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/cgroup.h>
  26 #include <linux/module.h>
  27 #include <linux/sort.h>
  28 #include <linux/sched/mm.h>
  29 #include <linux/sched/signal.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_work_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_work_dev_flush);
 247
 248 /* Flush any work that has been scheduled. When calling this, don't hold any
 249  * locks that are also used by the callback. */
 250 void vhost_poll_flush(struct vhost_poll *poll)
 251 {
 252         vhost_work_dev_flush(poll->dev);
 253 }
 254 EXPORT_SYMBOL_GPL(vhost_poll_flush);
 255
 256 void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work)
 257 {
 258         if (!dev->worker)
 259                 return;
 260
 261         if (!test_and_set_bit(VHOST_WORK_QUEUED, &work->flags)) {
 262                 /* We can only add the work to the list after we're
 263                  * sure it was not in the list.
 264                  * test_and_set_bit() implies a memory barrier.
 265                  */
 266                 llist_add(&work->node, &dev->work_list);
 267                 wake_up_process(dev->worker);
 268         }
 269 }
 270 EXPORT_SYMBOL_GPL(vhost_work_queue);
 271
 272 /* A lockless hint for busy polling code to exit the loop */
 273 bool vhost_has_work(struct vhost_dev *dev)
 274 {
 275         return !llist_empty(&dev->work_list);
 276 }
 277 EXPORT_SYMBOL_GPL(vhost_has_work);
 278
 279 void vhost_poll_queue(struct vhost_poll *poll)
 280 {
 281         vhost_work_queue(poll->dev, &poll->work);
 282 }
 283 EXPORT_SYMBOL_GPL(vhost_poll_queue);
 284
 285 static void __vhost_vq_meta_reset(struct vhost_virtqueue *vq)
 286 {
 287         int j;
 288
 289         for (j = 0; j < VHOST_NUM_ADDRS; j++)
 290                 vq->meta_iotlb[j] = NULL;
 291 }
 292
 293 static void vhost_vq_meta_reset(struct vhost_dev *d)
 294 {
 295         int i;
 296
 297         for (i = 0; i < d->nvqs; ++i)
 298                 __vhost_vq_meta_reset(d->vqs[i]);
 299 }
 300
 301 static void vhost_vring_call_reset(struct vhost_vring_call *call_ctx)
 302 {
 303         call_ctx->ctx = NULL;
 304         memset(&call_ctx->producer, 0x0, sizeof(struct irq_bypass_producer));
 305 }
 306
 307 bool vhost_vq_is_setup(struct vhost_virtqueue *vq)
 308 {
 309         return vq->avail && vq->desc && vq->used && vhost_vq_access_ok(vq);
 310 }
 311 EXPORT_SYMBOL_GPL(vhost_vq_is_setup);
 312
 313 static void vhost_vq_reset(struct vhost_dev *dev,
 314                            struct vhost_virtqueue *vq)
 315 {
 316         vq->num = 1;
 317         vq->desc = NULL;
 318         vq->avail = NULL;
 319         vq->used = NULL;
 320         vq->last_avail_idx = 0;
 321         vq->avail_idx = 0;
 322         vq->last_used_idx = 0;
 323         vq->signalled_used = 0;
 324         vq->signalled_used_valid = false;
 325         vq->used_flags = 0;
 326         vq->log_used = false;
 327         vq->log_addr = -1ull;
 328         vq->private_data = NULL;
 329         vq->acked_features = 0;
 330         vq->acked_backend_features = 0;
 331         vq->log_base = NULL;
 332         vq->error_ctx = NULL;
 333         vq->kick = NULL;
 334         vq->log_ctx = NULL;
 335         vhost_disable_cross_endian(vq);
 336         vhost_reset_is_le(vq);
 337         vq->busyloop_timeout = 0;
 338         vq->umem = NULL;
 339         vq->iotlb = NULL;
 340         vhost_vring_call_reset(&vq->call_ctx);
 341         __vhost_vq_meta_reset(vq);
 342 }
 343
 344 static int vhost_worker(void *data)
 345 {
 346         struct vhost_dev *dev = data;
 347         struct vhost_work *work, *work_next;
 348         struct llist_node *node;
 349
 350         kthread_use_mm(dev->mm);
 351
 352         for (;;) {
 353                 /* mb paired w/ kthread_stop */
 354                 set_current_state(TASK_INTERRUPTIBLE);
 355
 356                 if (kthread_should_stop()) {
 357                         __set_current_state(TASK_RUNNING);
 358                         break;
 359                 }
 360
 361                 node = llist_del_all(&dev->work_list);
 362                 if (!node)
 363                         schedule();
 364
 365                 node = llist_reverse_order(node);
 366                 /* make sure flag is seen after deletion */
 367                 smp_wmb();
 368                 llist_for_each_entry_safe(work, work_next, node, node) {
 369                         clear_bit(VHOST_WORK_QUEUED, &work->flags);
 370                         __set_current_state(TASK_RUNNING);
 371                         kcov_remote_start_common(dev->kcov_handle);
 372                         work->fn(work);
 373                         kcov_remote_stop();
 374                         if (need_resched())
 375                                 schedule();
 376                 }
 377         }
 378         kthread_unuse_mm(dev->mm);
 379         return 0;
 380 }
 381
 382 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
 383 {
 384         kfree(vq->indirect);
 385         vq->indirect = NULL;
 386         kfree(vq->log);
 387         vq->log = NULL;
 388         kfree(vq->heads);
 389         vq->heads = NULL;
 390 }
 391
 392 /* Helper to allocate iovec buffers for all vqs. */
 393 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
 394 {
 395         struct vhost_virtqueue *vq;
 396         int i;
 397
 398         for (i = 0; i < dev->nvqs; ++i) {
 399                 vq = dev->vqs[i];
 400                 vq->indirect = kmalloc_array(UIO_MAXIOV,
 401                                              sizeof(*vq->indirect),
 402                                              GFP_KERNEL);
 403                 vq->log = kmalloc_array(dev->iov_limit, sizeof(*vq->log),
 404                                         GFP_KERNEL);
 405                 vq->heads = kmalloc_array(dev->iov_limit, sizeof(*vq->heads),
 406                                           GFP_KERNEL);
 407                 if (!vq->indirect || !vq->log || !vq->heads)
 408                         goto err_nomem;
 409         }
 410         return 0;
 411
 412 err_nomem:
 413         for (; i >= 0; --i)
 414                 vhost_vq_free_iovecs(dev->vqs[i]);
 415         return -ENOMEM;
 416 }
 417
 418 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
 419 {
 420         int i;
 421
 422         for (i = 0; i < dev->nvqs; ++i)
 423                 vhost_vq_free_iovecs(dev->vqs[i]);
 424 }
 425
 426 bool vhost_exceeds_weight(struct vhost_virtqueue *vq,
 427                           int pkts, int total_len)
 428 {
 429         struct vhost_dev *dev = vq->dev;
 430
 431         if ((dev->byte_weight && total_len >= dev->byte_weight) ||
 432             pkts >= dev->weight) {
 433                 vhost_poll_queue(&vq->poll);
 434                 return true;
 435         }
 436
 437         return false;
 438 }
 439 EXPORT_SYMBOL_GPL(vhost_exceeds_weight);
 440
 441 static size_t vhost_get_avail_size(struct vhost_virtqueue *vq,
 442                                    unsigned int num)
 443 {
 444         size_t event __maybe_unused =
 445                vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
 446
 447         return sizeof(*vq->avail) +
 448                sizeof(*vq->avail->ring) * num + event;
 449 }
 450
 451 static size_t vhost_get_used_size(struct vhost_virtqueue *vq,
 452                                   unsigned int num)
 453 {
 454         size_t event __maybe_unused =
 455                vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
 456
 457         return sizeof(*vq->used) +
 458                sizeof(*vq->used->ring) * num + event;
 459 }
 460
 461 static size_t vhost_get_desc_size(struct vhost_virtqueue *vq,
 462                                   unsigned int num)
 463 {
 464         return sizeof(*vq->desc) * num;
 465 }
 466
 467 void vhost_dev_init(struct vhost_dev *dev,
 468                     struct vhost_virtqueue **vqs, int nvqs,
 469                     int iov_limit, int weight, int byte_weight,
 470                     bool use_worker,
 471                     int (*msg_handler)(struct vhost_dev *dev,
 472                                        struct vhost_iotlb_msg *msg))
 473 {
 474         struct vhost_virtqueue *vq;
 475         int i;
 476
 477         dev->vqs = vqs;
 478         dev->nvqs = nvqs;
 479         mutex_init(&dev->mutex);
 480         dev->log_ctx = NULL;
 481         dev->umem = NULL;
 482         dev->iotlb = NULL;
 483         dev->mm = NULL;
 484         dev->worker = NULL;
 485         dev->iov_limit = iov_limit;
 486         dev->weight = weight;
 487         dev->byte_weight = byte_weight;
 488         dev->use_worker = use_worker;
 489         dev->msg_handler = msg_handler;
 490         init_llist_head(&dev->work_list);
 491         init_waitqueue_head(&dev->wait);
 492         INIT_LIST_HEAD(&dev->read_list);
 493         INIT_LIST_HEAD(&dev->pending_list);
 494         spin_lock_init(&dev->iotlb_lock);
 495
 496
 497         for (i = 0; i < dev->nvqs; ++i) {
 498                 vq = dev->vqs[i];
 499                 vq->log = NULL;
 500                 vq->indirect = NULL;
 501                 vq->heads = NULL;
 502                 vq->dev = dev;
 503                 mutex_init(&vq->mutex);
 504                 vhost_vq_reset(dev, vq);
 505                 if (vq->handle_kick)
 506                         vhost_poll_init(&vq->poll, vq->handle_kick,
 507                                         EPOLLIN, dev);
 508         }
 509 }
 510 EXPORT_SYMBOL_GPL(vhost_dev_init);
 511
 512 /* Caller should have device mutex */
 513 long vhost_dev_check_owner(struct vhost_dev *dev)
 514 {
 515         /* Are you the owner? If not, I don't think you mean to do that */
 516         return dev->mm == current->mm ? 0 : -EPERM;
 517 }
 518 EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
 519
 520 struct vhost_attach_cgroups_struct {
 521         struct vhost_work work;
 522         struct task_struct *owner;
 523         int ret;
 524 };
 525
 526 static void vhost_attach_cgroups_work(struct vhost_work *work)
 527 {
 528         struct vhost_attach_cgroups_struct *s;
 529
 530         s = container_of(work, struct vhost_attach_cgroups_struct, work);
 531         s->ret = cgroup_attach_task_all(s->owner, current);
 532 }
 533
 534 static int vhost_attach_cgroups(struct vhost_dev *dev)
 535 {
 536         struct vhost_attach_cgroups_struct attach;
 537
 538         attach.owner = current;
 539         vhost_work_init(&attach.work, vhost_attach_cgroups_work);
 540         vhost_work_queue(dev, &attach.work);
 541         vhost_work_dev_flush(dev);
 542         return attach.ret;
 543 }
 544
 545 /* Caller should have device mutex */
 546 bool vhost_dev_has_owner(struct vhost_dev *dev)
 547 {
 548         return dev->mm;
 549 }
 550 EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
 551
 552 static void vhost_attach_mm(struct vhost_dev *dev)
 553 {
 554         /* No owner, become one */
 555         if (dev->use_worker) {
 556                 dev->mm = get_task_mm(current);
 557         } else {
 558                 /* vDPA device does not use worker thead, so there's
 559                  * no need to hold the address space for mm. This help
 560                  * to avoid deadlock in the case of mmap() which may
 561                  * held the refcnt of the file and depends on release
 562                  * method to remove vma.
