drivers/vhost/vhost.c

   1 /* Copyright (C) 2009 Red Hat, Inc.
   2  * Copyright (C) 2006 Rusty Russell IBM Corporation
   3  *
   4  * Author: Michael S. Tsirkin <mst@redhat.com>
   5  *
   6  * Inspiration, some code, and most witty comments come from
   7  * Documentation/virtual/lguest/lguest.c, by Rusty Russell
   8  *
   9  * This work is licensed under the terms of the GNU GPL, version 2.
  10  *
  11  * Generic code for virtio server in host kernel.
  12  */
  13
  14 #include <linux/eventfd.h>
  15 #include <linux/vhost.h>
  16 #include <linux/uio.h>
  17 #include <linux/mm.h>
  18 #include <linux/mmu_context.h>
  19 #include <linux/miscdevice.h>
  20 #include <linux/mutex.h>
  21 #include <linux/poll.h>
  22 #include <linux/file.h>
  23 #include <linux/highmem.h>
  24 #include <linux/slab.h>
  25 #include <linux/vmalloc.h>
  26 #include <linux/kthread.h>
  27 #include <linux/cgroup.h>
  28 #include <linux/module.h>
  29 #include <linux/sort.h>
  30 #include <linux/sched/mm.h>
  31 #include <linux/sched/signal.h>
  32 #include <linux/interval_tree_generic.h>
  33 #include <linux/nospec.h>
  34
  35 #include "vhost.h"
  36
  37 static ushort max_mem_regions = 64;
  38 module_param(max_mem_regions, ushort, 0444);
  39 MODULE_PARM_DESC(max_mem_regions,
  40         "Maximum number of memory regions in memory map. (default: 64)");
  41 static int max_iotlb_entries = 2048;
  42 module_param(max_iotlb_entries, int, 0444);
  43 MODULE_PARM_DESC(max_iotlb_entries,
  44         "Maximum number of iotlb entries. (default: 2048)");
  45
  46 enum {
  47         VHOST_MEMORY_F_LOG = 0x1,
  48 };
  49
  50 #define vhost_used_event(vq) ((__virtio16 __user *)&vq->avail->ring[vq->num])
  51 #define vhost_avail_event(vq) ((__virtio16 __user *)&vq->used->ring[vq->num])
  52
  53 INTERVAL_TREE_DEFINE(struct vhost_umem_node,
  54                      rb, __u64, __subtree_last,
  55                      START, LAST, static inline, vhost_umem_interval_tree);
  56
  57 #ifdef CONFIG_VHOST_CROSS_ENDIAN_LEGACY
  58 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
  59 {
  60         vq->user_be = !virtio_legacy_is_little_endian();
  61 }
  62
  63 static void vhost_enable_cross_endian_big(struct vhost_virtqueue *vq)
  64 {
  65         vq->user_be = true;
  66 }
  67
  68 static void vhost_enable_cross_endian_little(struct vhost_virtqueue *vq)
  69 {
  70         vq->user_be = false;
  71 }
  72
  73 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
  74 {
  75         struct vhost_vring_state s;
  76
  77         if (vq->private_data)
  78                 return -EBUSY;
  79
  80         if (copy_from_user(&s, argp, sizeof(s)))
  81                 return -EFAULT;
  82
  83         if (s.num != VHOST_VRING_LITTLE_ENDIAN &&
  84             s.num != VHOST_VRING_BIG_ENDIAN)
  85                 return -EINVAL;
  86
  87         if (s.num == VHOST_VRING_BIG_ENDIAN)
  88                 vhost_enable_cross_endian_big(vq);
  89         else
  90                 vhost_enable_cross_endian_little(vq);
  91
  92         return 0;
  93 }
  94
  95 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
  96                                    int __user *argp)
  97 {
  98         struct vhost_vring_state s = {
  99                 .index = idx,
 100                 .num = vq->user_be
 101         };
 102
 103         if (copy_to_user(argp, &s, sizeof(s)))
 104                 return -EFAULT;
 105
 106         return 0;
 107 }
 108
 109 static void vhost_init_is_le(struct vhost_virtqueue *vq)
 110 {
 111         /* Note for legacy virtio: user_be is initialized at reset time
 112          * according to the host endianness. If userspace does not set an
 113          * explicit endianness, the default behavior is native endian, as
 114          * expected by legacy virtio.
 115          */
 116         vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1) || !vq->user_be;
 117 }
 118 #else
 119 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
 120 {
 121 }
 122
 123 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
 124 {
 125         return -ENOIOCTLCMD;
 126 }
 127
 128 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
 129                                    int __user *argp)
 130 {
 131         return -ENOIOCTLCMD;
 132 }
 133
 134 static void vhost_init_is_le(struct vhost_virtqueue *vq)
 135 {
 136         vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1)
 137                 || virtio_legacy_is_little_endian();
 138 }
 139 #endif /* CONFIG_VHOST_CROSS_ENDIAN_LEGACY */
 140
 141 static void vhost_reset_is_le(struct vhost_virtqueue *vq)
 142 {
 143         vhost_init_is_le(vq);
 144 }
 145
 146 struct vhost_flush_struct {
 147         struct vhost_work work;
 148         struct completion wait_event;
 149 };
 150
 151 static void vhost_flush_work(struct vhost_work *work)
 152 {
 153         struct vhost_flush_struct *s;
 154
 155         s = container_of(work, struct vhost_flush_struct, work);
 156         complete(&s->wait_event);
 157 }
 158
 159 static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
 160                             poll_table *pt)
 161 {
 162         struct vhost_poll *poll;
 163
 164         poll = container_of(pt, struct vhost_poll, table);
 165         poll->wqh = wqh;
 166         add_wait_queue(wqh, &poll->wait);
 167 }
 168
 169 static int vhost_poll_wakeup(wait_queue_entry_t *wait, unsigned mode, int sync,
 170                              void *key)
 171 {
 172         struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
 173
 174         if (!(key_to_poll(key) & poll->mask))
 175                 return 0;
 176
 177         vhost_poll_queue(poll);
 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         int ret = 0;
 208
 209         if (poll->wqh)
 210                 return 0;
 211
 212         mask = vfs_poll(file, &poll->table);
 213         if (mask)
 214                 vhost_poll_wakeup(&poll->wait, 0, 0, poll_to_key(mask));
 215         if (mask & EPOLLERR) {
 216                 vhost_poll_stop(poll);
 217                 ret = -EINVAL;
 218         }
 219
 220         return ret;
 221 }
 222 EXPORT_SYMBOL_GPL(vhost_poll_start);
 223
 224 /* Stop polling a file. After this function returns, it becomes safe to drop the
 225  * file reference. You must also flush afterwards. */
 226 void vhost_poll_stop(struct vhost_poll *poll)
 227 {
 228         if (poll->wqh) {
 229                 remove_wait_queue(poll->wqh, &poll->wait);
 230                 poll->wqh = NULL;
 231         }
 232 }
 233 EXPORT_SYMBOL_GPL(vhost_poll_stop);
 234
 235 void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work)
 236 {
 237         struct vhost_flush_struct flush;
 238
 239         if (dev->worker) {
 240                 init_completion(&flush.wait_event);
 241                 vhost_work_init(&flush.work, vhost_flush_work);
 242
 243                 vhost_work_queue(dev, &flush.work);
 244                 wait_for_completion(&flush.wait_event);
 245         }
 246 }
 247 EXPORT_SYMBOL_GPL(vhost_work_flush);
 248
 249 /* Flush any work that has been scheduled. When calling this, don't hold any
 250  * locks that are also used by the callback. */
 251 void vhost_poll_flush(struct vhost_poll *poll)
 252 {
 253         vhost_work_flush(poll->dev, &poll->work);
 254 }
 255 EXPORT_SYMBOL_GPL(vhost_poll_flush);
 256
 257 void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work)
 258 {
 259         if (!dev->worker)
 260                 return;
 261
 262         if (!test_and_set_bit(VHOST_WORK_QUEUED, &work->flags)) {
 263                 /* We can only add the work to the list after we're
 264                  * sure it was not in the list.
 265                  * test_and_set_bit() implies a memory barrier.
