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 /* The possible results of a message handling function */
  75 enum vh_result {
  76         /* Message handling failed */
  77         VH_RESULT_ERR   = -1,
  78         /* Message handling successful */
  79         VH_RESULT_OK    =  0,
  80         /* Message handling successful and reply prepared */
  81         VH_RESULT_REPLY =  1,
  82 };
  83
  84 static uint64_t
  85 get_blk_size(int fd)
  86 {
  87         struct stat stat;
  88         int ret;
  89
  90         ret = fstat(fd, &stat);
  91         return ret == -1 ? (uint64_t)-1 : (uint64_t)stat.st_blksize;
  92 }
  93
  94 static void
  95 free_mem_region(struct virtio_net *dev)
  96 {
  97         uint32_t i;
  98         struct rte_vhost_mem_region *reg;
  99
 100         if (!dev || !dev->mem)
 101                 return;
 102
 103         for (i = 0; i < dev->mem->nregions; i++) {
 104                 reg = &dev->mem->regions[i];
 105                 if (reg->host_user_addr) {
 106                         munmap(reg->mmap_addr, reg->mmap_size);
 107                         close(reg->fd);
 108                 }
 109         }
 110 }
 111
 112 void
 113 vhost_backend_cleanup(struct virtio_net *dev)
 114 {
 115         if (dev->mem) {
 116                 free_mem_region(dev);
 117                 rte_free(dev->mem);
 118                 dev->mem = NULL;
 119         }
 120
 121         free(dev->guest_pages);
 122         dev->guest_pages = NULL;
 123
 124         if (dev->log_addr) {
 125                 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
 126                 dev->log_addr = 0;
 127         }
 128
 129         if (dev->slave_req_fd >= 0) {
 130                 close(dev->slave_req_fd);
 131                 dev->slave_req_fd = -1;
 132         }
 133 }
 134
 135 /*
 136  * This function just returns success at the moment unless
 137  * the device hasn't been initialised.
 138  */
 139 static int
 140 vhost_user_set_owner(struct virtio_net **pdev __rte_unused,
 141                 struct VhostUserMsg *msg __rte_unused)
 142 {
 143         return VH_RESULT_OK;
 144 }
 145
 146 static int
 147 vhost_user_reset_owner(struct virtio_net **pdev,
 148                 struct VhostUserMsg *msg __rte_unused)
 149 {
 150         struct virtio_net *dev = *pdev;
 151         vhost_destroy_device_notify(dev);
 152
 153         cleanup_device(dev, 0);
 154         reset_device(dev);
 155         return VH_RESULT_OK;
 156 }
 157
 158 /*
 159  * The features that we support are requested.
 160  */
 161 static int
 162 vhost_user_get_features(struct virtio_net **pdev, struct VhostUserMsg *msg)
 163 {
 164         struct virtio_net *dev = *pdev;
 165         uint64_t features = 0;
 166
 167         rte_vhost_driver_get_features(dev->ifname, &features);
 168
 169         msg->payload.u64 = features;
 170         msg->size = sizeof(msg->payload.u64);
 171
 172         return VH_RESULT_REPLY;
 173 }
 174
 175 /*
 176  * The queue number that we support are requested.
 177  */
 178 static int
 179 vhost_user_get_queue_num(struct virtio_net **pdev, struct VhostUserMsg *msg)
 180 {
 181         struct virtio_net *dev = *pdev;
 182         uint32_t queue_num = 0;
 183
 184         rte_vhost_driver_get_queue_num(dev->ifname, &queue_num);
 185
 186         msg->payload.u64 = (uint64_t)queue_num;
 187         msg->size = sizeof(msg->payload.u64);
 188
 189         return VH_RESULT_REPLY;
 190 }
 191
 192 /*
 193  * We receive the negotiated features supported by us and the virtio device.
 194  */
 195 static int
 196 vhost_user_set_features(struct virtio_net **pdev, struct VhostUserMsg *msg)
 197 {
 198         struct virtio_net *dev = *pdev;
 199         uint64_t features = msg->payload.u64;
 200         uint64_t vhost_features = 0;
 201         struct rte_vdpa_device *vdpa_dev;
 202         int did = -1;
 203
 204         rte_vhost_driver_get_features(dev->ifname, &vhost_features);
 205         if (features & ~vhost_features) {
 206                 RTE_LOG(ERR, VHOST_CONFIG,
 207                         "(%d) received invalid negotiated features.\n",
 208                         dev->vid);
 209                 return VH_RESULT_ERR;
 210         }
 211
 212         if (dev->flags & VIRTIO_DEV_RUNNING) {
 213                 if (dev->features == features)
 214                         return VH_RESULT_OK;
 215
 216                 /*
 217                  * Error out if master tries to change features while device is
 218                  * in running state. The exception being VHOST_F_LOG_ALL, which
 219                  * is enabled when the live-migration starts.
 220                  */
 221                 if ((dev->features ^ features) & ~(1ULL << VHOST_F_LOG_ALL)) {
 222                         RTE_LOG(ERR, VHOST_CONFIG,
 223                                 "(%d) features changed while device is running.\n",
 224                                 dev->vid);
 225                         return VH_RESULT_ERR;
 226                 }
 227
 228                 if (dev->notify_ops->features_changed)
 229                         dev->notify_ops->features_changed(dev->vid, features);
 230         }
 231
 232         dev->features = features;
 233         if (dev->features &
 234                 ((1 << VIRTIO_NET_F_MRG_RXBUF) | (1ULL << VIRTIO_F_VERSION_1))) {
 235                 dev->vhost_hlen = sizeof(struct virtio_net_hdr_mrg_rxbuf);
 236         } else {
 237                 dev->vhost_hlen = sizeof(struct virtio_net_hdr);
 238         }
 239         VHOST_LOG_DEBUG(VHOST_CONFIG,
 240                 "(%d) mergeable RX buffers %s, virtio 1 %s\n",
 241                 dev->vid,
 242                 (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ? "on" : "off",
 243                 (dev->features & (1ULL << VIRTIO_F_VERSION_1)) ? "on" : "off");
 244
 245         if ((dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET) &&
 246             !(dev->features & (1ULL << VIRTIO_NET_F_MQ))) {
 247                 /*
 248                  * Remove all but first queue pair if MQ hasn't been
 249                  * negotiated. This is safe because the device is not
 250                  * running at this stage.
 251                  */
 252                 while (dev->nr_vring > 2) {
 253                         struct vhost_virtqueue *vq;
 254
 255                         vq = dev->virtqueue[--dev->nr_vring];
 256                         if (!vq)
 257                                 continue;
 258
 259                         dev->virtqueue[dev->nr_vring] = NULL;
 260                         cleanup_vq(vq, 1);
 261                         free_vq(dev, vq);
 262                 }
 263         }
 264
 265         did = dev->vdpa_dev_id;
 266         vdpa_dev = rte_vdpa_get_device(did);
 267         if (vdpa_dev && vdpa_dev->ops->set_features)
 268                 vdpa_dev->ops->set_features(dev->vid);
 269
 270         return VH_RESULT_OK;
 271 }
 272
 273 /*
 274  * The virtio device sends us the size of the descriptor ring.
 275  */
 276 static int
 277 vhost_user_set_vring_num(struct virtio_net **pdev,
 278                          struct VhostUserMsg *msg)
 279 {
 280         struct virtio_net *dev = *pdev;
 281         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
 282
 283         vq->size = msg->payload.state.num;
 284
 285         /* VIRTIO 1.0, 2.4 Virtqueues says:
 286          *
 287          *   Queue Size value is always a power of 2. The maximum Queue Size
 288          *   value is 32768.
 289          */
 290         if ((vq->size & (vq->size - 1)) || vq->size > 32768) {
 291                 RTE_LOG(ERR, VHOST_CONFIG,
 292                         "invalid virtqueue size %u\n", vq->size);
 293                 return VH_RESULT_ERR;
 294         }
 295
 296         if (dev->dequeue_zero_copy) {
 297                 vq->nr_zmbuf = 0;
 298                 vq->last_zmbuf_idx = 0;
 299                 vq->zmbuf_size = vq->size;
 300                 vq->zmbufs = rte_zmalloc(NULL, vq->zmbuf_size *
 301                                          sizeof(struct zcopy_mbuf), 0);
 302                 if (vq->zmbufs == NULL) {
 303                         RTE_LOG(WARNING, VHOST_CONFIG,
 304                                 "failed to allocate mem for zero copy; "
 305                                 "zero copy is force disabled\n");
 306                         dev->dequeue_zero_copy = 0;
 307                 }
 308                 TAILQ_INIT(&vq->zmbuf_list);
 309         }
 310
 311         if (vq_is_packed(dev)) {
 312                 vq->shadow_used_packed = rte_malloc(NULL,
 313                                 vq->size *
 314                                 sizeof(struct vring_used_elem_packed),
 315                                 RTE_CACHE_LINE_SIZE);
 316                 if (!vq->shadow_used_packed) {
 317                         RTE_LOG(ERR, VHOST_CONFIG,
 318                                         "failed to allocate memory for shadow used ring.\n");
 319                         return VH_RESULT_ERR;
 320                 }
 321
 322         } else {
 323                 vq->shadow_used_split = rte_malloc(NULL,
 324                                 vq->size * sizeof(struct vring_used_elem),
 325                                 RTE_CACHE_LINE_SIZE);
 326                 if (!vq->shadow_used_split) {
 327                         RTE_LOG(ERR, VHOST_CONFIG,
 328                                         "failed to allocate memory for shadow used ring.\n");
 329                         return VH_RESULT_ERR;
 330                 }
 331         }
 332
 333         vq->batch_copy_elems = rte_malloc(NULL,
 334                                 vq->size * sizeof(struct batch_copy_elem),
 335                                 RTE_CACHE_LINE_SIZE);
 336         if (!vq->batch_copy_elems) {
 337                 RTE_LOG(ERR, VHOST_CONFIG,
 338                         "failed to allocate memory for batching copy.\n");
 339                 return VH_RESULT_ERR;
 340         }
 341
 342         return VH_RESULT_OK;
 343 }
 344
 345 /*
 346  * Reallocate virtio_dev and vhost_virtqueue data structure to make them on the
 347  * same numa node as the memory of vring descriptor.