 563                  */
 564                 dev->mm = current->mm;
 565                 mmgrab(dev->mm);
 566         }
 567 }
 568
 569 static void vhost_detach_mm(struct vhost_dev *dev)
 570 {
 571         if (!dev->mm)
 572                 return;
 573
 574         if (dev->use_worker)
 575                 mmput(dev->mm);
 576         else
 577                 mmdrop(dev->mm);
 578
 579         dev->mm = NULL;
 580 }
 581
 582 /* Caller should have device mutex */
 583 long vhost_dev_set_owner(struct vhost_dev *dev)
 584 {
 585         struct task_struct *worker;
 586         int err;
 587
 588         /* Is there an owner already? */
 589         if (vhost_dev_has_owner(dev)) {
 590                 err = -EBUSY;
 591                 goto err_mm;
 592         }
 593
 594         vhost_attach_mm(dev);
 595
 596         dev->kcov_handle = kcov_common_handle();
 597         if (dev->use_worker) {
 598                 worker = kthread_create(vhost_worker, dev,
 599                                         "vhost-%d", current->pid);
 600                 if (IS_ERR(worker)) {
 601                         err = PTR_ERR(worker);
 602                         goto err_worker;
 603                 }
 604
 605                 dev->worker = worker;
 606                 wake_up_process(worker); /* avoid contributing to loadavg */
 607
 608                 err = vhost_attach_cgroups(dev);
 609                 if (err)
 610                         goto err_cgroup;
 611         }
 612
 613         err = vhost_dev_alloc_iovecs(dev);
 614         if (err)
 615                 goto err_cgroup;
 616
 617         return 0;
 618 err_cgroup:
 619         if (dev->worker) {
 620                 kthread_stop(dev->worker);
 621                 dev->worker = NULL;
 622         }
 623 err_worker:
 624         vhost_detach_mm(dev);
 625         dev->kcov_handle = 0;
 626 err_mm:
 627         return err;
 628 }
 629 EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
 630
 631 static struct vhost_iotlb *iotlb_alloc(void)
 632 {
 633         return vhost_iotlb_alloc(max_iotlb_entries,
 634                                  VHOST_IOTLB_FLAG_RETIRE);
 635 }
 636
 637 struct vhost_iotlb *vhost_dev_reset_owner_prepare(void)
 638 {
 639         return iotlb_alloc();
 640 }
 641 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
 642
 643 /* Caller should have device mutex */
 644 void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_iotlb *umem)
 645 {
 646         int i;
 647
 648         vhost_dev_cleanup(dev);
 649
 650         dev->umem = umem;
 651         /* We don't need VQ locks below since vhost_dev_cleanup makes sure
 652          * VQs aren't running.
 653          */
 654         for (i = 0; i < dev->nvqs; ++i)
 655                 dev->vqs[i]->umem = umem;
 656 }
 657 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
 658
 659 void vhost_dev_stop(struct vhost_dev *dev)
 660 {
 661         int i;
 662
 663         for (i = 0; i < dev->nvqs; ++i) {
 664                 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick) {
 665                         vhost_poll_stop(&dev->vqs[i]->poll);
 666                         vhost_poll_flush(&dev->vqs[i]->poll);
 667                 }
 668         }
 669 }
 670 EXPORT_SYMBOL_GPL(vhost_dev_stop);
 671
 672 static void vhost_clear_msg(struct vhost_dev *dev)
 673 {
 674         struct vhost_msg_node *node, *n;
 675
 676         spin_lock(&dev->iotlb_lock);
 677
 678         list_for_each_entry_safe(node, n, &dev->read_list, node) {
 679                 list_del(&node->node);
 680                 kfree(node);
 681         }
 682
 683         list_for_each_entry_safe(node, n, &dev->pending_list, node) {
 684                 list_del(&node->node);
 685                 kfree(node);
 686         }
 687
 688         spin_unlock(&dev->iotlb_lock);
 689 }
 690
 691 void vhost_dev_cleanup(struct vhost_dev *dev)
 692 {
 693         int i;
 694
 695         for (i = 0; i < dev->nvqs; ++i) {
 696                 if (dev->vqs[i]->error_ctx)
 697                         eventfd_ctx_put(dev->vqs[i]->error_ctx);
 698                 if (dev->vqs[i]->kick)
 699                         fput(dev->vqs[i]->kick);
 700                 if (dev->vqs[i]->call_ctx.ctx)
 701                         eventfd_ctx_put(dev->vqs[i]->call_ctx.ctx);
 702                 vhost_vq_reset(dev, dev->vqs[i]);
 703         }
 704         vhost_dev_free_iovecs(dev);
 705         if (dev->log_ctx)
 706                 eventfd_ctx_put(dev->log_ctx);
 707         dev->log_ctx = NULL;
 708         /* No one will access memory at this point */
 709         vhost_iotlb_free(dev->umem);
 710         dev->umem = NULL;
 711         vhost_iotlb_free(dev->iotlb);
 712         dev->iotlb = NULL;
 713         vhost_clear_msg(dev);
 714         wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
 715         WARN_ON(!llist_empty(&dev->work_list));
 716         if (dev->worker) {
 717                 kthread_stop(dev->worker);
 718                 dev->worker = NULL;
 719                 dev->kcov_handle = 0;
 720         }
 721         vhost_detach_mm(dev);
 722 }
 723 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
 724
 725 static bool log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
 726 {
 727         u64 a = addr / VHOST_PAGE_SIZE / 8;
 728
 729         /* Make sure 64 bit math will not overflow. */
 730         if (a > ULONG_MAX - (unsigned long)log_base ||
 731             a + (unsigned long)log_base > ULONG_MAX)
 732                 return false;
 733
 734         return access_ok(log_base + a,
 735                          (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
 736 }
 737
 738 /* Make sure 64 bit math will not overflow. */
 739 static bool vhost_overflow(u64 uaddr, u64 size)
 740 {
 741         if (uaddr > ULONG_MAX || size > ULONG_MAX)
 742                 return true;
 743
 744         if (!size)
 745                 return false;
 746
 747         return uaddr > ULONG_MAX - size + 1;
 748 }
 749
 750 /* Caller should have vq mutex and device mutex. */
 751 static bool vq_memory_access_ok(void __user *log_base, struct vhost_iotlb *umem,
 752                                 int log_all)
 753 {
 754         struct vhost_iotlb_map *map;
 755
 756         if (!umem)
 757                 return false;
 758
 759         list_for_each_entry(map, &umem->list, link) {
 760                 unsigned long a = map->addr;
 761
 762                 if (vhost_overflow(map->addr, map->size))
 763                         return false;
 764
 765
 766                 if (!access_ok((void __user *)a, map->size))
 767                         return false;
 768                 else if (log_all && !log_access_ok(log_base,
 769                                                    map->start,
 770                                                    map->size))
 771                         return false;
 772         }
 773         return true;
 774 }
 775
 776 static inline void __user *vhost_vq_meta_fetch(struct vhost_virtqueue *vq,
 777                                                u64 addr, unsigned int size,
 778                                                int type)
 779 {
 780         const struct vhost_iotlb_map *map = vq->meta_iotlb[type];
 781
 782         if (!map)
 783                 return NULL;
 784
 785         return (void __user *)(uintptr_t)(map->addr + addr - map->start);
 786 }
 787
 788 /* Can we switch to this memory table? */
 789 /* Caller should have device mutex but not vq mutex */
 790 static bool memory_access_ok(struct vhost_dev *d, struct vhost_iotlb *umem,
 791                              int log_all)
 792 {
 793         int i;
 794
 795         for (i = 0; i < d->nvqs; ++i) {
 796                 bool ok;
 797                 bool log;
 798
 799                 mutex_lock(&d->vqs[i]->mutex);
 800                 log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
 801                 /* If ring is inactive, will check when it's enabled. */
 802                 if (d->vqs[i]->private_data)
 803                         ok = vq_memory_access_ok(d->vqs[i]->log_base,
 804                                                  umem, log);
 805                 else
 806                         ok = true;
 807                 mutex_unlock(&d->vqs[i]->mutex);
 808                 if (!ok)
 809                         return false;
 810         }
 811         return true;
 812 }
 813
 814 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
 815                           struct iovec iov[], int iov_size, int access);
 816
 817 static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
 818                               const void *from, unsigned size)
 819 {
 820         int ret;
 821
 822         if (!vq->iotlb)
 823                 return __copy_to_user(to, from, size);
 824         else {
 825                 /* This function should be called after iotlb
 826                  * prefetch, which means we're sure that all vq
 827                  * could be access through iotlb. So -EAGAIN should
 828                  * not happen in this case.
