drivers/vhost/vhost.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /* Copyright (C) 2009 Red Hat, Inc.
   3  * Copyright (C) 2006 Rusty Russell IBM Corporation
   4  *
   5  * Author: Michael S. Tsirkin <mst@redhat.com>
   6  *
   7  * Inspiration, some code, and most witty comments come from
   8  * Documentation/virtual/lguest/lguest.c, by Rusty Russell
   9  *
  10  * Generic code for virtio server in host kernel.
  11  */
  12
  13 #include <linux/eventfd.h>
  14 #include <linux/vhost.h>
  15 #include <linux/uio.h>
  16 #include <linux/mm.h>
  17 #include <linux/miscdevice.h>
  18 #include <linux/mutex.h>
  19 #include <linux/poll.h>
  20 #include <linux/file.h>
  21 #include <linux/highmem.h>
  22 #include <linux/slab.h>
  23 #include <linux/vmalloc.h>
  24 #include <linux/kthread.h>
  25 #include <linux/module.h>
  26 #include <linux/sort.h>
  27 #include <linux/sched/mm.h>
  28 #include <linux/sched/signal.h>
  29 #include <linux/sched/vhost_task.h>
  30 #include <linux/interval_tree_generic.h>
  31 #include <linux/nospec.h>
  32 #include <linux/kcov.h>
  33
  34 #include "vhost.h"
  35
  36 static ushort max_mem_regions = 64;
  37 module_param(max_mem_regions, ushort, 0444);
  38 MODULE_PARM_DESC(max_mem_regions,
  39         "Maximum number of memory regions in memory map. (default: 64)");
  40 static int max_iotlb_entries = 2048;
  41 module_param(max_iotlb_entries, int, 0444);
  42 MODULE_PARM_DESC(max_iotlb_entries,
  43         "Maximum number of iotlb entries. (default: 2048)");
  44
  45 enum {
  46         VHOST_MEMORY_F_LOG = 0x1,
  47 };
  48
  49 #define vhost_used_event(vq) ((__virtio16 __user *)&vq->avail->ring[vq->num])
  50 #define vhost_avail_event(vq) ((__virtio16 __user *)&vq->used->ring[vq->num])
  51
  52 #ifdef CONFIG_VHOST_CROSS_ENDIAN_LEGACY
  53 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
  54 {
  55         vq->user_be = !virtio_legacy_is_little_endian();
  56 }
  57
  58 static void vhost_enable_cross_endian_big(struct vhost_virtqueue *vq)
  59 {
  60         vq->user_be = true;
  61 }
  62
  63 static void vhost_enable_cross_endian_little(struct vhost_virtqueue *vq)
  64 {
  65         vq->user_be = false;
  66 }
  67
  68 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
  69 {
  70         struct vhost_vring_state s;
  71
  72         if (vq->private_data)
  73                 return -EBUSY;
  74
  75         if (copy_from_user(&s, argp, sizeof(s)))
  76                 return -EFAULT;
  77
  78         if (s.num != VHOST_VRING_LITTLE_ENDIAN &&
  79             s.num != VHOST_VRING_BIG_ENDIAN)
  80                 return -EINVAL;
  81
  82         if (s.num == VHOST_VRING_BIG_ENDIAN)
  83                 vhost_enable_cross_endian_big(vq);
  84         else
  85                 vhost_enable_cross_endian_little(vq);
  86
  87         return 0;
  88 }
  89
  90 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
  91                                    int __user *argp)
  92 {
  93         struct vhost_vring_state s = {
  94                 .index = idx,
  95                 .num = vq->user_be
  96         };
  97
  98         if (copy_to_user(argp, &s, sizeof(s)))
  99                 return -EFAULT;
 100
 101         return 0;
 102 }
 103
 104 static void vhost_init_is_le(struct vhost_virtqueue *vq)
 105 {
 106         /* Note for legacy virtio: user_be is initialized at reset time
 107          * according to the host endianness. If userspace does not set an
 108          * explicit endianness, the default behavior is native endian, as
 109          * expected by legacy virtio.
 110          */
 111         vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1) || !vq->user_be;
 112 }
 113 #else
 114 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
 115 {
 116 }
 117
 118 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
 119 {
 120         return -ENOIOCTLCMD;
 121 }
 122
 123 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
 124                                    int __user *argp)
 125 {
 126         return -ENOIOCTLCMD;
 127 }
 128
 129 static void vhost_init_is_le(struct vhost_virtqueue *vq)
 130 {
 131         vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1)
 132                 || virtio_legacy_is_little_endian();
 133 }
 134 #endif /* CONFIG_VHOST_CROSS_ENDIAN_LEGACY */
 135
 136 static void vhost_reset_is_le(struct vhost_virtqueue *vq)
 137 {
 138         vhost_init_is_le(vq);
 139 }
 140
 141 struct vhost_flush_struct {
 142         struct vhost_work work;
 143         struct completion wait_event;
 144 };
 145
 146 static void vhost_flush_work(struct vhost_work *work)
 147 {
 148         struct vhost_flush_struct *s;
 149
 150         s = container_of(work, struct vhost_flush_struct, work);
 151         complete(&s->wait_event);
 152 }
 153
 154 static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
 155                             poll_table *pt)
 156 {
 157         struct vhost_poll *poll;
 158
 159         poll = container_of(pt, struct vhost_poll, table);
 160         poll->wqh = wqh;
 161         add_wait_queue(wqh, &poll->wait);
 162 }
 163
 164 static int vhost_poll_wakeup(wait_queue_entry_t *wait, unsigned mode, int sync,
 165                              void *key)
 166 {
 167         struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
 168         struct vhost_work *work = &poll->work;
 169
 170         if (!(key_to_poll(key) & poll->mask))
 171                 return 0;
 172
 173         if (!poll->dev->use_worker)
 174                 work->fn(work);
 175         else
 176                 vhost_poll_queue(poll);
 177
 178         return 0;
 179 }
 180
 181 void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
 182 {
 183         clear_bit(VHOST_WORK_QUEUED, &work->flags);
 184         work->fn = fn;
 185 }
 186 EXPORT_SYMBOL_GPL(vhost_work_init);
 187
 188 /* Init poll structure */
 189 void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
 190                      __poll_t mask, struct vhost_dev *dev)
 191 {
 192         init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
 193         init_poll_funcptr(&poll->table, vhost_poll_func);
 194         poll->mask = mask;
 195         poll->dev = dev;
 196         poll->wqh = NULL;
 197
 198         vhost_work_init(&poll->work, fn);
 199 }
 200 EXPORT_SYMBOL_GPL(vhost_poll_init);
 201
 202 /* Start polling a file. We add ourselves to file's wait queue. The caller must
 203  * keep a reference to a file until after vhost_poll_stop is called. */
 204 int vhost_poll_start(struct vhost_poll *poll, struct file *file)
 205 {
 206         __poll_t mask;
 207
 208         if (poll->wqh)
 209                 return 0;
 210
 211         mask = vfs_poll(file, &poll->table);
 212         if (mask)
 213                 vhost_poll_wakeup(&poll->wait, 0, 0, poll_to_key(mask));
 214         if (mask & EPOLLERR) {
 215                 vhost_poll_stop(poll);
 216                 return -EINVAL;
 217         }
 218
 219         return 0;
 220 }
 221 EXPORT_SYMBOL_GPL(vhost_poll_start);
 222
 223 /* Stop polling a file. After this function returns, it becomes safe to drop the
 224  * file reference. You must also flush afterwards. */
 225 void vhost_poll_stop(struct vhost_poll *poll)
 226 {
 227         if (poll->wqh) {
 228                 remove_wait_queue(poll->wqh, &poll->wait);
 229                 poll->wqh = NULL;
 230         }
 231 }
 232 EXPORT_SYMBOL_GPL(vhost_poll_stop);
 233
 234 void vhost_dev_flush(struct vhost_dev *dev)
 235 {
 236         struct vhost_flush_struct flush;
 237
 238         if (dev->worker) {
 239                 init_completion(&flush.wait_event);
 240                 vhost_work_init(&flush.work, vhost_flush_work);
 241
 242                 vhost_work_queue(dev, &flush.work);
 243                 wait_for_completion(&flush.wait_event);
 244         }
 245 }
 246 EXPORT_SYMBOL_GPL(vhost_dev_flush);
 247
 248 void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work)
 249 {
 250         if (!dev->worker)
 251                 return;
 252
 253         if (!test_and_set_bit(VHOST_WORK_QUEUED, &work->flags)) {
 254                 /* We can only add the work to the list after we're
 255                  * sure it was not in the list.
 256                  * test_and_set_bit() implies a memory barrier.
