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 <[email protected]>
   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
  34 #include "vhost.h"
  35
  36 static ushort max_mem_regions = 64;
  37 module_param(max_mem_regions, ushort, 0444);
  38 MODULE_PARM_DESC(max_mem_regions,
  39         "Maximum number of memory regions in memory map. (default: 64)");
  40 static int max_iotlb_entries = 2048;
  41 module_param(max_iotlb_entries, int, 0444);
  42 MODULE_PARM_DESC(max_iotlb_entries,
  43         "Maximum number of iotlb entries. (default: 2048)");
  44
  45 enum {
  46         VHOST_MEMORY_F_LOG = 0x1,
  47 };
  48
  49 #define vhost_used_event(vq) ((__virtio16 __user *)&vq->avail->ring[vq->num])
  50 #define vhost_avail_event(vq) ((__virtio16 __user *)&vq->used->ring[vq->num])
  51
  52 INTERVAL_TREE_DEFINE(struct vhost_umem_node,
  53                      rb, __u64, __subtree_last,
  54                      START, LAST, static inline, vhost_umem_interval_tree);
  55
  56 #ifdef CONFIG_VHOST_CROSS_ENDIAN_LEGACY
  57 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
  58 {
  59         vq->user_be = !virtio_legacy_is_little_endian();
  60 }
  61
  62 static void vhost_enable_cross_endian_big(struct vhost_virtqueue *vq)
  63 {
  64         vq->user_be = true;
  65 }
  66
  67 static void vhost_enable_cross_endian_little(struct vhost_virtqueue *vq)
  68 {
  69         vq->user_be = false;
  70 }
  71
  72 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
  73 {
  74         struct vhost_vring_state s;
  75
  76         if (vq->private_data)
  77                 return -EBUSY;
  78
  79         if (copy_from_user(&s, argp, sizeof(s)))
  80                 return -EFAULT;
  81
  82         if (s.num != VHOST_VRING_LITTLE_ENDIAN &&
  83             s.num != VHOST_VRING_BIG_ENDIAN)
  84                 return -EINVAL;
  85
  86         if (s.num == VHOST_VRING_BIG_ENDIAN)
  87                 vhost_enable_cross_endian_big(vq);
  88         else
  89                 vhost_enable_cross_endian_little(vq);
  90
  91         return 0;
  92 }
  93
  94 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
  95                                    int __user *argp)
  96 {
  97         struct vhost_vring_state s = {
  98                 .index = idx,
  99                 .num = vq->user_be
 100         };
 101
 102         if (copy_to_user(argp, &s, sizeof(s)))
 103                 return -EFAULT;
 104
 105         return 0;
 106 }
 107
 108 static void vhost_init_is_le(struct vhost_virtqueue *vq)
 109 {
 110         /* Note for legacy virtio: user_be is initialized at reset time
 111          * according to the host endianness. If userspace does not set an
 112          * explicit endianness, the default behavior is native endian, as
 113          * expected by legacy virtio.
 114          */
 115         vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1) || !vq->user_be;
 116 }
 117 #else
 118 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
 119 {
 120 }
 121
 122 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
 123 {
 124         return -ENOIOCTLCMD;
 125 }
 126
 127 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
 128                                    int __user *argp)
 129 {
 130         return -ENOIOCTLCMD;
 131 }
 132
 133 static void vhost_init_is_le(struct vhost_virtqueue *vq)
 134 {
 135         vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1)
 136                 || virtio_legacy_is_little_endian();
 137 }
 138 #endif /* CONFIG_VHOST_CROSS_ENDIAN_LEGACY */
 139
 140 static void vhost_reset_is_le(struct vhost_virtqueue *vq)
 141 {
 142         vhost_init_is_le(vq);
 143 }
 144
 145 struct vhost_flush_struct {
 146         struct vhost_work work;
 147         struct completion wait_event;
 148 };
 149
 150 static void vhost_flush_work(struct vhost_work *work)
 151 {
 152         struct vhost_flush_struct *s;
 153
 154         s = container_of(work, struct vhost_flush_struct, work);
 155         complete(&s->wait_event);
 156 }
 157
 158 static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
 159                             poll_table *pt)
 160 {
 161         struct vhost_poll *poll;
 162
 163         poll = container_of(pt, struct vhost_poll, table);
 164         poll->wqh = wqh;
 165         add_wait_queue(wqh, &poll->wait);
 166 }
 167
 168 static int vhost_poll_wakeup(wait_queue_entry_t *wait, unsigned mode, int sync,
 169                              void *key)
 170 {
 171         struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
 172
 173         if (!(key_to_poll(key) & poll->mask))
 174                 return 0;
 175
 176         vhost_poll_queue(poll);
 177         return 0;
 178 }
 179
 180 void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
 181 {
 182         clear_bit(VHOST_WORK_QUEUED, &work->flags);
 183         work->fn = fn;
 184 }
 185 EXPORT_SYMBOL_GPL(vhost_work_init);
 186
 187 /* Init poll structure */
 188 void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
 189                      __poll_t mask, struct vhost_dev *dev)
 190 {
 191         init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
 192         init_poll_funcptr(&poll->table, vhost_poll_func);
 193         poll->mask = mask;
 194         poll->dev = dev;
 195         poll->wqh = NULL;
 196
 197         vhost_work_init(&poll->work, fn);
 198 }
 199 EXPORT_SYMBOL_GPL(vhost_poll_init);
 200
 201 /* Start polling a file. We add ourselves to file's wait queue. The caller must
 202  * keep a reference to a file until after vhost_poll_stop is called. */