 348  */
 349 #ifdef RTE_LIBRTE_VHOST_NUMA
 350 static struct virtio_net*
 351 numa_realloc(struct virtio_net *dev, int index)
 352 {
 353         int oldnode, newnode;
 354         struct virtio_net *old_dev;
 355         struct vhost_virtqueue *old_vq, *vq;
 356         struct zcopy_mbuf *new_zmbuf;
 357         struct vring_used_elem *new_shadow_used_split;
 358         struct vring_used_elem_packed *new_shadow_used_packed;
 359         struct batch_copy_elem *new_batch_copy_elems;
 360         int ret;
 361
 362         old_dev = dev;
 363         vq = old_vq = dev->virtqueue[index];
 364
 365         ret = get_mempolicy(&newnode, NULL, 0, old_vq->desc,
 366                             MPOL_F_NODE | MPOL_F_ADDR);
 367
 368         /* check if we need to reallocate vq */
 369         ret |= get_mempolicy(&oldnode, NULL, 0, old_vq,
 370                              MPOL_F_NODE | MPOL_F_ADDR);
 371         if (ret) {
 372                 RTE_LOG(ERR, VHOST_CONFIG,
 373                         "Unable to get vq numa information.\n");
 374                 return dev;
 375         }
 376         if (oldnode != newnode) {
 377                 RTE_LOG(INFO, VHOST_CONFIG,
 378                         "reallocate vq from %d to %d node\n", oldnode, newnode);
 379                 vq = rte_malloc_socket(NULL, sizeof(*vq), 0, newnode);
 380                 if (!vq)
 381                         return dev;
 382
 383                 memcpy(vq, old_vq, sizeof(*vq));
 384                 TAILQ_INIT(&vq->zmbuf_list);
 385
 386                 if (dev->dequeue_zero_copy) {
 387                         new_zmbuf = rte_malloc_socket(NULL, vq->zmbuf_size *
 388                                         sizeof(struct zcopy_mbuf), 0, newnode);
 389                         if (new_zmbuf) {
 390                                 rte_free(vq->zmbufs);
 391                                 vq->zmbufs = new_zmbuf;
 392                         }
 393                 }
 394
 395                 if (vq_is_packed(dev)) {
 396                         new_shadow_used_packed = rte_malloc_socket(NULL,
 397                                         vq->size *
 398                                         sizeof(struct vring_used_elem_packed),
 399                                         RTE_CACHE_LINE_SIZE,
 400                                         newnode);
 401                         if (new_shadow_used_packed) {
 402                                 rte_free(vq->shadow_used_packed);
 403                                 vq->shadow_used_packed = new_shadow_used_packed;
 404                         }
 405                 } else {
 406                         new_shadow_used_split = rte_malloc_socket(NULL,
 407                                         vq->size *
 408                                         sizeof(struct vring_used_elem),
 409                                         RTE_CACHE_LINE_SIZE,
 410                                         newnode);
 411                         if (new_shadow_used_split) {
 412                                 rte_free(vq->shadow_used_split);
 413                                 vq->shadow_used_split = new_shadow_used_split;
 414                         }
 415                 }
 416
 417                 new_batch_copy_elems = rte_malloc_socket(NULL,
 418                         vq->size * sizeof(struct batch_copy_elem),
 419                         RTE_CACHE_LINE_SIZE,
 420                         newnode);
 421                 if (new_batch_copy_elems) {
 422                         rte_free(vq->batch_copy_elems);
 423                         vq->batch_copy_elems = new_batch_copy_elems;
 424                 }
 425
 426                 rte_free(old_vq);
 427         }
 428
 429         /* check if we need to reallocate dev */
 430         ret = get_mempolicy(&oldnode, NULL, 0, old_dev,
 431                             MPOL_F_NODE | MPOL_F_ADDR);
 432         if (ret) {
 433                 RTE_LOG(ERR, VHOST_CONFIG,
 434                         "Unable to get dev numa information.\n");
 435                 goto out;
 436         }
 437         if (oldnode != newnode) {
 438                 RTE_LOG(INFO, VHOST_CONFIG,
 439                         "reallocate dev from %d to %d node\n",
 440                         oldnode, newnode);
 441                 dev = rte_malloc_socket(NULL, sizeof(*dev), 0, newnode);
 442                 if (!dev) {
 443                         dev = old_dev;
 444                         goto out;
 445                 }
 446
 447                 memcpy(dev, old_dev, sizeof(*dev));
 448                 rte_free(old_dev);
 449         }
 450
 451 out:
 452         dev->virtqueue[index] = vq;
 453         vhost_devices[dev->vid] = dev;
 454
 455         if (old_vq != vq)
 456                 vhost_user_iotlb_init(dev, index);
 457
 458         return dev;
 459 }
 460 #else
 461 static struct virtio_net*
 462 numa_realloc(struct virtio_net *dev, int index __rte_unused)
 463 {
 464         return dev;
 465 }
 466 #endif
 467
 468 /* Converts QEMU virtual address to Vhost virtual address. */
 469 static uint64_t
 470 qva_to_vva(struct virtio_net *dev, uint64_t qva, uint64_t *len)
 471 {
 472         struct rte_vhost_mem_region *r;
 473         uint32_t i;
 474
 475         /* Find the region where the address lives. */
 476         for (i = 0; i < dev->mem->nregions; i++) {
 477                 r = &dev->mem->regions[i];
 478
 479                 if (qva >= r->guest_user_addr &&
 480                     qva <  r->guest_user_addr + r->size) {
 481
 482                         if (unlikely(*len > r->guest_user_addr + r->size - qva))
 483                                 *len = r->guest_user_addr + r->size - qva;
 484
 485                         return qva - r->guest_user_addr +
 486                                r->host_user_addr;
 487                 }
 488         }
 489         *len = 0;
 490
 491         return 0;
 492 }
 493
 494
 495 /*
 496  * Converts ring address to Vhost virtual address.
 497  * If IOMMU is enabled, the ring address is a guest IO virtual address,
 498  * else it is a QEMU virtual address.
 499  */
 500 static uint64_t
 501 ring_addr_to_vva(struct virtio_net *dev, struct vhost_virtqueue *vq,
 502                 uint64_t ra, uint64_t *size)
 503 {
 504         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)) {
 505                 uint64_t vva;
 506
 507                 vva = vhost_user_iotlb_cache_find(vq, ra,
 508                                         size, VHOST_ACCESS_RW);
 509                 if (!vva)
 510                         vhost_user_iotlb_miss(dev, ra, VHOST_ACCESS_RW);
 511
 512                 return vva;
 513         }
 514
 515         return qva_to_vva(dev, ra, size);
 516 }
 517
 518 static struct virtio_net *
 519 translate_ring_addresses(struct virtio_net *dev, int vq_index)
 520 {
 521         struct vhost_virtqueue *vq = dev->virtqueue[vq_index];
 522         struct vhost_vring_addr *addr = &vq->ring_addrs;
 523         uint64_t len;
 524
 525         if (vq_is_packed(dev)) {
 526                 len = sizeof(struct vring_packed_desc) * vq->size;
 527                 vq->desc_packed = (struct vring_packed_desc *)(uintptr_t)
 528                         ring_addr_to_vva(dev, vq, addr->desc_user_addr, &len);
 529                 vq->log_guest_addr = 0;
 530                 if (vq->desc_packed == NULL ||
 531                                 len != sizeof(struct vring_packed_desc) *
 532                                 vq->size) {
 533                         RTE_LOG(DEBUG, VHOST_CONFIG,
 534                                 "(%d) failed to map desc_packed ring.\n",
 535                                 dev->vid);
 536                         return dev;
 537                 }
 538
 539                 dev = numa_realloc(dev, vq_index);
 540                 vq = dev->virtqueue[vq_index];
 541                 addr = &vq->ring_addrs;
 542
 543                 len = sizeof(struct vring_packed_desc_event);
 544                 vq->driver_event = (struct vring_packed_desc_event *)
 545                                         (uintptr_t)ring_addr_to_vva(dev,
 546                                         vq, addr->avail_user_addr, &len);
 547                 if (vq->driver_event == NULL ||
 548                                 len != sizeof(struct vring_packed_desc_event)) {
 549                         RTE_LOG(DEBUG, VHOST_CONFIG,
 550                                 "(%d) failed to find driver area address.\n",
 551                                 dev->vid);
 552                         return dev;
 553                 }
 554
 555                 len = sizeof(struct vring_packed_desc_event);
 556                 vq->device_event = (struct vring_packed_desc_event *)
 557                                         (uintptr_t)ring_addr_to_vva(dev,
 558                                         vq, addr->used_user_addr, &len);
 559                 if (vq->device_event == NULL ||
 560                                 len != sizeof(struct vring_packed_desc_event)) {
 561                         RTE_LOG(DEBUG, VHOST_CONFIG,
 562                                 "(%d) failed to find device area address.\n",
 563                                 dev->vid);
 564                         return dev;
 565                 }
 566
 567                 return dev;
 568         }
 569
 570         /* The addresses are converted from QEMU virtual to Vhost virtual. */
 571         if (vq->desc && vq->avail && vq->used)
 572                 return dev;
 573
 574         len = sizeof(struct vring_desc) * vq->size;
 575         vq->desc = (struct vring_desc *)(uintptr_t)ring_addr_to_vva(dev,
 576                         vq, addr->desc_user_addr, &len);
 577         if (vq->desc == 0 || len != sizeof(struct vring_desc) * vq->size) {
 578                 RTE_LOG(DEBUG, VHOST_CONFIG,
 579                         "(%d) failed to map desc ring.\n",
 580                         dev->vid);
 581                 return dev;