 829                  */
 830                 struct iov_iter t;
 831                 void __user *uaddr = vhost_vq_meta_fetch(vq,
 832                                      (u64)(uintptr_t)to, size,
 833                                      VHOST_ADDR_USED);
 834
 835                 if (uaddr)
 836                         return __copy_to_user(uaddr, from, size);
 837
 838                 ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
 839                                      ARRAY_SIZE(vq->iotlb_iov),
 840                                      VHOST_ACCESS_WO);
 841                 if (ret < 0)
 842                         goto out;
 843                 iov_iter_init(&t, WRITE, vq->iotlb_iov, ret, size);
 844                 ret = copy_to_iter(from, size, &t);
 845                 if (ret == size)
 846                         ret = 0;
 847         }
 848 out:
 849         return ret;
 850 }
 851
 852 static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
 853                                 void __user *from, unsigned size)
 854 {
 855         int ret;
 856
 857         if (!vq->iotlb)
 858                 return __copy_from_user(to, from, size);
 859         else {
 860                 /* This function should be called after iotlb
 861                  * prefetch, which means we're sure that vq
 862                  * could be access through iotlb. So -EAGAIN should
 863                  * not happen in this case.
 864                  */
 865                 void __user *uaddr = vhost_vq_meta_fetch(vq,
 866                                      (u64)(uintptr_t)from, size,
 867                                      VHOST_ADDR_DESC);
 868                 struct iov_iter f;
 869
 870                 if (uaddr)
 871                         return __copy_from_user(to, uaddr, size);
 872
 873                 ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
 874                                      ARRAY_SIZE(vq->iotlb_iov),
 875                                      VHOST_ACCESS_RO);
 876                 if (ret < 0) {
 877                         vq_err(vq, "IOTLB translation failure: uaddr "
 878                                "%p size 0x%llx\n", from,
 879                                (unsigned long long) size);
 880                         goto out;
 881                 }
 882                 iov_iter_init(&f, READ, vq->iotlb_iov, ret, size);
 883                 ret = copy_from_iter(to, size, &f);
 884                 if (ret == size)
 885                         ret = 0;
 886         }
 887
 888 out:
 889         return ret;
 890 }
 891
 892 static void __user *__vhost_get_user_slow(struct vhost_virtqueue *vq,
 893                                           void __user *addr, unsigned int size,
 894                                           int type)
 895 {
 896         int ret;
 897
 898         ret = translate_desc(vq, (u64)(uintptr_t)addr, size, vq->iotlb_iov,
 899                              ARRAY_SIZE(vq->iotlb_iov),
 900                              VHOST_ACCESS_RO);
 901         if (ret < 0) {
 902                 vq_err(vq, "IOTLB translation failure: uaddr "
 903                         "%p size 0x%llx\n", addr,
 904                         (unsigned long long) size);
 905                 return NULL;
 906         }
 907
 908         if (ret != 1 || vq->iotlb_iov[0].iov_len != size) {
 909                 vq_err(vq, "Non atomic userspace memory access: uaddr "
 910                         "%p size 0x%llx\n", addr,
 911                         (unsigned long long) size);
 912                 return NULL;
 913         }
 914
 915         return vq->iotlb_iov[0].iov_base;
 916 }
 917
 918 /* This function should be called after iotlb
 919  * prefetch, which means we're sure that vq
 920  * could be access through iotlb. So -EAGAIN should
 921  * not happen in this case.
 922  */
 923 static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
 924                                             void __user *addr, unsigned int size,
 925                                             int type)
 926 {
 927         void __user *uaddr = vhost_vq_meta_fetch(vq,
 928                              (u64)(uintptr_t)addr, size, type);
 929         if (uaddr)
 930                 return uaddr;
 931
 932         return __vhost_get_user_slow(vq, addr, size, type);
 933 }
 934
 935 #define vhost_put_user(vq, x, ptr)              \
 936 ({ \
 937         int ret; \
 938         if (!vq->iotlb) { \
 939                 ret = __put_user(x, ptr); \
 940         } else { \
 941                 __typeof__(ptr) to = \
 942                         (__typeof__(ptr)) __vhost_get_user(vq, ptr,     \
 943                                           sizeof(*ptr), VHOST_ADDR_USED); \
 944                 if (to != NULL) \
 945                         ret = __put_user(x, to); \
 946                 else \
 947                         ret = -EFAULT;  \
 948         } \
 949         ret; \
 950 })
 951
 952 static inline int vhost_put_avail_event(struct vhost_virtqueue *vq)
 953 {
 954         return vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
 955                               vhost_avail_event(vq));
 956 }
 957
 958 static inline int vhost_put_used(struct vhost_virtqueue *vq,
 959                                  struct vring_used_elem *head, int idx,
 960                                  int count)
 961 {
 962         return vhost_copy_to_user(vq, vq->used->ring + idx, head,
 963                                   count * sizeof(*head));
 964 }
 965
 966 static inline int vhost_put_used_flags(struct vhost_virtqueue *vq)
 967
 968 {
 969         return vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
 970                               &vq->used->flags);
 971 }
 972
 973 static inline int vhost_put_used_idx(struct vhost_virtqueue *vq)
 974
 975 {
 976         return vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
 977                               &vq->used->idx);
 978 }
 979
 980 #define vhost_get_user(vq, x, ptr, type)                \
 981 ({ \
 982         int ret; \
 983         if (!vq->iotlb) { \
 984                 ret = __get_user(x, ptr); \
 985         } else { \
 986                 __typeof__(ptr) from = \
 987                         (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
 988                                                            sizeof(*ptr), \
 989                                                            type); \
 990                 if (from != NULL) \
 991                         ret = __get_user(x, from); \
 992                 else \
 993                         ret = -EFAULT; \
 994         } \
 995         ret; \
 996 })
 997
 998 #define vhost_get_avail(vq, x, ptr) \
 999         vhost_get_user(vq, x, ptr, VHOST_ADDR_AVAIL)
1000
1001 #define vhost_get_used(vq, x, ptr) \
1002         vhost_get_user(vq, x, ptr, VHOST_ADDR_USED)
1003
1004 static void vhost_dev_lock_vqs(struct vhost_dev *d)
1005 {
1006         int i = 0;
1007         for (i = 0; i < d->nvqs; ++i)
1008                 mutex_lock_nested(&d->vqs[i]->mutex, i);
1009 }
1010
1011 static void vhost_dev_unlock_vqs(struct vhost_dev *d)
1012 {
1013         int i = 0;
1014         for (i = 0; i < d->nvqs; ++i)
1015                 mutex_unlock(&d->vqs[i]->mutex);
1016 }
1017
1018 static inline int vhost_get_avail_idx(struct vhost_virtqueue *vq,
1019                                       __virtio16 *idx)
1020 {
1021         return vhost_get_avail(vq, *idx, &vq->avail->idx);
1022 }
1023
1024 static inline int vhost_get_avail_head(struct vhost_virtqueue *vq,
1025                                        __virtio16 *head, int idx)
1026 {
1027         return vhost_get_avail(vq, *head,
1028                                &vq->avail->ring[idx & (vq->num - 1)]);
1029 }
1030
1031 static inline int vhost_get_avail_flags(struct vhost_virtqueue *vq,
1032                                         __virtio16 *flags)
1033 {
1034         return vhost_get_avail(vq, *flags, &vq->avail->flags);
1035 }
1036
1037 static inline int vhost_get_used_event(struct vhost_virtqueue *vq,
1038                                        __virtio16 *event)
1039 {
1040         return vhost_get_avail(vq, *event, vhost_used_event(vq));
1041 }
1042
1043 static inline int vhost_get_used_idx(struct vhost_virtqueue *vq,
1044                                      __virtio16 *idx)
1045 {
1046         return vhost_get_used(vq, *idx, &vq->used->idx);
1047 }
1048
1049 static inline int vhost_get_desc(struct vhost_virtqueue *vq,
1050                                  struct vring_desc *desc, int idx)
1051 {
1052         return vhost_copy_from_user(vq, desc, vq->desc + idx, sizeof(*desc));
1053 }
1054
1055 static void vhost_iotlb_notify_vq(struct vhost_dev *d,
1056                                   struct vhost_iotlb_msg *msg)
1057 {
1058         struct vhost_msg_node *node, *n;
1059
1060         spin_lock(&d->iotlb_lock);
1061
1062         list_for_each_entry_safe(node, n, &d->pending_list, node) {
1063                 struct vhost_iotlb_msg *vq_msg = &node->msg.iotlb;
1064                 if (msg->iova <= vq_msg->iova &&
1065                     msg->iova + msg->size - 1 >= vq_msg->iova &&
1066                     vq_msg->type == VHOST_IOTLB_MISS) {
1067                         vhost_poll_queue(&node->vq->poll);
1068                         list_del(&node->node);
1069                         kfree(node);
1070                 }
1071         }
1072
1073         spin_unlock(&d->iotlb_lock);
1074 }
1075
1076 static bool umem_access_ok(u64 uaddr, u64 size, int access)
1077 {
1078         unsigned long a = uaddr;
1079
1080         /* Make sure 64 bit math will not overflow. */
1081         if (vhost_overflow(uaddr, size))
1082                 return false;
1083
1084         if ((access & VHOST_ACCESS_RO) &&
1085             !access_ok((void __user *)a, size))
1086                 return false;
1087         if ((access & VHOST_ACCESS_WO) &&
1088             !access_ok((void __user *)a, size))
1089                 return false;
1090         return true;
1091 }
1092
1093 static int vhost_process_iotlb_msg(struct vhost_dev *dev,
1094                                    struct vhost_iotlb_msg *msg)
1095 {
1096         int ret = 0;
1097
1098         mutex_lock(&dev->mutex);
1099         vhost_dev_lock_vqs(dev);
1100         switch (msg->type) {
1101         case VHOST_IOTLB_UPDATE:
1102                 if (!dev->iotlb) {
1103                         ret = -EFAULT;
1104                         break;
1105                 }
1106                 if (!umem_access_ok(msg->uaddr, msg->size, msg->perm)) {
1107                         ret = -EFAULT;
1108                         break;
1109                 }
1110                 vhost_vq_meta_reset(dev);
1111                 if (vhost_iotlb_add_range(dev->iotlb, msg->iova,
1112                                           msg->iova + msg->size - 1,
1113                                           msg->uaddr, msg->perm)) {
1114                         ret = -ENOMEM;
1115                         break;
1116                 }
1117                 vhost_iotlb_notify_vq(dev, msg);
1118                 break;
1119         case VHOST_IOTLB_INVALIDATE:
1120                 if (!dev->iotlb) {
1121                         ret = -EFAULT;
1122                         break;
1123                 }
1124                 vhost_vq_meta_reset(dev);
1125                 vhost_iotlb_del_range(dev->iotlb, msg->iova,
1126                                       msg->iova + msg->size - 1);
1127                 break;
1128         default:
1129                 ret = -EINVAL;
1130                 break;
1131         }
1132
1133         vhost_dev_unlock_vqs(dev);
1134         mutex_unlock(&dev->mutex);
1135
1136         return ret;
1137 }
1138 ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
1139                              struct iov_iter *from)
1140 {
1141         struct vhost_iotlb_msg msg;
1142         size_t offset;
1143         int type, ret;
1144
1145         ret = copy_from_iter(&type, sizeof(type), from);
1146         if (ret != sizeof(type)) {
1147                 ret = -EINVAL;
1148                 goto done;
1149         }
1150
1151         switch (type) {
1152         case VHOST_IOTLB_MSG:
1153                 /* There maybe a hole after type for V1 message type,
1154                  * so skip it here.