 257                  */
 258                 llist_add(&work->node, &dev->worker->work_list);
 259                 wake_up_process(dev->worker->vtsk->task);
 260         }
 261 }
 262 EXPORT_SYMBOL_GPL(vhost_work_queue);
 263
 264 /* A lockless hint for busy polling code to exit the loop */
 265 bool vhost_has_work(struct vhost_dev *dev)
 266 {
 267         return dev->worker && !llist_empty(&dev->worker->work_list);
 268 }
 269 EXPORT_SYMBOL_GPL(vhost_has_work);
 270
 271 void vhost_poll_queue(struct vhost_poll *poll)
 272 {
 273         vhost_work_queue(poll->dev, &poll->work);
 274 }
 275 EXPORT_SYMBOL_GPL(vhost_poll_queue);
 276
 277 static void __vhost_vq_meta_reset(struct vhost_virtqueue *vq)
 278 {
 279         int j;
 280
 281         for (j = 0; j < VHOST_NUM_ADDRS; j++)
 282                 vq->meta_iotlb[j] = NULL;
 283 }
 284
 285 static void vhost_vq_meta_reset(struct vhost_dev *d)
 286 {
 287         int i;
 288
 289         for (i = 0; i < d->nvqs; ++i)
 290                 __vhost_vq_meta_reset(d->vqs[i]);
 291 }
 292
 293 static void vhost_vring_call_reset(struct vhost_vring_call *call_ctx)
 294 {
 295         call_ctx->ctx = NULL;
 296         memset(&call_ctx->producer, 0x0, sizeof(struct irq_bypass_producer));
 297 }
 298
 299 bool vhost_vq_is_setup(struct vhost_virtqueue *vq)
 300 {
 301         return vq->avail && vq->desc && vq->used && vhost_vq_access_ok(vq);
 302 }
 303 EXPORT_SYMBOL_GPL(vhost_vq_is_setup);
 304
 305 static void vhost_vq_reset(struct vhost_dev *dev,
 306                            struct vhost_virtqueue *vq)
 307 {
 308         vq->num = 1;
 309         vq->desc = NULL;
 310         vq->avail = NULL;
 311         vq->used = NULL;
 312         vq->last_avail_idx = 0;
 313         vq->avail_idx = 0;
 314         vq->last_used_idx = 0;
 315         vq->signalled_used = 0;
 316         vq->signalled_used_valid = false;
 317         vq->used_flags = 0;
 318         vq->log_used = false;
 319         vq->log_addr = -1ull;
 320         vq->private_data = NULL;
 321         vq->acked_features = 0;
 322         vq->acked_backend_features = 0;
 323         vq->log_base = NULL;
 324         vq->error_ctx = NULL;
 325         vq->kick = NULL;
 326         vq->log_ctx = NULL;
 327         vhost_disable_cross_endian(vq);
 328         vhost_reset_is_le(vq);
 329         vq->busyloop_timeout = 0;
 330         vq->umem = NULL;
 331         vq->iotlb = NULL;
 332         vhost_vring_call_reset(&vq->call_ctx);
 333         __vhost_vq_meta_reset(vq);
 334 }
 335
 336 static int vhost_worker(void *data)
 337 {
 338         struct vhost_worker *worker = data;
 339         struct vhost_work *work, *work_next;
 340         struct llist_node *node;
 341
 342         for (;;) {
 343                 /* mb paired w/ kthread_stop */
 344                 set_current_state(TASK_INTERRUPTIBLE);
 345
 346                 if (vhost_task_should_stop(worker->vtsk)) {
 347                         __set_current_state(TASK_RUNNING);
 348                         break;
 349                 }
 350
 351                 node = llist_del_all(&worker->work_list);
 352                 if (!node)
 353                         schedule();
 354
 355                 node = llist_reverse_order(node);
 356                 /* make sure flag is seen after deletion */
 357                 smp_wmb();
 358                 llist_for_each_entry_safe(work, work_next, node, node) {
 359                         clear_bit(VHOST_WORK_QUEUED, &work->flags);
 360                         __set_current_state(TASK_RUNNING);
 361                         kcov_remote_start_common(worker->kcov_handle);
 362                         work->fn(work);
 363                         kcov_remote_stop();
 364                         cond_resched();
 365                 }
 366         }
 367
 368         return 0;
 369 }
 370
 371 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
 372 {
 373         kfree(vq->indirect);
 374         vq->indirect = NULL;
 375         kfree(vq->log);
 376         vq->log = NULL;
 377         kfree(vq->heads);
 378         vq->heads = NULL;
 379 }
 380
 381 /* Helper to allocate iovec buffers for all vqs. */
 382 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
 383 {
 384         struct vhost_virtqueue *vq;
 385         int i;
 386
 387         for (i = 0; i < dev->nvqs; ++i) {
 388                 vq = dev->vqs[i];
 389                 vq->indirect = kmalloc_array(UIO_MAXIOV,
 390                                              sizeof(*vq->indirect),
 391                                              GFP_KERNEL);
 392                 vq->log = kmalloc_array(dev->iov_limit, sizeof(*vq->log),
 393                                         GFP_KERNEL);
 394                 vq->heads = kmalloc_array(dev->iov_limit, sizeof(*vq->heads),
 395                                           GFP_KERNEL);
 396                 if (!vq->indirect || !vq->log || !vq->heads)
 397                         goto err_nomem;
 398         }
 399         return 0;
 400
 401 err_nomem:
 402         for (; i >= 0; --i)
 403                 vhost_vq_free_iovecs(dev->vqs[i]);
 404         return -ENOMEM;
 405 }
 406
 407 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
 408 {
 409         int i;
 410
 411         for (i = 0; i < dev->nvqs; ++i)
 412                 vhost_vq_free_iovecs(dev->vqs[i]);
 413 }
 414
 415 bool vhost_exceeds_weight(struct vhost_virtqueue *vq,
 416                           int pkts, int total_len)
 417 {
 418         struct vhost_dev *dev = vq->dev;
 419
 420         if ((dev->byte_weight && total_len >= dev->byte_weight) ||
 421             pkts >= dev->weight) {
 422                 vhost_poll_queue(&vq->poll);
 423                 return true;
 424         }
 425
 426         return false;
 427 }
 428 EXPORT_SYMBOL_GPL(vhost_exceeds_weight);
 429
 430 static size_t vhost_get_avail_size(struct vhost_virtqueue *vq,
 431                                    unsigned int num)
 432 {
 433         size_t event __maybe_unused =
 434                vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
 435
 436         return size_add(struct_size(vq->avail, ring, num), event);
 437 }
 438
 439 static size_t vhost_get_used_size(struct vhost_virtqueue *vq,
 440                                   unsigned int num)
 441 {
 442         size_t event __maybe_unused =
 443                vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
 444
 445         return size_add(struct_size(vq->used, ring, num), event);
 446 }
 447
 448 static size_t vhost_get_desc_size(struct vhost_virtqueue *vq,
 449                                   unsigned int num)
 450 {
 451         return sizeof(*vq->desc) * num;
 452 }
 453
 454 void vhost_dev_init(struct vhost_dev *dev,
 455                     struct vhost_virtqueue **vqs, int nvqs,
 456                     int iov_limit, int weight, int byte_weight,
 457                     bool use_worker,
 458                     int (*msg_handler)(struct vhost_dev *dev, u32 asid,
 459                                        struct vhost_iotlb_msg *msg))
 460 {
 461         struct vhost_virtqueue *vq;
 462         int i;
 463
 464         dev->vqs = vqs;
 465         dev->nvqs = nvqs;
 466         mutex_init(&dev->mutex);
 467         dev->log_ctx = NULL;
 468         dev->umem = NULL;
 469         dev->iotlb = NULL;
 470         dev->mm = NULL;
 471         dev->worker = NULL;
 472         dev->iov_limit = iov_limit;
 473         dev->weight = weight;
 474         dev->byte_weight = byte_weight;
 475         dev->use_worker = use_worker;
 476         dev->msg_handler = msg_handler;
 477         init_waitqueue_head(&dev->wait);
 478         INIT_LIST_HEAD(&dev->read_list);
 479         INIT_LIST_HEAD(&dev->pending_list);
 480         spin_lock_init(&dev->iotlb_lock);
 481
 482
 483         for (i = 0; i < dev->nvqs; ++i) {
 484                 vq = dev->vqs[i];
 485                 vq->log = NULL;
 486                 vq->indirect = NULL;
 487                 vq->heads = NULL;
 488                 vq->dev = dev;
 489                 mutex_init(&vq->mutex);
 490                 vhost_vq_reset(dev, vq);
 491                 if (vq->handle_kick)
 492                         vhost_poll_init(&vq->poll, vq->handle_kick,
 493                                         EPOLLIN, dev);
 494         }
 495 }
 496 EXPORT_SYMBOL_GPL(vhost_dev_init);
 497
 498 /* Caller should have device mutex */
 499 long vhost_dev_check_owner(struct vhost_dev *dev)
 500 {
 501         /* Are you the owner? If not, I don't think you mean to do that */
 502         return dev->mm == current->mm ? 0 : -EPERM;
 503 }
 504 EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
 505
 506 /* Caller should have device mutex */
 507 bool vhost_dev_has_owner(struct vhost_dev *dev)
 508 {
 509         return dev->mm;
 510 }
 511 EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
 512
 513 static void vhost_attach_mm(struct vhost_dev *dev)
 514 {
 515         /* No owner, become one */
 516         if (dev->use_worker) {
 517                 dev->mm = get_task_mm(current);
 518         } else {
 519                 /* vDPA device does not use worker thead, so there's
 520                  * no need to hold the address space for mm. This help
 521                  * to avoid deadlock in the case of mmap() which may
 522                  * held the refcnt of the file and depends on release
 523                  * method to remove vma.
 524                  */
 525                 dev->mm = current->mm;
 526                 mmgrab(dev->mm);
 527         }
 528 }
 529
 530 static void vhost_detach_mm(struct vhost_dev *dev)
 531 {
 532         if (!dev->mm)
 533                 return;
 534
 535         if (dev->use_worker)
 536                 mmput(dev->mm);
 537         else
 538                 mmdrop(dev->mm);
 539
 540         dev->mm = NULL;
 541 }
 542
 543 static void vhost_worker_free(struct vhost_dev *dev)
 544 {
 545         struct vhost_worker *worker = dev->worker;
 546
 547         if (!worker)
 548                 return;
 549
 550         dev->worker = NULL;
 551         WARN_ON(!llist_empty(&worker->work_list));
 552         vhost_task_stop(worker->vtsk);
 553         kfree(worker);
 554 }
 555
 556 static int vhost_worker_create(struct vhost_dev *dev)
 557 {
 558         struct vhost_worker *worker;
 559         struct vhost_task *vtsk;
 560         char name[TASK_COMM_LEN];
 561         int ret;
 562
 563         worker = kzalloc(sizeof(*worker), GFP_KERNEL_ACCOUNT);
 564         if (!worker)
 565                 return -ENOMEM;
 566
 567         dev->worker = worker;
 568         worker->kcov_handle = kcov_common_handle();
 569         init_llist_head(&worker->work_list);
 570         snprintf(name, sizeof(name), "vhost-%d", current->pid);
 571
 572         vtsk = vhost_task_create(vhost_worker, worker, name);
 573         if (!vtsk) {
 574                 ret = -ENOMEM;
 575                 goto free_worker;
 576         }
 577
 578         worker->vtsk = vtsk;
 579         vhost_task_start(vtsk);
 580         return 0;
 581
 582 free_worker:
 583         kfree(worker);
 584         dev->worker = NULL;
 585         return ret;
 586 }
 587
 588 /* Caller should have device mutex */
 589 long vhost_dev_set_owner(struct vhost_dev *dev)
 590 {
 591         int err;
 592
 593         /* Is there an owner already? */
 594         if (vhost_dev_has_owner(dev)) {
 595                 err = -EBUSY;
 596                 goto err_mm;
 597         }
 598
 599         vhost_attach_mm(dev);
 600
 601         if (dev->use_worker) {
 602                 err = vhost_worker_create(dev);
 603                 if (err)
 604                         goto err_worker;
 605         }
 606
 607         err = vhost_dev_alloc_iovecs(dev);
 608         if (err)
 609                 goto err_iovecs;
 610
 611         return 0;
 612 err_iovecs:
 613         vhost_worker_free(dev);
 614 err_worker:
 615         vhost_detach_mm(dev);
 616 err_mm:
 617         return err;
 618 }
 619 EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
 620
 621 static struct vhost_iotlb *iotlb_alloc(void)
 622 {
 623         return vhost_iotlb_alloc(max_iotlb_entries,
 624                                  VHOST_IOTLB_FLAG_RETIRE);
 625 }
 626
 627 struct vhost_iotlb *vhost_dev_reset_owner_prepare(void)
 628 {
 629         return iotlb_alloc();
 630 }
 631 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
 632
 633 /* Caller should have device mutex */
 634 void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_iotlb *umem)
 635 {
 636         int i;
 637
 638         vhost_dev_cleanup(dev);
 639
 640         dev->umem = umem;
 641         /* We don't need VQ locks below since vhost_dev_cleanup makes sure
 642          * VQs aren't running.