 582         }
 583
 584         dev = numa_realloc(dev, vq_index);
 585         vq = dev->virtqueue[vq_index];
 586         addr = &vq->ring_addrs;
 587
 588         len = sizeof(struct vring_avail) + sizeof(uint16_t) * vq->size;
 589         vq->avail = (struct vring_avail *)(uintptr_t)ring_addr_to_vva(dev,
 590                         vq, addr->avail_user_addr, &len);
 591         if (vq->avail == 0 ||
 592                         len != sizeof(struct vring_avail) +
 593                         sizeof(uint16_t) * vq->size) {
 594                 RTE_LOG(DEBUG, VHOST_CONFIG,
 595                         "(%d) failed to map avail ring.\n",
 596                         dev->vid);
 597                 return dev;
 598         }
 599
 600         len = sizeof(struct vring_used) +
 601                 sizeof(struct vring_used_elem) * vq->size;
 602         vq->used = (struct vring_used *)(uintptr_t)ring_addr_to_vva(dev,
 603                         vq, addr->used_user_addr, &len);
 604         if (vq->used == 0 || len != sizeof(struct vring_used) +
 605                         sizeof(struct vring_used_elem) * vq->size) {
 606                 RTE_LOG(DEBUG, VHOST_CONFIG,
 607                         "(%d) failed to map used ring.\n",
 608                         dev->vid);
 609                 return dev;
 610         }
 611
 612         if (vq->last_used_idx != vq->used->idx) {
 613                 RTE_LOG(WARNING, VHOST_CONFIG,
 614                         "last_used_idx (%u) and vq->used->idx (%u) mismatches; "
 615                         "some packets maybe resent for Tx and dropped for Rx\n",
 616                         vq->last_used_idx, vq->used->idx);
 617                 vq->last_used_idx  = vq->used->idx;
 618                 vq->last_avail_idx = vq->used->idx;
 619         }
 620
 621         vq->log_guest_addr = addr->log_guest_addr;
 622
 623         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address desc: %p\n",
 624                         dev->vid, vq->desc);
 625         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address avail: %p\n",
 626                         dev->vid, vq->avail);
 627         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) mapped address used: %p\n",
 628                         dev->vid, vq->used);
 629         VHOST_LOG_DEBUG(VHOST_CONFIG, "(%d) log_guest_addr: %" PRIx64 "\n",
 630                         dev->vid, vq->log_guest_addr);
 631
 632         return dev;
 633 }
 634
 635 /*
 636  * The virtio device sends us the desc, used and avail ring addresses.
 637  * This function then converts these to our address space.
 638  */
 639 static int
 640 vhost_user_set_vring_addr(struct virtio_net **pdev, struct VhostUserMsg *msg)
 641 {
 642         struct virtio_net *dev = *pdev;
 643         struct vhost_virtqueue *vq;
 644         struct vhost_vring_addr *addr = &msg->payload.addr;
 645
 646         if (dev->mem == NULL)
 647                 return VH_RESULT_ERR;
 648
 649         /* addr->index refers to the queue index. The txq 1, rxq is 0. */
 650         vq = dev->virtqueue[msg->payload.addr.index];
 651
 652         /*
 653          * Rings addresses should not be interpreted as long as the ring is not
 654          * started and enabled
 655          */
 656         memcpy(&vq->ring_addrs, addr, sizeof(*addr));
 657
 658         vring_invalidate(dev, vq);
 659
 660         if (vq->enabled && (dev->features &
 661                                 (1ULL << VHOST_USER_F_PROTOCOL_FEATURES))) {
 662                 dev = translate_ring_addresses(dev, msg->payload.addr.index);
 663                 if (!dev)
 664                         return VH_RESULT_ERR;
 665
 666                 *pdev = dev;
 667         }
 668
 669         return VH_RESULT_OK;
 670 }
 671
 672 /*
 673  * The virtio device sends us the available ring last used index.
 674  */
 675 static int
 676 vhost_user_set_vring_base(struct virtio_net **pdev,
 677                           struct VhostUserMsg *msg)
 678 {
 679         struct virtio_net *dev = *pdev;
 680         dev->virtqueue[msg->payload.state.index]->last_used_idx  =
 681                         msg->payload.state.num;
 682         dev->virtqueue[msg->payload.state.index]->last_avail_idx =
 683                         msg->payload.state.num;
 684
 685         return VH_RESULT_OK;
 686 }
 687
 688 static int
 689 add_one_guest_page(struct virtio_net *dev, uint64_t guest_phys_addr,
 690                    uint64_t host_phys_addr, uint64_t size)
 691 {
 692         struct guest_page *page, *last_page;
 693
 694         if (dev->nr_guest_pages == dev->max_guest_pages) {
 695                 dev->max_guest_pages *= 2;
 696                 dev->guest_pages = realloc(dev->guest_pages,
 697                                         dev->max_guest_pages * sizeof(*page));
 698                 if (!dev->guest_pages) {
 699                         RTE_LOG(ERR, VHOST_CONFIG, "cannot realloc guest_pages\n");
 700                         return -1;
 701                 }
 702         }
 703
 704         if (dev->nr_guest_pages > 0) {
 705                 last_page = &dev->guest_pages[dev->nr_guest_pages - 1];
 706                 /* merge if the two pages are continuous */
 707                 if (host_phys_addr == last_page->host_phys_addr +
 708                                       last_page->size) {
 709                         last_page->size += size;
 710                         return 0;
 711                 }
 712         }
 713
 714         page = &dev->guest_pages[dev->nr_guest_pages++];
 715         page->guest_phys_addr = guest_phys_addr;
 716         page->host_phys_addr  = host_phys_addr;
 717         page->size = size;
 718
 719         return 0;
 720 }
 721
 722 static int
 723 add_guest_pages(struct virtio_net *dev, struct rte_vhost_mem_region *reg,
 724                 uint64_t page_size)
 725 {
 726         uint64_t reg_size = reg->size;
 727         uint64_t host_user_addr  = reg->host_user_addr;
 728         uint64_t guest_phys_addr = reg->guest_phys_addr;
 729         uint64_t host_phys_addr;
 730         uint64_t size;
 731
 732         host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)host_user_addr);
 733         size = page_size - (guest_phys_addr & (page_size - 1));
 734         size = RTE_MIN(size, reg_size);
 735
 736         if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr, size) < 0)
 737                 return -1;
 738
 739         host_user_addr  += size;
 740         guest_phys_addr += size;
 741         reg_size -= size;
 742
 743         while (reg_size > 0) {
 744                 size = RTE_MIN(reg_size, page_size);
 745                 host_phys_addr = rte_mem_virt2iova((void *)(uintptr_t)
 746                                                   host_user_addr);
 747                 if (add_one_guest_page(dev, guest_phys_addr, host_phys_addr,
 748                                 size) < 0)
 749                         return -1;
 750
 751                 host_user_addr  += size;
 752                 guest_phys_addr += size;
 753                 reg_size -= size;
 754         }
 755
 756         return 0;
 757 }
 758
 759 #ifdef RTE_LIBRTE_VHOST_DEBUG
 760 /* TODO: enable it only in debug mode? */
 761 static void
 762 dump_guest_pages(struct virtio_net *dev)
 763 {
 764         uint32_t i;
 765         struct guest_page *page;
 766
 767         for (i = 0; i < dev->nr_guest_pages; i++) {
 768                 page = &dev->guest_pages[i];
 769
 770                 RTE_LOG(INFO, VHOST_CONFIG,
 771                         "guest physical page region %u\n"
 772                         "\t guest_phys_addr: %" PRIx64 "\n"
 773                         "\t host_phys_addr : %" PRIx64 "\n"
 774                         "\t size           : %" PRIx64 "\n",
 775                         i,
 776                         page->guest_phys_addr,
 777                         page->host_phys_addr,
 778                         page->size);
 779         }
 780 }
 781 #else
 782 #define dump_guest_pages(dev)
 783 #endif
 784
 785 static bool
 786 vhost_memory_changed(struct VhostUserMemory *new,
 787                      struct rte_vhost_memory *old)
 788 {
 789         uint32_t i;
 790
 791         if (new->nregions != old->nregions)
 792                 return true;
 793
 794         for (i = 0; i < new->nregions; ++i) {
 795                 VhostUserMemoryRegion *new_r = &new->regions[i];
 796                 struct rte_vhost_mem_region *old_r = &old->regions[i];
 797
 798                 if (new_r->guest_phys_addr != old_r->guest_phys_addr)
 799                         return true;
 800                 if (new_r->memory_size != old_r->size)
 801                         return true;
 802                 if (new_r->userspace_addr != old_r->guest_user_addr)
 803                         return true;
 804         }
 805
 806         return false;
 807 }
 808
 809 static int
 810 vhost_user_set_mem_table(struct virtio_net **pdev, struct VhostUserMsg *msg)
 811 {
 812         struct virtio_net *dev = *pdev;
 813         struct VhostUserMemory memory = msg->payload.memory;
 814         struct rte_vhost_mem_region *reg;
 815         void *mmap_addr;
 816         uint64_t mmap_size;
 817         uint64_t mmap_offset;
 818         uint64_t alignment;
 819         uint32_t i;
 820         int populate;
 821         int fd;
 822
 823         if (memory.nregions > VHOST_MEMORY_MAX_NREGIONS) {
 824                 RTE_LOG(ERR, VHOST_CONFIG,
 825                         "too many memory regions (%u)\n", memory.nregions);
 826                 return VH_RESULT_ERR;
 827         }
 828
 829         if (dev->mem && !vhost_memory_changed(&memory, dev->mem)) {
 830                 RTE_LOG(INFO, VHOST_CONFIG,
 831                         "(%d) memory regions not changed\n", dev->vid);
 832
 833                 for (i = 0; i < memory.nregions; i++)
 834                         close(msg->fds[i]);
 835
 836                 return VH_RESULT_OK;