 266                  */
 267                 llist_add(&work->node, &dev->work_list);
 268                 wake_up_process(dev->worker);
 269         }
 270 }
 271 EXPORT_SYMBOL_GPL(vhost_work_queue);
 272
 273 /* A lockless hint for busy polling code to exit the loop */
 274 bool vhost_has_work(struct vhost_dev *dev)
 275 {
 276         return !llist_empty(&dev->work_list);
 277 }
 278 EXPORT_SYMBOL_GPL(vhost_has_work);
 279
 280 void vhost_poll_queue(struct vhost_poll *poll)
 281 {
 282         vhost_work_queue(poll->dev, &poll->work);
 283 }
 284 EXPORT_SYMBOL_GPL(vhost_poll_queue);
 285
 286 static void __vhost_vq_meta_reset(struct vhost_virtqueue *vq)
 287 {
 288         int j;
 289
 290         for (j = 0; j < VHOST_NUM_ADDRS; j++)
 291                 vq->meta_iotlb[j] = NULL;
 292 }
 293
 294 static void vhost_vq_meta_reset(struct vhost_dev *d)
 295 {
 296         int i;
 297
 298         for (i = 0; i < d->nvqs; ++i)
 299                 __vhost_vq_meta_reset(d->vqs[i]);
 300 }
 301
 302 static void vhost_vq_reset(struct vhost_dev *dev,
 303                            struct vhost_virtqueue *vq)
 304 {
 305         vq->num = 1;
 306         vq->desc = NULL;
 307         vq->avail = NULL;
 308         vq->used = NULL;
 309         vq->last_avail_idx = 0;
 310         vq->avail_idx = 0;
 311         vq->last_used_idx = 0;
 312         vq->signalled_used = 0;
 313         vq->signalled_used_valid = false;
 314         vq->used_flags = 0;
 315         vq->log_used = false;
 316         vq->log_addr = -1ull;
 317         vq->private_data = NULL;
 318         vq->acked_features = 0;
 319         vq->acked_backend_features = 0;
 320         vq->log_base = NULL;
 321         vq->error_ctx = NULL;
 322         vq->kick = NULL;
 323         vq->call_ctx = NULL;
 324         vq->log_ctx = NULL;
 325         vhost_reset_is_le(vq);
 326         vhost_disable_cross_endian(vq);
 327         vq->busyloop_timeout = 0;
 328         vq->umem = NULL;
 329         vq->iotlb = NULL;
 330         __vhost_vq_meta_reset(vq);
 331 }
 332
 333 static int vhost_worker(void *data)
 334 {
 335         struct vhost_dev *dev = data;
 336         struct vhost_work *work, *work_next;
 337         struct llist_node *node;
 338         mm_segment_t oldfs = get_fs();
 339
 340         set_fs(USER_DS);
 341         use_mm(dev->mm);
 342
 343         for (;;) {
 344                 /* mb paired w/ kthread_stop */
 345                 set_current_state(TASK_INTERRUPTIBLE);
 346
 347                 if (kthread_should_stop()) {
 348                         __set_current_state(TASK_RUNNING);
 349                         break;
 350                 }
 351
 352                 node = llist_del_all(&dev->work_list);
 353                 if (!node)
 354                         schedule();
 355
 356                 node = llist_reverse_order(node);
 357                 /* make sure flag is seen after deletion */
 358                 smp_wmb();
 359                 llist_for_each_entry_safe(work, work_next, node, node) {
 360                         clear_bit(VHOST_WORK_QUEUED, &work->flags);
 361                         __set_current_state(TASK_RUNNING);
 362                         work->fn(work);
 363                         if (need_resched())
 364                                 schedule();
 365                 }
 366         }
 367         unuse_mm(dev->mm);
 368         set_fs(oldfs);
 369         return 0;
 370 }
 371
 372 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
 373 {
 374         kfree(vq->indirect);
 375         vq->indirect = NULL;
 376         kfree(vq->log);
 377         vq->log = NULL;
 378         kfree(vq->heads);
 379         vq->heads = NULL;
 380 }
 381
 382 /* Helper to allocate iovec buffers for all vqs. */
 383 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
 384 {
 385         struct vhost_virtqueue *vq;
 386         int i;
 387
 388         for (i = 0; i < dev->nvqs; ++i) {
 389                 vq = dev->vqs[i];
 390                 vq->indirect = kmalloc_array(UIO_MAXIOV,
 391                                              sizeof(*vq->indirect),
 392                                              GFP_KERNEL);
 393                 vq->log = kmalloc_array(UIO_MAXIOV, sizeof(*vq->log),
 394                                         GFP_KERNEL);
 395                 vq->heads = kmalloc_array(UIO_MAXIOV, sizeof(*vq->heads),
 396                                           GFP_KERNEL);
 397                 if (!vq->indirect || !vq->log || !vq->heads)
 398                         goto err_nomem;
 399         }
 400         return 0;
 401
 402 err_nomem:
 403         for (; i >= 0; --i)
 404                 vhost_vq_free_iovecs(dev->vqs[i]);
 405         return -ENOMEM;
 406 }
 407
 408 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
 409 {
 410         int i;
 411
 412         for (i = 0; i < dev->nvqs; ++i)
 413                 vhost_vq_free_iovecs(dev->vqs[i]);
 414 }
 415
 416 void vhost_dev_init(struct vhost_dev *dev,
 417                     struct vhost_virtqueue **vqs, int nvqs)
 418 {
 419         struct vhost_virtqueue *vq;
 420         int i;
 421
 422         dev->vqs = vqs;
 423         dev->nvqs = nvqs;
 424         mutex_init(&dev->mutex);
 425         dev->log_ctx = NULL;
 426         dev->umem = NULL;
 427         dev->iotlb = NULL;
 428         dev->mm = NULL;
 429         dev->worker = NULL;
 430         init_llist_head(&dev->work_list);
 431         init_waitqueue_head(&dev->wait);
 432         INIT_LIST_HEAD(&dev->read_list);
 433         INIT_LIST_HEAD(&dev->pending_list);
 434         spin_lock_init(&dev->iotlb_lock);
 435
 436
 437         for (i = 0; i < dev->nvqs; ++i) {
 438                 vq = dev->vqs[i];
 439                 vq->log = NULL;
 440                 vq->indirect = NULL;
 441                 vq->heads = NULL;
 442                 vq->dev = dev;
 443                 mutex_init(&vq->mutex);
 444                 vhost_vq_reset(dev, vq);
 445                 if (vq->handle_kick)
 446                         vhost_poll_init(&vq->poll, vq->handle_kick,
 447                                         EPOLLIN, dev);
 448         }
 449 }
 450 EXPORT_SYMBOL_GPL(vhost_dev_init);
 451
 452 /* Caller should have device mutex */
 453 long vhost_dev_check_owner(struct vhost_dev *dev)
 454 {
 455         /* Are you the owner? If not, I don't think you mean to do that */
 456         return dev->mm == current->mm ? 0 : -EPERM;
 457 }
 458 EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
 459
 460 struct vhost_attach_cgroups_struct {
 461         struct vhost_work work;
 462         struct task_struct *owner;
 463         int ret;
 464 };
 465
 466 static void vhost_attach_cgroups_work(struct vhost_work *work)
 467 {
 468         struct vhost_attach_cgroups_struct *s;
 469
 470         s = container_of(work, struct vhost_attach_cgroups_struct, work);
 471         s->ret = cgroup_attach_task_all(s->owner, current);
 472 }
 473
 474 static int vhost_attach_cgroups(struct vhost_dev *dev)
 475 {
 476         struct vhost_attach_cgroups_struct attach;
 477
 478         attach.owner = current;
 479         vhost_work_init(&attach.work, vhost_attach_cgroups_work);
 480         vhost_work_queue(dev, &attach.work);
 481         vhost_work_flush(dev, &attach.work);
 482         return attach.ret;
 483 }
 484
 485 /* Caller should have device mutex */
 486 bool vhost_dev_has_owner(struct vhost_dev *dev)
 487 {
 488         return dev->mm;
 489 }
 490 EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
 491
 492 /* Caller should have device mutex */
 493 long vhost_dev_set_owner(struct vhost_dev *dev)
 494 {
 495         struct task_struct *worker;
 496         int err;
 497
 498         /* Is there an owner already? */
 499         if (vhost_dev_has_owner(dev)) {
 500                 err = -EBUSY;
 501                 goto err_mm;
 502         }
 503
 504         /* No owner, become one */
 505         dev->mm = get_task_mm(current);
 506         worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
 507         if (IS_ERR(worker)) {
 508                 err = PTR_ERR(worker);
 509                 goto err_worker;
 510         }
 511
 512         dev->worker = worker;
 513         wake_up_process(worker);        /* avoid contributing to loadavg */
 514
 515         err = vhost_attach_cgroups(dev);
 516         if (err)
 517                 goto err_cgroup;
 518
 519         err = vhost_dev_alloc_iovecs(dev);
 520         if (err)
 521                 goto err_cgroup;
 522
 523         return 0;
 524 err_cgroup:
 525         kthread_stop(worker);
 526         dev->worker = NULL;
 527 err_worker:
 528         if (dev->mm)
 529                 mmput(dev->mm);
 530         dev->mm = NULL;
 531 err_mm:
 532         return err;
 533 }
 534 EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
 535
 536 struct vhost_umem *vhost_dev_reset_owner_prepare(void)
 537 {
 538         return kvzalloc(sizeof(struct vhost_umem), GFP_KERNEL);
 539 }
 540 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
 541
 542 /* Caller should have device mutex */
 543 void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_umem *umem)
 544 {
 545         int i;
 546
 547         vhost_dev_cleanup(dev);
 548
 549         /* Restore memory to default empty mapping. */
 550         INIT_LIST_HEAD(&umem->umem_list);
 551         dev->umem = umem;
 552         /* We don't need VQ locks below since vhost_dev_cleanup makes sure
 553          * VQs aren't running.