 643          */
 644         for (i = 0; i < dev->nvqs; ++i)
 645                 dev->vqs[i]->umem = umem;
 646 }
 647 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
 648
 649 void vhost_dev_stop(struct vhost_dev *dev)
 650 {
 651         int i;
 652
 653         for (i = 0; i < dev->nvqs; ++i) {
 654                 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick)
 655                         vhost_poll_stop(&dev->vqs[i]->poll);
 656         }
 657
 658         vhost_dev_flush(dev);
 659 }
 660 EXPORT_SYMBOL_GPL(vhost_dev_stop);
 661
 662 void vhost_clear_msg(struct vhost_dev *dev)
 663 {
 664         struct vhost_msg_node *node, *n;
 665
 666         spin_lock(&dev->iotlb_lock);
 667
 668         list_for_each_entry_safe(node, n, &dev->read_list, node) {
 669                 list_del(&node->node);
 670                 kfree(node);
 671         }
 672
 673         list_for_each_entry_safe(node, n, &dev->pending_list, node) {
 674                 list_del(&node->node);
 675                 kfree(node);
 676         }
 677
 678         spin_unlock(&dev->iotlb_lock);
 679 }
 680 EXPORT_SYMBOL_GPL(vhost_clear_msg);
 681
 682 void vhost_dev_cleanup(struct vhost_dev *dev)
 683 {
 684         int i;
 685
 686         for (i = 0; i < dev->nvqs; ++i) {
 687                 if (dev->vqs[i]->error_ctx)
 688                         eventfd_ctx_put(dev->vqs[i]->error_ctx);
 689                 if (dev->vqs[i]->kick)
 690                         fput(dev->vqs[i]->kick);
 691                 if (dev->vqs[i]->call_ctx.ctx)
 692                         eventfd_ctx_put(dev->vqs[i]->call_ctx.ctx);
 693                 vhost_vq_reset(dev, dev->vqs[i]);
 694         }
 695         vhost_dev_free_iovecs(dev);
 696         if (dev->log_ctx)
 697                 eventfd_ctx_put(dev->log_ctx);
 698         dev->log_ctx = NULL;
 699         /* No one will access memory at this point */
 700         vhost_iotlb_free(dev->umem);
 701         dev->umem = NULL;
 702         vhost_iotlb_free(dev->iotlb);
 703         dev->iotlb = NULL;
 704         vhost_clear_msg(dev);
 705         wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
 706         vhost_worker_free(dev);
 707         vhost_detach_mm(dev);
 708 }
 709 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
 710
 711 static bool log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
 712 {
 713         u64 a = addr / VHOST_PAGE_SIZE / 8;
 714
 715         /* Make sure 64 bit math will not overflow. */
 716         if (a > ULONG_MAX - (unsigned long)log_base ||
 717             a + (unsigned long)log_base > ULONG_MAX)
 718                 return false;
 719
 720         return access_ok(log_base + a,
 721                          (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
 722 }
 723
 724 /* Make sure 64 bit math will not overflow. */
 725 static bool vhost_overflow(u64 uaddr, u64 size)
 726 {
 727         if (uaddr > ULONG_MAX || size > ULONG_MAX)
 728                 return true;
 729
 730         if (!size)
 731                 return false;
 732
 733         return uaddr > ULONG_MAX - size + 1;
 734 }
 735
 736 /* Caller should have vq mutex and device mutex. */
 737 static bool vq_memory_access_ok(void __user *log_base, struct vhost_iotlb *umem,
 738                                 int log_all)
 739 {
 740         struct vhost_iotlb_map *map;
 741
 742         if (!umem)
 743                 return false;
 744
 745         list_for_each_entry(map, &umem->list, link) {
 746                 unsigned long a = map->addr;
 747
 748                 if (vhost_overflow(map->addr, map->size))
 749                         return false;
 750
 751
 752                 if (!access_ok((void __user *)a, map->size))
 753                         return false;
 754                 else if (log_all && !log_access_ok(log_base,
 755                                                    map->start,
 756                                                    map->size))
 757                         return false;
 758         }
 759         return true;
 760 }
 761
 762 static inline void __user *vhost_vq_meta_fetch(struct vhost_virtqueue *vq,
 763                                                u64 addr, unsigned int size,
 764                                                int type)
 765 {
 766         const struct vhost_iotlb_map *map = vq->meta_iotlb[type];
 767
 768         if (!map)
 769                 return NULL;
 770
 771         return (void __user *)(uintptr_t)(map->addr + addr - map->start);
 772 }
 773
 774 /* Can we switch to this memory table? */
 775 /* Caller should have device mutex but not vq mutex */
 776 static bool memory_access_ok(struct vhost_dev *d, struct vhost_iotlb *umem,
 777                              int log_all)
 778 {
 779         int i;
 780
 781         for (i = 0; i < d->nvqs; ++i) {
 782                 bool ok;
 783                 bool log;
 784
 785                 mutex_lock(&d->vqs[i]->mutex);
 786                 log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
 787                 /* If ring is inactive, will check when it's enabled. */
 788                 if (d->vqs[i]->private_data)
 789                         ok = vq_memory_access_ok(d->vqs[i]->log_base,
 790                                                  umem, log);
 791                 else
 792                         ok = true;
 793                 mutex_unlock(&d->vqs[i]->mutex);
 794                 if (!ok)
 795                         return false;
 796         }
 797         return true;
 798 }
 799
 800 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
 801                           struct iovec iov[], int iov_size, int access);
 802
 803 static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
 804                               const void *from, unsigned size)
 805 {
 806         int ret;
 807
 808         if (!vq->iotlb)
 809                 return __copy_to_user(to, from, size);
 810         else {
 811                 /* This function should be called after iotlb
 812                  * prefetch, which means we're sure that all vq
 813                  * could be access through iotlb. So -EAGAIN should
 814                  * not happen in this case.
 815                  */
 816                 struct iov_iter t;
 817                 void __user *uaddr = vhost_vq_meta_fetch(vq,
 818                                      (u64)(uintptr_t)to, size,
 819                                      VHOST_ADDR_USED);
 820
 821                 if (uaddr)
 822                         return __copy_to_user(uaddr, from, size);
 823
 824                 ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
 825                                      ARRAY_SIZE(vq->iotlb_iov),
 826                                      VHOST_ACCESS_WO);
 827                 if (ret < 0)
 828                         goto out;
 829                 iov_iter_init(&t, ITER_DEST, vq->iotlb_iov, ret, size);
 830                 ret = copy_to_iter(from, size, &t);
 831                 if (ret == size)
 832                         ret = 0;
 833         }
 834 out:
 835         return ret;
 836 }
 837
 838 static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
 839                                 void __user *from, unsigned size)
 840 {
 841         int ret;
 842
 843         if (!vq->iotlb)
 844                 return __copy_from_user(to, from, size);
 845         else {
 846                 /* This function should be called after iotlb
 847                  * prefetch, which means we're sure that vq
 848                  * could be access through iotlb. So -EAGAIN should
 849                  * not happen in this case.
 850                  */
 851                 void __user *uaddr = vhost_vq_meta_fetch(vq,
 852                                      (u64)(uintptr_t)from, size,
 853                                      VHOST_ADDR_DESC);
 854                 struct iov_iter f;
 855
 856                 if (uaddr)
 857                         return __copy_from_user(to, uaddr, size);
 858
 859                 ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
 860                                      ARRAY_SIZE(vq->iotlb_iov),
 861                                      VHOST_ACCESS_RO);
 862                 if (ret < 0) {
 863                         vq_err(vq, "IOTLB translation failure: uaddr "
 864                                "%p size 0x%llx\n", from,
 865                                (unsigned long long) size);
 866                         goto out;
 867                 }
 868                 iov_iter_init(&f, ITER_SOURCE, vq->iotlb_iov, ret, size);
 869                 ret = copy_from_iter(to, size, &f);
 870                 if (ret == size)
 871                         ret = 0;
 872         }
 873
 874 out:
 875         return ret;
 876 }
 877
 878 static void __user *__vhost_get_user_slow(struct vhost_virtqueue *vq,
 879                                           void __user *addr, unsigned int size,
 880                                           int type)
 881 {
 882         int ret;
 883
 884         ret = translate_desc(vq, (u64)(uintptr_t)addr, size, vq->iotlb_iov,
 885                              ARRAY_SIZE(vq->iotlb_iov),
 886                              VHOST_ACCESS_RO);
 887         if (ret < 0) {
 888                 vq_err(vq, "IOTLB translation failure: uaddr "
 889                         "%p size 0x%llx\n", addr,
 890                         (unsigned long long) size);
 891                 return NULL;
 892         }
 893
 894         if (ret != 1 || vq->iotlb_iov[0].iov_len != size) {
 895                 vq_err(vq, "Non atomic userspace memory access: uaddr "
 896                         "%p size 0x%llx\n", addr,
 897                         (unsigned long long) size);
 898                 return NULL;
 899         }
 900
 901         return vq->iotlb_iov[0].iov_base;
 902 }
 903
 904 /* This function should be called after iotlb
 905  * prefetch, which means we're sure that vq
 906  * could be access through iotlb. So -EAGAIN should
 907  * not happen in this case.
