725         case VHOST_USER_SET_VRING_NUM:
1726                 ret = vhost_user_set_vring_num(dev, &msg);
1727                 break;
1728         case VHOST_USER_SET_VRING_ADDR:
1729                 ret = vhost_user_set_vring_addr(&dev, &msg);
1730                 break;
1731         case VHOST_USER_SET_VRING_BASE:
1732                 ret = vhost_user_set_vring_base(dev, &msg);
1733                 break;
1734
1735         case VHOST_USER_GET_VRING_BASE:
1736                 ret = vhost_user_get_vring_base(dev, &msg);
1737                 if (ret)
1738                         goto skip_to_reply;
1739                 send_vhost_reply(fd, &msg);
1740                 break;
1741
1742         case VHOST_USER_SET_VRING_KICK:
1743                 ret = vhost_user_set_vring_kick(&dev, &msg);
1744                 break;
1745         case VHOST_USER_SET_VRING_CALL:
1746                 vhost_user_set_vring_call(dev, &msg);
1747                 break;
1748
1749         case VHOST_USER_SET_VRING_ERR:
1750                 vhost_user_set_vring_err(&dev, &msg);
1751                 break;
1752
1753         case VHOST_USER_GET_QUEUE_NUM:
1754                 vhost_user_get_queue_num(dev, &msg);
1755                 send_vhost_reply(fd, &msg);
1756                 break;
1757
1758         case VHOST_USER_SET_VRING_ENABLE:
1759                 ret = vhost_user_set_vring_enable(dev, &msg);
1760                 break;
1761         case VHOST_USER_SEND_RARP:
1762                 ret = vhost_user_send_rarp(dev, &msg);
1763                 break;
1764
1765         case VHOST_USER_NET_SET_MTU:
1766                 ret = vhost_user_net_set_mtu(dev, &msg);
1767                 break;
1768
1769         case VHOST_USER_SET_SLAVE_REQ_FD:
1770                 ret = vhost_user_set_req_fd(dev, &msg);
1771                 break;
1772
1773         case VHOST_USER_IOTLB_MSG:
1774                 ret = vhost_user_iotlb_msg(&dev, &msg);
1775                 break;
1776
1777         default:
1778                 ret = -1;
1779                 break;
1780         }
1781
1782 skip_to_post_handle:
1783         if (!ret && dev->extern_ops.post_msg_handle) {
1784                 uint32_t need_reply;
1785
1786                 ret = (*dev->extern_ops.post_msg_handle)(
1787                                 dev->vid, (void *)&msg, &need_reply);
1788                 if (ret < 0)
1789                         goto skip_to_reply;
1790
1791                 if (need_reply)
1792                         send_vhost_reply(fd, &msg);
1793         }
1794
1795 skip_to_reply:
1796         if (unlock_required)
1797                 vhost_user_unlock_all_queue_pairs(dev);
1798
1799         if (msg.flags & VHOST_USER_NEED_REPLY) {
1800                 msg.payload.u64 = !!ret;
1801                 msg.size = sizeof(msg.payload.u64);
1802                 send_vhost_reply(fd, &msg);
1803         } else if (ret) {
1804                 RTE_LOG(ERR, VHOST_CONFIG,
1805                         "vhost message handling failed.\n");
1806                 return -1;
1807         }
1808
1809         if (!(dev->flags & VIRTIO_DEV_RUNNING) && virtio_is_ready(dev)) {
1810                 dev->flags |= VIRTIO_DEV_READY;
1811
1812                 if (!(dev->flags & VIRTIO_DEV_RUNNING)) {
1813                         if (dev->dequeue_zero_copy) {
1814                                 RTE_LOG(INFO, VHOST_CONFIG,
1815                                                 "dequeue zero copy is enabled\n");
1816                         }
1817
1818                         if (dev->notify_ops->new_device(dev->vid) == 0)
1819                                 dev->flags |= VIRTIO_DEV_RUNNING;
1820                 }
1821         }
1822
1823         did = dev->vdpa_dev_id;
1824         vdpa_dev = rte_vdpa_get_device(did);
1825         if (vdpa_dev && virtio_is_ready(dev) &&
1826                         !(dev->flags & VIRTIO_DEV_VDPA_CONFIGURED) &&
1827                         msg.request.master == VHOST_USER_SET_VRING_ENABLE) {
1828                 if (vdpa_dev->ops->dev_conf)
1829                         vdpa_dev->ops->dev_conf(dev->vid);
1830                 dev->flags |= VIRTIO_DEV_VDPA_CONFIGURED;
1831                 if (vhost_user_host_notifier_ctrl(dev->vid, true) != 0) {
1832                         RTE_LOG(INFO, VHOST_CONFIG,
1833                                 "(%d) software relay is used for vDPA, performance may be low.\n",
1834                                 dev->vid);
1835                 }
1836         }
1837
1838         return 0;
1839 }
1840
1841 static int process_slave_message_reply(struct virtio_net *dev,
1842                                        const struct VhostUserMsg *msg)
1843 {
1844         struct VhostUserMsg msg_reply;
1845         int ret;
1846
1847         if ((msg->flags & VHOST_USER_NEED_REPLY) == 0)
1848                 return 0;
1849
1850         if (read_vhost_message(dev->slave_req_fd, &msg_reply) < 0) {
1851                 ret = -1;
1852                 goto out;
1853         }
1854
1855         if (msg_reply.request.slave != msg->request.slave) {
1856                 RTE_LOG(ERR, VHOST_CONFIG,
1857                         "Received unexpected msg type (%u), expected %u\n",