 554          */
 555         for (i = 0; i < dev->nvqs; ++i)
 556                 dev->vqs[i]->umem = umem;
 557 }
 558 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
 559
 560 void vhost_dev_stop(struct vhost_dev *dev)
 561 {
 562         int i;
 563
 564         for (i = 0; i < dev->nvqs; ++i) {
 565                 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick) {
 566                         vhost_poll_stop(&dev->vqs[i]->poll);
 567                         vhost_poll_flush(&dev->vqs[i]->poll);
 568                 }
 569         }
 570 }
 571 EXPORT_SYMBOL_GPL(vhost_dev_stop);
 572
 573 static void vhost_umem_free(struct vhost_umem *umem,
 574                             struct vhost_umem_node *node)
 575 {
 576         vhost_umem_interval_tree_remove(node, &umem->umem_tree);
 577         list_del(&node->link);
 578         kfree(node);
 579         umem->numem--;
 580 }
 581
 582 static void vhost_umem_clean(struct vhost_umem *umem)
 583 {
 584         struct vhost_umem_node *node, *tmp;
 585
 586         if (!umem)
 587                 return;
 588
 589         list_for_each_entry_safe(node, tmp, &umem->umem_list, link)
 590                 vhost_umem_free(umem, node);
 591
 592         kvfree(umem);
 593 }
 594
 595 static void vhost_clear_msg(struct vhost_dev *dev)
 596 {
 597         struct vhost_msg_node *node, *n;
 598
 599         spin_lock(&dev->iotlb_lock);
 600
 601         list_for_each_entry_safe(node, n, &dev->read_list, node) {
 602                 list_del(&node->node);
 603                 kfree(node);
 604         }
 605
 606         list_for_each_entry_safe(node, n, &dev->pending_list, node) {
 607                 list_del(&node->node);
 608                 kfree(node);
 609         }
 610
 611         spin_unlock(&dev->iotlb_lock);
 612 }
 613
 614 void vhost_dev_cleanup(struct vhost_dev *dev)
 615 {
 616         int i;
 617
 618         for (i = 0; i < dev->nvqs; ++i) {
 619                 if (dev->vqs[i]->error_ctx)
 620                         eventfd_ctx_put(dev->vqs[i]->error_ctx);
 621                 if (dev->vqs[i]->kick)
 622                         fput(dev->vqs[i]->kick);
 623                 if (dev->vqs[i]->call_ctx)
 624                         eventfd_ctx_put(dev->vqs[i]->call_ctx);
 625                 vhost_vq_reset(dev, dev->vqs[i]);
 626         }
 627         vhost_dev_free_iovecs(dev);
 628         if (dev->log_ctx)
 629                 eventfd_ctx_put(dev->log_ctx);
 630         dev->log_ctx = NULL;
 631         /* No one will access memory at this point */
 632         vhost_umem_clean(dev->umem);
 633         dev->umem = NULL;
 634         vhost_umem_clean(dev->iotlb);
 635         dev->iotlb = NULL;
 636         vhost_clear_msg(dev);
 637         wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
 638         WARN_ON(!llist_empty(&dev->work_list));
 639         if (dev->worker) {
 640                 kthread_stop(dev->worker);
 641                 dev->worker = NULL;
 642         }
 643         if (dev->mm)
 644                 mmput(dev->mm);
 645         dev->mm = NULL;
 646 }
 647 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
 648
 649 static bool log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
 650 {
 651         u64 a = addr / VHOST_PAGE_SIZE / 8;
 652
 653         /* Make sure 64 bit math will not overflow. */
 654         if (a > ULONG_MAX - (unsigned long)log_base ||
 655             a + (unsigned long)log_base > ULONG_MAX)
 656                 return false;
 657
 658         return access_ok(log_base + a,
 659                          (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
 660 }
 661
 662 static bool vhost_overflow(u64 uaddr, u64 size)
 663 {
 664         /* Make sure 64 bit math will not overflow. */
 665         return uaddr > ULONG_MAX || size > ULONG_MAX || uaddr > ULONG_MAX - size;
 666 }
 667
 668 /* Caller should have vq mutex and device mutex. */
 669 static bool vq_memory_access_ok(void __user *log_base, struct vhost_umem *umem,
 670                                 int log_all)
 671 {
 672         struct vhost_umem_node *node;
 673
 674         if (!umem)
 675                 return false;
 676
 677         list_for_each_entry(node, &umem->umem_list, link) {
 678                 unsigned long a = node->userspace_addr;
 679
 680                 if (vhost_overflow(node->userspace_addr, node->size))
 681                         return false;
 682
 683
 684                 if (!access_ok((void __user *)a,
 685                                     node->size))
 686                         return false;
 687                 else if (log_all && !log_access_ok(log_base,
 688                                                    node->start,
 689                                                    node->size))
 690                         return false;
 691         }
 692         return true;
 693 }
 694
 695 static inline void __user *vhost_vq_meta_fetch(struct vhost_virtqueue *vq,
 696                                                u64 addr, unsigned int size,
 697                                                int type)
 698 {
 699         const struct vhost_umem_node *node = vq->meta_iotlb[type];
 700
 701         if (!node)
 702                 return NULL;
 703
 704         return (void *)(uintptr_t)(node->userspace_addr + addr - node->start);
 705 }
 706
 707 /* Can we switch to this memory table? */
 708 /* Caller should have device mutex but not vq mutex */
 709 static bool memory_access_ok(struct vhost_dev *d, struct vhost_umem *umem,
 710                              int log_all)
 711 {
 712         int i;
 713
 714         for (i = 0; i < d->nvqs; ++i) {
 715                 bool ok;
 716                 bool log;
 717
 718                 mutex_lock(&d->vqs[i]->mutex);
 719                 log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
 720                 /* If ring is inactive, will check when it's enabled. */
 721                 if (d->vqs[i]->private_data)
 722                         ok = vq_memory_access_ok(d->vqs[i]->log_base,
 723                                                  umem, log);
 724                 else
 725                         ok = true;
 726                 mutex_unlock(&d->vqs[i]->mutex);
 727                 if (!ok)
 728                         return false;
 729         }
 730         return true;
 731 }
 732
 733 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
 734                           struct iovec iov[], int iov_size, int access);
 735
 736 static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
 737                               const void *from, unsigned size)
 738 {
 739         int ret;
 740
 741         if (!vq->iotlb)
 742                 return __copy_to_user(to, from, size);
 743         else {
 744                 /* This function should be called after iotlb
 745                  * prefetch, which means we're sure that all vq
 746                  * could be access through iotlb. So -EAGAIN should
 747                  * not happen in this case.
 748                  */
 749                 struct iov_iter t;
 750                 void __user *uaddr = vhost_vq_meta_fetch(vq,
 751                                      (u64)(uintptr_t)to, size,
 752                                      VHOST_ADDR_USED);
 753
 754                 if (uaddr)
 755                         return __copy_to_user(uaddr, from, size);
 756
 757                 ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
 758                                      ARRAY_SIZE(vq->iotlb_iov),
 759                                      VHOST_ACCESS_WO);
 760                 if (ret < 0)
 761                         goto out;
 762                 iov_iter_init(&t, WRITE, vq->iotlb_iov, ret, size);
 763                 ret = copy_to_iter(from, size, &t);
 764                 if (ret == size)
 765                         ret = 0;
 766         }
 767 out:
 768         return ret;
 769 }
 770
 771 static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
 772                                 void __user *from, unsigned size)
 773 {
 774         int ret;
 775
 776         if (!vq->iotlb)
 777                 return __copy_from_user(to, from, size);
 778         else {
 779                 /* This function should be called after iotlb
 780                  * prefetch, which means we're sure that vq
 781                  * could be access through iotlb. So -EAGAIN should
 782                  * not happen in this case.
 783                  */
 784                 void __user *uaddr = vhost_vq_meta_fetch(vq,
 785                                      (u64)(uintptr_t)from, size,
 786                                      VHOST_ADDR_DESC);
 787                 struct iov_iter f;
 788
 789                 if (uaddr)
 790                         return __copy_from_user(to, uaddr, size);
 791
 792                 ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
 793                                      ARRAY_SIZE(vq->iotlb_iov),
 794                                      VHOST_ACCESS_RO);
 795                 if (ret < 0) {
 796                         vq_err(vq, "IOTLB translation failure: uaddr "
 797                                "%p size 0x%llx\n", from,
 798                                (unsigned long long) size);
 799                         goto out;
 800                 }
 801                 iov_iter_init(&f, READ, vq->iotlb_iov, ret, size);
 802                 ret = copy_from_iter(to, size, &f);
 803                 if (ret == size)
 804                         ret = 0;
 805         }
 806
 807 out:
 808         return ret;
 809 }
 810
 811 static void __user *__vhost_get_user_slow(struct vhost_virtqueue *vq,
 812                                           void __user *addr, unsigned int size,
 813                                           int type)
 814 {
 815         int ret;
 816
 817         ret = translate_desc(vq, (u64)(uintptr_t)addr, size, vq->iotlb_iov,
 818                              ARRAY_SIZE(vq->iotlb_iov),
 819                              VHOST_ACCESS_RO);
 820         if (ret < 0) {
 821                 vq_err(vq, "IOTLB translation failure: uaddr "
 822                         "%p size 0x%llx\n", addr,
 823                         (unsigned long long) size);
 824                 return NULL;
 825         }
 826
 827         if (ret != 1 || vq->iotlb_iov[0].iov_len != size) {
 828                 vq_err(vq, "Non atomic userspace memory access: uaddr "
 829                         "%p size 0x%llx\n", addr,
 830                         (unsigned long long) size);
 831                 return NULL;
 832         }
 833
 834         return vq->iotlb_iov[0].iov_base;
 835 }
 836
 837 /* This function should be called after iotlb
 838  * prefetch, which means we're sure that vq
 839  * could be access through iotlb. So -EAGAIN should
 840  * not happen in this case.
 841  */
 842 static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
 843                                             void *addr, unsigned int size,
 844                                             int type)
 845 {
 846         void __user *uaddr = vhost_vq_meta_fetch(vq,
 847                              (u64)(uintptr_t)addr, size, type);
 848         if (uaddr)
 849                 return uaddr;
 850
 851         return __vhost_get_user_slow(vq, addr, size, type);
 852 }
 853
 854 #define vhost_put_user(vq, x, ptr)              \
 855 ({ \
 856         int ret = -EFAULT; \
 857         if (!vq->iotlb) { \
 858                 ret = __put_user(x, ptr); \
 859         } else { \
 860                 __typeof__(ptr) to = \
 861                         (__typeof__(ptr)) __vhost_get_user(vq, ptr,     \
 862                                           sizeof(*ptr), VHOST_ADDR_USED); \
 863                 if (to != NULL) \
 864                         ret = __put_user(x, to); \
 865                 else \
 866                         ret = -EFAULT;  \
 867         } \
 868         ret; \
 869 })
 870
 871 #define vhost_get_user(vq, x, ptr, type)                \
 872 ({ \
 873         int ret; \
 874         if (!vq->iotlb) { \
 875                 ret = __get_user(x, ptr); \
 876         } else { \
 877                 __typeof__(ptr) from = \
 878                         (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
 879                                                            sizeof(*ptr), \
 880                                                            type); \
 881                 if (from != NULL) \