 908  */
 909 static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
 910                                             void __user *addr, unsigned int size,
 911                                             int type)
 912 {
 913         void __user *uaddr = vhost_vq_meta_fetch(vq,
 914                              (u64)(uintptr_t)addr, size, type);
 915         if (uaddr)
 916                 return uaddr;
 917
 918         return __vhost_get_user_slow(vq, addr, size, type);
 919 }
 920
 921 #define vhost_put_user(vq, x, ptr)              \
 922 ({ \
 923         int ret; \
 924         if (!vq->iotlb) { \
 925                 ret = __put_user(x, ptr); \
 926         } else { \
 927                 __typeof__(ptr) to = \
 928                         (__typeof__(ptr)) __vhost_get_user(vq, ptr,     \
 929                                           sizeof(*ptr), VHOST_ADDR_USED); \
 930                 if (to != NULL) \
 931                         ret = __put_user(x, to); \
 932                 else \
 933                         ret = -EFAULT;  \
 934         } \
 935         ret; \
 936 })
 937
 938 static inline int vhost_put_avail_event(struct vhost_virtqueue *vq)
 939 {
 940         return vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
 941                               vhost_avail_event(vq));
 942 }
 943
 944 static inline int vhost_put_used(struct vhost_virtqueue *vq,
 945                                  struct vring_used_elem *head, int idx,
 946                                  int count)
 947 {
 948         return vhost_copy_to_user(vq, vq->used->ring + idx, head,
 949                                   count * sizeof(*head));
 950 }
 951
 952 static inline int vhost_put_used_flags(struct vhost_virtqueue *vq)
 953
 954 {
 955         return vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
 956                               &vq->used->flags);
 957 }
 958
 959 static inline int vhost_put_used_idx(struct vhost_virtqueue *vq)
 960
 961 {
 962         return vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
 963                               &vq->used->idx);
 964 }
 965
 966 #define vhost_get_user(vq, x, ptr, type)                \
 967 ({ \
 968         int ret; \
 969         if (!vq->iotlb) { \
 970                 ret = __get_user(x, ptr); \
 971         } else { \
 972                 __typeof__(ptr) from = \
 973                         (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
 974                                                            sizeof(*ptr), \
 975                                                            type); \
 976                 if (from != NULL) \
 977                         ret = __get_user(x, from); \
 978                 else \
 979                         ret = -EFAULT; \
 980         } \
 981         ret; \
 982 })
 983
 984 #define vhost_get_avail(vq, x, ptr) \
 985         vhost_get_user(vq, x, ptr, VHOST_ADDR_AVAIL)
 986
 987 #define vhost_get_used(vq, x, ptr) \
 988         vhost_get_user(vq, x, ptr, VHOST_ADDR_USED)
 989
 990 static void vhost_dev_lock_vqs(struct vhost_dev *d)
 991 {
 992         int i = 0;
 993         for (i = 0; i < d->nvqs; ++i)
 994                 mutex_lock_nested(&d->vqs[i]->mutex, i);
 995 }
 996
 997 static void vhost_dev_unlock_vqs(struct vhost_dev *d)
 998 {
 999         int i = 0;
1000         for (i = 0; i < d->nvqs; ++i)
1001                 mutex_unlock(&d->vqs[i]->mutex);
1002 }
1003
1004 static inline int vhost_get_avail_idx(struct vhost_virtqueue *vq,
1005                                       __virtio16 *idx)
1006 {
1007         return vhost_get_avail(vq, *idx, &vq->avail->idx);
1008 }
1009
1010 static inline int vhost_get_avail_head(struct vhost_virtqueue *vq,
1011                                        __virtio16 *head, int idx)
1012 {
1013         return vhost_get_avail(vq, *head,
1014                                &vq->avail->ring[idx & (vq->num - 1)]);
1015 }
1016
1017 static inline int vhost_get_avail_flags(struct vhost_virtqueue *vq,
1018                                         __virtio16 *flags)
1019 {
1020         return vhost_get_avail(vq, *flags, &vq->avail->flags);
1021 }
1022
1023 static inline int vhost_get_used_event(struct vhost_virtqueue *vq,
1024                                        __virtio16 *event)
1025 {
1026         return vhost_get_avail(vq, *event, vhost_used_event(vq));
1027 }
1028
1029 static inline int vhost_get_used_idx(struct vhost_virtqueue *vq,
1030                                      __virtio16 *idx)
1031 {
1032         return vhost_get_used(vq, *idx, &vq->used->idx);
1033 }
1034
1035 static inline int vhost_get_desc(struct vhost_virtqueue *vq,
1036                                  struct vring_desc *desc, int idx)
1037 {
1038         return vhost_copy_from_user(vq, desc, vq->desc + idx, sizeof(*desc));
1039 }
1040
1041 static void vhost_iotlb_notify_vq(struct vhost_dev *d,
1042                                   struct vhost_iotlb_msg *msg)
1043 {
1044         struct vhost_msg_node *node, *n;
1045
1046         spin_lock(&d->iotlb_lock);
1047
1048         list_for_each_entry_safe(node, n, &d->pending_list, node) {
1049                 struct vhost_iotlb_msg *vq_msg = &node->msg.iotlb;
1050                 if (msg->iova <= vq_msg->iova &&
1051                     msg->iova + msg->size - 1 >= vq_msg->iova &&
1052                     vq_msg->type == VHOST_IOTLB_MISS) {
1053                         vhost_poll_queue(&node->vq->poll);
1054                         list_del(&node->node);
1055                         kfree(node);
1056                 }
1057         }
1058
1059         spin_unlock(&d->iotlb_lock);
1060 }
1061
1062 static bool umem_access_ok(u64 uaddr, u64 size, int access)
1063 {
1064         unsigned long a = uaddr;
1065
1066         /* Make sure 64 bit math will not overflow. */
1067         if (vhost_overflow(uaddr, size))
1068                 return false;
1069
1070         if ((access & VHOST_ACCESS_RO) &&
1071             !access_ok((void __user *)a, size))
1072                 return false;
1073         if ((access & VHOST_ACCESS_WO) &&
1074             !access_ok((void __user *)a, size))
1075                 return false;
1076         return true;
1077 }
1078
1079 static int vhost_process_iotlb_msg(struct vhost_dev *dev, u32 asid,
1080                                    struct vhost_iotlb_msg *msg)
1081 {
1082         int ret = 0;
1083
1084         if (asid != 0)
1085                 return -EINVAL;
1086
1087         mutex_lock(&dev->mutex);
1088         vhost_dev_lock_vqs(dev);
1089         switch (msg->type) {
1090         case VHOST_IOTLB_UPDATE:
1091                 if (!dev->iotlb) {
1092                         ret = -EFAULT;
1093                         break;
1094                 }
1095                 if (!umem_access_ok(msg->uaddr, msg->size, msg->perm)) {
1096                         ret = -EFAULT;
1097                         break;
1098                 }
1099                 vhost_vq_meta_reset(dev);
1100                 if (vhost_iotlb_add_range(dev->iotlb, msg->iova,
1101                                           msg->iova + msg->size - 1,
1102                                           msg->uaddr, msg->perm)) {
1103                         ret = -ENOMEM;
1104                         break;
1105                 }
1106                 vhost_iotlb_notify_vq(dev, msg);
1107                 break;
1108         case VHOST_IOTLB_INVALIDATE:
1109                 if (!dev->iotlb) {
1110                         ret = -EFAULT;
1111                         break;
1112                 }
1113                 vhost_vq_meta_reset(dev);
1114                 vhost_iotlb_del_range(dev->iotlb, msg->iova,
1115                                       msg->iova + msg->size - 1);
1116                 break;
1117         default:
1118                 ret = -EINVAL;
1119                 break;
1120         }
1121
1122         vhost_dev_unlock_vqs(dev);
1123         mutex_unlock(&dev->mutex);
1124
1125         return ret;
1126 }
1127 ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
1128                              struct iov_iter *from)
1129 {
1130         struct vhost_iotlb_msg msg;
1131         size_t offset;
1132         int type, ret;
1133         u32 asid = 0;
1134
1135         ret = copy_from_iter(&type, sizeof(type), from);
1136         if (ret != sizeof(type)) {
1137                 ret = -EINVAL;
1138                 goto done;
1139         }
1140
1141         switch (type) {
1142         case VHOST_IOTLB_MSG:
1143                 /* There maybe a hole after type for V1 message type,
1144                  * so skip it here.
1145                  */
1146                 offset = offsetof(struct vhost_msg, iotlb) - sizeof(int);
1147                 break;
1148         case VHOST_IOTLB_MSG_V2:
1149                 if (vhost_backend_has_feature(dev->vqs[0],
1150                                               VHOST_BACKEND_F_IOTLB_ASID)) {
1151                         ret = copy_from_iter(&asid, sizeof(asid), from);
1152                         if (ret != sizeof(asid)) {
1153                                 ret = -EINVAL;
1154                                 goto done;
1155                         }
1156                         offset = 0;
1157                 } else
1158                         offset = sizeof(__u32);
1159                 break;
1160         default:
1161                 ret = -EINVAL;
1162                 goto done;
1163         }
1164
1165         iov_iter_advance(from, offset);
1166         ret = copy_from_iter(&msg, sizeof(msg), from);
1167         if (ret != sizeof(msg)) {
1168                 ret = -EINVAL;
1169                 goto done;
1170         }
1171
1172         if ((msg.type == VHOST_IOTLB_UPDATE ||
1173              msg.type == VHOST_IOTLB_INVALIDATE) &&
1174              msg.size == 0) {
1175                 ret = -EINVAL;
1176                 goto done;
1177         }
1178
1179         if (dev->msg_handler)
1180                 ret = dev->msg_handler(dev, asid, &msg);
1181         else
1182                 ret = vhost_process_iotlb_msg(dev, asid, &msg);
1183         if (ret) {
1184                 ret = -EFAULT;
1185                 goto done;
1186         }
1187
1188         ret = (type == VHOST_IOTLB_MSG) ? sizeof(struct vhost_msg) :
1189               sizeof(struct vhost_msg_v2);
1190 done:
1191         return ret;
1192 }
1193 EXPORT_SYMBOL(vhost_chr_write_iter);
1194
1195 __poll_t vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1196                             poll_table *wait)
1197 {
1198         __poll_t mask = 0;
1199
1200         poll_wait(file, &dev->wait, wait);
1201
1202         if (!list_empty(&dev->read_list))
1203                 mask |= EPOLLIN | EPOLLRDNORM;
1204
1205         return mask;
1206 }
1207 EXPORT_SYMBOL(vhost_chr_poll);
1208
1209 ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1210                             int noblock)
1211 {
1212         DEFINE_WAIT(wait);
1213         struct vhost_msg_node *node;
1214         ssize_t ret = 0;
1215         unsigned size = sizeof(struct vhost_msg);
1216
1217         if (iov_iter_count(to) < size)
1218                 return 0;
1219
1220         while (1) {
1221                 if (!noblock)
1222                         prepare_to_wait(&dev->wait, &wait,
1223                                         TASK_INTERRUPTIBLE);
1224
1225                 node = vhost_dequeue_msg(dev, &dev->read_list);
1226                 if (node)
1227                         break;
1228                 if (noblock) {
1229                         ret = -EAGAIN;
1230                         break;
1231                 }
1232                 if (signal_pending(current)) {
1233                         ret = -ERESTARTSYS;
1234                         break;
1235                 }
1236                 if (!dev->iotlb) {
1237                         ret = -EBADFD;
1238                         break;
1239                 }
1240
1241                 schedule();
1242         }
1243
1244         if (!noblock)
1245                 finish_wait(&dev->wait, &wait);
1246
1247         if (node) {
1248                 struct vhost_iotlb_msg *msg;
1249                 void *start = &node->msg;
1250
1251                 switch (node->msg.type) {
1252                 case VHOST_IOTLB_MSG:
1253                         size = sizeof(node->msg);
1254                         msg = &node->msg.iotlb;
1255                         break;
1256                 case VHOST_IOTLB_MSG_V2:
1257                         size = sizeof(node->msg_v2);
1258                         msg = &node->msg_v2.iotlb;
1259                         break;
1260                 default:
1261                         BUG();
1262                         break;
1263                 }
1264
1265                 ret = copy_to_iter(start, size, to);
1266                 if (ret != size || msg->type != VHOST_IOTLB_MISS) {
1267                         kfree(node);
1268                         return ret;
1269                 }
1270                 vhost_enqueue_msg(dev, &dev->pending_list, node);
1271         }
1272
1273         return ret;
1274 }
1275 EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1276
1277 static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1278 {
1279         struct vhost_dev *dev = vq->dev;
1280         struct vhost_msg_node *node;
1281         struct vhost_iotlb_msg *msg;
1282         bool v2 = vhost_backend_has_feature(vq, VHOST_BACKEND_F_IOTLB_MSG_V2);
1283
1284         node = vhost_new_msg(vq, v2 ? VHOST_IOTLB_MSG_V2 : VHOST_IOTLB_MSG);
1285         if (!node)
1286                 return -ENOMEM;
1287
1288         if (v2) {
1289                 node->msg_v2.type = VHOST_IOTLB_MSG_V2;
1290                 msg = &node->msg_v2.iotlb;
1291         } else {
1292                 msg = &node->msg.iotlb;
1293         }
1294
1295         msg->type = VHOST_IOTLB_MISS;
1296         msg->iova = iova;
1297         msg->perm = access;
1298
1299         vhost_enqueue_msg(dev, &dev->read_list, node);
1300
1301         return 0;
1302 }
1303
1304 static bool vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1305                          vring_desc_t __user *desc,
1306                          vring_avail_t __user *avail,
1307                          vring_used_t __user *used)
1308
1309 {
1310         /* If an IOTLB device is present, the vring addresses are
1311          * GIOVAs. Access validation occurs at prefetch time. */
1312         if (vq->iotlb)
1313                 return true;
1314
1315         return access_ok(desc, vhost_get_desc_size(vq, num)) &&
1316                access_ok(avail, vhost_get_avail_size(vq, num)) &&
1317                access_ok(used, vhost_get_used_size(vq, num));
1318 }
1319
1320 static void vhost_vq_meta_update(struct vhost_virtqueue *vq,
1321                                  const struct vhost_iotlb_map *map,
1322                                  int type)
1323 {
1324         int access = (type == VHOST_ADDR_USED) ?