1155                  */
1156                 offset = offsetof(struct vhost_msg, iotlb) - sizeof(int);
1157                 break;
1158         case VHOST_IOTLB_MSG_V2:
1159                 offset = sizeof(__u32);
1160                 break;
1161         default:
1162                 ret = -EINVAL;
1163                 goto done;
1164         }
1165
1166         iov_iter_advance(from, offset);
1167         ret = copy_from_iter(&msg, sizeof(msg), from);
1168         if (ret != sizeof(msg)) {
1169                 ret = -EINVAL;
1170                 goto done;
1171         }
1172
1173         if ((msg.type == VHOST_IOTLB_UPDATE ||
1174              msg.type == VHOST_IOTLB_INVALIDATE) &&
1175              msg.size == 0) {
1176                 ret = -EINVAL;
1177                 goto done;
1178         }
1179
1180         if (dev->msg_handler)
1181                 ret = dev->msg_handler(dev, &msg);
1182         else
1183                 ret = vhost_process_iotlb_msg(dev, &msg);
1184         if (ret) {
1185                 ret = -EFAULT;
1186                 goto done;
1187         }
1188
1189         ret = (type == VHOST_IOTLB_MSG) ? sizeof(struct vhost_msg) :
1190               sizeof(struct vhost_msg_v2);
1191 done:
1192         return ret;
1193 }
1194 EXPORT_SYMBOL(vhost_chr_write_iter);
1195
1196 __poll_t vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1197                             poll_table *wait)
1198 {
1199         __poll_t mask = 0;
1200
1201         poll_wait(file, &dev->wait, wait);
1202
1203         if (!list_empty(&dev->read_list))
1204                 mask |= EPOLLIN | EPOLLRDNORM;
1205
1206         return mask;
1207 }
1208 EXPORT_SYMBOL(vhost_chr_poll);
1209
1210 ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1211                             int noblock)
1212 {
1213         DEFINE_WAIT(wait);
1214         struct vhost_msg_node *node;
1215         ssize_t ret = 0;
1216         unsigned size = sizeof(struct vhost_msg);
1217
1218         if (iov_iter_count(to) < size)
1219                 return 0;
1220
1221         while (1) {
1222                 if (!noblock)
1223                         prepare_to_wait(&dev->wait, &wait,
1224                                         TASK_INTERRUPTIBLE);
1225
1226                 node = vhost_dequeue_msg(dev, &dev->read_list);
1227                 if (node)
1228                         break;
1229                 if (noblock) {
1230                         ret = -EAGAIN;
1231                         break;
1232                 }
1233                 if (signal_pending(current)) {
1234                         ret = -ERESTARTSYS;
1235                         break;
1236                 }
1237                 if (!dev->iotlb) {
1238                         ret = -EBADFD;
1239                         break;
1240                 }
1241
1242                 schedule();
1243         }
1244
1245         if (!noblock)
1246                 finish_wait(&dev->wait, &wait);
1247
1248         if (node) {
1249                 struct vhost_iotlb_msg *msg;
1250                 void *start = &node->msg;
1251
1252                 switch (node->msg.type) {
1253                 case VHOST_IOTLB_MSG:
1254                         size = sizeof(node->msg);
1255                         msg = &node->msg.iotlb;
1256                         break;
1257                 case VHOST_IOTLB_MSG_V2:
1258                         size = sizeof(node->msg_v2);
1259                         msg = &node->msg_v2.iotlb;
1260                         break;
1261                 default:
1262                         BUG();
1263                         break;
1264                 }
1265
1266                 ret = copy_to_iter(start, size, to);
1267                 if (ret != size || msg->type != VHOST_IOTLB_MISS) {
1268                         kfree(node);
1269                         return ret;
1270                 }
1271                 vhost_enqueue_msg(dev, &dev->pending_list, node);
1272         }
1273
1274         return ret;
1275 }
1276 EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1277
1278 static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1279 {
1280         struct vhost_dev *dev = vq->dev;
1281         struct vhost_msg_node *node;
1282         struct vhost_iotlb_msg *msg;
1283         bool v2 = vhost_backend_has_feature(vq, VHOST_BACKEND_F_IOTLB_MSG_V2);
1284
1285         node = vhost_new_msg(vq, v2 ? VHOST_IOTLB_MSG_V2 : VHOST_IOTLB_MSG);
1286         if (!node)
1287                 return -ENOMEM;
1288
1289         if (v2) {
1290                 node->msg_v2.type = VHOST_IOTLB_MSG_V2;
1291                 msg = &node->msg_v2.iotlb;
1292         } else {
1293                 msg = &node->msg.iotlb;
1294         }
1295
1296         msg->type = VHOST_IOTLB_MISS;
1297         msg->iova = iova;
1298         msg->perm = access;
1299
1300         vhost_enqueue_msg(dev, &dev->read_list, node);
1301
1302         return 0;
1303 }
1304
1305 static bool vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1306                          vring_desc_t __user *desc,
1307                          vring_avail_t __user *avail,
1308                          vring_used_t __user *used)
1309
1310 {
1311         /* If an IOTLB device is present, the vring addresses are
1312          * GIOVAs. Access validation occurs at prefetch time. */
1313         if (vq->iotlb)
1314                 return true;
1315
1316         return access_ok(desc, vhost_get_desc_size(vq, num)) &&
1317                access_ok(avail, vhost_get_avail_size(vq, num)) &&
1318                access_ok(used, vhost_get_used_size(vq, num));
1319 }
1320
1321 static void vhost_vq_meta_update(struct vhost_virtqueue *vq,
1322                                  const struct vhost_iotlb_map *map,
1323                                  int type)
1324 {
1325         int access = (type == VHOST_ADDR_USED) ?
1326                      VHOST_ACCESS_WO : VHOST_ACCESS_RO;
1327
1328         if (likely(map->perm & access))
1329                 vq->meta_iotlb[type] = map;
1330 }
1331
1332 static bool iotlb_access_ok(struct vhost_virtqueue *vq,
1333                             int access, u64 addr, u64 len, int type)
1334 {
1335         const struct vhost_iotlb_map *map;
1336         struct vhost_iotlb *umem = vq->iotlb;
1337         u64 s = 0, size, orig_addr = addr, last = addr + len - 1;
1338
1339         if (vhost_vq_meta_fetch(vq, addr, len, type))
1340                 return true;
1341
1342         while (len > s) {
1343                 map = vhost_iotlb_itree_first(umem, addr, last);
1344                 if (map == NULL || map->start > addr) {
1345                         vhost_iotlb_miss(vq, addr, access);
1346                         return false;
1347                 } else if (!(map->perm & access)) {
1348                         /* Report the possible access violation by
1349                          * request another translation from userspace.
1350                          */
1351                         return false;
1352                 }
1353
1354                 size = map->size - addr + map->start;
1355
1356                 if (orig_addr == addr && size >= len)
1357                         vhost_vq_meta_update(vq, map, type);
1358
1359                 s += size;
1360                 addr += size;
1361         }
1362
1363         return true;
1364 }
1365
1366 int vq_meta_prefetch(struct vhost_virtqueue *vq)
1367 {
1368         unsigned int num = vq->num;
1369
1370         if (!vq->iotlb)
1371                 return 1;
1372
1373         return iotlb_access_ok(vq, VHOST_MAP_RO, (u64)(uintptr_t)vq->desc,
1374                                vhost_get_desc_size(vq, num), VHOST_ADDR_DESC) &&
1375                iotlb_access_ok(vq, VHOST_MAP_RO, (u64)(uintptr_t)vq->avail,
1376                                vhost_get_avail_size(vq, num),
1377                                VHOST_ADDR_AVAIL) &&
1378                iotlb_access_ok(vq, VHOST_MAP_WO, (u64)(uintptr_t)vq->used,
1379                                vhost_get_used_size(vq, num), VHOST_ADDR_USED);
1380 }
1381 EXPORT_SYMBOL_GPL(vq_meta_prefetch);
1382
1383 /* Can we log writes? */
1384 /* Caller should have device mutex but not vq mutex */
1385 bool vhost_log_access_ok(struct vhost_dev *dev)
1386 {
1387         return memory_access_ok(dev, dev->umem, 1);
1388 }
1389 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1390
1391 static bool vq_log_used_access_ok(struct vhost_virtqueue *vq,
1392                                   void __user *log_base,
1393                                   bool log_used,
1394                                   u64 log_addr)
1395 {
1396         /* If an IOTLB device is present, log_addr is a GIOVA that
1397          * will never be logged by log_used(). */
1398         if (vq->iotlb)
1399                 return true;
1400
1401         return !log_used || log_access_ok(log_base, log_addr,
1402                                           vhost_get_used_size(vq, vq->num));
1403 }
1404
1405 /* Verify access for write logging. */
1406 /* Caller should have vq mutex and device mutex */
1407 static bool vq_log_access_ok(struct vhost_virtqueue *vq,
1408                              void __user *log_base)
1409 {
1410         return vq_memory_access_ok(log_base, vq->umem,
1411                                    vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1412                 vq_log_used_access_ok(vq, log_base, vq->log_used, vq->log_addr);
1413 }
1414
1415 /* Can we start vq? */
1416 /* Caller should have vq mutex and device mutex */
1417 bool vhost_vq_access_ok(struct vhost_virtqueue *vq)
1418 {
1419         if (!vq_log_access_ok(vq, vq->log_base))
1420                 return false;
1421
1422         return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used);
1423 }
1424 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1425
1426 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1427 {
1428         struct vhost_memory mem, *newmem;
1429         struct vhost_memory_region *region;
1430         struct vhost_iotlb *newumem, *oldumem;
1431         unsigned long size = offsetof(struct vhost_memory, regions);
1432         int i;
1433
1434         if (copy_from_user(&mem, m, size))
1435                 return -EFAULT;
1436         if (mem.padding)
1437                 return -EOPNOTSUPP;
1438         if (mem.nregions > max_mem_regions)
1439                 return -E2BIG;
1440         newmem = kvzalloc(struct_size(newmem, regions, mem.nregions),
1441                         GFP_KERNEL);
1442         if (!newmem)
1443                 return -ENOMEM;
1444
1445         memcpy(newmem, &mem, size);
1446         if (copy_from_user(newmem->regions, m->regions,
1447                            flex_array_size(newmem, regions, mem.nregions))) {
1448                 kvfree(newmem);
1449                 return -EFAULT;
1450         }
1451
1452         newumem = iotlb_alloc();
1453         if (!newumem) {
1454                 kvfree(newmem);
1455                 return -ENOMEM;
1456         }
1457
1458         for (region = newmem->regions;
1459              region < newmem->regions + mem.nregions;
1460              region++) {
1461                 if (vhost_iotlb_add_range(newumem,
1462                                           region->guest_phys_addr,
1463                                           region->guest_phys_addr +
1464                                           region->memory_size - 1,
1465                                           region->userspace_addr,
1466                                           VHOST_MAP_RW))
1467                         goto err;
1468         }
1469
1470         if (!memory_access_ok(d, newumem, 0))
1471                 goto err;
1472
1473         oldumem = d->umem;
1474         d->umem = newumem;
1475
1476         /* All memory accesses are done under some VQ mutex. */
1477         for (i = 0; i < d->nvqs; ++i) {
1478                 mutex_lock(&d->vqs[i]->mutex);
1479                 d->vqs[i]->umem = newumem;
1480                 mutex_unlock(&d->vqs[i]->mutex);
1481         }
1482
1483         kvfree(newmem);
1484         vhost_iotlb_free(oldumem);
1485         return 0;
1486
1487 err:
1488         vhost_iotlb_free(newumem);
1489         kvfree(newmem);
1490         return -EFAULT;
1491 }
1492
1493 static long vhost_vring_set_num(struct vhost_dev *d,
1494                                 struct vhost_virtqueue *vq,
1495                                 void __user *argp)
1496 {
1497         struct vhost_vring_state s;
1498
1499         /* Resizing ring with an active backend?