 837         }
 838
 839         if (dev->mem) {
 840                 free_mem_region(dev);
 841                 rte_free(dev->mem);
 842                 dev->mem = NULL;
 843         }
 844
 845         /* Flush IOTLB cache as previous HVAs are now invalid */
 846         if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 847                 for (i = 0; i < dev->nr_vring; i++)
 848                         vhost_user_iotlb_flush_all(dev->virtqueue[i]);
 849
 850         dev->nr_guest_pages = 0;
 851         if (!dev->guest_pages) {
 852                 dev->max_guest_pages = 8;
 853                 dev->guest_pages = malloc(dev->max_guest_pages *
 854                                                 sizeof(struct guest_page));
 855                 if (dev->guest_pages == NULL) {
 856                         RTE_LOG(ERR, VHOST_CONFIG,
 857                                 "(%d) failed to allocate memory "
 858                                 "for dev->guest_pages\n",
 859                                 dev->vid);
 860                         return VH_RESULT_ERR;
 861                 }
 862         }
 863
 864         dev->mem = rte_zmalloc("vhost-mem-table", sizeof(struct rte_vhost_memory) +
 865                 sizeof(struct rte_vhost_mem_region) * memory.nregions, 0);
 866         if (dev->mem == NULL) {
 867                 RTE_LOG(ERR, VHOST_CONFIG,
 868                         "(%d) failed to allocate memory for dev->mem\n",
 869                         dev->vid);
 870                 return VH_RESULT_ERR;
 871         }
 872         dev->mem->nregions = memory.nregions;
 873
 874         for (i = 0; i < memory.nregions; i++) {
 875                 fd  = msg->fds[i];
 876                 reg = &dev->mem->regions[i];
 877
 878                 reg->guest_phys_addr = memory.regions[i].guest_phys_addr;
 879                 reg->guest_user_addr = memory.regions[i].userspace_addr;
 880                 reg->size            = memory.regions[i].memory_size;
 881                 reg->fd              = fd;
 882
 883                 mmap_offset = memory.regions[i].mmap_offset;
 884
 885                 /* Check for memory_size + mmap_offset overflow */
 886                 if (mmap_offset >= -reg->size) {
 887                         RTE_LOG(ERR, VHOST_CONFIG,
 888                                 "mmap_offset (%#"PRIx64") and memory_size "
 889                                 "(%#"PRIx64") overflow\n",
 890                                 mmap_offset, reg->size);
 891                         goto err_mmap;
 892                 }
 893
 894                 mmap_size = reg->size + mmap_offset;
 895
 896                 /* mmap() without flag of MAP_ANONYMOUS, should be called
 897                  * with length argument aligned with hugepagesz at older
 898                  * longterm version Linux, like 2.6.32 and 3.2.72, or
 899                  * mmap() will fail with EINVAL.
 900                  *
 901                  * to avoid failure, make sure in caller to keep length
 902                  * aligned.
 903                  */
 904                 alignment = get_blk_size(fd);
 905                 if (alignment == (uint64_t)-1) {
 906                         RTE_LOG(ERR, VHOST_CONFIG,
 907                                 "couldn't get hugepage size through fstat\n");
 908                         goto err_mmap;
 909                 }
 910                 mmap_size = RTE_ALIGN_CEIL(mmap_size, alignment);
 911
 912                 populate = (dev->dequeue_zero_copy) ? MAP_POPULATE : 0;
 913                 mmap_addr = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE,
 914                                  MAP_SHARED | populate, fd, 0);
 915
 916                 if (mmap_addr == MAP_FAILED) {
 917                         RTE_LOG(ERR, VHOST_CONFIG,
 918                                 "mmap region %u failed.\n", i);
 919                         goto err_mmap;
 920                 }
 921
 922                 reg->mmap_addr = mmap_addr;
 923                 reg->mmap_size = mmap_size;
 924                 reg->host_user_addr = (uint64_t)(uintptr_t)mmap_addr +
 925                                       mmap_offset;
 926
 927                 if (dev->dequeue_zero_copy)
 928                         if (add_guest_pages(dev, reg, alignment) < 0) {
 929                                 RTE_LOG(ERR, VHOST_CONFIG,
 930                                         "adding guest pages to region %u failed.\n",
 931                                         i);
 932                                 goto err_mmap;
 933                         }
 934
 935                 RTE_LOG(INFO, VHOST_CONFIG,
 936                         "guest memory region %u, size: 0x%" PRIx64 "\n"
 937                         "\t guest physical addr: 0x%" PRIx64 "\n"
 938                         "\t guest virtual  addr: 0x%" PRIx64 "\n"
 939                         "\t host  virtual  addr: 0x%" PRIx64 "\n"
 940                         "\t mmap addr : 0x%" PRIx64 "\n"
 941                         "\t mmap size : 0x%" PRIx64 "\n"
 942                         "\t mmap align: 0x%" PRIx64 "\n"
 943                         "\t mmap off  : 0x%" PRIx64 "\n",
 944                         i, reg->size,
 945                         reg->guest_phys_addr,
 946                         reg->guest_user_addr,
 947                         reg->host_user_addr,
 948                         (uint64_t)(uintptr_t)mmap_addr,
 949                         mmap_size,
 950                         alignment,
 951                         mmap_offset);
 952         }
 953
 954         for (i = 0; i < dev->nr_vring; i++) {
 955                 struct vhost_virtqueue *vq = dev->virtqueue[i];
 956
 957                 if (vq->desc || vq->avail || vq->used) {
 958                         /*
 959                          * If the memory table got updated, the ring addresses
 960                          * need to be translated again as virtual addresses have
 961                          * changed.
 962                          */
 963                         vring_invalidate(dev, vq);
 964
 965                         dev = translate_ring_addresses(dev, i);
 966                         if (!dev)
 967                                 return VH_RESULT_ERR;
 968
 969                         *pdev = dev;
 970                 }
 971         }
 972
 973         dump_guest_pages(dev);
 974
 975         return VH_RESULT_OK;
 976
 977 err_mmap:
 978         free_mem_region(dev);
 979         rte_free(dev->mem);
 980         dev->mem = NULL;
 981         return VH_RESULT_ERR;
 982 }
 983
 984 static bool
 985 vq_is_ready(struct virtio_net *dev, struct vhost_virtqueue *vq)
 986 {
 987         bool rings_ok;
 988
 989         if (!vq)
 990                 return false;
 991
 992         if (vq_is_packed(dev))
 993                 rings_ok = !!vq->desc_packed;
 994         else
 995                 rings_ok = vq->desc && vq->avail && vq->used;
 996
 997         return rings_ok &&
 998                vq->kickfd != VIRTIO_UNINITIALIZED_EVENTFD &&
 999                vq->callfd != VIRTIO_UNINITIALIZED_EVENTFD;
1000 }
1001
1002 static int
1003 virtio_is_ready(struct virtio_net *dev)
1004 {
1005         struct vhost_virtqueue *vq;
1006         uint32_t i;
1007
1008         if (dev->nr_vring == 0)
1009                 return 0;
1010
1011         for (i = 0; i < dev->nr_vring; i++) {
1012                 vq = dev->virtqueue[i];
1013
1014                 if (!vq_is_ready(dev, vq))
1015                         return 0;
1016         }
1017
1018         RTE_LOG(INFO, VHOST_CONFIG,
1019                 "virtio is now ready for processing.\n");
1020         return 1;
1021 }
1022
1023 static int
1024 vhost_user_set_vring_call(struct virtio_net **pdev, struct VhostUserMsg *msg)
1025 {
1026         struct virtio_net *dev = *pdev;
1027         struct vhost_vring_file file;
1028         struct vhost_virtqueue *vq;
1029
1030         file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1031         if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1032                 file.fd = VIRTIO_INVALID_EVENTFD;
1033         else
1034                 file.fd = msg->fds[0];
1035         RTE_LOG(INFO, VHOST_CONFIG,
1036                 "vring call idx:%d file:%d\n", file.index, file.fd);
1037
1038         vq = dev->virtqueue[file.index];
1039         if (vq->callfd >= 0)
1040                 close(vq->callfd);
1041
1042         vq->callfd = file.fd;
1043
1044         return VH_RESULT_OK;
1045 }
1046
1047 static int vhost_user_set_vring_err(struct virtio_net **pdev __rte_unused,
1048                         struct VhostUserMsg *msg)
1049 {
1050         if (!(msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK))
1051                 close(msg->fds[0]);
1052         RTE_LOG(INFO, VHOST_CONFIG, "not implemented\n");
1053
1054         return VH_RESULT_OK;
1055 }
1056
1057 static int
1058 vhost_user_set_vring_kick(struct virtio_net **pdev, struct VhostUserMsg *msg)
1059 {
1060         struct virtio_net *dev = *pdev;
1061         struct vhost_vring_file file;
1062         struct vhost_virtqueue *vq;
1063
1064         file.index = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1065         if (msg->payload.u64 & VHOST_USER_VRING_NOFD_MASK)
1066                 file.fd = VIRTIO_INVALID_EVENTFD;
1067         else
1068                 file.fd = msg->fds[0];
1069         RTE_LOG(INFO, VHOST_CONFIG,
1070                 "vring kick idx:%d file:%d\n", file.index, file.fd);
1071
1072         /* Interpret ring addresses only when ring is started. */
1073         dev = translate_ring_addresses(dev, file.index);
1074         if (!dev)
1075                 return VH_RESULT_ERR;
1076
1077         *pdev = dev;
1078
1079         vq = dev->virtqueue[file.index];
1080
1081         /*
1082          * When VHOST_USER_F_PROTOCOL_FEATURES is not negotiated,
1083          * the ring starts already enabled. Otherwise, it is enabled via
1084          * the SET_VRING_ENABLE message.