1325                      VHOST_ACCESS_WO : VHOST_ACCESS_RO;
1326
1327         if (likely(map->perm & access))
1328                 vq->meta_iotlb[type] = map;
1329 }
1330
1331 static bool iotlb_access_ok(struct vhost_virtqueue *vq,
1332                             int access, u64 addr, u64 len, int type)
1333 {
1334         const struct vhost_iotlb_map *map;
1335         struct vhost_iotlb *umem = vq->iotlb;
1336         u64 s = 0, size, orig_addr = addr, last = addr + len - 1;
1337
1338         if (vhost_vq_meta_fetch(vq, addr, len, type))
1339                 return true;
1340
1341         while (len > s) {
1342                 map = vhost_iotlb_itree_first(umem, addr, last);
1343                 if (map == NULL || map->start > addr) {
1344                         vhost_iotlb_miss(vq, addr, access);
1345                         return false;
1346                 } else if (!(map->perm & access)) {
1347                         /* Report the possible access violation by
1348                          * request another translation from userspace.
1349                          */
1350                         return false;
1351                 }
1352
1353                 size = map->size - addr + map->start;
1354
1355                 if (orig_addr == addr && size >= len)
1356                         vhost_vq_meta_update(vq, map, type);
1357
1358                 s += size;
1359                 addr += size;
1360         }
1361
1362         return true;
1363 }
1364
1365 int vq_meta_prefetch(struct vhost_virtqueue *vq)
1366 {
1367         unsigned int num = vq->num;
1368
1369         if (!vq->iotlb)
1370                 return 1;
1371
1372         return iotlb_access_ok(vq, VHOST_MAP_RO, (u64)(uintptr_t)vq->desc,
1373                                vhost_get_desc_size(vq, num), VHOST_ADDR_DESC) &&
1374                iotlb_access_ok(vq, VHOST_MAP_RO, (u64)(uintptr_t)vq->avail,
1375                                vhost_get_avail_size(vq, num),
1376                                VHOST_ADDR_AVAIL) &&
1377                iotlb_access_ok(vq, VHOST_MAP_WO, (u64)(uintptr_t)vq->used,
1378                                vhost_get_used_size(vq, num), VHOST_ADDR_USED);
1379 }
1380 EXPORT_SYMBOL_GPL(vq_meta_prefetch);
1381
1382 /* Can we log writes? */
1383 /* Caller should have device mutex but not vq mutex */
1384 bool vhost_log_access_ok(struct vhost_dev *dev)
1385 {
1386         return memory_access_ok(dev, dev->umem, 1);
1387 }
1388 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1389
1390 static bool vq_log_used_access_ok(struct vhost_virtqueue *vq,
1391                                   void __user *log_base,
1392                                   bool log_used,
1393                                   u64 log_addr)
1394 {
1395         /* If an IOTLB device is present, log_addr is a GIOVA that
1396          * will never be logged by log_used(). */
1397         if (vq->iotlb)
1398                 return true;
1399
1400         return !log_used || log_access_ok(log_base, log_addr,
1401                                           vhost_get_used_size(vq, vq->num));
1402 }
1403
1404 /* Verify access for write logging. */
1405 /* Caller should have vq mutex and device mutex */
1406 static bool vq_log_access_ok(struct vhost_virtqueue *vq,
1407                              void __user *log_base)
1408 {
1409         return vq_memory_access_ok(log_base, vq->umem,
1410                                    vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1411                 vq_log_used_access_ok(vq, log_base, vq->log_used, vq->log_addr);
1412 }
1413
1414 /* Can we start vq? */
1415 /* Caller should have vq mutex and device mutex */
1416 bool vhost_vq_access_ok(struct vhost_virtqueue *vq)
1417 {
1418         if (!vq_log_access_ok(vq, vq->log_base))
1419                 return false;
1420
1421         return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used);
1422 }
1423 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1424
1425 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1426 {
1427         struct vhost_memory mem, *newmem;
1428         struct vhost_memory_region *region;
1429         struct vhost_iotlb *newumem, *oldumem;
1430         unsigned long size = offsetof(struct vhost_memory, regions);
1431         int i;
1432
1433         if (copy_from_user(&mem, m, size))
1434                 return -EFAULT;
1435         if (mem.padding)
1436                 return -EOPNOTSUPP;
1437         if (mem.nregions > max_mem_regions)
1438                 return -E2BIG;
1439         newmem = kvzalloc(struct_size(newmem, regions, mem.nregions),
1440                         GFP_KERNEL);
1441         if (!newmem)
1442                 return -ENOMEM;
1443
1444         memcpy(newmem, &mem, size);
1445         if (copy_from_user(newmem->regions, m->regions,
1446                            flex_array_size(newmem, regions, mem.nregions))) {
1447                 kvfree(newmem);
1448                 return -EFAULT;
1449         }
1450
1451         newumem = iotlb_alloc();
1452         if (!newumem) {
1453                 kvfree(newmem);
1454                 return -ENOMEM;
1455         }
1456
1457         for (region = newmem->regions;
1458              region < newmem->regions + mem.nregions;
1459              region++) {
1460                 if (vhost_iotlb_add_range(newumem,
1461                                           region->guest_phys_addr,
1462                                           region->guest_phys_addr +
1463                                           region->memory_size - 1,
1464                                           region->userspace_addr,
1465                                           VHOST_MAP_RW))
1466                         goto err;
1467         }
1468
1469         if (!memory_access_ok(d, newumem, 0))
1470                 goto err;
1471
1472         oldumem = d->umem;
1473         d->umem = newumem;
1474
1475         /* All memory accesses are done under some VQ mutex. */
1476         for (i = 0; i < d->nvqs; ++i) {
1477                 mutex_lock(&d->vqs[i]->mutex);
1478                 d->vqs[i]->umem = newumem;
1479                 mutex_unlock(&d->vqs[i]->mutex);
1480         }
1481
1482         kvfree(newmem);
1483         vhost_iotlb_free(oldumem);
1484         return 0;
1485
1486 err:
1487         vhost_iotlb_free(newumem);
1488         kvfree(newmem);
1489         return -EFAULT;
1490 }
1491
1492 static long vhost_vring_set_num(struct vhost_dev *d,
1493                                 struct vhost_virtqueue *vq,
1494                                 void __user *argp)
1495 {
1496         struct vhost_vring_state s;
1497
1498         /* Resizing ring with an active backend?
1499          * You don't want to do that. */
1500         if (vq->private_data)
1501                 return -EBUSY;
1502
1503         if (copy_from_user(&s, argp, sizeof s))
1504                 return -EFAULT;
1505
1506         if (!s.num || s.num > 0xffff || (s.num & (s.num - 1)))
1507                 return -EINVAL;
1508         vq->num = s.num;
1509
1510         return 0;
1511 }
1512
1513 static long vhost_vring_set_addr(struct vhost_dev *d,
1514                                  struct vhost_virtqueue *vq,
1515                                  void __user *argp)
1516 {
1517         struct vhost_vring_addr a;
1518
1519         if (copy_from_user(&a, argp, sizeof a))
1520                 return -EFAULT;
1521         if (a.flags & ~(0x1 << VHOST_VRING_F_LOG))
1522                 return -EOPNOTSUPP;
1523
1524         /* For 32bit, verify that the top 32bits of the user
1525            data are set to zero. */
1526         if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1527             (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1528             (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr)
1529                 return -EFAULT;
1530
1531         /* Make sure it's safe to cast pointers to vring types. */
1532         BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1533         BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1534         if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1535             (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1536             (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1)))
1537                 return -EINVAL;
1538
1539         /* We only verify access here if backend is configured.
1540          * If it is not, we don't as size might not have been setup.
1541          * We will verify when backend is configured. */
1542         if (vq->private_data) {
1543                 if (!vq_access_ok(vq, vq->num,
1544                         (void __user *)(unsigned long)a.desc_user_addr,
1545                         (void __user *)(unsigned long)a.avail_user_addr,
1546                         (void __user *)(unsigned long)a.used_user_addr))
1547                         return -EINVAL;
1548
1549                 /* Also validate log access for used ring if enabled. */
1550                 if (!vq_log_used_access_ok(vq, vq->log_base,
1551                                 a.flags & (0x1 << VHOST_VRING_F_LOG),
1552                                 a.log_guest_addr))
1553                         return -EINVAL;
1554         }
1555
1556         vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1557         vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1558         vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1559         vq->log_addr = a.log_guest_addr;
1560         vq->used = (void __user *)(unsigned long)a.used_user_addr;
1561
1562         return 0;
1563 }
1564
1565 static long vhost_vring_set_num_addr(struct vhost_dev *d,
1566                                      struct vhost_virtqueue *vq,
1567                                      unsigned int ioctl,
1568                                      void __user *argp)
1569 {
1570         long r;
1571
1572         mutex_lock(&vq->mutex);
1573
1574         switch (ioctl) {
1575         case VHOST_SET_VRING_NUM:
1576                 r = vhost_vring_set_num(d, vq, argp);
1577                 break;
1578         case VHOST_SET_VRING_ADDR:
1579                 r = vhost_vring_set_addr(d, vq, argp);
1580                 break;
1581         default:
1582                 BUG();
1583         }
1584
1585         mutex_unlock(&vq->mutex);
1586
1587         return r;
1588 }
1589 long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1590 {
1591         struct file *eventfp, *filep = NULL;
1592         bool pollstart = false, pollstop = false;
1593         struct eventfd_ctx *ctx = NULL;
1594         u32 __user *idxp = argp;
1595         struct vhost_virtqueue *vq;
1596         struct vhost_vring_state s;
1597         struct vhost_vring_file f;
1598         u32 idx;
1599         long r;
1600
1601         r = get_user(idx, idxp);
1602         if (r < 0)
1603                 return r;
1604         if (idx >= d->nvqs)
1605                 return -ENOBUFS;
1606
1607         idx = array_index_nospec(idx, d->nvqs);
1608         vq = d->vqs[idx];
1609
1610         if (ioctl == VHOST_SET_VRING_NUM ||
1611             ioctl == VHOST_SET_VRING_ADDR) {
1612                 return vhost_vring_set_num_addr(d, vq, ioctl, argp);
1613         }
1614
1615         mutex_lock(&vq->mutex);
1616
1617         switch (ioctl) {
1618         case VHOST_SET_VRING_BASE:
1619                 /* Moving base with an active backend?