 882                         ret = __get_user(x, from); \
 883                 else \
 884                         ret = -EFAULT; \
 885         } \
 886         ret; \
 887 })
 888
 889 #define vhost_get_avail(vq, x, ptr) \
 890         vhost_get_user(vq, x, ptr, VHOST_ADDR_AVAIL)
 891
 892 #define vhost_get_used(vq, x, ptr) \
 893         vhost_get_user(vq, x, ptr, VHOST_ADDR_USED)
 894
 895 static void vhost_dev_lock_vqs(struct vhost_dev *d)
 896 {
 897         int i = 0;
 898         for (i = 0; i < d->nvqs; ++i)
 899                 mutex_lock_nested(&d->vqs[i]->mutex, i);
 900 }
 901
 902 static void vhost_dev_unlock_vqs(struct vhost_dev *d)
 903 {
 904         int i = 0;
 905         for (i = 0; i < d->nvqs; ++i)
 906                 mutex_unlock(&d->vqs[i]->mutex);
 907 }
 908
 909 static int vhost_new_umem_range(struct vhost_umem *umem,
 910                                 u64 start, u64 size, u64 end,
 911                                 u64 userspace_addr, int perm)
 912 {
 913         struct vhost_umem_node *tmp, *node = kmalloc(sizeof(*node), GFP_ATOMIC);
 914
 915         if (!node)
 916                 return -ENOMEM;
 917
 918         if (umem->numem == max_iotlb_entries) {
 919                 tmp = list_first_entry(&umem->umem_list, typeof(*tmp), link);
 920                 vhost_umem_free(umem, tmp);
 921         }
 922
 923         node->start = start;
 924         node->size = size;
 925         node->last = end;
 926         node->userspace_addr = userspace_addr;
 927         node->perm = perm;
 928         INIT_LIST_HEAD(&node->link);
 929         list_add_tail(&node->link, &umem->umem_list);
 930         vhost_umem_interval_tree_insert(node, &umem->umem_tree);
 931         umem->numem++;
 932
 933         return 0;
 934 }
 935
 936 static void vhost_del_umem_range(struct vhost_umem *umem,
 937                                  u64 start, u64 end)
 938 {
 939         struct vhost_umem_node *node;
 940
 941         while ((node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
 942                                                            start, end)))
 943                 vhost_umem_free(umem, node);
 944 }
 945
 946 static void vhost_iotlb_notify_vq(struct vhost_dev *d,
 947                                   struct vhost_iotlb_msg *msg)
 948 {
 949         struct vhost_msg_node *node, *n;
 950
 951         spin_lock(&d->iotlb_lock);
 952
 953         list_for_each_entry_safe(node, n, &d->pending_list, node) {
 954                 struct vhost_iotlb_msg *vq_msg = &node->msg.iotlb;
 955                 if (msg->iova <= vq_msg->iova &&
 956                     msg->iova + msg->size - 1 >= vq_msg->iova &&
 957                     vq_msg->type == VHOST_IOTLB_MISS) {
 958                         vhost_poll_queue(&node->vq->poll);
 959                         list_del(&node->node);
 960                         kfree(node);
 961                 }
 962         }
 963
 964         spin_unlock(&d->iotlb_lock);
 965 }
 966
 967 static bool umem_access_ok(u64 uaddr, u64 size, int access)
 968 {
 969         unsigned long a = uaddr;
 970
 971         /* Make sure 64 bit math will not overflow. */
 972         if (vhost_overflow(uaddr, size))
 973                 return false;
 974
 975         if ((access & VHOST_ACCESS_RO) &&
 976             !access_ok((void __user *)a, size))
 977                 return false;
 978         if ((access & VHOST_ACCESS_WO) &&
 979             !access_ok((void __user *)a, size))
 980                 return false;
 981         return true;
 982 }
 983
 984 static int vhost_process_iotlb_msg(struct vhost_dev *dev,
 985                                    struct vhost_iotlb_msg *msg)
 986 {
 987         int ret = 0;
 988
 989         mutex_lock(&dev->mutex);
 990         vhost_dev_lock_vqs(dev);
 991         switch (msg->type) {
 992         case VHOST_IOTLB_UPDATE:
 993                 if (!dev->iotlb) {
 994                         ret = -EFAULT;
 995                         break;
 996                 }
 997                 if (!umem_access_ok(msg->uaddr, msg->size, msg->perm)) {
 998                         ret = -EFAULT;
 999                         break;
1000                 }
1001                 vhost_vq_meta_reset(dev);
1002                 if (vhost_new_umem_range(dev->iotlb, msg->iova, msg->size,
1003                                          msg->iova + msg->size - 1,
1004                                          msg->uaddr, msg->perm)) {
1005                         ret = -ENOMEM;
1006                         break;
1007                 }
1008                 vhost_iotlb_notify_vq(dev, msg);
1009                 break;
1010         case VHOST_IOTLB_INVALIDATE:
1011                 if (!dev->iotlb) {
1012                         ret = -EFAULT;
1013                         break;
1014                 }
1015                 vhost_vq_meta_reset(dev);
1016                 vhost_del_umem_range(dev->iotlb, msg->iova,
1017                                      msg->iova + msg->size - 1);
1018                 break;
1019         default:
1020                 ret = -EINVAL;
1021                 break;
1022         }
1023
1024         vhost_dev_unlock_vqs(dev);
1025         mutex_unlock(&dev->mutex);
1026
1027         return ret;
1028 }
1029 ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
1030                              struct iov_iter *from)
1031 {
1032         struct vhost_iotlb_msg msg;
1033         size_t offset;
1034         int type, ret;
1035
1036         ret = copy_from_iter(&type, sizeof(type), from);
1037         if (ret != sizeof(type)) {
1038                 ret = -EINVAL;
1039                 goto done;
1040         }
1041
1042         switch (type) {
1043         case VHOST_IOTLB_MSG:
1044                 /* There maybe a hole after type for V1 message type,
1045                  * so skip it here.
1046                  */
1047                 offset = offsetof(struct vhost_msg, iotlb) - sizeof(int);
1048                 break;
1049         case VHOST_IOTLB_MSG_V2:
1050                 offset = sizeof(__u32);
1051                 break;
1052         default:
1053                 ret = -EINVAL;
1054                 goto done;
1055         }
1056
1057         iov_iter_advance(from, offset);
1058         ret = copy_from_iter(&msg, sizeof(msg), from);
1059         if (ret != sizeof(msg)) {
1060                 ret = -EINVAL;
1061                 goto done;
1062         }
1063         if (vhost_process_iotlb_msg(dev, &msg)) {
1064                 ret = -EFAULT;
1065                 goto done;
1066         }
1067
1068         ret = (type == VHOST_IOTLB_MSG) ? sizeof(struct vhost_msg) :
1069               sizeof(struct vhost_msg_v2);
1070 done:
1071         return ret;
1072 }
1073 EXPORT_SYMBOL(vhost_chr_write_iter);
1074
1075 __poll_t vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1076                             poll_table *wait)
1077 {
1078         __poll_t mask = 0;
1079
1080         poll_wait(file, &dev->wait, wait);
1081
1082         if (!list_empty(&dev->read_list))
1083                 mask |= EPOLLIN | EPOLLRDNORM;
1084
1085         return mask;
1086 }
1087 EXPORT_SYMBOL(vhost_chr_poll);
1088
1089 ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1090                             int noblock)
1091 {
1092         DEFINE_WAIT(wait);
1093         struct vhost_msg_node *node;
1094         ssize_t ret = 0;
1095         unsigned size = sizeof(struct vhost_msg);
1096
1097         if (iov_iter_count(to) < size)
1098                 return 0;
1099
1100         while (1) {
1101                 if (!noblock)
1102                         prepare_to_wait(&dev->wait, &wait,
1103                                         TASK_INTERRUPTIBLE);
1104
1105                 node = vhost_dequeue_msg(dev, &dev->read_list);
1106                 if (node)
1107                         break;
1108                 if (noblock) {
1109                         ret = -EAGAIN;
1110                         break;
1111                 }
1112                 if (signal_pending(current)) {
1113                         ret = -ERESTARTSYS;
1114                         break;
1115                 }
1116                 if (!dev->iotlb) {
1117                         ret = -EBADFD;
1118                         break;
1119                 }
1120
1121                 schedule();
1122         }
1123
1124         if (!noblock)
1125                 finish_wait(&dev->wait, &wait);
1126
1127         if (node) {
1128                 struct vhost_iotlb_msg *msg;
1129                 void *start = &node->msg;
1130
1131                 switch (node->msg.type) {
1132                 case VHOST_IOTLB_MSG:
1133                         size = sizeof(node->msg);
1134                         msg = &node->msg.iotlb;
1135                         break;
1136                 case VHOST_IOTLB_MSG_V2:
1137                         size = sizeof(node->msg_v2);
1138                         msg = &node->msg_v2.iotlb;
1139                         break;
1140                 default:
1141                         BUG();
1142                         break;
1143                 }
1144
1145                 ret = copy_to_iter(start, size, to);
1146                 if (ret != size || msg->type != VHOST_IOTLB_MISS) {
1147                         kfree(node);
1148                         return ret;
1149                 }
1150                 vhost_enqueue_msg(dev, &dev->pending_list, node);
1151         }
1152
1153         return ret;
1154 }
1155 EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1156
1157 static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1158 {
1159         struct vhost_dev *dev = vq->dev;
1160         struct vhost_msg_node *node;
1161         struct vhost_iotlb_msg *msg;
1162         bool v2 = vhost_backend_has_feature(vq, VHOST_BACKEND_F_IOTLB_MSG_V2);
1163
1164         node = vhost_new_msg(vq, v2 ? VHOST_IOTLB_MSG_V2 : VHOST_IOTLB_MSG);
1165         if (!node)
1166                 return -ENOMEM;
1167
1168         if (v2) {
1169                 node->msg_v2.type = VHOST_IOTLB_MSG_V2;
1170                 msg = &node->msg_v2.iotlb;
1171         } else {
1172                 msg = &node->msg.iotlb;
1173         }
1174
1175         msg->type = VHOST_IOTLB_MISS;
1176         msg->iova = iova;
1177         msg->perm = access;
1178
1179         vhost_enqueue_msg(dev, &dev->read_list, node);
1180
1181         return 0;
1182 }
1183
1184 static bool vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1185                          struct vring_desc __user *desc,
1186                          struct vring_avail __user *avail,
1187                          struct vring_used __user *used)
1188
1189 {
1190         size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1191
1192         return access_ok(desc, num * sizeof *desc) &&
1193                access_ok(avail,
1194                          sizeof *avail + num * sizeof *avail->ring + s) &&
1195                access_ok(used,
1196                         sizeof *used + num * sizeof *used->ring + s);
1197 }
1198
1199 static void vhost_vq_meta_update(struct vhost_virtqueue *vq,
1200                                  const struct vhost_umem_node *node,
1201                                  int type)
1202 {
1203         int access = (type == VHOST_ADDR_USED) ?
1204                      VHOST_ACCESS_WO : VHOST_ACCESS_RO;
1205
1206         if (likely(node->perm & access))
1207                 vq->meta_iotlb[type] = node;
1208 }
1209
1210 static bool iotlb_access_ok(struct vhost_virtqueue *vq,
1211                             int access, u64 addr, u64 len, int type)
1212 {
1213         const struct vhost_umem_node *node;
1214         struct vhost_umem *umem = vq->iotlb;
1215         u64 s = 0, size, orig_addr = addr, last = addr + len - 1;
1216
1217         if (vhost_vq_meta_fetch(vq, addr, len, type))
1218                 return true;
1219
1220         while (len > s) {
1221                 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1222                                                            addr,
1223                                                            last);
1224                 if (node == NULL || node->start > addr) {
1225                         vhost_iotlb_miss(vq, addr, access);
1226                         return false;
1227                 } else if (!(node->perm & access)) {
1228                         /* Report the possible access violation by
1229                          * request another translation from userspace.
