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