1620                  * You don't want to do that. */
1621                 if (vq->private_data) {
1622                         r = -EBUSY;
1623                         break;
1624                 }
1625                 if (copy_from_user(&s, argp, sizeof s)) {
1626                         r = -EFAULT;
1627                         break;
1628                 }
1629                 if (s.num > 0xffff) {
1630                         r = -EINVAL;
1631                         break;
1632                 }
1633                 vq->last_avail_idx = s.num;
1634                 /* Forget the cached index value. */
1635                 vq->avail_idx = vq->last_avail_idx;
1636                 break;
1637         case VHOST_GET_VRING_BASE:
1638                 s.index = idx;
1639                 s.num = vq->last_avail_idx;
1640                 if (copy_to_user(argp, &s, sizeof s))
1641                         r = -EFAULT;
1642                 break;
1643         case VHOST_SET_VRING_KICK:
1644                 if (copy_from_user(&f, argp, sizeof f)) {
1645                         r = -EFAULT;
1646                         break;
1647                 }
1648                 eventfp = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_fget(f.fd);
1649                 if (IS_ERR(eventfp)) {
1650                         r = PTR_ERR(eventfp);
1651                         break;
1652                 }
1653                 if (eventfp != vq->kick) {
1654                         pollstop = (filep = vq->kick) != NULL;
1655                         pollstart = (vq->kick = eventfp) != NULL;
1656                 } else
1657                         filep = eventfp;
1658                 break;
1659         case VHOST_SET_VRING_CALL:
1660                 if (copy_from_user(&f, argp, sizeof f)) {
1661                         r = -EFAULT;
1662                         break;
1663                 }
1664                 ctx = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(f.fd);
1665                 if (IS_ERR(ctx)) {
1666                         r = PTR_ERR(ctx);
1667                         break;
1668                 }
1669
1670                 swap(ctx, vq->call_ctx.ctx);
1671                 break;
1672         case VHOST_SET_VRING_ERR:
1673                 if (copy_from_user(&f, argp, sizeof f)) {
1674                         r = -EFAULT;
1675                         break;
1676                 }
1677                 ctx = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(f.fd);
1678                 if (IS_ERR(ctx)) {
1679                         r = PTR_ERR(ctx);
1680                         break;
1681                 }
1682                 swap(ctx, vq->error_ctx);
1683                 break;
1684         case VHOST_SET_VRING_ENDIAN:
1685                 r = vhost_set_vring_endian(vq, argp);
1686                 break;
1687         case VHOST_GET_VRING_ENDIAN:
1688                 r = vhost_get_vring_endian(vq, idx, argp);
1689                 break;
1690         case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
1691                 if (copy_from_user(&s, argp, sizeof(s))) {
1692                         r = -EFAULT;
1693                         break;
1694                 }
1695                 vq->busyloop_timeout = s.num;
1696                 break;
1697         case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
1698                 s.index = idx;
1699                 s.num = vq->busyloop_timeout;
1700                 if (copy_to_user(argp, &s, sizeof(s)))
1701                         r = -EFAULT;
1702                 break;
1703         default:
1704                 r = -ENOIOCTLCMD;
1705         }
1706
1707         if (pollstop && vq->handle_kick)
1708                 vhost_poll_stop(&vq->poll);
1709
1710         if (!IS_ERR_OR_NULL(ctx))
1711                 eventfd_ctx_put(ctx);
1712         if (filep)
1713                 fput(filep);
1714
1715         if (pollstart && vq->handle_kick)
1716                 r = vhost_poll_start(&vq->poll, vq->kick);
1717
1718         mutex_unlock(&vq->mutex);
1719
1720         if (pollstop && vq->handle_kick)
1721                 vhost_dev_flush(vq->poll.dev);
1722         return r;
1723 }
1724 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
1725
1726 int vhost_init_device_iotlb(struct vhost_dev *d)
1727 {
1728         struct vhost_iotlb *niotlb, *oiotlb;
1729         int i;
1730
1731         niotlb = iotlb_alloc();
1732         if (!niotlb)
1733                 return -ENOMEM;
1734
1735         oiotlb = d->iotlb;
1736         d->iotlb = niotlb;
1737
1738         for (i = 0; i < d->nvqs; ++i) {
1739                 struct vhost_virtqueue *vq = d->vqs[i];
1740
1741                 mutex_lock(&vq->mutex);
1742                 vq->iotlb = niotlb;
1743                 __vhost_vq_meta_reset(vq);
1744                 mutex_unlock(&vq->mutex);
1745         }
1746
1747         vhost_iotlb_free(oiotlb);
1748
1749         return 0;
1750 }
1751 EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
1752
1753 /* Caller must have device mutex */
1754 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1755 {
1756         struct eventfd_ctx *ctx;
1757         u64 p;
1758         long r;
1759         int i, fd;
1760
1761         /* If you are not the owner, you can become one */
1762         if (ioctl == VHOST_SET_OWNER) {
1763                 r = vhost_dev_set_owner(d);
1764                 goto done;
1765         }
1766
1767         /* You must be the owner to do anything else */
1768         r = vhost_dev_check_owner(d);
1769         if (r)
1770                 goto done;
1771
1772         switch (ioctl) {
1773         case VHOST_SET_MEM_TABLE:
1774                 r = vhost_set_memory(d, argp);
1775                 break;
1776         case VHOST_SET_LOG_BASE:
1777                 if (copy_from_user(&p, argp, sizeof p)) {
1778                         r = -EFAULT;
1779                         break;
1780                 }
1781                 if ((u64)(unsigned long)p != p) {
1782                         r = -EFAULT;
1783                         break;
1784                 }
1785                 for (i = 0; i < d->nvqs; ++i) {
1786                         struct vhost_virtqueue *vq;
1787                         void __user *base = (void __user *)(unsigned long)p;
1788                         vq = d->vqs[i];
1789                         mutex_lock(&vq->mutex);
1790                         /* If ring is inactive, will check when it's enabled. */
1791                         if (vq->private_data && !vq_log_access_ok(vq, base))
1792                                 r = -EFAULT;
1793                         else
1794                                 vq->log_base = base;
1795                         mutex_unlock(&vq->mutex);
1796                 }
1797                 break;
1798         case VHOST_SET_LOG_FD:
1799                 r = get_user(fd, (int __user *)argp);
1800                 if (r < 0)
1801                         break;
1802                 ctx = fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(fd);
1803                 if (IS_ERR(ctx)) {
1804                         r = PTR_ERR(ctx);
1805                         break;
1806                 }
1807                 swap(ctx, d->log_ctx);
1808                 for (i = 0; i < d->nvqs; ++i) {
1809                         mutex_lock(&d->vqs[i]->mutex);
1810                         d->vqs[i]->log_ctx = d->log_ctx;
1811                         mutex_unlock(&d->vqs[i]->mutex);
1812                 }
1813                 if (ctx)
1814                         eventfd_ctx_put(ctx);
1815                 break;
1816         default:
1817                 r = -ENOIOCTLCMD;
1818                 break;
1819         }
1820 done:
1821         return r;
1822 }
1823 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1824
1825 /* TODO: This is really inefficient.  We need something like get_user()
1826  * (instruction directly accesses the data, with an exception table entry
1827  * returning -EFAULT). See Documentation/arch/x86/exception-tables.rst.
1828  */
1829 static int set_bit_to_user(int nr, void __user *addr)
1830 {
1831         unsigned long log = (unsigned long)addr;
1832         struct page *page;
1833         void *base;
1834         int bit = nr + (log % PAGE_SIZE) * 8;
1835         int r;
1836
1837         r = pin_user_pages_fast(log, 1, FOLL_WRITE, &page);
1838         if (r < 0)
1839                 return r;
1840         BUG_ON(r != 1);
1841         base = kmap_atomic(page);
1842         set_bit(bit, base);
1843         kunmap_atomic(base);
1844         unpin_user_pages_dirty_lock(&page, 1, true);
1845         return 0;
1846 }
1847
1848 static int log_write(void __user *log_base,
1849                      u64 write_address, u64 write_length)
1850 {
1851         u64 write_page = write_address / VHOST_PAGE_SIZE;
1852         int r;
1853
1854         if (!write_length)
1855                 return 0;
1856         write_length += write_address % VHOST_PAGE_SIZE;
1857         for (;;) {
1858                 u64 base = (u64)(unsigned long)log_base;
1859                 u64 log = base + write_page / 8;
1860                 int bit = write_page % 8;
1861                 if ((u64)(unsigned long)log != log)
1862                         return -EFAULT;
1863                 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1864                 if (r < 0)
1865                         return r;
1866                 if (write_length <= VHOST_PAGE_SIZE)
1867                         break;
1868                 write_length -= VHOST_PAGE_SIZE;
1869                 write_page += 1;
1870         }
1871         return r;
1872 }
1873
1874 static int log_write_hva(struct vhost_virtqueue *vq, u64 hva, u64 len)
1875 {
1876         struct vhost_iotlb *umem = vq->umem;
1877         struct vhost_iotlb_map *u;
1878         u64 start, end, l, min;
1879         int r;
1880         bool hit = false;
1881
1882         while (len) {
1883                 min = len;
1884                 /* More than one GPAs can be mapped into a single HVA. So
1885                  * iterate all possible umems here to be safe.