1230                          */
1231                         return false;
1232                 }
1233
1234                 size = node->size - addr + node->start;
1235
1236                 if (orig_addr == addr && size >= len)
1237                         vhost_vq_meta_update(vq, node, type);
1238
1239                 s += size;
1240                 addr += size;
1241         }
1242
1243         return true;
1244 }
1245
1246 int vq_iotlb_prefetch(struct vhost_virtqueue *vq)
1247 {
1248         size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1249         unsigned int num = vq->num;
1250
1251         if (!vq->iotlb)
1252                 return 1;
1253
1254         return iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->desc,
1255                                num * sizeof(*vq->desc), VHOST_ADDR_DESC) &&
1256                iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->avail,
1257                                sizeof *vq->avail +
1258                                num * sizeof(*vq->avail->ring) + s,
1259                                VHOST_ADDR_AVAIL) &&
1260                iotlb_access_ok(vq, VHOST_ACCESS_WO, (u64)(uintptr_t)vq->used,
1261                                sizeof *vq->used +
1262                                num * sizeof(*vq->used->ring) + s,
1263                                VHOST_ADDR_USED);
1264 }
1265 EXPORT_SYMBOL_GPL(vq_iotlb_prefetch);
1266
1267 /* Can we log writes? */
1268 /* Caller should have device mutex but not vq mutex */
1269 bool vhost_log_access_ok(struct vhost_dev *dev)
1270 {
1271         return memory_access_ok(dev, dev->umem, 1);
1272 }
1273 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1274
1275 /* Verify access for write logging. */
1276 /* Caller should have vq mutex and device mutex */
1277 static bool vq_log_access_ok(struct vhost_virtqueue *vq,
1278                              void __user *log_base)
1279 {
1280         size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1281
1282         return vq_memory_access_ok(log_base, vq->umem,
1283                                    vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1284                 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
1285                                         sizeof *vq->used +
1286                                         vq->num * sizeof *vq->used->ring + s));
1287 }
1288
1289 /* Can we start vq? */
1290 /* Caller should have vq mutex and device mutex */
1291 bool vhost_vq_access_ok(struct vhost_virtqueue *vq)
1292 {
1293         if (!vq_log_access_ok(vq, vq->log_base))
1294                 return false;
1295
1296         /* Access validation occurs at prefetch time with IOTLB */
1297         if (vq->iotlb)
1298                 return true;
1299
1300         return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used);
1301 }
1302 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1303
1304 static struct vhost_umem *vhost_umem_alloc(void)
1305 {
1306         struct vhost_umem *umem = kvzalloc(sizeof(*umem), GFP_KERNEL);
1307
1308         if (!umem)
1309                 return NULL;
1310
1311         umem->umem_tree = RB_ROOT_CACHED;
1312         umem->numem = 0;
1313         INIT_LIST_HEAD(&umem->umem_list);
1314
1315         return umem;
1316 }
1317
1318 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1319 {
1320         struct vhost_memory mem, *newmem;
1321         struct vhost_memory_region *region;
1322         struct vhost_umem *newumem, *oldumem;
1323         unsigned long size = offsetof(struct vhost_memory, regions);
1324         int i;
1325
1326         if (copy_from_user(&mem, m, size))
1327                 return -EFAULT;
1328         if (mem.padding)
1329                 return -EOPNOTSUPP;
1330         if (mem.nregions > max_mem_regions)
1331                 return -E2BIG;
1332         newmem = kvzalloc(struct_size(newmem, regions, mem.nregions),
1333                         GFP_KERNEL);
1334         if (!newmem)
1335                 return -ENOMEM;
1336
1337         memcpy(newmem, &mem, size);
1338         if (copy_from_user(newmem->regions, m->regions,
1339                            mem.nregions * sizeof *m->regions)) {
1340                 kvfree(newmem);
1341                 return -EFAULT;
1342         }
1343
1344         newumem = vhost_umem_alloc();
1345         if (!newumem) {
1346                 kvfree(newmem);
1347                 return -ENOMEM;
1348         }
1349
1350         for (region = newmem->regions;
1351              region < newmem->regions + mem.nregions;
1352              region++) {
1353                 if (vhost_new_umem_range(newumem,
1354                                          region->guest_phys_addr,
1355                                          region->memory_size,
1356                                          region->guest_phys_addr +
1357                                          region->memory_size - 1,
1358                                          region->userspace_addr,
1359                                          VHOST_ACCESS_RW))
1360                         goto err;
1361         }
1362
1363         if (!memory_access_ok(d, newumem, 0))
1364                 goto err;
1365
1366         oldumem = d->umem;
1367         d->umem = newumem;
1368
1369         /* All memory accesses are done under some VQ mutex. */
1370         for (i = 0; i < d->nvqs; ++i) {
1371                 mutex_lock(&d->vqs[i]->mutex);
1372                 d->vqs[i]->umem = newumem;
1373                 mutex_unlock(&d->vqs[i]->mutex);
1374         }
1375
1376         kvfree(newmem);
1377         vhost_umem_clean(oldumem);
1378         return 0;
1379
1380 err:
1381         vhost_umem_clean(newumem);
1382         kvfree(newmem);
1383         return -EFAULT;
1384 }
1385
1386 long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1387 {
1388         struct file *eventfp, *filep = NULL;
1389         bool pollstart = false, pollstop = false;
1390         struct eventfd_ctx *ctx = NULL;
1391         u32 __user *idxp = argp;
1392         struct vhost_virtqueue *vq;
1393         struct vhost_vring_state s;
1394         struct vhost_vring_file f;
1395         struct vhost_vring_addr a;
1396         u32 idx;
1397         long r;
1398
1399         r = get_user(idx, idxp);
1400         if (r < 0)
1401                 return r;
1402         if (idx >= d->nvqs)
1403                 return -ENOBUFS;
1404
1405         idx = array_index_nospec(idx, d->nvqs);
1406         vq = d->vqs[idx];
1407
1408         mutex_lock(&vq->mutex);
1409
1410         switch (ioctl) {
1411         case VHOST_SET_VRING_NUM:
1412                 /* Resizing ring with an active backend?
1413                  * You don't want to do that. */
1414                 if (vq->private_data) {
1415                         r = -EBUSY;
1416                         break;
1417                 }
1418                 if (copy_from_user(&s, argp, sizeof s)) {
1419                         r = -EFAULT;
1420                         break;
1421                 }
1422                 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
1423                         r = -EINVAL;
1424                         break;
1425                 }
1426                 vq->num = s.num;
1427                 break;
1428         case VHOST_SET_VRING_BASE:
1429                 /* Moving base with an active backend?
1430                  * You don't want to do that. */
1431                 if (vq->private_data) {
1432                         r = -EBUSY;
1433                         break;
1434                 }
1435                 if (copy_from_user(&s, argp, sizeof s)) {
1436                         r = -EFAULT;
1437                         break;
1438                 }
1439                 if (s.num > 0xffff) {
1440                         r = -EINVAL;
1441                         break;
1442                 }
1443                 vq->last_avail_idx = s.num;
1444                 /* Forget the cached index value. */
1445                 vq->avail_idx = vq->last_avail_idx;
1446                 break;
1447         case VHOST_GET_VRING_BASE:
1448                 s.index = idx;
1449                 s.num = vq->last_avail_idx;
1450                 if (copy_to_user(argp, &s, sizeof s))
1451                         r = -EFAULT;
1452                 break;
1453         case VHOST_SET_VRING_ADDR:
1454                 if (copy_from_user(&a, argp, sizeof a)) {
1455                         r = -EFAULT;
1456                         break;
1457                 }
1458                 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
1459                         r = -EOPNOTSUPP;
1460                         break;
1461                 }
1462                 /* For 32bit, verify that the top 32bits of the user
1463                    data are set to zero. */
1464                 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1465                     (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1466                     (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
1467                         r = -EFAULT;
1468                         break;
1469                 }
1470
1471                 /* Make sure it's safe to cast pointers to vring types. */
1472                 BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1473                 BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1474                 if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1475                     (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1476                     (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1))) {
1477                         r = -EINVAL;
1478                         break;
1479                 }
1480
1481                 /* We only verify access here if backend is configured.
1482                  * If it is not, we don't as size might not have been setup.
1483                  * We will verify when backend is configured. */
1484                 if (vq->private_data) {
1485                         if (!vq_access_ok(vq, vq->num,
1486                                 (void __user *)(unsigned long)a.desc_user_addr,
1487                                 (void __user *)(unsigned long)a.avail_user_addr,
1488                                 (void __user *)(unsigned long)a.used_user_addr)) {
1489                                 r = -EINVAL;
1490                                 break;
1491                         }
1492
1493                         /* Also validate log access for used ring if enabled. */
1494                         if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
1495                             !log_access_ok(vq->log_base, a.log_guest_addr,
1496                                            sizeof *vq->used +
1497                                            vq->num * sizeof *vq->used->ring)) {
1498                                 r = -EINVAL;
1499                                 break;
1500                         }
1501                 }
1502
1503                 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1504                 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1505                 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1506                 vq->log_addr = a.log_guest_addr;
1507                 vq->used = (void __user *)(unsigned long)a.used_user_addr;
1508                 break;
1509         case VHOST_SET_VRING_KICK:
1510                 if (copy_from_user(&f, argp, sizeof f)) {
1511                         r = -EFAULT;
1512                         break;
1513                 }
1514                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1515                 if (IS_ERR(eventfp)) {
1516                         r = PTR_ERR(eventfp);
1517                         break;
1518                 }
1519                 if (eventfp != vq->kick) {
1520                         pollstop = (filep = vq->kick) != NULL;