1085          */
1086         if (!(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES)))
1087                 vq->enabled = 1;
1088
1089         if (vq->kickfd >= 0)
1090                 close(vq->kickfd);
1091         vq->kickfd = file.fd;
1092
1093         return VH_RESULT_OK;
1094 }
1095
1096 static void
1097 free_zmbufs(struct vhost_virtqueue *vq)
1098 {
1099         struct zcopy_mbuf *zmbuf, *next;
1100
1101         for (zmbuf = TAILQ_FIRST(&vq->zmbuf_list);
1102              zmbuf != NULL; zmbuf = next) {
1103                 next = TAILQ_NEXT(zmbuf, next);
1104
1105                 rte_pktmbuf_free(zmbuf->mbuf);
1106                 TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
1107         }
1108
1109         rte_free(vq->zmbufs);
1110 }
1111
1112 /*
1113  * when virtio is stopped, qemu will send us the GET_VRING_BASE message.
1114  */
1115 static int
1116 vhost_user_get_vring_base(struct virtio_net **pdev,
1117                           struct VhostUserMsg *msg)
1118 {
1119         struct virtio_net *dev = *pdev;
1120         struct vhost_virtqueue *vq = dev->virtqueue[msg->payload.state.index];
1121
1122         /* We have to stop the queue (virtio) if it is running. */
1123         vhost_destroy_device_notify(dev);
1124
1125         dev->flags &= ~VIRTIO_DEV_READY;
1126         dev->flags &= ~VIRTIO_DEV_VDPA_CONFIGURED;
1127
1128         /* Here we are safe to get the last avail index */
1129         msg->payload.state.num = vq->last_avail_idx;
1130
1131         RTE_LOG(INFO, VHOST_CONFIG,
1132                 "vring base idx:%d file:%d\n", msg->payload.state.index,
1133                 msg->payload.state.num);
1134         /*
1135          * Based on current qemu vhost-user implementation, this message is
1136          * sent and only sent in vhost_vring_stop.
1137          * TODO: cleanup the vring, it isn't usable since here.
1138          */
1139         if (vq->kickfd >= 0)
1140                 close(vq->kickfd);
1141
1142         vq->kickfd = VIRTIO_UNINITIALIZED_EVENTFD;
1143
1144         if (vq->callfd >= 0)
1145                 close(vq->callfd);
1146
1147         vq->callfd = VIRTIO_UNINITIALIZED_EVENTFD;
1148
1149         if (dev->dequeue_zero_copy)
1150                 free_zmbufs(vq);
1151         if (vq_is_packed(dev)) {
1152                 rte_free(vq->shadow_used_packed);
1153                 vq->shadow_used_packed = NULL;
1154         } else {
1155                 rte_free(vq->shadow_used_split);
1156                 vq->shadow_used_split = NULL;
1157         }
1158
1159         rte_free(vq->batch_copy_elems);
1160         vq->batch_copy_elems = NULL;
1161
1162         msg->size = sizeof(msg->payload.state);
1163
1164         return VH_RESULT_OK;
1165 }
1166
1167 /*
1168  * when virtio queues are ready to work, qemu will send us to
1169  * enable the virtio queue pair.
1170  */
1171 static int
1172 vhost_user_set_vring_enable(struct virtio_net **pdev,
1173                             struct VhostUserMsg *msg)
1174 {
1175         struct virtio_net *dev = *pdev;
1176         int enable = (int)msg->payload.state.num;
1177         int index = (int)msg->payload.state.index;
1178         struct rte_vdpa_device *vdpa_dev;
1179         int did = -1;
1180
1181         RTE_LOG(INFO, VHOST_CONFIG,
1182                 "set queue enable: %d to qp idx: %d\n",
1183                 enable, index);
1184
1185         did = dev->vdpa_dev_id;
1186         vdpa_dev = rte_vdpa_get_device(did);
1187         if (vdpa_dev && vdpa_dev->ops->set_vring_state)
1188                 vdpa_dev->ops->set_vring_state(dev->vid, index, enable);
1189
1190         if (dev->notify_ops->vring_state_changed)
1191                 dev->notify_ops->vring_state_changed(dev->vid,
1192                                 index, enable);
1193
1194         dev->virtqueue[index]->enabled = enable;
1195
1196         return VH_RESULT_OK;
1197 }
1198
1199 static int
1200 vhost_user_get_protocol_features(struct virtio_net **pdev,
1201                                  struct VhostUserMsg *msg)
1202 {
1203         struct virtio_net *dev = *pdev;
1204         uint64_t features, protocol_features;
1205
1206         rte_vhost_driver_get_features(dev->ifname, &features);
1207         rte_vhost_driver_get_protocol_features(dev->ifname, &protocol_features);
1208
1209         /*
1210          * REPLY_ACK protocol feature is only mandatory for now
1211          * for IOMMU feature. If IOMMU is explicitly disabled by the
1212          * application, disable also REPLY_ACK feature for older buggy
1213          * Qemu versions (from v2.7.0 to v2.9.0).
1214          */
1215         if (!(features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)))
1216                 protocol_features &= ~(1ULL << VHOST_USER_PROTOCOL_F_REPLY_ACK);
1217
1218         msg->payload.u64 = protocol_features;
1219         msg->size = sizeof(msg->payload.u64);
1220
1221         return VH_RESULT_OK;
1222 }
1223
1224 static int
1225 vhost_user_set_protocol_features(struct virtio_net **pdev,
1226                                  struct VhostUserMsg *msg)
1227 {
1228         struct virtio_net *dev = *pdev;
1229         uint64_t protocol_features = msg->payload.u64;
1230         if (protocol_features & ~VHOST_USER_PROTOCOL_FEATURES) {
1231                 RTE_LOG(ERR, VHOST_CONFIG,
1232                         "(%d) received invalid protocol features.\n",
1233                         dev->vid);
1234                 return VH_RESULT_ERR;
1235         }
1236
1237         dev->protocol_features = protocol_features;
1238
1239         return VH_RESULT_OK;
1240 }
1241
1242 static int
1243 vhost_user_set_log_base(struct virtio_net **pdev, struct VhostUserMsg *msg)
1244 {
1245         struct virtio_net *dev = *pdev;
1246         int fd = msg->fds[0];
1247         uint64_t size, off;
1248         void *addr;
1249
1250         if (fd < 0) {
1251                 RTE_LOG(ERR, VHOST_CONFIG, "invalid log fd: %d\n", fd);
1252                 return VH_RESULT_ERR;
1253         }
1254
1255         if (msg->size != sizeof(VhostUserLog)) {
1256                 RTE_LOG(ERR, VHOST_CONFIG,
1257                         "invalid log base msg size: %"PRId32" != %d\n",
1258                         msg->size, (int)sizeof(VhostUserLog));
1259                 return VH_RESULT_ERR;
1260         }
1261
1262         size = msg->payload.log.mmap_size;
1263         off  = msg->payload.log.mmap_offset;
1264
1265         /* Don't allow mmap_offset to point outside the mmap region */
1266         if (off > size) {
1267                 RTE_LOG(ERR, VHOST_CONFIG,
1268                         "log offset %#"PRIx64" exceeds log size %#"PRIx64"\n",
1269                         off, size);
1270                 return VH_RESULT_ERR;
1271         }
1272
1273         RTE_LOG(INFO, VHOST_CONFIG,
1274                 "log mmap size: %"PRId64", offset: %"PRId64"\n",
1275                 size, off);
1276
1277         /*
1278          * mmap from 0 to workaround a hugepage mmap bug: mmap will
1279          * fail when offset is not page size aligned.
1280          */
1281         addr = mmap(0, size + off, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
1282         close(fd);
1283         if (addr == MAP_FAILED) {
1284                 RTE_LOG(ERR, VHOST_CONFIG, "mmap log base failed!\n");
1285                 return VH_RESULT_ERR;
1286         }
1287
1288         /*
1289          * Free previously mapped log memory on occasionally
1290          * multiple VHOST_USER_SET_LOG_BASE.