1500          * You don't want to do that. */
1501         if (vq->private_data)
1502                 return -EBUSY;
1503
1504         if (copy_from_user(&s, argp, sizeof s))
1505                 return -EFAULT;
1506
1507         if (!s.num || s.num > 0xffff || (s.num & (s.num - 1)))
1508                 return -EINVAL;
1509         vq->num = s.num;
1510
1511         return 0;
1512 }
1513
1514 static long vhost_vring_set_addr(struct vhost_dev *d,
1515                                  struct vhost_virtqueue *vq,
1516                                  void __user *argp)
1517 {
1518         struct vhost_vring_addr a;
1519
1520         if (copy_from_user(&a, argp, sizeof a))
1521                 return -EFAULT;
1522         if (a.flags & ~(0x1 << VHOST_VRING_F_LOG))
1523                 return -EOPNOTSUPP;
1524
1525         /* For 32bit, verify that the top 32bits of the user
1526            data are set to zero. */
1527         if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1528             (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1529             (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr)
1530                 return -EFAULT;
1531
1532         /* Make sure it's safe to cast pointers to vring types. */
1533         BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1534         BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1535         if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1536             (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1537             (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1)))
1538                 return -EINVAL;
1539
1540         /* We only verify access here if backend is configured.
1541          * If it is not, we don't as size might not have been setup.
1542          * We will verify when backend is configured. */
1543         if (vq->private_data) {
1544                 if (!vq_access_ok(vq, vq->num,
1545                         (void __user *)(unsigned long)a.desc_user_addr,
1546                         (void __user *)(unsigned long)a.avail_user_addr,
1547                         (void __user *)(unsigned long)a.used_user_addr))
1548                         return -EINVAL;
1549
1550                 /* Also validate log access for used ring if enabled. */
1551                 if (!vq_log_used_access_ok(vq, vq->log_base,
1552                                 a.flags & (0x1 << VHOST_VRING_F_LOG),
1553                                 a.log_guest_addr))
1554                         return -EINVAL;
1555         }
1556
1557         vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1558         vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1559         vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1560         vq->log_addr = a.log_guest_addr;
1561         vq->used = (void __user *)(unsigned long)a.used_user_addr;
1562
1563         return 0;
1564 }
1565
1566 static long vhost_vring_set_num_addr(struct vhost_dev *d,
1567                                      struct vhost_virtqueue *vq,
1568                                      unsigned int ioctl,
1569                                      void __user *argp)
1570 {
1571         long r;
1572
1573         mutex_lock(&vq->mutex);
1574
1575         switch (ioctl) {
1576         case VHOST_SET_VRING_NUM:
1577                 r = vhost_vring_set_num(d, vq, argp);
1578                 break;
1579         case VHOST_SET_VRING_ADDR:
1580                 r = vhost_vring_set_addr(d, vq, argp);
1581                 break;
1582         default:
1583                 BUG();
1584         }
1585
1586         mutex_unlock(&vq->mutex);
1587
1588         return r;
1589 }
1590 long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1591 {
1592         struct file *eventfp, *filep = NULL;
1593         bool pollstart = false, pollstop = false;
1594         struct eventfd_ctx *ctx = NULL;
1595         u32 __user *idxp = argp;
1596         struct vhost_virtqueue *vq;
1597         struct vhost_vring_state s;
1598         struct vhost_vring_file f;
1599         u32 idx;
1600         long r;
1601
1602         r = get_user(idx, idxp);
1603         if (r < 0)
1604                 return r;
1605         if (idx >= d->nvqs)
1606                 return -ENOBUFS;
1607
1608         idx = array_index_nospec(idx, d->nvqs);
1609         vq = d->vqs[idx];
1610
1611         if (ioctl == VHOST_SET_VRING_NUM ||
1612             ioctl == VHOST_SET_VRING_ADDR) {
1613                 return vhost_vring_set_num_addr(d, vq, ioctl, argp);
1614         }
1615
1616         mutex_lock(&vq->mutex);
1617
1618         switch (ioctl) {
1619         case VHOST_SET_VRING_BASE:
1620                 /* Moving base with an active backend?
1621                  * You don't want to do that. */
1622                 if (vq->private_data) {
1623                         r = -EBUSY;
1624                         break;
1625                 }
1626                 if (copy_from_user(&s, argp, sizeof s)) {
1627                         r = -EFAULT;
1628                         break;
1629                 }
1630                 if (s.num > 0xffff) {
1631                         r = -EINVAL;
1632                         break;
1633                 }
1634                 vq->last_avail_idx = s.num;
1635                 /* Forget the cached index value. */
1636                 vq->avail_idx = vq->last_avail_idx;
1637                 break;
1638         case VHOST_GET_VRING_BASE:
1639                 s.index = idx;
1640                 s.num = vq->last_avail_idx;
1641                 if (copy_to_user(argp, &s, sizeof s))
1642                         r = -EFAULT;
1643                 break;
1644         case VHOST_SET_VRING_KICK:
1645                 if (copy_from_user(&f, argp, sizeof f)) {
1646                         r = -EFAULT;
1647                         break;
1648                 }
1649                 eventfp = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_fget(f.fd);
1650                 if (IS_ERR(eventfp)) {
1651                         r = PTR_ERR(eventfp);
1652                         break;
1653                 }
1654                 if (eventfp != vq->kick) {
1655                         pollstop = (filep = vq->kick) != NULL;
1656                         pollstart = (vq->kick = eventfp) != NULL;
1657                 } else
1658                         filep = eventfp;
1659                 break;
1660         case VHOST_SET_VRING_CALL:
1661                 if (copy_from_user(&f, argp, sizeof f)) {
1662                         r = -EFAULT;
1663                         break;
1664                 }
1665                 ctx = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(f.fd);
1666                 if (IS_ERR(ctx)) {
1667                         r = PTR_ERR(ctx);
1668                         break;
1669                 }
1670
1671                 swap(ctx, vq->call_ctx.ctx);
1672                 break;
1673         case VHOST_SET_VRING_ERR:
1674                 if (copy_from_user(&f, argp, sizeof f)) {
1675                         r = -EFAULT;
1676                         break;
1677                 }
1678                 ctx = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(f.fd);
1679                 if (IS_ERR(ctx)) {
1680                         r = PTR_ERR(ctx);
1681                         break;
1682                 }
1683                 swap(ctx, vq->error_ctx);
1684                 break;
1685         case VHOST_SET_VRING_ENDIAN:
1686                 r = vhost_set_vring_endian(vq, argp);
1687                 break;
1688         case VHOST_GET_VRING_ENDIAN:
1689                 r = vhost_get_vring_endian(vq, idx, argp);
1690                 break;
1691         case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
1692                 if (copy_from_user(&s, argp, sizeof(s))) {
1693                         r = -EFAULT;
1694                         break;
1695                 }
1696                 vq->busyloop_timeout = s.num;
1697                 break;
1698         case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
1699                 s.index = idx;
1700                 s.num = vq->busyloop_timeout;
1701                 if (copy_to_user(argp, &s, sizeof(s)))
1702                         r = -EFAULT;
1703                 break;
1704         default:
1705                 r = -ENOIOCTLCMD;
1706         }
1707
1708         if (pollstop && vq->handle_kick)
1709                 vhost_poll_stop(&vq->poll);
1710
1711         if (!IS_ERR_OR_NULL(ctx))
1712                 eventfd_ctx_put(ctx);
1713         if (filep)
1714                 fput(filep);
1715
1716         if (pollstart && vq->handle_kick)
1717                 r = vhost_poll_start(&vq->poll, vq->kick);
1718
1719         mutex_unlock(&vq->mutex);
1720
1721         if (pollstop && vq->handle_kick)
1722                 vhost_poll_flush(&vq->poll);
1723         return r;
1724 }
1725 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
1726
1727 int vhost_init_device_iotlb(struct vhost_dev *d, bool enabled)
1728 {
1729         struct vhost_iotlb *niotlb, *oiotlb;
1730         int i;
1731
1732         niotlb = iotlb_alloc();
1733         if (!niotlb)
1734                 return -ENOMEM;
1735
1736         oiotlb = d->iotlb;
1737         d->iotlb = niotlb;
1738
1739         for (i = 0; i < d->nvqs; ++i) {
1740                 struct vhost_virtqueue *vq = d->vqs[i];
1741
1742                 mutex_lock(&vq->mutex);
1743                 vq->iotlb = niotlb;
1744                 __vhost_vq_meta_reset(vq);
1745                 mutex_unlock(&vq->mutex);
1746         }
1747
1748         vhost_iotlb_free(oiotlb);
1749
1750         return 0;
1751 }
1752 EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
1753
1754 /* Caller must have device mutex */
1755 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1756 {