1521                         pollstart = (vq->kick = eventfp) != NULL;
1522                 } else
1523                         filep = eventfp;
1524                 break;
1525         case VHOST_SET_VRING_CALL:
1526                 if (copy_from_user(&f, argp, sizeof f)) {
1527                         r = -EFAULT;
1528                         break;
1529                 }
1530                 ctx = f.fd == -1 ? NULL : eventfd_ctx_fdget(f.fd);
1531                 if (IS_ERR(ctx)) {
1532                         r = PTR_ERR(ctx);
1533                         break;
1534                 }
1535                 swap(ctx, vq->call_ctx);
1536                 break;
1537         case VHOST_SET_VRING_ERR:
1538                 if (copy_from_user(&f, argp, sizeof f)) {
1539                         r = -EFAULT;
1540                         break;
1541                 }
1542                 ctx = f.fd == -1 ? NULL : eventfd_ctx_fdget(f.fd);
1543                 if (IS_ERR(ctx)) {
1544                         r = PTR_ERR(ctx);
1545                         break;
1546                 }
1547                 swap(ctx, vq->error_ctx);
1548                 break;
1549         case VHOST_SET_VRING_ENDIAN:
1550                 r = vhost_set_vring_endian(vq, argp);
1551                 break;
1552         case VHOST_GET_VRING_ENDIAN:
1553                 r = vhost_get_vring_endian(vq, idx, argp);
1554                 break;
1555         case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
1556                 if (copy_from_user(&s, argp, sizeof(s))) {
1557                         r = -EFAULT;
1558                         break;
1559                 }
1560                 vq->busyloop_timeout = s.num;
1561                 break;
1562         case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
1563                 s.index = idx;
1564                 s.num = vq->busyloop_timeout;
1565                 if (copy_to_user(argp, &s, sizeof(s)))
1566                         r = -EFAULT;
1567                 break;
1568         default:
1569                 r = -ENOIOCTLCMD;
1570         }
1571
1572         if (pollstop && vq->handle_kick)
1573                 vhost_poll_stop(&vq->poll);
1574
1575         if (!IS_ERR_OR_NULL(ctx))
1576                 eventfd_ctx_put(ctx);
1577         if (filep)
1578                 fput(filep);
1579
1580         if (pollstart && vq->handle_kick)
1581                 r = vhost_poll_start(&vq->poll, vq->kick);
1582
1583         mutex_unlock(&vq->mutex);
1584
1585         if (pollstop && vq->handle_kick)
1586                 vhost_poll_flush(&vq->poll);
1587         return r;
1588 }
1589 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
1590
1591 int vhost_init_device_iotlb(struct vhost_dev *d, bool enabled)
1592 {
1593         struct vhost_umem *niotlb, *oiotlb;
1594         int i;
1595
1596         niotlb = vhost_umem_alloc();
1597         if (!niotlb)
1598                 return -ENOMEM;
1599
1600         oiotlb = d->iotlb;
1601         d->iotlb = niotlb;
1602
1603         for (i = 0; i < d->nvqs; ++i) {
1604                 struct vhost_virtqueue *vq = d->vqs[i];
1605
1606                 mutex_lock(&vq->mutex);
1607                 vq->iotlb = niotlb;
1608                 __vhost_vq_meta_reset(vq);
1609                 mutex_unlock(&vq->mutex);
1610         }
1611
1612         vhost_umem_clean(oiotlb);
1613
1614         return 0;
1615 }
1616 EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
1617
1618 /* Caller must have device mutex */
1619 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1620 {
1621         struct eventfd_ctx *ctx;
1622         u64 p;
1623         long r;
1624         int i, fd;
1625
1626         /* If you are not the owner, you can become one */
1627         if (ioctl == VHOST_SET_OWNER) {
1628                 r = vhost_dev_set_owner(d);
1629                 goto done;
1630         }
1631
1632         /* You must be the owner to do anything else */
1633         r = vhost_dev_check_owner(d);
1634         if (r)
1635                 goto done;
1636
1637         switch (ioctl) {
1638         case VHOST_SET_MEM_TABLE:
1639                 r = vhost_set_memory(d, argp);
1640                 break;
1641         case VHOST_SET_LOG_BASE:
1642                 if (copy_from_user(&p, argp, sizeof p)) {
1643                         r = -EFAULT;
1644                         break;
1645                 }
1646                 if ((u64)(unsigned long)p != p) {
1647                         r = -EFAULT;
1648                         break;
1649                 }
1650                 for (i = 0; i < d->nvqs; ++i) {
1651                         struct vhost_virtqueue *vq;
1652                         void __user *base = (void __user *)(unsigned long)p;
1653                         vq = d->vqs[i];
1654                         mutex_lock(&vq->mutex);
1655                         /* If ring is inactive, will check when it's enabled. */
1656                         if (vq->private_data && !vq_log_access_ok(vq, base))
1657                                 r = -EFAULT;
1658                         else
1659                                 vq->log_base = base;
1660                         mutex_unlock(&vq->mutex);
1661                 }
1662                 break;
1663         case VHOST_SET_LOG_FD:
1664                 r = get_user(fd, (int __user *)argp);
1665                 if (r < 0)
1666                         break;
1667                 ctx = fd == -1 ? NULL : eventfd_ctx_fdget(fd);
1668                 if (IS_ERR(ctx)) {
1669                         r = PTR_ERR(ctx);
1670                         break;
1671                 }
1672                 swap(ctx, d->log_ctx);
1673                 for (i = 0; i < d->nvqs; ++i) {
1674                         mutex_lock(&d->vqs[i]->mutex);
1675                         d->vqs[i]->log_ctx = d->log_ctx;
1676                         mutex_unlock(&d->vqs[i]->mutex);
1677                 }
1678                 if (ctx)
1679                         eventfd_ctx_put(ctx);
1680                 break;
1681         default:
1682                 r = -ENOIOCTLCMD;
1683                 break;
1684         }
1685 done:
1686         return r;
1687 }
1688 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1689
1690 /* TODO: This is really inefficient.  We need something like get_user()
1691  * (instruction directly accesses the data, with an exception table entry
1692  * returning -EFAULT). See Documentation/x86/exception-tables.txt.
1693  */
1694 static int set_bit_to_user(int nr, void __user *addr)
1695 {
1696         unsigned long log = (unsigned long)addr;
1697         struct page *page;
1698         void *base;
1699         int bit = nr + (log % PAGE_SIZE) * 8;
1700         int r;
1701
1702         r = get_user_pages_fast(log, 1, 1, &page);
1703         if (r < 0)
1704                 return r;
1705         BUG_ON(r != 1);
1706         base = kmap_atomic(page);
1707         set_bit(bit, base);
1708         kunmap_atomic(base);
1709         set_page_dirty_lock(page);
1710         put_page(page);
1711         return 0;
1712 }
1713
1714 static int log_write(void __user *log_base,
1715                      u64 write_address, u64 write_length)
1716 {
1717         u64 write_page = write_address / VHOST_PAGE_SIZE;
1718         int r;
1719
1720         if (!write_length)
1721                 return 0;
1722         write_length += write_address % VHOST_PAGE_SIZE;
1723         for (;;) {
1724                 u64 base = (u64)(unsigned long)log_base;
1725                 u64 log = base + write_page / 8;
1726                 int bit = write_page % 8;
1727                 if ((u64)(unsigned long)log != log)
1728                         return -EFAULT;
1729                 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1730                 if (r < 0)
1731                         return r;
1732                 if (write_length <= VHOST_PAGE_SIZE)
1733                         break;
1734                 write_length -= VHOST_PAGE_SIZE;
1735                 write_page += 1;
1736         }
1737         return r;
1738 }
1739
1740 static int log_write_hva(struct vhost_virtqueue *vq, u64 hva, u64 len)
1741 {
1742         struct vhost_umem *umem = vq->umem;
1743         struct vhost_umem_node *u;
1744         u64 start, end, l, min;
1745         int r;
1746         bool hit = false;
1747
1748         while (len) {
1749                 min = len;
1750                 /* More than one GPAs can be mapped into a single HVA. So
1751                  * iterate all possible umems here to be safe.
1752                  */
1753                 list_for_each_entry(u, &umem->umem_list, link) {
1754                         if (u->userspace_addr > hva - 1 + len ||
1755                             u->userspace_addr - 1 + u->size < hva)
1756                                 continue;
1757                         start = max(u->userspace_addr, hva);
1758                         end = min(u->userspace_addr - 1 + u->size,
1759                                   hva - 1 + len);
1760                         l = end - start + 1;
1761                         r = log_write(vq->log_base,
1762                                       u->start + start - u->userspace_addr,
1763                                       l);
1764                         if (r < 0)
1765                                 return r;
1766                         hit = true;
1767                         min = min(l, min);
1768                 }
1769
1770                 if (!hit)
1771                         return -EFAULT;
1772
1773                 len -= min;
1774                 hva += min;
1775         }
1776
1777         return 0;
1778 }
1779
1780 static int log_used(struct vhost_virtqueue *vq, u64 used_offset, u64 len)
1781 {
1782         struct iovec iov[64];
1783         int i, ret;
1784
1785         if (!vq->iotlb)
1786                 return log_write(vq->log_base, vq->log_addr + used_offset, len);
1787
1788         ret = translate_desc(vq, (uintptr_t)vq->used + used_offset,
1789                              len, iov, 64, VHOST_ACCESS_WO);
1790         if (ret)
1791                 return ret;
1792
1793         for (i = 0; i < ret; i++) {
1794                 ret = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1795                                     iov[i].iov_len);
1796                 if (ret)
1797                         return ret;
1798         }
1799
1800         return 0;
1801 }
1802
1803 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1804                     unsigned int log_num, u64 len, struct iovec *iov, int count)
1805 {
1806         int i, r;
1807
1808         /* Make sure data written is seen before log. */
1809         smp_wmb();
1810
1811         if (vq->iotlb) {
1812                 for (i = 0; i < count; i++) {
1813                         r = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1814                                           iov[i].iov_len);
1815                         if (r < 0)
1816                                 return r;
1817                 }
1818                 return 0;
1819         }
1820
1821         for (i = 0; i < log_num; ++i) {
1822                 u64 l = min(log[i].len, len);
1823                 r = log_write(vq->log_base, log[i].addr, l);
1824                 if (r < 0)
1825                         return r;
1826                 len -= l;
1827                 if (!len) {
1828                         if (vq->log_ctx)
1829                                 eventfd_signal(vq->log_ctx, 1);
1830                         return 0;
1831                 }
1832         }
1833         /* Length written exceeds what we have stored. This is a bug. */
1834         BUG();
1835         return 0;
1836 }
1837 EXPORT_SYMBOL_GPL(vhost_log_write);
1838
1839 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1840 {
1841         void __user *used;