1291          */
1292         if (dev->log_addr) {
1293                 munmap((void *)(uintptr_t)dev->log_addr, dev->log_size);
1294         }
1295         dev->log_addr = (uint64_t)(uintptr_t)addr;
1296         dev->log_base = dev->log_addr + off;
1297         dev->log_size = size;
1298
1299         msg->size = sizeof(msg->payload.u64);
1300
1301         return VH_RESULT_OK;
1302 }
1303
1304 static int vhost_user_set_log_fd(struct virtio_net **pdev __rte_unused,
1305                         struct VhostUserMsg *msg)
1306 {
1307         close(msg->fds[0]);
1308         RTE_LOG(INFO, VHOST_CONFIG, "not implemented.\n");
1309
1310         return VH_RESULT_OK;
1311 }
1312
1313 /*
1314  * An rarp packet is constructed and broadcasted to notify switches about
1315  * the new location of the migrated VM, so that packets from outside will
1316  * not be lost after migration.
1317  *
1318  * However, we don't actually "send" a rarp packet here, instead, we set
1319  * a flag 'broadcast_rarp' to let rte_vhost_dequeue_burst() inject it.
1320  */
1321 static int
1322 vhost_user_send_rarp(struct virtio_net **pdev, struct VhostUserMsg *msg)
1323 {
1324         struct virtio_net *dev = *pdev;
1325         uint8_t *mac = (uint8_t *)&msg->payload.u64;
1326         struct rte_vdpa_device *vdpa_dev;
1327         int did = -1;
1328
1329         RTE_LOG(DEBUG, VHOST_CONFIG,
1330                 ":: mac: %02x:%02x:%02x:%02x:%02x:%02x\n",
1331                 mac[0], mac[1], mac[2], mac[3], mac[4], mac[5]);
1332         memcpy(dev->mac.addr_bytes, mac, 6);
1333
1334         /*
1335          * Set the flag to inject a RARP broadcast packet at
1336          * rte_vhost_dequeue_burst().
1337          *
1338          * rte_smp_wmb() is for making sure the mac is copied
1339          * before the flag is set.
1340          */
1341         rte_smp_wmb();
1342         rte_atomic16_set(&dev->broadcast_rarp, 1);
1343         did = dev->vdpa_dev_id;
1344         vdpa_dev = rte_vdpa_get_device(did);
1345         if (vdpa_dev && vdpa_dev->ops->migration_done)
1346                 vdpa_dev->ops->migration_done(dev->vid);
1347
1348         return VH_RESULT_OK;
1349 }
1350
1351 static int
1352 vhost_user_net_set_mtu(struct virtio_net **pdev, struct VhostUserMsg *msg)
1353 {
1354         struct virtio_net *dev = *pdev;
1355         if (msg->payload.u64 < VIRTIO_MIN_MTU ||
1356                         msg->payload.u64 > VIRTIO_MAX_MTU) {
1357                 RTE_LOG(ERR, VHOST_CONFIG, "Invalid MTU size (%"PRIu64")\n",
1358                                 msg->payload.u64);
1359
1360                 return VH_RESULT_ERR;
1361         }
1362
1363         dev->mtu = msg->payload.u64;
1364
1365         return VH_RESULT_OK;
1366 }
1367
1368 static int
1369 vhost_user_set_req_fd(struct virtio_net **pdev, struct VhostUserMsg *msg)
1370 {
1371         struct virtio_net *dev = *pdev;
1372         int fd = msg->fds[0];
1373
1374         if (fd < 0) {
1375                 RTE_LOG(ERR, VHOST_CONFIG,
1376                                 "Invalid file descriptor for slave channel (%d)\n",
1377                                 fd);
1378                 return VH_RESULT_ERR;
1379         }
1380
1381         dev->slave_req_fd = fd;
1382
1383         return VH_RESULT_OK;
1384 }
1385
1386 static int
1387 is_vring_iotlb_update(struct vhost_virtqueue *vq, struct vhost_iotlb_msg *imsg)
1388 {
1389         struct vhost_vring_addr *ra;
1390         uint64_t start, end;
1391
1392         start = imsg->iova;
1393         end = start + imsg->size;
1394
1395         ra = &vq->ring_addrs;
1396         if (ra->desc_user_addr >= start && ra->desc_user_addr < end)
1397                 return 1;
1398         if (ra->avail_user_addr >= start && ra->avail_user_addr < end)
1399                 return 1;
1400         if (ra->used_user_addr >= start && ra->used_user_addr < end)
1401                 return 1;
1402
1403         return 0;
1404 }
1405
1406 static int
1407 is_vring_iotlb_invalidate(struct vhost_virtqueue *vq,
1408                                 struct vhost_iotlb_msg *imsg)
1409 {
1410         uint64_t istart, iend, vstart, vend;
1411
1412         istart = imsg->iova;
1413         iend = istart + imsg->size - 1;
1414
1415         vstart = (uintptr_t)vq->desc;
1416         vend = vstart + sizeof(struct vring_desc) * vq->size - 1;
1417         if (vstart <= iend && istart <= vend)
1418                 return 1;
1419
1420         vstart = (uintptr_t)vq->avail;
1421         vend = vstart + sizeof(struct vring_avail);
1422         vend += sizeof(uint16_t) * vq->size - 1;
1423         if (vstart <= iend && istart <= vend)
1424                 return 1;
1425
1426         vstart = (uintptr_t)vq->used;
1427         vend = vstart + sizeof(struct vring_used);
1428         vend += sizeof(struct vring_used_elem) * vq->size - 1;
1429         if (vstart <= iend && istart <= vend)
1430                 return 1;
1431
1432         return 0;
1433 }
1434
1435 static int
1436 vhost_user_iotlb_msg(struct virtio_net **pdev, struct VhostUserMsg *msg)
1437 {
1438         struct virtio_net *dev = *pdev;
1439         struct vhost_iotlb_msg *imsg = &msg->payload.iotlb;
1440         uint16_t i;
1441         uint64_t vva, len;
1442
1443         switch (imsg->type) {
1444         case VHOST_IOTLB_UPDATE:
1445                 len = imsg->size;
1446                 vva = qva_to_vva(dev, imsg->uaddr, &len);
1447                 if (!vva)
1448                         return VH_RESULT_ERR;
1449
1450                 for (i = 0; i < dev->nr_vring; i++) {
1451                         struct vhost_virtqueue *vq = dev->virtqueue[i];
1452
1453                         vhost_user_iotlb_cache_insert(vq, imsg->iova, vva,
1454                                         len, imsg->perm);
1455
1456                         if (is_vring_iotlb_update(vq, imsg))
1457                                 *pdev = dev = translate_ring_addresses(dev, i);
1458                 }
1459                 break;
1460         case VHOST_IOTLB_INVALIDATE:
1461                 for (i = 0; i < dev->nr_vring; i++) {
1462                         struct vhost_virtqueue *vq = dev->virtqueue[i];
1463
1464                         vhost_user_iotlb_cache_remove(vq, imsg->iova,
1465                                         imsg->size);
1466
1467                         if (is_vring_iotlb_invalidate(vq, imsg))
1468                                 vring_invalidate(dev, vq);
1469                 }
1470                 break;
1471         default:
1472                 RTE_LOG(ERR, VHOST_CONFIG, "Invalid IOTLB message type (%d)\n",
1473                                 imsg->type);
1474                 return VH_RESULT_ERR;
1475         }
1476
1477         return VH_RESULT_OK;
1478 }
1479
1480 /* return bytes# of read on success or negative val on failure. */
1481 static int
1482 read_vhost_message(int sockfd, struct VhostUserMsg *msg)
1483 {
1484         int ret;
1485
1486         ret = read_fd_message(sockfd, (char *)msg, VHOST_USER_HDR_SIZE,
1487                 msg->fds, VHOST_MEMORY_MAX_NREGIONS);
1488         if (ret <= 0)
1489                 return ret;
1490
1491         if (msg && msg->size) {
1492                 if (msg->size > sizeof(msg->payload)) {
1493                         RTE_LOG(ERR, VHOST_CONFIG,
1494                                 "invalid msg size: %d\n", msg->size);
1495                         return -1;
1496                 }
1497                 ret = read(sockfd, &msg->payload, msg->size);
1498                 if (ret <= 0)
1499                         return ret;
1500                 if (ret != (int)msg->size) {
1501                         RTE_LOG(ERR, VHOST_CONFIG,
1502                                 "read control message failed\n");
1503                         return -1;
1504                 }
1505         }
1506
1507         return ret;
1508 }
1509
1510 static int
1511 send_vhost_message(int sockfd, struct VhostUserMsg *msg, int *fds, int fd_num)
1512 {
1513         if (!msg)
1514                 return 0;
1515
1516         return send_fd_message(sockfd, (char *)msg,
1517                 VHOST_USER_HDR_SIZE + msg->size, fds, fd_num);
1518 }
1519
1520 static int
1521 send_vhost_reply(int sockfd, struct VhostUserMsg *msg)
1522 {
1523         if (!msg)
1524                 return 0;
1525
1526         msg->flags &= ~VHOST_USER_VERSION_MASK;
1527         msg->flags &= ~VHOST_USER_NEED_REPLY;
1528         msg->flags |= VHOST_USER_VERSION;
1529         msg->flags |= VHOST_USER_REPLY_MASK;
1530
1531         return send_vhost_message(sockfd, msg, NULL, 0);
1532 }
1533
1534 static int
1535 send_vhost_slave_message(struct virtio_net *dev, struct VhostUserMsg *msg,
1536                          int *fds, int fd_num)
1537 {
1538         int ret;
1539
1540         if (msg->flags & VHOST_USER_NEED_REPLY)
1541                 rte_spinlock_lock(&dev->slave_req_lock);
1542
1543         ret = send_vhost_message(dev->slave_req_fd, msg, fds, fd_num);
1544         if (ret < 0 && (msg->flags & VHOST_USER_NEED_REPLY))
1545                 rte_spinlock_unlock(&dev->slave_req_lock);
1546
1547         return ret;
1548 }
1549
1550 /*