1886                  */
1887                 list_for_each_entry(u, &umem->list, link) {
1888                         if (u->addr > hva - 1 + len ||
1889                             u->addr - 1 + u->size < hva)
1890                                 continue;
1891                         start = max(u->addr, hva);
1892                         end = min(u->addr - 1 + u->size, hva - 1 + len);
1893                         l = end - start + 1;
1894                         r = log_write(vq->log_base,
1895                                       u->start + start - u->addr,
1896                                       l);
1897                         if (r < 0)
1898                                 return r;
1899                         hit = true;
1900                         min = min(l, min);
1901                 }
1902
1903                 if (!hit)
1904                         return -EFAULT;
1905
1906                 len -= min;
1907                 hva += min;
1908         }
1909
1910         return 0;
1911 }
1912
1913 static int log_used(struct vhost_virtqueue *vq, u64 used_offset, u64 len)
1914 {
1915         struct iovec *iov = vq->log_iov;
1916         int i, ret;
1917
1918         if (!vq->iotlb)
1919                 return log_write(vq->log_base, vq->log_addr + used_offset, len);
1920
1921         ret = translate_desc(vq, (uintptr_t)vq->used + used_offset,
1922                              len, iov, 64, VHOST_ACCESS_WO);
1923         if (ret < 0)
1924                 return ret;
1925
1926         for (i = 0; i < ret; i++) {
1927                 ret = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1928                                     iov[i].iov_len);
1929                 if (ret)
1930                         return ret;
1931         }
1932
1933         return 0;
1934 }
1935
1936 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1937                     unsigned int log_num, u64 len, struct iovec *iov, int count)
1938 {
1939         int i, r;
1940
1941         /* Make sure data written is seen before log. */
1942         smp_wmb();
1943
1944         if (vq->iotlb) {
1945                 for (i = 0; i < count; i++) {
1946                         r = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1947                                           iov[i].iov_len);
1948                         if (r < 0)
1949                                 return r;
1950                 }
1951                 return 0;
1952         }
1953
1954         for (i = 0; i < log_num; ++i) {
1955                 u64 l = min(log[i].len, len);
1956                 r = log_write(vq->log_base, log[i].addr, l);
1957                 if (r < 0)
1958                         return r;
1959                 len -= l;
1960                 if (!len) {
1961                         if (vq->log_ctx)
1962                                 eventfd_signal(vq->log_ctx, 1);
1963                         return 0;
1964                 }
1965         }
1966         /* Length written exceeds what we have stored. This is a bug. */
1967         BUG();
1968         return 0;
1969 }
1970 EXPORT_SYMBOL_GPL(vhost_log_write);
1971
1972 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1973 {
1974         void __user *used;
1975         if (vhost_put_used_flags(vq))
1976                 return -EFAULT;
1977         if (unlikely(vq->log_used)) {
1978                 /* Make sure the flag is seen before log. */
1979                 smp_wmb();
1980                 /* Log used flag write. */
1981                 used = &vq->used->flags;
1982                 log_used(vq, (used - (void __user *)vq->used),
1983                          sizeof vq->used->flags);
1984                 if (vq->log_ctx)
1985                         eventfd_signal(vq->log_ctx, 1);
1986         }
1987         return 0;
1988 }
1989
1990 static int vhost_update_avail_event(struct vhost_virtqueue *vq)
1991 {
1992         if (vhost_put_avail_event(vq))
1993                 return -EFAULT;
1994         if (unlikely(vq->log_used)) {
1995                 void __user *used;
1996                 /* Make sure the event is seen before log. */
1997                 smp_wmb();
1998                 /* Log avail event write */
1999                 used = vhost_avail_event(vq);
2000                 log_used(vq, (used - (void __user *)vq->used),
2001                          sizeof *vhost_avail_event(vq));
2002                 if (vq->log_ctx)
2003                         eventfd_signal(vq->log_ctx, 1);
2004         }
2005         return 0;
2006 }
2007
2008 int vhost_vq_init_access(struct vhost_virtqueue *vq)
2009 {
2010         __virtio16 last_used_idx;
2011         int r;
2012         bool is_le = vq->is_le;
2013
2014         if (!vq->private_data)
2015                 return 0;
2016
2017         vhost_init_is_le(vq);
2018
2019         r = vhost_update_used_flags(vq);
2020         if (r)
2021                 goto err;
2022         vq->signalled_used_valid = false;
2023         if (!vq->iotlb &&
2024             !access_ok(&vq->used->idx, sizeof vq->used->idx)) {
2025                 r = -EFAULT;
2026                 goto err;
2027         }
2028         r = vhost_get_used_idx(vq, &last_used_idx);
2029         if (r) {
2030                 vq_err(vq, "Can't access used idx at %p\n",
2031                        &vq->used->idx);
2032                 goto err;
2033         }
2034         vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
2035         return 0;
2036
2037 err:
2038         vq->is_le = is_le;
2039         return r;
2040 }
2041 EXPORT_SYMBOL_GPL(vhost_vq_init_access);
2042
2043 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
2044                           struct iovec iov[], int iov_size, int access)
2045 {
2046         const struct vhost_iotlb_map *map;
2047         struct vhost_dev *dev = vq->dev;
2048         struct vhost_iotlb *umem = dev->iotlb ? dev->iotlb : dev->umem;
2049         struct iovec *_iov;
2050         u64 s = 0, last = addr + len - 1;
2051         int ret = 0;
2052
2053         while ((u64)len > s) {
2054                 u64 size;
2055                 if (unlikely(ret >= iov_size)) {
2056                         ret = -ENOBUFS;
2057                         break;
2058                 }
2059
2060                 map = vhost_iotlb_itree_first(umem, addr, last);
2061                 if (map == NULL || map->start > addr) {
2062                         if (umem != dev->iotlb) {
2063                                 ret = -EFAULT;
2064                                 break;
2065                         }
2066                         ret = -EAGAIN;
2067                         break;
2068                 } else if (!(map->perm & access)) {
2069                         ret = -EPERM;
2070                         break;
2071                 }
2072
2073                 _iov = iov + ret;
2074                 size = map->size - addr + map->start;
2075                 _iov->iov_len = min((u64)len - s, size);
2076                 _iov->iov_base = (void __user *)(unsigned long)
2077                                  (map->addr + addr - map->start);
2078                 s += size;
2079                 addr += size;
2080                 ++ret;
2081         }
2082
2083         if (ret == -EAGAIN)
2084                 vhost_iotlb_miss(vq, addr, access);
2085         return ret;
2086 }
2087
2088 /* Each buffer in the virtqueues is actually a chain of descriptors.  This
2089  * function returns the next descriptor in the chain,
2090  * or -1U if we're at the end. */
2091 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
2092 {
2093         unsigned int next;
2094
2095         /* If this descriptor says it doesn't chain, we're done. */
2096         if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
2097                 return -1U;
2098
2099         /* Check they're not leading us off end of descriptors. */
2100         next = vhost16_to_cpu(vq, READ_ONCE(desc->next));
2101         return next;
2102 }
2103
2104 static int get_indirect(struct vhost_virtqueue *vq,
2105                         struct iovec iov[], unsigned int iov_size,
2106                         unsigned int *out_num, unsigned int *in_num,
2107                         struct vhost_log *log, unsigned int *log_num,
2108                         struct vring_desc *indirect)
2109 {
2110         struct vring_desc desc;
2111         unsigned int i = 0, count, found = 0;
2112         u32 len = vhost32_to_cpu(vq, indirect->len);
2113         struct iov_iter from;
2114         int ret, access;
2115
2116         /* Sanity check */
2117         if (unlikely(len % sizeof desc)) {
2118                 vq_err(vq, "Invalid length in indirect descriptor: "
2119                        "len 0x%llx not multiple of 0x%zx\n",
2120                        (unsigned long long)len,
2121                        sizeof desc);
2122                 return -EINVAL;
2123         }
2124
2125         ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
2126                              UIO_MAXIOV, VHOST_ACCESS_RO);
2127         if (unlikely(ret < 0)) {
2128                 if (ret != -EAGAIN)
2129                         vq_err(vq, "Translation failure %d in indirect.\n", ret);
2130                 return ret;
2131         }
2132         iov_iter_init(&from, ITER_SOURCE, vq->indirect, ret, len);
2133         count = len / sizeof desc;
2134         /* Buffers are chained via a 16 bit next field, so
2135          * we can have at most 2^16 of these. */
2136         if (unlikely(count > USHRT_MAX + 1)) {
2137                 vq_err(vq, "Indirect buffer length too big: %d\n",
2138                        indirect->len);
2139                 return -E2BIG;
2140         }
2141
2142         do {
2143                 unsigned iov_count = *in_num + *out_num;
2144                 if (unlikely(++found > count)) {
2145                         vq_err(vq, "Loop detected: last one at %u "
2146                                "indirect size %u\n",
2147                                i, count);
2148                         return -EINVAL;
2149                 }
2150                 if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
2151                         vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
2152                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2153                         return -EINVAL;
2154                 }
2155                 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
2156                         vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
2157                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2158                         return -EINVAL;
2159                 }
2160
2161                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2162                         access = VHOST_ACCESS_WO;
2163                 else
2164                         access = VHOST_ACCESS_RO;
2165
2166                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2167                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
2168                                      iov_size - iov_count, access);
2169                 if (unlikely(ret < 0)) {
2170                         if (ret != -EAGAIN)
2171                                 vq_err(vq, "Translation failure %d indirect idx %d\n",
2172                                         ret, i);
2173                         return ret;
2174                 }
2175                 /* If this is an input descriptor, increment that count. */
2176                 if (access == VHOST_ACCESS_WO) {
2177                         *in_num += ret;
2178                         if (unlikely(log && ret)) {
2179                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2180                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2181                                 ++*log_num;
2182                         }
2183                 } else {
2184                         /* If it's an output descriptor, they're all supposed
2185                          * to come before any input descriptors. */
2186                         if (unlikely(*in_num)) {
2187                                 vq_err(vq, "Indirect descriptor "
2188                                        "has out after in: idx %d\n", i);
2189                                 return -EINVAL;
2190                         }
2191                         *out_num += ret;
2192                 }
2193         } while ((i = next_desc(vq, &desc)) != -1);
2194         return 0;
2195 }
2196
2197 /* This looks in the virtqueue and for the first available buffer, and converts
2198  * it to an iovec for convenient access.  Since descriptors consist of some
2199  * number of output then some number of input descriptors, it's actually two
2200  * iovecs, but we pack them into one and note how many of each there were.
2201  *
2202  * This function returns the descriptor number found, or vq->num (which is
2203  * never a valid descriptor number) if none was found.  A negative code is
2204  * returned on error. */
2205 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
2206                       struct iovec iov[], unsigned int iov_size,
2207                       unsigned int *out_num, unsigned int *in_num,
2208                       struct vhost_log *log, unsigned int *log_num)
2209 {
2210         struct vring_desc desc;
2211         unsigned int i, head, found = 0;
2212         u16 last_avail_idx;
2213         __virtio16 avail_idx;
2214         __virtio16 ring_head;
2215         int ret, access;
2216
2217         /* Check it isn't doing very strange things with descriptor numbers. */
2218         last_avail_idx = vq->last_avail_idx;
2219
2220         if (vq->avail_idx == vq->last_avail_idx) {
2221                 if (unlikely(vhost_get_avail_idx(vq, &avail_idx))) {
2222                         vq_err(vq, "Failed to access avail idx at %p\n",
2223                                 &vq->avail->idx);
2224                         return -EFAULT;
2225                 }
2226                 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2227
2228                 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
2229                         vq_err(vq, "Guest moved used index from %u to %u",
2230                                 last_avail_idx, vq->avail_idx);
2231                         return -EFAULT;
2232                 }
2233
2234                 /* If there's nothing new since last we looked, return
2235                  * invalid.
2236                  */
2237                 if (vq->avail_idx == last_avail_idx)
2238                         return vq->num;
2239
2240                 /* Only get avail ring entries after they have been
2241                  * exposed by guest.