1858                         msg_reply.request.slave, msg->request.slave);
1859                 ret = -1;
1860                 goto out;
1861         }
1862
1863         ret = msg_reply.payload.u64 ? -1 : 0;
1864
1865 out:
1866         rte_spinlock_unlock(&dev->slave_req_lock);
1867         return ret;
1868 }
1869
1870 int
1871 vhost_user_iotlb_miss(struct virtio_net *dev, uint64_t iova, uint8_t perm)
1872 {
1873         int ret;
1874         struct VhostUserMsg msg = {
1875                 .request.slave = VHOST_USER_SLAVE_IOTLB_MSG,
1876                 .flags = VHOST_USER_VERSION,
1877                 .size = sizeof(msg.payload.iotlb),
1878                 .payload.iotlb = {
1879                         .iova = iova,
1880                         .perm = perm,
1881                         .type = VHOST_IOTLB_MISS,
1882                 },
1883         };
1884
1885         ret = send_vhost_message(dev->slave_req_fd, &msg, NULL, 0);
1886         if (ret < 0) {
1887                 RTE_LOG(ERR, VHOST_CONFIG,
1888                                 "Failed to send IOTLB miss message (%d)\n",
1889                                 ret);
1890                 return ret;
1891         }
1892
1893         return 0;
1894 }
1895
1896 static int vhost_user_slave_set_vring_host_notifier(struct virtio_net *dev,
1897                                                     int index, int fd,
1898                                                     uint64_t offset,
1899                                                     uint64_t size)
1900 {
1901         int *fdp = NULL;
1902         size_t fd_num = 0;
1903         int ret;
1904         struct VhostUserMsg msg = {
1905                 .request.slave = VHOST_USER_SLAVE_VRING_HOST_NOTIFIER_MSG,
1906                 .flags = VHOST_USER_VERSION | VHOST_USER_NEED_REPLY,
1907                 .size = sizeof(msg.payload.area),
1908                 .payload.area = {
1909                         .u64 = index & VHOST_USER_VRING_IDX_MASK,
1910                         .size = size,
1911                         .offset = offset,
1912                 },
1913         };
1914
1915         if (fd < 0)
1916                 msg.payload.area.u64 |= VHOST_USER_VRING_NOFD_MASK;
1917         else {
1918                 fdp = &fd;
1919                 fd_num = 1;
1920         }
1921
1922         ret = send_vhost_slave_message(dev, &msg, fdp, fd_num);
1923         if (ret < 0) {
1924                 RTE_LOG(ERR, VHOST_CONFIG,
1925                         "Failed to set host notifier (%d)\n", ret);
1926                 return ret;
1927         }
1928
1929         return process_slave_message_reply(dev, &msg);
1930 }
1931
1932 int vhost_user_host_notifier_ctrl(int vid, bool enable)
1933 {
1934         struct virtio_net *dev;
1935         struct rte_vdpa_device *vdpa_dev;
1936         int vfio_device_fd, did, ret = 0;
1937         uint64_t offset, size;
1938         unsigned int i;
1939
1940         dev = get_device(vid);
1941         if (!dev)
1942                 return -ENODEV;
1943
1944         did = dev->vdpa_dev_id;
1945         if (did < 0)
1946                 return -EINVAL;
1947
1948         if (!(dev->features & (1ULL << VIRTIO_F_VERSION_1)) ||
1949             !(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES)) ||
1950             !(dev->protocol_features &
1951                         (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_REQ)) ||
1952             !(dev->protocol_features &
1953                         (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_SEND_FD)) ||
1954             !(dev->protocol_features &
1955                         (1ULL << VHOST_USER_PROTOCOL_F_HOST_NOTIFIER)))
1956                 return -ENOTSUP;
1957
1958         vdpa_dev = rte_vdpa_get_device(did);
1959         if (!vdpa_dev)
1960                 return -ENODEV;
1961
1962         RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_vfio_device_fd, -ENOTSUP);
1963         RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_notify_area, -ENOTSUP);
1964
1965         vfio_device_fd = vdpa_dev->ops->get_vfio_device_fd(vid);
1966         if (vfio_device_fd < 0)
1967                 return -ENOTSUP;
1968
1969         if (enable) {
1970                 for (i = 0; i < dev->nr_vring; i++) {
1971                         if (vdpa_dev->ops->get_notify_area(vid, i, &offset,
1972                                         &size) < 0) {
1973                                 ret = -ENOTSUP;
1974                                 goto disable;
1975                         }
1976
1977                         if (vhost_user_slave_set_vring_host_notifier(dev, i,
1978                                         vfio_device_fd, offset, size) < 0) {
1979                                 ret = -EFAULT;
1980                                 goto disable;
1981                         }
1982                 }
1983         } else {
1984 disable:
1985                 for (i = 0; i < dev->nr_vring; i++) {
1986                         vhost_user_slave_set_vring_host_notifier(dev, i, -1,
1987                                         0, 0);
1988                 }
1989         }
1990
1991         return ret;
1992 }