1757         struct eventfd_ctx *ctx;
1758         u64 p;
1759         long r;
1760         int i, fd;
1761
1762         /* If you are not the owner, you can become one */
1763         if (ioctl == VHOST_SET_OWNER) {
1764                 r = vhost_dev_set_owner(d);
1765                 goto done;
1766         }
1767
1768         /* You must be the owner to do anything else */
1769         r = vhost_dev_check_owner(d);
1770         if (r)
1771                 goto done;
1772
1773         switch (ioctl) {
1774         case VHOST_SET_MEM_TABLE:
1775                 r = vhost_set_memory(d, argp);
1776                 break;
1777         case VHOST_SET_LOG_BASE:
1778                 if (copy_from_user(&p, argp, sizeof p)) {
1779                         r = -EFAULT;
1780                         break;
1781                 }
1782                 if ((u64)(unsigned long)p != p) {
1783                         r = -EFAULT;
1784                         break;
1785                 }
1786                 for (i = 0; i < d->nvqs; ++i) {
1787                         struct vhost_virtqueue *vq;
1788                         void __user *base = (void __user *)(unsigned long)p;
1789                         vq = d->vqs[i];
1790                         mutex_lock(&vq->mutex);
1791                         /* If ring is inactive, will check when it's enabled. */
1792                         if (vq->private_data && !vq_log_access_ok(vq, base))
1793                                 r = -EFAULT;
1794                         else
1795                                 vq->log_base = base;
1796                         mutex_unlock(&vq->mutex);
1797                 }
1798                 break;
1799         case VHOST_SET_LOG_FD:
1800                 r = get_user(fd, (int __user *)argp);
1801                 if (r < 0)
1802                         break;
1803                 ctx = fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(fd);
1804                 if (IS_ERR(ctx)) {
1805                         r = PTR_ERR(ctx);
1806                         break;
1807                 }
1808                 swap(ctx, d->log_ctx);
1809                 for (i = 0; i < d->nvqs; ++i) {
1810                         mutex_lock(&d->vqs[i]->mutex);
1811                         d->vqs[i]->log_ctx = d->log_ctx;
1812                         mutex_unlock(&d->vqs[i]->mutex);
1813                 }
1814                 if (ctx)
1815                         eventfd_ctx_put(ctx);
1816                 break;
1817         default:
1818                 r = -ENOIOCTLCMD;
1819                 break;
1820         }
1821 done:
1822         return r;
1823 }
1824 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1825
1826 /* TODO: This is really inefficient.  We need something like get_user()
1827  * (instruction directly accesses the data, with an exception table entry
1828  * returning -EFAULT). See Documentation/x86/exception-tables.rst.
1829  */
1830 static int set_bit_to_user(int nr, void __user *addr)
1831 {
1832         unsigned long log = (unsigned long)addr;
1833         struct page *page;
1834         void *base;
1835         int bit = nr + (log % PAGE_SIZE) * 8;
1836         int r;
1837
1838         r = pin_user_pages_fast(log, 1, FOLL_WRITE, &page);
1839         if (r < 0)
1840                 return r;
1841         BUG_ON(r != 1);
1842         base = kmap_atomic(page);
1843         set_bit(bit, base);
1844         kunmap_atomic(base);
1845         unpin_user_pages_dirty_lock(&page, 1, true);
1846         return 0;
1847 }
1848
1849 static int log_write(void __user *log_base,
1850                      u64 write_address, u64 write_length)
1851 {
1852         u64 write_page = write_address / VHOST_PAGE_SIZE;
1853         int r;
1854
1855         if (!write_length)
1856                 return 0;
1857         write_length += write_address % VHOST_PAGE_SIZE;
1858         for (;;) {
1859                 u64 base = (u64)(unsigned long)log_base;
1860                 u64 log = base + write_page / 8;
1861                 int bit = write_page % 8;
1862                 if ((u64)(unsigned long)log != log)
1863                         return -EFAULT;
1864                 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1865                 if (r < 0)
1866                         return r;
1867                 if (write_length <= VHOST_PAGE_SIZE)
1868                         break;
1869                 write_length -= VHOST_PAGE_SIZE;
1870                 write_page += 1;
1871         }
1872         return r;
1873 }
1874
1875 static int log_write_hva(struct vhost_virtqueue *vq, u64 hva, u64 len)
1876 {
1877         struct vhost_iotlb *umem = vq->umem;
1878         struct vhost_iotlb_map *u;
1879         u64 start, end, l, min;
1880         int r;
1881         bool hit = false;
1882
1883         while (len) {
1884                 min = len;
1885                 /* More than one GPAs can be mapped into a single HVA. So
1886                  * iterate all possible umems here to be safe.
1887                  */
1888                 list_for_each_entry(u, &umem->list, link) {
1889                         if (u->addr > hva - 1 + len ||
1890                             u->addr - 1 + u->size < hva)
1891                                 continue;
1892                         start = max(u->addr, hva);
1893                         end = min(u->addr - 1 + u->size, hva - 1 + len);
1894                         l = end - start + 1;
1895                         r = log_write(vq->log_base,
1896                                       u->start + start - u->addr,
1897                                       l);
1898                         if (r < 0)
1899                                 return r;
1900                         hit = true;
1901                         min = min(l, min);
1902                 }
1903
1904                 if (!hit)
1905                         return -EFAULT;
1906
1907                 len -= min;
1908                 hva += min;
1909         }
1910
1911         return 0;
1912 }
1913
1914 static int log_used(struct vhost_virtqueue *vq, u64 used_offset, u64 len)
1915 {
1916         struct iovec *iov = vq->log_iov;
1917         int i, ret;
1918
1919         if (!vq->iotlb)
1920                 return log_write(vq->log_base, vq->log_addr + used_offset, len);
1921
1922         ret = translate_desc(vq, (uintptr_t)vq->used + used_offset,
1923                              len, iov, 64, VHOST_ACCESS_WO);
1924         if (ret < 0)
1925                 return ret;
1926
1927         for (i = 0; i < ret; i++) {
1928                 ret = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1929                                     iov[i].iov_len);
1930                 if (ret)
1931                         return ret;
1932         }
1933
1934         return 0;
1935 }
1936
1937 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1938                     unsigned int log_num, u64 len, struct iovec *iov, int count)
1939 {
1940         int i, r;
1941
1942         /* Make sure data written is seen before log. */
1943         smp_wmb();
1944
1945         if (vq->iotlb) {
1946                 for (i = 0; i < count; i++) {
1947                         r = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1948                                           iov[i].iov_len);
1949                         if (r < 0)
1950                                 return r;
1951                 }
1952                 return 0;
1953         }
1954
1955         for (i = 0; i < log_num; ++i) {
1956                 u64 l = min(log[i].len, len);
1957                 r = log_write(vq->log_base, log[i].addr, l);
1958                 if (r < 0)
1959                         return r;
1960                 len -= l;
1961                 if (!len) {
1962                         if (vq->log_ctx)
1963                                 eventfd_signal(vq->log_ctx, 1);
1964                         return 0;
1965                 }
1966         }
1967         /* Length written exceeds what we have stored. This is a bug. */
1968         BUG();
1969         return 0;
1970 }
1971 EXPORT_SYMBOL_GPL(vhost_log_write);
1972
1973 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1974 {
1975         void __user *used;
1976         if (vhost_put_used_flags(vq))
1977                 return -EFAULT;
1978         if (unlikely(vq->log_used)) {
1979                 /* Make sure the flag is seen before log. */
1980                 smp_wmb();
1981                 /* Log used flag write. */
1982                 used = &vq->used->flags;
1983                 log_used(vq, (used - (void __user *)vq->used),
1984                          sizeof vq->used->flags);
1985                 if (vq->log_ctx)
1986                         eventfd_signal(vq->log_ctx, 1);
1987         }
1988         return 0;
1989 }
1990
1991 static int vhost_update_avail_event(struct vhost_virtqueue *vq)
1992 {
1993         if (vhost_put_avail_event(vq))
1994                 return -EFAULT;
1995         if (unlikely(vq->log_used)) {
1996                 void __user *used;
1997                 /* Make sure the event is seen before log. */
1998                 smp_wmb();
1999                 /* Log avail event write */
2000                 used = vhost_avail_event(vq);
2001                 log_used(vq, (used - (void __user *)vq->used),
2002                          sizeof *vhost_avail_event(vq));
2003                 if (vq->log_ctx)
2004                         eventfd_signal(vq->log_ctx, 1);
2005         }
2006         return 0;
2007 }
2008
2009 int vhost_vq_init_access(struct vhost_virtqueue *vq)
2010 {
2011         __virtio16 last_used_idx;
2012         int r;
2013         bool is_le = vq->is_le;
2014
2015         if (!vq->private_data)
2016                 return 0;
2017
2018         vhost_init_is_le(vq);
2019
2020         r = vhost_update_used_flags(vq);