2242                  */
2243                 smp_rmb();
2244         }
2245
2246         /* Grab the next descriptor number they're advertising, and increment
2247          * the index we've seen. */
2248         if (unlikely(vhost_get_avail_head(vq, &ring_head, last_avail_idx))) {
2249                 vq_err(vq, "Failed to read head: idx %d address %p\n",
2250                        last_avail_idx,
2251                        &vq->avail->ring[last_avail_idx % vq->num]);
2252                 return -EFAULT;
2253         }
2254
2255         head = vhost16_to_cpu(vq, ring_head);
2256
2257         /* If their number is silly, that's an error. */
2258         if (unlikely(head >= vq->num)) {
2259                 vq_err(vq, "Guest says index %u > %u is available",
2260                        head, vq->num);
2261                 return -EINVAL;
2262         }
2263
2264         /* When we start there are none of either input nor output. */
2265         *out_num = *in_num = 0;
2266         if (unlikely(log))
2267                 *log_num = 0;
2268
2269         i = head;
2270         do {
2271                 unsigned iov_count = *in_num + *out_num;
2272                 if (unlikely(i >= vq->num)) {
2273                         vq_err(vq, "Desc index is %u > %u, head = %u",
2274                                i, vq->num, head);
2275                         return -EINVAL;
2276                 }
2277                 if (unlikely(++found > vq->num)) {
2278                         vq_err(vq, "Loop detected: last one at %u "
2279                                "vq size %u head %u\n",
2280                                i, vq->num, head);
2281                         return -EINVAL;
2282                 }
2283                 ret = vhost_get_desc(vq, &desc, i);
2284                 if (unlikely(ret)) {
2285                         vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
2286                                i, vq->desc + i);
2287                         return -EFAULT;
2288                 }
2289                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2290                         ret = get_indirect(vq, iov, iov_size,
2291                                            out_num, in_num,
2292                                            log, log_num, &desc);
2293                         if (unlikely(ret < 0)) {
2294                                 if (ret != -EAGAIN)
2295                                         vq_err(vq, "Failure detected "
2296                                                 "in indirect descriptor at idx %d\n", i);
2297                                 return ret;
2298                         }
2299                         continue;
2300                 }
2301
2302                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2303                         access = VHOST_ACCESS_WO;
2304                 else
2305                         access = VHOST_ACCESS_RO;
2306                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2307                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
2308                                      iov_size - iov_count, access);
2309                 if (unlikely(ret < 0)) {
2310                         if (ret != -EAGAIN)
2311                                 vq_err(vq, "Translation failure %d descriptor idx %d\n",
2312                                         ret, i);
2313                         return ret;
2314                 }
2315                 if (access == VHOST_ACCESS_WO) {
2316                         /* If this is an input descriptor,
2317                          * increment that count. */
2318                         *in_num += ret;
2319                         if (unlikely(log && ret)) {
2320                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2321                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2322                                 ++*log_num;
2323                         }
2324                 } else {
2325                         /* If it's an output descriptor, they're all supposed
2326                          * to come before any input descriptors. */
2327                         if (unlikely(*in_num)) {
2328                                 vq_err(vq, "Descriptor has out after in: "
2329                                        "idx %d\n", i);
2330                                 return -EINVAL;
2331                         }
2332                         *out_num += ret;
2333                 }
2334         } while ((i = next_desc(vq, &desc)) != -1);
2335
2336         /* On success, increment avail index. */
2337         vq->last_avail_idx++;
2338
2339         /* Assume notifications from guest are disabled at this point,
2340          * if they aren't we would need to update avail_event index. */
2341         BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2342         return head;
2343 }
2344 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2345
2346 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2347 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2348 {
2349         vq->last_avail_idx -= n;
2350 }
2351 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2352
2353 /* After we've used one of their buffers, we tell them about it.  We'll then
2354  * want to notify the guest, using eventfd. */
2355 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2356 {
2357         struct vring_used_elem heads = {
2358                 cpu_to_vhost32(vq, head),
2359                 cpu_to_vhost32(vq, len)
2360         };
2361
2362         return vhost_add_used_n(vq, &heads, 1);
2363 }
2364 EXPORT_SYMBOL_GPL(vhost_add_used);
2365
2366 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2367                             struct vring_used_elem *heads,
2368                             unsigned count)
2369 {
2370         vring_used_elem_t __user *used;
2371         u16 old, new;
2372         int start;
2373
2374         start = vq->last_used_idx & (vq->num - 1);
2375         used = vq->used->ring + start;
2376         if (vhost_put_used(vq, heads, start, count)) {
2377                 vq_err(vq, "Failed to write used");
2378                 return -EFAULT;
2379         }
2380         if (unlikely(vq->log_used)) {
2381                 /* Make sure data is seen before log. */
2382                 smp_wmb();
2383                 /* Log used ring entry write. */
2384                 log_used(vq, ((void __user *)used - (void __user *)vq->used),
2385                          count * sizeof *used);
2386         }
2387         old = vq->last_used_idx;
2388         new = (vq->last_used_idx += count);
2389         /* If the driver never bothers to signal in a very long while,
2390          * used index might wrap around. If that happens, invalidate
2391          * signalled_used index we stored. TODO: make sure driver
2392          * signals at least once in 2^16 and remove this. */
2393         if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2394                 vq->signalled_used_valid = false;
2395         return 0;
2396 }
2397
2398 /* After we've used one of their buffers, we tell them about it.  We'll then
2399  * want to notify the guest, using eventfd. */
2400 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2401                      unsigned count)
2402 {
2403         int start, n, r;
2404
2405         start = vq->last_used_idx & (vq->num - 1);
2406         n = vq->num - start;
2407         if (n < count) {
2408                 r = __vhost_add_used_n(vq, heads, n);
2409                 if (r < 0)
2410                         return r;
2411                 heads += n;
2412                 count -= n;
2413         }
2414         r = __vhost_add_used_n(vq, heads, count);
2415
2416         /* Make sure buffer is written before we update index. */
2417         smp_wmb();
2418         if (vhost_put_used_idx(vq)) {
2419                 vq_err(vq, "Failed to increment used idx");
2420                 return -EFAULT;
2421         }
2422         if (unlikely(vq->log_used)) {
2423                 /* Make sure used idx is seen before log. */
2424                 smp_wmb();
2425                 /* Log used index update. */
2426                 log_used(vq, offsetof(struct vring_used, idx),
2427                          sizeof vq->used->idx);
2428                 if (vq->log_ctx)
2429                         eventfd_signal(vq->log_ctx, 1);
2430         }
2431         return r;
2432 }
2433 EXPORT_SYMBOL_GPL(vhost_add_used_n);
2434
2435 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2436 {
2437         __u16 old, new;
2438         __virtio16 event;
2439         bool v;
2440         /* Flush out used index updates. This is paired
2441          * with the barrier that the Guest executes when enabling
2442          * interrupts. */
2443         smp_mb();
2444
2445         if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2446             unlikely(vq->avail_idx == vq->last_avail_idx))
2447                 return true;
2448
2449         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2450                 __virtio16 flags;
2451                 if (vhost_get_avail_flags(vq, &flags)) {
2452                         vq_err(vq, "Failed to get flags");
2453                         return true;
2454                 }
2455                 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2456         }
2457         old = vq->signalled_used;
2458         v = vq->signalled_used_valid;
2459         new = vq->signalled_used = vq->last_used_idx;
2460         vq->signalled_used_valid = true;
2461
2462         if (unlikely(!v))
2463                 return true;
2464
2465         if (vhost_get_used_event(vq, &event)) {
2466                 vq_err(vq, "Failed to get used event idx");
2467                 return true;
2468         }
2469         return vring_need_event(vhost16_to_cpu(vq, event), new, old);
2470 }
2471
2472 /* This actually signals the guest, using eventfd. */
2473 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2474 {
2475         /* Signal the Guest tell them we used something up. */
2476         if (vq->call_ctx.ctx && vhost_notify(dev, vq))
2477                 eventfd_signal(vq->call_ctx.ctx, 1);
2478 }
2479 EXPORT_SYMBOL_GPL(vhost_signal);
2480
2481 /* And here's the combo meal deal.  Supersize me! */
2482 void vhost_add_used_and_signal(struct vhost_dev *dev,
2483                                struct vhost_virtqueue *vq,
2484                                unsigned int head, int len)
2485 {
2486         vhost_add_used(vq, head, len);
2487         vhost_signal(dev, vq);
2488 }
2489 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2490
2491 /* multi-buffer version of vhost_add_used_and_signal */
2492 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2493                                  struct vhost_virtqueue *vq,
2494                                  struct vring_used_elem *heads, unsigned count)
2495 {
2496         vhost_add_used_n(vq, heads, count);
2497         vhost_signal(dev, vq);
2498 }
2499 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2500
2501 /* return true if we're sure that avaiable ring is empty */
2502 bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2503 {
2504         __virtio16 avail_idx;
2505         int r;
2506
2507         if (vq->avail_idx != vq->last_avail_idx)
2508                 return false;
2509
2510         r = vhost_get_avail_idx(vq, &avail_idx);
2511         if (unlikely(r))
2512                 return false;
2513         vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2514
2515         return vq->avail_idx == vq->last_avail_idx;
2516 }
2517 EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2518
2519 /* OK, now we need to know about added descriptors. */
2520 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2521 {
2522         __virtio16 avail_idx;
2523         int r;
2524
2525         if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2526                 return false;
2527         vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2528         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2529                 r = vhost_update_used_flags(vq);
2530                 if (r) {
2531                         vq_err(vq, "Failed to enable notification at %p: %d\n",
2532                                &vq->used->flags, r);
2533                         return false;
2534                 }
2535         } else {
2536                 r = vhost_update_avail_event(vq);
2537                 if (r) {
2538                         vq_err(vq, "Failed to update avail event index at %p: %d\n",
2539                                vhost_avail_event(vq), r);
2540                         return false;
2541                 }
2542         }
2543         /* They could have slipped one in as we were doing that: make
2544          * sure it's written, then check again. */
2545         smp_mb();
2546         r = vhost_get_avail_idx(vq, &avail_idx);
2547         if (r) {
2548                 vq_err(vq, "Failed to check avail idx at %p: %d\n",
2549                        &vq->avail->idx, r);
2550                 return false;
2551         }
2552         vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2553
2554         return vq->avail_idx != vq->last_avail_idx;
2555 }
2556 EXPORT_SYMBOL_GPL(vhost_enable_notify);
2557
2558 /* We don't need to be notified again. */
2559 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2560 {
2561         int r;
2562
2563         if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2564                 return;
2565         vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2566         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2567                 r = vhost_update_used_flags(vq);
2568                 if (r)
2569                         vq_err(vq, "Failed to disable notification at %p: %d\n",
2570                                &vq->used->flags, r);
2571         }
2572 }
2573 EXPORT_SYMBOL_GPL(vhost_disable_notify);
2574
2575 /* Create a new message. */
2576 struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2577 {
2578         struct vhost_msg_node *node = kmalloc(sizeof *node, GFP_KERNEL);
2579         if (!node)
2580                 return NULL;
2581
2582         /* Make sure all padding within the structure is initialized. */
2583         memset(&node->msg, 0, sizeof node->msg);
2584         node->vq = vq;
2585         node->msg.type = type;
2586         return node;
2587 }
2588 EXPORT_SYMBOL_GPL(vhost_new_msg);
2589
2590 void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2591                        struct vhost_msg_node *node)
2592 {
2593         spin_lock(&dev->iotlb_lock);
2594         list_add_tail(&node->node, head);
2595         spin_unlock(&dev->iotlb_lock);
2596
2597         wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
2598 }
2599 EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2600
2601 struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2602                                          struct list_head *head)
2603 {
2604         struct vhost_msg_node *node = NULL;
2605
2606         spin_lock(&dev->iotlb_lock);
2607         if (!list_empty(head)) {
2608                 node = list_first_entry(head, struct vhost_msg_node,
2609                                         node);
2610                 list_del(&node->node);
2611         }
2612         spin_unlock(&dev->iotlb_lock);
2613
2614         return node;
2615 }
2616 EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2617
2618 void vhost_set_backend_features(struct vhost_dev *dev, u64 features)
2619 {
2620         struct vhost_virtqueue *vq;
2621         int i;
2622
2623         mutex_lock(&dev->mutex);
2624         for (i = 0; i < dev->nvqs; ++i) {
2625                 vq = dev->vqs[i];
2626                 mutex_lock(&vq->mutex);
2627                 vq->acked_backend_features = features;
2628                 mutex_unlock(&vq->mutex);
2629         }
2630         mutex_unlock(&dev->mutex);
2631 }
2632 EXPORT_SYMBOL_GPL(vhost_set_backend_features);
2633
2634 static int __init vhost_init(void)
2635 {
2636         return 0;
2637 }
2638
2639 static void __exit vhost_exit(void)
2640 {
2641 }
2642
2643 module_init(vhost_init);
2644 module_exit(vhost_exit);
2645
2646 MODULE_VERSION("0.0.1");
2647 MODULE_LICENSE("GPL v2");
2648 MODULE_AUTHOR("Michael S. Tsirkin");
2649 MODULE_DESCRIPTION("Host kernel accelerator for virtio");