1842         if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
1843                            &vq->used->flags) < 0)
1844                 return -EFAULT;
1845         if (unlikely(vq->log_used)) {
1846                 /* Make sure the flag is seen before log. */
1847                 smp_wmb();
1848                 /* Log used flag write. */
1849                 used = &vq->used->flags;
1850                 log_used(vq, (used - (void __user *)vq->used),
1851                          sizeof vq->used->flags);
1852                 if (vq->log_ctx)
1853                         eventfd_signal(vq->log_ctx, 1);
1854         }
1855         return 0;
1856 }
1857
1858 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1859 {
1860         if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
1861                            vhost_avail_event(vq)))
1862                 return -EFAULT;
1863         if (unlikely(vq->log_used)) {
1864                 void __user *used;
1865                 /* Make sure the event is seen before log. */
1866                 smp_wmb();
1867                 /* Log avail event write */
1868                 used = vhost_avail_event(vq);
1869                 log_used(vq, (used - (void __user *)vq->used),
1870                          sizeof *vhost_avail_event(vq));
1871                 if (vq->log_ctx)
1872                         eventfd_signal(vq->log_ctx, 1);
1873         }
1874         return 0;
1875 }
1876
1877 int vhost_vq_init_access(struct vhost_virtqueue *vq)
1878 {
1879         __virtio16 last_used_idx;
1880         int r;
1881         bool is_le = vq->is_le;
1882
1883         if (!vq->private_data)
1884                 return 0;
1885
1886         vhost_init_is_le(vq);
1887
1888         r = vhost_update_used_flags(vq);
1889         if (r)
1890                 goto err;
1891         vq->signalled_used_valid = false;
1892         if (!vq->iotlb &&
1893             !access_ok(&vq->used->idx, sizeof vq->used->idx)) {
1894                 r = -EFAULT;
1895                 goto err;
1896         }
1897         r = vhost_get_used(vq, last_used_idx, &vq->used->idx);
1898         if (r) {
1899                 vq_err(vq, "Can't access used idx at %p\n",
1900                        &vq->used->idx);
1901                 goto err;
1902         }
1903         vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
1904         return 0;
1905
1906 err:
1907         vq->is_le = is_le;
1908         return r;
1909 }
1910 EXPORT_SYMBOL_GPL(vhost_vq_init_access);
1911
1912 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
1913                           struct iovec iov[], int iov_size, int access)
1914 {
1915         const struct vhost_umem_node *node;
1916         struct vhost_dev *dev = vq->dev;
1917         struct vhost_umem *umem = dev->iotlb ? dev->iotlb : dev->umem;
1918         struct iovec *_iov;
1919         u64 s = 0;
1920         int ret = 0;
1921
1922         while ((u64)len > s) {
1923                 u64 size;
1924                 if (unlikely(ret >= iov_size)) {
1925                         ret = -ENOBUFS;
1926                         break;
1927                 }
1928
1929                 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1930                                                         addr, addr + len - 1);
1931                 if (node == NULL || node->start > addr) {
1932                         if (umem != dev->iotlb) {
1933                                 ret = -EFAULT;
1934                                 break;
1935                         }
1936                         ret = -EAGAIN;
1937                         break;
1938                 } else if (!(node->perm & access)) {
1939                         ret = -EPERM;
1940                         break;
1941                 }
1942
1943                 _iov = iov + ret;
1944                 size = node->size - addr + node->start;
1945                 _iov->iov_len = min((u64)len - s, size);
1946                 _iov->iov_base = (void __user *)(unsigned long)
1947                         (node->userspace_addr + addr - node->start);
1948                 s += size;
1949                 addr += size;
1950                 ++ret;
1951         }
1952
1953         if (ret == -EAGAIN)
1954                 vhost_iotlb_miss(vq, addr, access);
1955         return ret;
1956 }
1957
1958 /* Each buffer in the virtqueues is actually a chain of descriptors.  This
1959  * function returns the next descriptor in the chain,
1960  * or -1U if we're at the end. */
1961 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
1962 {
1963         unsigned int next;
1964
1965         /* If this descriptor says it doesn't chain, we're done. */
1966         if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
1967                 return -1U;
1968
1969         /* Check they're not leading us off end of descriptors. */
1970         next = vhost16_to_cpu(vq, READ_ONCE(desc->next));
1971         return next;
1972 }
1973
1974 static int get_indirect(struct vhost_virtqueue *vq,
1975                         struct iovec iov[], unsigned int iov_size,
1976                         unsigned int *out_num, unsigned int *in_num,
1977                         struct vhost_log *log, unsigned int *log_num,
1978                         struct vring_desc *indirect)
1979 {
1980         struct vring_desc desc;
1981         unsigned int i = 0, count, found = 0;
1982         u32 len = vhost32_to_cpu(vq, indirect->len);
1983         struct iov_iter from;
1984         int ret, access;
1985
1986         /* Sanity check */
1987         if (unlikely(len % sizeof desc)) {
1988                 vq_err(vq, "Invalid length in indirect descriptor: "
1989                        "len 0x%llx not multiple of 0x%zx\n",
1990                        (unsigned long long)len,
1991                        sizeof desc);
1992                 return -EINVAL;
1993         }
1994
1995         ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
1996                              UIO_MAXIOV, VHOST_ACCESS_RO);
1997         if (unlikely(ret < 0)) {
1998                 if (ret != -EAGAIN)
1999                         vq_err(vq, "Translation failure %d in indirect.\n", ret);
2000                 return ret;
2001         }
2002         iov_iter_init(&from, READ, vq->indirect, ret, len);
2003
2004         /* We will use the result as an address to read from, so most
2005          * architectures only need a compiler barrier here. */
2006         read_barrier_depends();
2007
2008         count = len / sizeof desc;
2009         /* Buffers are chained via a 16 bit next field, so
2010          * we can have at most 2^16 of these. */
2011         if (unlikely(count > USHRT_MAX + 1)) {
2012                 vq_err(vq, "Indirect buffer length too big: %d\n",
2013                        indirect->len);
2014                 return -E2BIG;
2015         }
2016
2017         do {
2018                 unsigned iov_count = *in_num + *out_num;
2019                 if (unlikely(++found > count)) {
2020                         vq_err(vq, "Loop detected: last one at %u "
2021                                "indirect size %u\n",
2022                                i, count);
2023                         return -EINVAL;
2024                 }
2025                 if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
2026                         vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
2027                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2028                         return -EINVAL;
2029                 }
2030                 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
2031                         vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
2032                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2033                         return -EINVAL;
2034                 }
2035
2036                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2037                         access = VHOST_ACCESS_WO;
2038                 else
2039                         access = VHOST_ACCESS_RO;
2040
2041                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2042                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
2043                                      iov_size - iov_count, access);
2044                 if (unlikely(ret < 0)) {
2045                         if (ret != -EAGAIN)
2046                                 vq_err(vq, "Translation failure %d indirect idx %d\n",
2047                                         ret, i);
2048                         return ret;
2049                 }
2050                 /* If this is an input descriptor, increment that count. */
2051                 if (access == VHOST_ACCESS_WO) {
2052                         *in_num += ret;
2053                         if (unlikely(log)) {
2054                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2055                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2056                                 ++*log_num;
2057                         }
2058                 } else {
2059                         /* If it's an output descriptor, they're all supposed
2060                          * to come before any input descriptors. */
2061                         if (unlikely(*in_num)) {
2062                                 vq_err(vq, "Indirect descriptor "
2063                                        "has out after in: idx %d\n", i);
2064                                 return -EINVAL;
2065                         }
2066                         *out_num += ret;
2067                 }
2068         } while ((i = next_desc(vq, &desc)) != -1);
2069         return 0;
2070 }
2071
2072 /* This looks in the virtqueue and for the first available buffer, and converts
2073  * it to an iovec for convenient access.  Since descriptors consist of some
2074  * number of output then some number of input descriptors, it's actually two
2075  * iovecs, but we pack them into one and note how many of each there were.
2076  *
2077  * This function returns the descriptor number found, or vq->num (which is
2078  * never a valid descriptor number) if none was found.  A negative code is
2079  * returned on error. */
2080 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
2081                       struct iovec iov[], unsigned int iov_size,
2082                       unsigned int *out_num, unsigned int *in_num,
2083                       struct vhost_log *log, unsigned int *log_num)
2084 {
2085         struct vring_desc desc;
2086         unsigned int i, head, found = 0;
2087         u16 last_avail_idx;
2088         __virtio16 avail_idx;
2089         __virtio16 ring_head;
2090         int ret, access;
2091
2092         /* Check it isn't doing very strange things with descriptor numbers. */
2093         last_avail_idx = vq->last_avail_idx;
2094
2095         if (vq->avail_idx == vq->last_avail_idx) {
2096                 if (unlikely(vhost_get_avail(vq, avail_idx, &vq->avail->idx))) {
2097                         vq_err(vq, "Failed to access avail idx at %p\n",
2098                                 &vq->avail->idx);
2099                         return -EFAULT;
2100                 }
2101                 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2102
2103                 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
2104                         vq_err(vq, "Guest moved used index from %u to %u",
2105                                 last_avail_idx, vq->avail_idx);
2106                         return -EFAULT;
2107                 }
2108
2109                 /* If there's nothing new since last we looked, return
2110                  * invalid.
2111                  */
2112                 if (vq->avail_idx == last_avail_idx)
2113                         return vq->num;
2114
2115                 /* Only get avail ring entries after they have been
2116                  * exposed by guest.