1551  * Allocate a queue pair if it hasn't been allocated yet
1552  */
1553 static int
1554 vhost_user_check_and_alloc_queue_pair(struct virtio_net *dev,
1555                         struct VhostUserMsg *msg)
1556 {
1557         uint16_t vring_idx;
1558
1559         switch (msg->request.master) {
1560         case VHOST_USER_SET_VRING_KICK:
1561         case VHOST_USER_SET_VRING_CALL:
1562         case VHOST_USER_SET_VRING_ERR:
1563                 vring_idx = msg->payload.u64 & VHOST_USER_VRING_IDX_MASK;
1564                 break;
1565         case VHOST_USER_SET_VRING_NUM:
1566         case VHOST_USER_SET_VRING_BASE:
1567         case VHOST_USER_SET_VRING_ENABLE:
1568                 vring_idx = msg->payload.state.index;
1569                 break;
1570         case VHOST_USER_SET_VRING_ADDR:
1571                 vring_idx = msg->payload.addr.index;
1572                 break;
1573         default:
1574                 return 0;
1575         }
1576
1577         if (vring_idx >= VHOST_MAX_VRING) {
1578                 RTE_LOG(ERR, VHOST_CONFIG,
1579                         "invalid vring index: %u\n", vring_idx);
1580                 return -1;
1581         }
1582
1583         if (dev->virtqueue[vring_idx])
1584                 return 0;
1585
1586         return alloc_vring_queue(dev, vring_idx);
1587 }
1588
1589 static void
1590 vhost_user_lock_all_queue_pairs(struct virtio_net *dev)
1591 {
1592         unsigned int i = 0;
1593         unsigned int vq_num = 0;
1594
1595         while (vq_num < dev->nr_vring) {
1596                 struct vhost_virtqueue *vq = dev->virtqueue[i];
1597
1598                 if (vq) {
1599                         rte_spinlock_lock(&vq->access_lock);
1600                         vq_num++;
1601                 }
1602                 i++;
1603         }
1604 }
1605
1606 static void
1607 vhost_user_unlock_all_queue_pairs(struct virtio_net *dev)
1608 {
1609         unsigned int i = 0;
1610         unsigned int vq_num = 0;
1611
1612         while (vq_num < dev->nr_vring) {
1613                 struct vhost_virtqueue *vq = dev->virtqueue[i];
1614
1615                 if (vq) {
1616                         rte_spinlock_unlock(&vq->access_lock);
1617                         vq_num++;
1618                 }
1619                 i++;
1620         }
1621 }
1622
1623 int
1624 vhost_user_msg_handler(int vid, int fd)
1625 {
1626         struct virtio_net *dev;
1627         struct VhostUserMsg msg;
1628         struct rte_vdpa_device *vdpa_dev;
1629         int did = -1;
1630         int ret;
1631         int unlock_required = 0;
1632         uint32_t skip_master = 0;
1633
1634         dev = get_device(vid);
1635         if (dev == NULL)
1636                 return -1;
1637
1638         if (!dev->notify_ops) {
1639                 dev->notify_ops = vhost_driver_callback_get(dev->ifname);
1640                 if (!dev->notify_ops) {
1641                         RTE_LOG(ERR, VHOST_CONFIG,
1642                                 "failed to get callback ops for driver %s\n",
1643                                 dev->ifname);
1644                         return -1;
1645                 }
1646         }
1647
1648         ret = read_vhost_message(fd, &msg);
1649         if (ret <= 0 || msg.request.master >= VHOST_USER_MAX) {
1650                 if (ret < 0)
1651                         RTE_LOG(ERR, VHOST_CONFIG,
1652                                 "vhost read message failed\n");
1653                 else if (ret == 0)
1654                         RTE_LOG(INFO, VHOST_CONFIG,
1655                                 "vhost peer closed\n");
1656                 else
1657                         RTE_LOG(ERR, VHOST_CONFIG,
1658                                 "vhost read incorrect message\n");
1659
1660                 return -1;
1661         }
1662
1663         ret = 0;
1664         if (msg.request.master != VHOST_USER_IOTLB_MSG)
1665                 RTE_LOG(INFO, VHOST_CONFIG, "read message %s\n",
1666                         vhost_message_str[msg.request.master]);
1667         else
1668                 RTE_LOG(DEBUG, VHOST_CONFIG, "read message %s\n",
1669                         vhost_message_str[msg.request.master]);
1670
1671         ret = vhost_user_check_and_alloc_queue_pair(dev, &msg);
1672         if (ret < 0) {
1673                 RTE_LOG(ERR, VHOST_CONFIG,
1674                         "failed to alloc queue\n");
1675                 return -1;
1676         }
1677
1678         /*
1679          * Note: we don't lock all queues on VHOST_USER_GET_VRING_BASE
1680          * and VHOST_USER_RESET_OWNER, since it is sent when virtio stops
1681          * and device is destroyed. destroy_device waits for queues to be
1682          * inactive, so it is safe. Otherwise taking the access_lock
1683          * would cause a dead lock.
1684          */
1685         switch (msg.request.master) {
1686         case VHOST_USER_SET_FEATURES:
1687         case VHOST_USER_SET_PROTOCOL_FEATURES:
1688         case VHOST_USER_SET_OWNER:
1689         case VHOST_USER_SET_MEM_TABLE:
1690         case VHOST_USER_SET_LOG_BASE:
1691         case VHOST_USER_SET_LOG_FD:
1692         case VHOST_USER_SET_VRING_NUM:
1693         case VHOST_USER_SET_VRING_ADDR:
1694         case VHOST_USER_SET_VRING_BASE:
1695         case VHOST_USER_SET_VRING_KICK:
1696         case VHOST_USER_SET_VRING_CALL:
1697         case VHOST_USER_SET_VRING_ERR:
1698         case VHOST_USER_SET_VRING_ENABLE:
1699         case VHOST_USER_SEND_RARP:
1700         case VHOST_USER_NET_SET_MTU:
1701         case VHOST_USER_SET_SLAVE_REQ_FD:
1702                 vhost_user_lock_all_queue_pairs(dev);
1703                 unlock_required = 1;
1704                 break;
1705         default:
1706                 break;
1707
1708         }
1709
1710         if (dev->extern_ops.pre_msg_handle) {
1711                 uint32_t need_reply;
1712
1713                 ret = (*dev->extern_ops.pre_msg_handle)(dev->vid,
1714                                 (void *)&msg, &need_reply, &skip_master);
1715                 if (ret < 0)
1716                         goto skip_to_reply;
1717
1718                 if (need_reply)
1719                         send_vhost_reply(fd, &msg);
1720
1721                 if (skip_master)
1722                         goto skip_to_post_handle;
1723         }
1724
1725         switch (msg.request.master) {
1726         case VHOST_USER_GET_FEATURES:
1727                 ret = vhost_user_get_features(&dev, &msg);
1728                 send_vhost_reply(fd, &msg);
1729                 break;
1730         case VHOST_USER_SET_FEATURES:
1731                 ret = vhost_user_set_features(&dev, &msg);
1732                 break;
1733
1734         case VHOST_USER_GET_PROTOCOL_FEATURES:
1735                 ret = vhost_user_get_protocol_features(&dev, &msg);
1736                 send_vhost_reply(fd, &msg);
1737                 break;
1738         case VHOST_USER_SET_PROTOCOL_FEATURES:
1739                 ret = vhost_user_set_protocol_features(&dev, &msg);
1740                 break;
1741
1742         case VHOST_USER_SET_OWNER:
1743                 ret = vhost_user_set_owner(&dev, &msg);
1744                 break;
1745         case VHOST_USER_RESET_OWNER:
1746                 ret = vhost_user_reset_owner(&dev, &msg);
1747                 break;
1748
1749         case VHOST_USER_SET_MEM_TABLE:
1750                 ret = vhost_user_set_mem_table(&dev, &msg);
1751                 break;
1752
1753         case VHOST_USER_SET_LOG_BASE:
1754                 ret = vhost_user_set_log_base(&dev, &msg);
1755                 if (ret)
1756                         goto skip_to_reply;
1757                 /* it needs a reply */
1758                 send_vhost_reply(fd, &msg);
1759                 break;
1760         case VHOST_USER_SET_LOG_FD:
1761                 ret = vhost_user_set_log_fd(&dev, &msg);
1762                 break;
1763
1764         case VHOST_USER_SET_VRING_NUM:
1765                 ret = vhost_user_set_vring_num(&dev, &msg);
1766                 break;
1767         case VHOST_USER_SET_VRING_ADDR:
1768                 ret = vhost_user_set_vring_addr(&dev, &msg);
1769                 break;
1770         case VHOST_USER_SET_VRING_BASE:
1771                 ret = vhost_user_set_vring_base(&dev, &msg);
1772                 break;
1773
1774         case VHOST_USER_GET_VRING_BASE:
1775                 ret = vhost_user_get_vring_base(&dev, &msg);
1776                 if (ret)
1777                         goto skip_to_reply;
1778                 send_vhost_reply(fd, &msg);
1779                 break;
1780
1781         case VHOST_USER_SET_VRING_KICK:
1782                 ret = vhost_user_set_vring_kick(&dev, &msg);
1783                 break;
1784         case VHOST_USER_SET_VRING_CALL:
1785                 ret = vhost_user_set_vring_call(&dev, &msg);
1786                 break;
1787
1788         case VHOST_USER_SET_VRING_ERR:
1789                 ret = vhost_user_set_vring_err(&dev, &msg);
1790                 break;
1791
1792         case VHOST_USER_GET_QUEUE_NUM:
1793                 ret = vhost_user_get_queue_num(&dev, &msg);
1794                 send_vhost_reply(fd, &msg);