2021         if (r)
2022                 goto err;
2023         vq->signalled_used_valid = false;
2024         if (!vq->iotlb &&
2025             !access_ok(&vq->used->idx, sizeof vq->used->idx)) {
2026                 r = -EFAULT;
2027                 goto err;
2028         }
2029         r = vhost_get_used_idx(vq, &last_used_idx);
2030         if (r) {
2031                 vq_err(vq, "Can't access used idx at %p\n",
2032                        &vq->used->idx);
2033                 goto err;
2034         }
2035         vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
2036         return 0;
2037
2038 err:
2039         vq->is_le = is_le;
2040         return r;
2041 }
2042 EXPORT_SYMBOL_GPL(vhost_vq_init_access);
2043
2044 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
2045                           struct iovec iov[], int iov_size, int access)
2046 {
2047         const struct vhost_iotlb_map *map;
2048         struct vhost_dev *dev = vq->dev;
2049         struct vhost_iotlb *umem = dev->iotlb ? dev->iotlb : dev->umem;
2050         struct iovec *_iov;
2051         u64 s = 0;
2052         int ret = 0;
2053
2054         while ((u64)len > s) {
2055                 u64 size;
2056                 if (unlikely(ret >= iov_size)) {
2057                         ret = -ENOBUFS;
2058                         break;
2059                 }
2060
2061                 map = vhost_iotlb_itree_first(umem, addr, addr + len - 1);
2062                 if (map == NULL || map->start > addr) {
2063                         if (umem != dev->iotlb) {
2064                                 ret = -EFAULT;
2065                                 break;
2066                         }
2067                         ret = -EAGAIN;
2068                         break;
2069                 } else if (!(map->perm & access)) {
2070                         ret = -EPERM;
2071                         break;
2072                 }
2073
2074                 _iov = iov + ret;
2075                 size = map->size - addr + map->start;
2076                 _iov->iov_len = min((u64)len - s, size);
2077                 _iov->iov_base = (void __user *)(unsigned long)
2078                                  (map->addr + addr - map->start);
2079                 s += size;
2080                 addr += size;
2081                 ++ret;
2082         }
2083
2084         if (ret == -EAGAIN)
2085                 vhost_iotlb_miss(vq, addr, access);
2086         return ret;
2087 }
2088
2089 /* Each buffer in the virtqueues is actually a chain of descriptors.  This
2090  * function returns the next descriptor in the chain,
2091  * or -1U if we're at the end. */
2092 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
2093 {
2094         unsigned int next;
2095
2096         /* If this descriptor says it doesn't chain, we're done. */
2097         if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
2098                 return -1U;
2099
2100         /* Check they're not leading us off end of descriptors. */
2101         next = vhost16_to_cpu(vq, READ_ONCE(desc->next));
2102         return next;
2103 }
2104
2105 static int get_indirect(struct vhost_virtqueue *vq,
2106                         struct iovec iov[], unsigned int iov_size,
2107                         unsigned int *out_num, unsigned int *in_num,
2108                         struct vhost_log *log, unsigned int *log_num,
2109                         struct vring_desc *indirect)
2110 {
2111         struct vring_desc desc;
2112         unsigned int i = 0, count, found = 0;
2113         u32 len = vhost32_to_cpu(vq, indirect->len);
2114         struct iov_iter from;
2115         int ret, access;
2116
2117         /* Sanity check */
2118         if (unlikely(len % sizeof desc)) {
2119                 vq_err(vq, "Invalid length in indirect descriptor: "
2120                        "len 0x%llx not multiple of 0x%zx\n",
2121                        (unsigned long long)len,
2122                        sizeof desc);
2123                 return -EINVAL;
2124         }
2125
2126         ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
2127                              UIO_MAXIOV, VHOST_ACCESS_RO);
2128         if (unlikely(ret < 0)) {
2129                 if (ret != -EAGAIN)
2130                         vq_err(vq, "Translation failure %d in indirect.\n", ret);
2131                 return ret;
2132         }
2133         iov_iter_init(&from, READ, vq->indirect, ret, len);
2134         count = len / sizeof desc;
2135         /* Buffers are chained via a 16 bit next field, so
2136          * we can have at most 2^16 of these. */
2137         if (unlikely(count > USHRT_MAX + 1)) {
2138                 vq_err(vq, "Indirect buffer length too big: %d\n",
2139                        indirect->len);
2140                 return -E2BIG;
2141         }
2142
2143         do {
2144                 unsigned iov_count = *in_num + *out_num;
2145                 if (unlikely(++found > count)) {
2146                         vq_err(vq, "Loop detected: last one at %u "
2147                                "indirect size %u\n",
2148                                i, count);
2149                         return -EINVAL;
2150                 }
2151                 if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
2152                         vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
2153                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2154                         return -EINVAL;
2155                 }
2156                 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
2157                         vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
2158                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2159                         return -EINVAL;
2160                 }
2161
2162                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2163                         access = VHOST_ACCESS_WO;
2164                 else
2165                         access = VHOST_ACCESS_RO;
2166
2167                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2168                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
2169                                      iov_size - iov_count, access);
2170                 if (unlikely(ret < 0)) {
2171                         if (ret != -EAGAIN)
2172                                 vq_err(vq, "Translation failure %d indirect idx %d\n",
2173                                         ret, i);
2174                         return ret;
2175                 }
2176                 /* If this is an input descriptor, increment that count. */
2177                 if (access == VHOST_ACCESS_WO) {
2178                         *in_num += ret;
2179                         if (unlikely(log && ret)) {
2180                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2181                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2182                                 ++*log_num;
2183                         }
2184                 } else {
2185                         /* If it's an output descriptor, they're all supposed
2186                          * to come before any input descriptors. */
2187                         if (unlikely(*in_num)) {
2188                                 vq_err(vq, "Indirect descriptor "
2189                                        "has out after in: idx %d\n", i);
2190                                 return -EINVAL;
2191                         }
2192                         *out_num += ret;
2193                 }
2194         } while ((i = next_desc(vq, &desc)) != -1);
2195         return 0;
2196 }
2197
2198 /* This looks in the virtqueue and for the first available buffer, and converts
2199  * it to an iovec for convenient access.  Since descriptors consist of some
2200  * number of output then some number of input descriptors, it's actually two
2201  * iovecs, but we pack them into one and note how many of each there were.
2202  *
2203  * This function returns the descriptor number found, or vq->num (which is
2204  * never a valid descriptor number) if none was found.  A negative code is
2205  * returned on error. */
2206 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
2207                       struct iovec iov[], unsigned int iov_size,
2208                       unsigned int *out_num, unsigned int *in_num,
2209                       struct vhost_log *log, unsigned int *log_num)
2210 {
2211         struct vring_desc desc;
2212         unsigned int i, head, found = 0;
2213         u16 last_avail_idx;
2214         __virtio16 avail_idx;
2215         __virtio16 ring_head;
2216         int ret, access;
2217
2218         /* Check it isn't doing very strange things with descriptor numbers. */
2219         last_avail_idx = vq->last_avail_idx;
2220
2221         if (vq->avail_idx == vq->last_avail_idx) {
2222                 if (unlikely(vhost_get_avail_idx(vq, &avail_idx))) {
2223                         vq_err(vq, "Failed to access avail idx at %p\n",
2224                                 &vq->avail->idx);
2225                         return -EFAULT;
2226                 }
2227                 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2228
2229                 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
2230                         vq_err(vq, "Guest moved used index from %u to %u",
2231                                 last_avail_idx, vq->avail_idx);
2232                         return -EFAULT;
2233                 }
2234
2235                 /* If there's nothing new since last we looked, return
2236                  * invalid.
2237                  */
2238                 if (vq->avail_idx == last_avail_idx)
2239                         return vq->num;
2240
2241                 /* Only get avail ring entries after they have been
2242                  * exposed by guest.