2117                  */
2118                 smp_rmb();
2119         }
2120
2121         /* Grab the next descriptor number they're advertising, and increment
2122          * the index we've seen. */
2123         if (unlikely(vhost_get_avail(vq, ring_head,
2124                      &vq->avail->ring[last_avail_idx & (vq->num - 1)]))) {
2125                 vq_err(vq, "Failed to read head: idx %d address %p\n",
2126                        last_avail_idx,
2127                        &vq->avail->ring[last_avail_idx % vq->num]);
2128                 return -EFAULT;
2129         }
2130
2131         head = vhost16_to_cpu(vq, ring_head);
2132
2133         /* If their number is silly, that's an error. */
2134         if (unlikely(head >= vq->num)) {
2135                 vq_err(vq, "Guest says index %u > %u is available",
2136                        head, vq->num);
2137                 return -EINVAL;
2138         }
2139
2140         /* When we start there are none of either input nor output. */
2141         *out_num = *in_num = 0;
2142         if (unlikely(log))
2143                 *log_num = 0;
2144
2145         i = head;
2146         do {
2147                 unsigned iov_count = *in_num + *out_num;
2148                 if (unlikely(i >= vq->num)) {
2149                         vq_err(vq, "Desc index is %u > %u, head = %u",
2150                                i, vq->num, head);
2151                         return -EINVAL;
2152                 }
2153                 if (unlikely(++found > vq->num)) {
2154                         vq_err(vq, "Loop detected: last one at %u "
2155                                "vq size %u head %u\n",
2156                                i, vq->num, head);
2157                         return -EINVAL;
2158                 }
2159                 ret = vhost_copy_from_user(vq, &desc, vq->desc + i,
2160                                            sizeof desc);
2161                 if (unlikely(ret)) {
2162                         vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
2163                                i, vq->desc + i);
2164                         return -EFAULT;
2165                 }
2166                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2167                         ret = get_indirect(vq, iov, iov_size,
2168                                            out_num, in_num,
2169                                            log, log_num, &desc);
2170                         if (unlikely(ret < 0)) {
2171                                 if (ret != -EAGAIN)
2172                                         vq_err(vq, "Failure detected "
2173                                                 "in indirect descriptor at idx %d\n", i);
2174                                 return ret;
2175                         }
2176                         continue;
2177                 }
2178
2179                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2180                         access = VHOST_ACCESS_WO;
2181                 else
2182                         access = VHOST_ACCESS_RO;
2183                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2184                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
2185                                      iov_size - iov_count, access);
2186                 if (unlikely(ret < 0)) {
2187                         if (ret != -EAGAIN)
2188                                 vq_err(vq, "Translation failure %d descriptor idx %d\n",
2189                                         ret, i);
2190                         return ret;
2191                 }
2192                 if (access == VHOST_ACCESS_WO) {
2193                         /* If this is an input descriptor,
2194                          * increment that count. */
2195                         *in_num += ret;
2196                         if (unlikely(log)) {
2197                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2198                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2199                                 ++*log_num;
2200                         }
2201                 } else {
2202                         /* If it's an output descriptor, they're all supposed
2203                          * to come before any input descriptors. */
2204                         if (unlikely(*in_num)) {
2205                                 vq_err(vq, "Descriptor has out after in: "
2206                                        "idx %d\n", i);
2207                                 return -EINVAL;
2208                         }
2209                         *out_num += ret;
2210                 }
2211         } while ((i = next_desc(vq, &desc)) != -1);
2212
2213         /* On success, increment avail index. */
2214         vq->last_avail_idx++;
2215
2216         /* Assume notifications from guest are disabled at this point,
2217          * if they aren't we would need to update avail_event index. */
2218         BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2219         return head;
2220 }
2221 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2222
2223 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2224 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2225 {
2226         vq->last_avail_idx -= n;
2227 }
2228 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2229
2230 /* After we've used one of their buffers, we tell them about it.  We'll then
2231  * want to notify the guest, using eventfd. */
2232 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2233 {
2234         struct vring_used_elem heads = {
2235                 cpu_to_vhost32(vq, head),
2236                 cpu_to_vhost32(vq, len)
2237         };
2238
2239         return vhost_add_used_n(vq, &heads, 1);
2240 }
2241 EXPORT_SYMBOL_GPL(vhost_add_used);
2242
2243 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2244                             struct vring_used_elem *heads,
2245                             unsigned count)
2246 {
2247         struct vring_used_elem __user *used;
2248         u16 old, new;
2249         int start;
2250
2251         start = vq->last_used_idx & (vq->num - 1);
2252         used = vq->used->ring + start;
2253         if (count == 1) {
2254                 if (vhost_put_user(vq, heads[0].id, &used->id)) {
2255                         vq_err(vq, "Failed to write used id");
2256                         return -EFAULT;
2257                 }
2258                 if (vhost_put_user(vq, heads[0].len, &used->len)) {
2259                         vq_err(vq, "Failed to write used len");
2260                         return -EFAULT;
2261                 }
2262         } else if (vhost_copy_to_user(vq, used, heads, count * sizeof *used)) {
2263                 vq_err(vq, "Failed to write used");
2264                 return -EFAULT;
2265         }
2266         if (unlikely(vq->log_used)) {
2267                 /* Make sure data is seen before log. */
2268                 smp_wmb();
2269                 /* Log used ring entry write. */
2270                 log_used(vq, ((void __user *)used - (void __user *)vq->used),
2271                          count * sizeof *used);
2272         }
2273         old = vq->last_used_idx;
2274         new = (vq->last_used_idx += count);
2275         /* If the driver never bothers to signal in a very long while,
2276          * used index might wrap around. If that happens, invalidate
2277          * signalled_used index we stored. TODO: make sure driver
2278          * signals at least once in 2^16 and remove this. */
2279         if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2280                 vq->signalled_used_valid = false;
2281         return 0;
2282 }
2283
2284 /* After we've used one of their buffers, we tell them about it.  We'll then
2285  * want to notify the guest, using eventfd. */
2286 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2287                      unsigned count)
2288 {
2289         int start, n, r;
2290
2291         start = vq->last_used_idx & (vq->num - 1);
2292         n = vq->num - start;
2293         if (n < count) {
2294                 r = __vhost_add_used_n(vq, heads, n);
2295                 if (r < 0)
2296                         return r;
2297                 heads += n;
2298                 count -= n;
2299         }
2300         r = __vhost_add_used_n(vq, heads, count);
2301
2302         /* Make sure buffer is written before we update index. */
2303         smp_wmb();
2304         if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
2305                            &vq->used->idx)) {
2306                 vq_err(vq, "Failed to increment used idx");
2307                 return -EFAULT;
2308         }
2309         if (unlikely(vq->log_used)) {
2310                 /* Make sure used idx is seen before log. */
2311                 smp_wmb();
2312                 /* Log used index update. */
2313                 log_used(vq, offsetof(struct vring_used, idx),
2314                          sizeof vq->used->idx);
2315                 if (vq->log_ctx)
2316                         eventfd_signal(vq->log_ctx, 1);
2317         }
2318         return r;
2319 }
2320 EXPORT_SYMBOL_GPL(vhost_add_used_n);
2321
2322 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2323 {
2324         __u16 old, new;
2325         __virtio16 event;
2326         bool v;
2327         /* Flush out used index updates. This is paired
2328          * with the barrier that the Guest executes when enabling
2329          * interrupts. */
2330         smp_mb();
2331
2332         if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2333             unlikely(vq->avail_idx == vq->last_avail_idx))
2334                 return true;
2335
2336         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2337                 __virtio16 flags;
2338                 if (vhost_get_avail(vq, flags, &vq->avail->flags)) {
2339                         vq_err(vq, "Failed to get flags");
2340                         return true;
2341                 }
2342                 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2343         }
2344         old = vq->signalled_used;
2345         v = vq->signalled_used_valid;
2346         new = vq->signalled_used = vq->last_used_idx;
2347         vq->signalled_used_valid = true;
2348
2349         if (unlikely(!v))
2350                 return true;
2351
2352         if (vhost_get_avail(vq, event, vhost_used_event(vq))) {
2353                 vq_err(vq, "Failed to get used event idx");
2354                 return true;
2355         }
2356         return vring_need_event(vhost16_to_cpu(vq, event), new, old);
2357 }
2358
2359 /* This actually signals the guest, using eventfd. */
2360 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2361 {
2362         /* Signal the Guest tell them we used something up. */
2363         if (vq->call_ctx && vhost_notify(dev, vq))
2364                 eventfd_signal(vq->call_ctx, 1);
2365 }
2366 EXPORT_SYMBOL_GPL(vhost_signal);
2367
2368 /* And here's the combo meal deal.  Supersize me! */
2369 void vhost_add_used_and_signal(struct vhost_dev *dev,
2370                                struct vhost_virtqueue *vq,
2371                                unsigned int head, int len)
2372 {
2373         vhost_add_used(vq, head, len);
2374         vhost_signal(dev, vq);
2375 }
2376 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2377
2378 /* multi-buffer version of vhost_add_used_and_signal */
2379 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2380                                  struct vhost_virtqueue *vq,
2381                                  struct vring_used_elem *heads, unsigned count)
2382 {
2383         vhost_add_used_n(vq, heads, count);
2384         vhost_signal(dev, vq);
2385 }
2386 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2387
2388 /* return true if we're sure that avaiable ring is empty */
2389 bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2390 {
2391         __virtio16 avail_idx;
2392         int r;
2393
2394         if (vq->avail_idx != vq->last_avail_idx)
2395                 return false;
2396
2397         r = vhost_get_avail(vq, avail_idx, &vq->avail->idx);
2398         if (unlikely(r))
2399                 return false;
2400         vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2401
2402         return vq->avail_idx == vq->last_avail_idx;
2403 }
2404 EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2405
2406 /* OK, now we need to know about added descriptors. */
2407 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2408 {
2409         __virtio16 avail_idx;
2410         int r;
2411
2412         if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2413                 return false;
2414         vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2415         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2416                 r = vhost_update_used_flags(vq);
2417                 if (r) {
2418                         vq_err(vq, "Failed to enable notification at %p: %d\n",
2419                                &vq->used->flags, r);
2420                         return false;
2421                 }
2422         } else {
2423                 r = vhost_update_avail_event(vq, vq->avail_idx);
2424                 if (r) {
2425                         vq_err(vq, "Failed to update avail event index at %p: %d\n",
2426                                vhost_avail_event(vq), r);
2427                         return false;
2428                 }
2429         }
2430         /* They could have slipped one in as we were doing that: make
2431          * sure it's written, then check again. */
2432         smp_mb();
2433         r = vhost_get_avail(vq, avail_idx, &vq->avail->idx);
2434         if (r) {
2435                 vq_err(vq, "Failed to check avail idx at %p: %d\n",
2436                        &vq->avail->idx, r);
2437                 return false;
2438         }
2439
2440         return vhost16_to_cpu(vq, avail_idx) != vq->avail_idx;
2441 }
2442 EXPORT_SYMBOL_GPL(vhost_enable_notify);
2443
2444 /* We don't need to be notified again. */
2445 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2446 {
2447         int r;
2448
2449         if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2450                 return;
2451         vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2452         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2453                 r = vhost_update_used_flags(vq);
2454                 if (r)
2455                         vq_err(vq, "Failed to enable notification at %p: %d\n",
2456                                &vq->used->flags, r);
2457         }
2458 }
2459 EXPORT_SYMBOL_GPL(vhost_disable_notify);
2460
2461 /* Create a new message. */
2462 struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2463 {
2464         struct vhost_msg_node *node = kmalloc(sizeof *node, GFP_KERNEL);
2465         if (!node)
2466                 return NULL;
2467
2468         /* Make sure all padding within the structure is initialized. */
2469         memset(&node->msg, 0, sizeof node->msg);
2470         node->vq = vq;
2471         node->msg.type = type;
2472         return node;
2473 }
2474 EXPORT_SYMBOL_GPL(vhost_new_msg);
2475
2476 void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2477                        struct vhost_msg_node *node)
2478 {
2479         spin_lock(&dev->iotlb_lock);
2480         list_add_tail(&node->node, head);
2481         spin_unlock(&dev->iotlb_lock);
2482
2483         wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
2484 }
2485 EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2486
2487 struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2488                                          struct list_head *head)
2489 {
2490         struct vhost_msg_node *node = NULL;
2491
2492         spin_lock(&dev->iotlb_lock);
2493         if (!list_empty(head)) {
2494                 node = list_first_entry(head, struct vhost_msg_node,
2495                                         node);
2496                 list_del(&node->node);
2497         }
2498         spin_unlock(&dev->iotlb_lock);
2499
2500         return node;
2501 }
2502 EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2503
2504
2505 static int __init vhost_init(void)
2506 {
2507         return 0;
2508 }
2509
2510 static void __exit vhost_exit(void)
2511 {
2512 }
2513
2514 module_init(vhost_init);
2515 module_exit(vhost_exit);
2516
2517 MODULE_VERSION("0.0.1");
2518 MODULE_LICENSE("GPL v2");
2519 MODULE_AUTHOR("Michael S. Tsirkin");
2520 MODULE_DESCRIPTION("Host kernel accelerator for virtio");