1795                 break;
1796
1797         case VHOST_USER_SET_VRING_ENABLE:
1798                 ret = vhost_user_set_vring_enable(&dev, &msg);
1799                 break;
1800         case VHOST_USER_SEND_RARP:
1801                 ret = vhost_user_send_rarp(&dev, &msg);
1802                 break;
1803
1804         case VHOST_USER_NET_SET_MTU:
1805                 ret = vhost_user_net_set_mtu(&dev, &msg);
1806                 break;
1807
1808         case VHOST_USER_SET_SLAVE_REQ_FD:
1809                 ret = vhost_user_set_req_fd(&dev, &msg);
1810                 break;
1811
1812         case VHOST_USER_IOTLB_MSG:
1813                 ret = vhost_user_iotlb_msg(&dev, &msg);
1814                 break;
1815
1816         default:
1817                 ret = -1;
1818                 break;
1819         }
1820
1821 skip_to_post_handle:
1822         if (!ret && dev->extern_ops.post_msg_handle) {
1823                 uint32_t need_reply;
1824
1825                 ret = (*dev->extern_ops.post_msg_handle)(
1826                                 dev->vid, (void *)&msg, &need_reply);
1827                 if (ret < 0)
1828                         goto skip_to_reply;
1829
1830                 if (need_reply)
1831                         send_vhost_reply(fd, &msg);
1832         }
1833
1834 skip_to_reply:
1835         if (unlock_required)
1836                 vhost_user_unlock_all_queue_pairs(dev);
1837
1838         if (msg.flags & VHOST_USER_NEED_REPLY) {
1839                 msg.payload.u64 = !!ret;
1840                 msg.size = sizeof(msg.payload.u64);
1841                 send_vhost_reply(fd, &msg);
1842         } else if (ret) {
1843                 RTE_LOG(ERR, VHOST_CONFIG,
1844                         "vhost message handling failed.\n");
1845                 return -1;
1846         }
1847
1848         if (!(dev->flags & VIRTIO_DEV_RUNNING) && virtio_is_ready(dev)) {
1849                 dev->flags |= VIRTIO_DEV_READY;
1850
1851                 if (!(dev->flags & VIRTIO_DEV_RUNNING)) {
1852                         if (dev->dequeue_zero_copy) {
1853                                 RTE_LOG(INFO, VHOST_CONFIG,
1854                                                 "dequeue zero copy is enabled\n");
1855                         }
1856
1857                         if (dev->notify_ops->new_device(dev->vid) == 0)
1858                                 dev->flags |= VIRTIO_DEV_RUNNING;
1859                 }
1860         }
1861
1862         did = dev->vdpa_dev_id;
1863         vdpa_dev = rte_vdpa_get_device(did);
1864         if (vdpa_dev && virtio_is_ready(dev) &&
1865                         !(dev->flags & VIRTIO_DEV_VDPA_CONFIGURED) &&
1866                         msg.request.master == VHOST_USER_SET_VRING_ENABLE) {
1867                 if (vdpa_dev->ops->dev_conf)
1868                         vdpa_dev->ops->dev_conf(dev->vid);
1869                 dev->flags |= VIRTIO_DEV_VDPA_CONFIGURED;
1870                 if (vhost_user_host_notifier_ctrl(dev->vid, true) != 0) {
1871                         RTE_LOG(INFO, VHOST_CONFIG,
1872                                 "(%d) software relay is used for vDPA, performance may be low.\n",
1873                                 dev->vid);
1874                 }
1875         }
1876
1877         return 0;
1878 }
1879
1880 static int process_slave_message_reply(struct virtio_net *dev,
1881                                        const struct VhostUserMsg *msg)
1882 {
1883         struct VhostUserMsg msg_reply;
1884         int ret;
1885
1886         if ((msg->flags & VHOST_USER_NEED_REPLY) == 0)
1887                 return 0;
1888
1889         if (read_vhost_message(dev->slave_req_fd, &msg_reply) < 0) {
1890                 ret = -1;
1891                 goto out;
1892         }
1893
1894         if (msg_reply.request.slave != msg->request.slave) {
1895                 RTE_LOG(ERR, VHOST_CONFIG,
1896                         "Received unexpected msg type (%u), expected %u\n",
1897                         msg_reply.request.slave, msg->request.slave);
1898                 ret = -1;
1899                 goto out;
1900         }
1901
1902         ret = msg_reply.payload.u64 ? -1 : 0;
1903
1904 out:
1905         rte_spinlock_unlock(&dev->slave_req_lock);
1906         return ret;
1907 }
1908
1909 int
1910 vhost_user_iotlb_miss(struct virtio_net *dev, uint64_t iova, uint8_t perm)
1911 {
1912         int ret;
1913         struct VhostUserMsg msg = {
1914                 .request.slave = VHOST_USER_SLAVE_IOTLB_MSG,
1915                 .flags = VHOST_USER_VERSION,
1916                 .size = sizeof(msg.payload.iotlb),
1917                 .payload.iotlb = {
1918                         .iova = iova,
1919                         .perm = perm,
1920                         .type = VHOST_IOTLB_MISS,
1921                 },
1922         };
1923
1924         ret = send_vhost_message(dev->slave_req_fd, &msg, NULL, 0);
1925         if (ret < 0) {
1926                 RTE_LOG(ERR, VHOST_CONFIG,
1927                                 "Failed to send IOTLB miss message (%d)\n",
1928                                 ret);
1929                 return ret;
1930         }
1931
1932         return 0;
1933 }
1934
1935 static int vhost_user_slave_set_vring_host_notifier(struct virtio_net *dev,
1936                                                     int index, int fd,
1937                                                     uint64_t offset,
1938                                                     uint64_t size)
1939 {
1940         int *fdp = NULL;
1941         size_t fd_num = 0;
1942         int ret;
1943         struct VhostUserMsg msg = {
1944                 .request.slave = VHOST_USER_SLAVE_VRING_HOST_NOTIFIER_MSG,
1945                 .flags = VHOST_USER_VERSION | VHOST_USER_NEED_REPLY,
1946                 .size = sizeof(msg.payload.area),
1947                 .payload.area = {
1948                         .u64 = index & VHOST_USER_VRING_IDX_MASK,
1949                         .size = size,
1950                         .offset = offset,
1951                 },
1952         };
1953
1954         if (fd < 0)
1955                 msg.payload.area.u64 |= VHOST_USER_VRING_NOFD_MASK;
1956         else {
1957                 fdp = &fd;
1958                 fd_num = 1;
1959         }
1960
1961         ret = send_vhost_slave_message(dev, &msg, fdp, fd_num);
1962         if (ret < 0) {
1963                 RTE_LOG(ERR, VHOST_CONFIG,
1964                         "Failed to set host notifier (%d)\n", ret);
1965                 return ret;
1966         }
1967
1968         return process_slave_message_reply(dev, &msg);
1969 }
1970
1971 int vhost_user_host_notifier_ctrl(int vid, bool enable)
1972 {
1973         struct virtio_net *dev;
1974         struct rte_vdpa_device *vdpa_dev;
1975         int vfio_device_fd, did, ret = 0;
1976         uint64_t offset, size;
1977         unsigned int i;
1978
1979         dev = get_device(vid);
1980         if (!dev)
1981                 return -ENODEV;
1982
1983         did = dev->vdpa_dev_id;
1984         if (did < 0)
1985                 return -EINVAL;
1986
1987         if (!(dev->features & (1ULL << VIRTIO_F_VERSION_1)) ||
1988             !(dev->features & (1ULL << VHOST_USER_F_PROTOCOL_FEATURES)) ||
1989             !(dev->protocol_features &
1990                         (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_REQ)) ||
1991             !(dev->protocol_features &
1992                         (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_SEND_FD)) ||
1993             !(dev->protocol_features &
1994                         (1ULL << VHOST_USER_PROTOCOL_F_HOST_NOTIFIER)))
1995                 return -ENOTSUP;
1996
1997         vdpa_dev = rte_vdpa_get_device(did);
1998         if (!vdpa_dev)
1999                 return -ENODEV;
2000
2001         RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_vfio_device_fd, -ENOTSUP);
2002         RTE_FUNC_PTR_OR_ERR_RET(vdpa_dev->ops->get_notify_area, -ENOTSUP);
2003
2004         vfio_device_fd = vdpa_dev->ops->get_vfio_device_fd(vid);
2005         if (vfio_device_fd < 0)
2006                 return -ENOTSUP;
2007
2008         if (enable) {
2009                 for (i = 0; i < dev->nr_vring; i++) {
2010                         if (vdpa_dev->ops->get_notify_area(vid, i, &offset,
2011                                         &size) < 0) {
2012                                 ret = -ENOTSUP;
2013                                 goto disable;
2014                         }
2015
2016                         if (vhost_user_slave_set_vring_host_notifier(dev, i,
2017                                         vfio_device_fd, offset, size) < 0) {
2018                                 ret = -EFAULT;
2019                                 goto disable;
2020                         }
2021                 }
2022         } else {
2023 disable:
2024                 for (i = 0; i < dev->nr_vring; i++) {
2025                         vhost_user_slave_set_vring_host_notifier(dev, i, -1,
2026                                         0, 0);
2027                 }
2028         }
2029
2030         return ret;
2031 }