2243                  */
2244                 smp_rmb();
2245         }
2246
2247         /* Grab the next descriptor number they're advertising, and increment
2248          * the index we've seen. */
2249         if (unlikely(vhost_get_avail_head(vq, &ring_head, last_avail_idx))) {
2250                 vq_err(vq, "Failed to read head: idx %d address %p\n",
2251                        last_avail_idx,
2252                        &vq->avail->ring[last_avail_idx % vq->num]);
2253                 return -EFAULT;
2254         }
2255
2256         head = vhost16_to_cpu(vq, ring_head);
2257
2258         /* If their number is silly, that's an error. */
2259         if (unlikely(head >= vq->num)) {
2260                 vq_err(vq, "Guest says index %u > %u is available",
2261                        head, vq->num);
2262                 return -EINVAL;
2263         }
2264
2265         /* When we start there are none of either input nor output. */
2266         *out_num = *in_num = 0;
2267         if (unlikely(log))
2268                 *log_num = 0;
2269
2270         i = head;
2271         do {
2272                 unsigned iov_count = *in_num + *out_num;
2273                 if (unlikely(i >= vq->num)) {
2274                         vq_err(vq, "Desc index is %u > %u, head = %u",
2275                                i, vq->num, head);
2276                         return -EINVAL;
2277                 }
2278                 if (unlikely(++found > vq->num)) {
2279                         vq_err(vq, "Loop detected: last one at %u "
2280                                "vq size %u head %u\n",
2281                                i, vq->num, head);
2282                         return -EINVAL;
2283                 }
2284                 ret = vhost_get_desc(vq, &desc, i);
2285                 if (unlikely(ret)) {
2286                         vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
2287                                i, vq->desc + i);
2288                         return -EFAULT;
2289                 }
2290                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2291                         ret = get_indirect(vq, iov, iov_size,
2292                                            out_num, in_num,
2293                                            log, log_num, &desc);
2294                         if (unlikely(ret < 0)) {
2295                                 if (ret != -EAGAIN)
2296                                         vq_err(vq, "Failure detected "
2297                                                 "in indirect descriptor at idx %d\n", i);
2298                                 return ret;
2299                         }
2300                         continue;
2301                 }
2302
2303                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2304                         access = VHOST_ACCESS_WO;
2305                 else
2306                         access = VHOST_ACCESS_RO;
2307                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2308                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
2309                                      iov_size - iov_count, access);
2310                 if (unlikely(ret < 0)) {
2311                         if (ret != -EAGAIN)
2312                                 vq_err(vq, "Translation failure %d descriptor idx %d\n",
2313                                         ret, i);
2314                         return ret;
2315                 }
2316                 if (access == VHOST_ACCESS_WO) {
2317                         /* If this is an input descriptor,
2318                          * increment that count. */
2319                         *in_num += ret;
2320                         if (unlikely(log && ret)) {
2321                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2322                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2323                                 ++*log_num;
2324                         }
2325                 } else {
2326                         /* If it's an output descriptor, they're all supposed
2327                          * to come before any input descriptors. */
2328                         if (unlikely(*in_num)) {
2329                                 vq_err(vq, "Descriptor has out after in: "
2330                                        "idx %d\n", i);
2331                                 return -EINVAL;
2332                         }
2333                         *out_num += ret;
2334                 }
2335         } while ((i = next_desc(vq, &desc)) != -1);
2336
2337         /* On success, increment avail index. */
2338         vq->last_avail_idx++;
2339
2340         /* Assume notifications from guest are disabled at this point,
2341          * if they aren't we would need to update avail_event index. */
2342         BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2343         return head;
2344 }
2345 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2346
2347 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2348 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2349 {
2350         vq->last_avail_idx -= n;
2351 }
2352 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2353
2354 /* After we've used one of their buffers, we tell them about it.  We'll then
2355  * want to notify the guest, using eventfd. */
2356 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2357 {
2358         struct vring_used_elem heads = {
2359                 cpu_to_vhost32(vq, head),
2360                 cpu_to_vhost32(vq, len)
2361         };
2362
2363         return vhost_add_used_n(vq, &heads, 1);
2364 }
2365 EXPORT_SYMBOL_GPL(vhost_add_used);
2366
2367 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2368                             struct vring_used_elem *heads,
2369                             unsigned count)
2370 {
2371         vring_used_elem_t __user *used;
2372         u16 old, new;
2373         int start;
2374
2375         start = vq->last_used_idx & (vq->num - 1);
2376         used = vq->used->ring + start;
2377         if (vhost_put_used(vq, heads, start, count)) {
2378                 vq_err(vq, "Failed to write used");
2379                 return -EFAULT;
2380         }
2381         if (unlikely(vq->log_used)) {
2382                 /* Make sure data is seen before log. */
2383                 smp_wmb();
2384                 /* Log used ring entry write. */
2385                 log_used(vq, ((void __user *)used - (void __user *)vq->used),
2386                          count * sizeof *used);
2387         }
2388         old = vq->last_used_idx;
2389         new = (vq->last_used_idx += count);
2390         /* If the driver never bothers to signal in a very long while,
2391          * used index might wrap around. If that happens, invalidate
2392          * signalled_used index we stored. TODO: make sure driver
2393          * signals at least once in 2^16 and remove this. */
2394         if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2395                 vq->signalled_used_valid = false;
2396         return 0;
2397 }
2398
2399 /* After we've used one of their buffers, we tell them about it.  We'll then
2400  * want to notify the guest, using eventfd. */
2401 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2402                      unsigned count)
2403 {
2404         int start, n, r;
2405
2406         start = vq->last_used_idx & (vq->num - 1);
2407         n = vq->num - start;
2408         if (n < count) {
2409                 r = __vhost_add_used_n(vq, heads, n);
2410                 if (r < 0)
2411                         return r;
2412                 heads += n;
2413                 count -= n;
2414         }
2415         r = __vhost_add_used_n(vq, heads, count);
2416
2417         /* Make sure buffer is written before we update index. */
2418         smp_wmb();
2419         if (vhost_put_used_idx(vq)) {
2420                 vq_err(vq, "Failed to increment used idx");
2421                 return -EFAULT;
2422         }
2423         if (unlikely(vq->log_used)) {
2424                 /* Make sure used idx is seen before log. */
2425                 smp_wmb();
2426                 /* Log used index update. */
2427                 log_used(vq, offsetof(struct vring_used, idx),
2428                          sizeof vq->used->idx);
2429                 if (vq->log_ctx)
2430                         eventfd_signal(vq->log_ctx, 1);
2431         }
2432         return r;
2433 }
2434 EXPORT_SYMBOL_GPL(vhost_add_used_n);
2435
2436 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2437 {
2438         __u16 old, new;
2439         __virtio16 event;
2440         bool v;
2441         /* Flush out used index updates. This is paired
2442          * with the barrier that the Guest executes when enabling
2443          * interrupts. */
2444         smp_mb();
2445
2446         if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2447             unlikely(vq->avail_idx == vq->last_avail_idx))
2448                 return true;
2449
2450         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2451                 __virtio16 flags;
2452                 if (vhost_get_avail_flags(vq, &flags)) {
2453                         vq_err(vq, "Failed to get flags");
2454                         return true;
2455                 }
2456                 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2457         }
2458         old = vq->signalled_used;
2459         v = vq->signalled_used_valid;
2460         new = vq->signalled_used = vq->last_used_idx;
2461         vq->signalled_used_valid = true;
2462
2463         if (unlikely(!v))
2464                 return true;
2465
2466         if (vhost_get_used_event(vq, &event)) {
2467                 vq_err(vq, "Failed to get used event idx");
2468                 return true;
2469         }
2470         return vring_need_event(vhost16_to_cpu(vq, event), new, old);
2471 }
2472
2473 /* This actually signals the guest, using eventfd. */
2474 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2475 {
2476         /* Signal the Guest tell them we used something up. */
2477         if (vq->call_ctx.ctx && vhost_notify(dev, vq))
2478                 eventfd_signal(vq->call_ctx.ctx, 1);
2479 }
2480 EXPORT_SYMBOL_GPL(vhost_signal);
2481
2482 /* And here's the combo meal deal.  Supersize me! */
2483 void vhost_add_used_and_signal(struct vhost_dev *dev,
2484                                struct vhost_virtqueue *vq,
2485                                unsigned int head, int len)
2486 {
2487         vhost_add_used(vq, head, len);
2488         vhost_signal(dev, vq);
2489 }
2490 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2491
2492 /* multi-buffer version of vhost_add_used_and_signal */
2493 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2494                                  struct vhost_virtqueue *vq,
2495                                  struct vring_used_elem *heads, unsigned count)
2496 {
2497         vhost_add_used_n(vq, heads, count);
2498         vhost_signal(dev, vq);
2499 }
2500 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2501
2502 /* return true if we're sure that avaiable ring is empty */
2503 bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2504 {
2505         __virtio16 avail_idx;
2506         int r;
2507
2508         if (vq->avail_idx != vq->last_avail_idx)
2509                 return false;
2510
2511         r = vhost_get_avail_idx(vq, &avail_idx);
2512         if (unlikely(r))
2513                 return false;
2514         vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2515
2516         return vq->avail_idx == vq->last_avail_idx;
2517 }
2518 EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2519
2520 /* OK, now we need to know about added descriptors. */
2521 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2522 {
2523         __virtio16 avail_idx;
2524         int r;
2525
2526         if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2527                 return false;
2528         vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2529         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2530                 r = vhost_update_used_flags(vq);
2531                 if (r) {
2532                         vq_err(vq, "Failed to enable notification at %p: %d\n",
2533                                &vq->used->flags, r);
2534                         return false;
2535                 }
2536         } else {
2537                 r = vhost_update_avail_event(vq);
2538                 if (r) {
2539                         vq_err(vq, "Failed to update avail event index at %p: %d\n",
2540                                vhost_avail_event(vq), r);
2541                         return false;
2542                 }
2543         }
2544         /* They could have slipped one in as we were doing that: make
2545          * sure it's written, then check again. */
2546         smp_mb();
2547         r = vhost_get_avail_idx(vq, &avail_idx);
2548         if (r) {
2549                 vq_err(vq, "Failed to check avail idx at %p: %d\n",
2550                        &vq->avail->idx, r);
2551                 return false;
2552         }
2553         vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2554
2555         return vq->avail_idx != vq->last_avail_idx;
2556 }
2557 EXPORT_SYMBOL_GPL(vhost_enable_notify);
2558
2559 /* We don't need to be notified again. */
2560 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2561 {
2562         int r;
2563
2564         if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2565                 return;
2566         vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2567         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2568                 r = vhost_update_used_flags(vq);
2569                 if (r)
2570                         vq_err(vq, "Failed to disable notification at %p: %d\n",
2571                                &vq->used->flags, r);
2572         }
2573 }
2574 EXPORT_SYMBOL_GPL(vhost_disable_notify);
2575
2576 /* Create a new message. */
2577 struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2578 {
2579         struct vhost_msg_node *node = kmalloc(sizeof *node, GFP_KERNEL);
2580         if (!node)
2581                 return NULL;
2582
2583         /* Make sure all padding within the structure is initialized. */
2584         memset(&node->msg, 0, sizeof node->msg);
2585         node->vq = vq;
2586         node->msg.type = type;
2587         return node;
2588 }
2589 EXPORT_SYMBOL_GPL(vhost_new_msg);
2590
2591 void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2592                        struct vhost_msg_node *node)
2593 {
2594         spin_lock(&dev->iotlb_lock);
2595         list_add_tail(&node->node, head);
2596         spin_unlock(&dev->iotlb_lock);
2597
2598         wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
2599 }
2600 EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2601
2602 struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2603                                          struct list_head *head)
2604 {
2605         struct vhost_msg_node *node = NULL;
2606
2607         spin_lock(&dev->iotlb_lock);
2608         if (!list_empty(head)) {
2609                 node = list_first_entry(head, struct vhost_msg_node,
2610                                         node);
2611                 list_del(&node->node);
2612         }
2613         spin_unlock(&dev->iotlb_lock);
2614
2615         return node;
2616 }
2617 EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2618
2619 void vhost_set_backend_features(struct vhost_dev *dev, u64 features)
2620 {
2621         struct vhost_virtqueue *vq;
2622         int i;
2623
2624         mutex_lock(&dev->mutex);
2625         for (i = 0; i < dev->nvqs; ++i) {
2626                 vq = dev->vqs[i];
2627                 mutex_lock(&vq->mutex);
2628                 vq->acked_backend_features = features;
2629                 mutex_unlock(&vq->mutex);
2630         }
2631         mutex_unlock(&dev->mutex);
2632 }
2633 EXPORT_SYMBOL_GPL(vhost_set_backend_features);
2634
2635 static int __init vhost_init(void)
2636 {
2637         return 0;
2638 }
2639
2640 static void __exit vhost_exit(void)
2641 {
2642 }
2643
2644 module_init(vhost_init);
2645 module_exit(vhost_exit);
2646
2647 MODULE_VERSION("0.0.1");
2648 MODULE_LICENSE("GPL v2");
2649 MODULE_AUTHOR("Michael S. Tsirkin");
2650 MODULE_DESCRIPTION("Host kernel accelerator for virtio");