 203 int vhost_poll_start(struct vhost_poll *poll, struct file *file)
 204 {
 205         __poll_t mask;
 206         int ret = 0;
 207
 208         if (poll->wqh)
 209                 return 0;
 210
 211         mask = file->f_op->poll(file, &poll->table);
 212         if (mask)
 213                 vhost_poll_wakeup(&poll->wait, 0, 0, poll_to_key(mask));
 214         if (mask & EPOLLERR) {
 215                 if (poll->wqh)
 216                         remove_wait_queue(poll->wqh, &poll->wait);
 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->log_base = NULL;
 320         vq->error_ctx = NULL;
 321         vq->kick = NULL;
 322         vq->call_ctx = NULL;
 323         vq->log_ctx = NULL;
 324         vhost_reset_is_le(vq);
 325         vhost_disable_cross_endian(vq);
 326         vq->busyloop_timeout = 0;
 327         vq->umem = NULL;
 328         vq->iotlb = NULL;
 329         __vhost_vq_meta_reset(vq);
 330 }
 331
 332 static int vhost_worker(void *data)
 333 {
 334         struct vhost_dev *dev = data;
 335         struct vhost_work *work, *work_next;
 336         struct llist_node *node;
 337         mm_segment_t oldfs = get_fs();
 338
 339         set_fs(USER_DS);
 340         use_mm(dev->mm);
 341
 342         for (;;) {
 343                 /* mb paired w/ kthread_stop */
 344                 set_current_state(TASK_INTERRUPTIBLE);
 345
 346                 if (kthread_should_stop()) {
 347                         __set_current_state(TASK_RUNNING);
 348                         break;
 349                 }
 350
 351                 node = llist_del_all(&dev->work_list);
 352                 if (!node)
 353                         schedule();
 354
 355                 node = llist_reverse_order(node);
 356                 /* make sure flag is seen after deletion */
 357                 smp_wmb();
 358                 llist_for_each_entry_safe(work, work_next, node, node) {
 359                         clear_bit(VHOST_WORK_QUEUED, &work->flags);
 360                         __set_current_state(TASK_RUNNING);
 361                         work->fn(work);
 362                         if (need_resched())
 363                                 schedule();
 364                 }
 365         }
 366         unuse_mm(dev->mm);
 367         set_fs(oldfs);
 368         return 0;
 369 }
 370
 371 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
 372 {
 373         kfree(vq->indirect);
 374         vq->indirect = NULL;
 375         kfree(vq->log);
 376         vq->log = NULL;
 377         kfree(vq->heads);
 378         vq->heads = NULL;
 379 }
 380
 381 /* Helper to allocate iovec buffers for all vqs. */
 382 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
 383 {
 384         struct vhost_virtqueue *vq;
 385         int i;
 386
 387         for (i = 0; i < dev->nvqs; ++i) {
 388                 vq = dev->vqs[i];
 389                 vq->indirect = kmalloc(sizeof *vq->indirect * UIO_MAXIOV,
 390                                        GFP_KERNEL);
 391                 vq->log = kmalloc(sizeof *vq->log * UIO_MAXIOV, GFP_KERNEL);
 392                 vq->heads = kmalloc(sizeof *vq->heads * UIO_MAXIOV, GFP_KERNEL);
 393                 if (!vq->indirect || !vq->log || !vq->heads)
 394                         goto err_nomem;
 395         }
 396         return 0;
 397
 398 err_nomem:
 399         for (; i >= 0; --i)
 400                 vhost_vq_free_iovecs(dev->vqs[i]);
 401         return -ENOMEM;
 402 }
 403
 404 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
 405 {
 406         int i;
 407
 408         for (i = 0; i < dev->nvqs; ++i)
 409                 vhost_vq_free_iovecs(dev->vqs[i]);
 410 }
 411
 412 void vhost_dev_init(struct vhost_dev *dev,
 413                     struct vhost_virtqueue **vqs, int nvqs)
 414 {
 415         struct vhost_virtqueue *vq;
 416         int i;
 417
 418         dev->vqs = vqs;
 419         dev->nvqs = nvqs;
 420         mutex_init(&dev->mutex);
 421         dev->log_ctx = NULL;
 422         dev->umem = NULL;
 423         dev->iotlb = NULL;
 424         dev->mm = NULL;
 425         dev->worker = NULL;
 426         init_llist_head(&dev->work_list);
 427         init_waitqueue_head(&dev->wait);
 428         INIT_LIST_HEAD(&dev->read_list);
 429         INIT_LIST_HEAD(&dev->pending_list);
 430         spin_lock_init(&dev->iotlb_lock);
 431
 432
 433         for (i = 0; i < dev->nvqs; ++i) {
 434                 vq = dev->vqs[i];
 435                 vq->log = NULL;
 436                 vq->indirect = NULL;
 437                 vq->heads = NULL;
 438                 vq->dev = dev;
 439                 mutex_init(&vq->mutex);
 440                 vhost_vq_reset(dev, vq);
 441                 if (vq->handle_kick)
 442                         vhost_poll_init(&vq->poll, vq->handle_kick,
 443                                         EPOLLIN, dev);
 444         }
 445 }
 446 EXPORT_SYMBOL_GPL(vhost_dev_init);
 447
 448 /* Caller should have device mutex */
 449 long vhost_dev_check_owner(struct vhost_dev *dev)
 450 {
 451         /* Are you the owner? If not, I don't think you mean to do that */
 452         return dev->mm == current->mm ? 0 : -EPERM;
 453 }
 454 EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
 455
 456 struct vhost_attach_cgroups_struct {
 457         struct vhost_work work;
 458         struct task_struct *owner;
 459         int ret;
 460 };
 461
 462 static void vhost_attach_cgroups_work(struct vhost_work *work)
 463 {
 464         struct vhost_attach_cgroups_struct *s;
 465
 466         s = container_of(work, struct vhost_attach_cgroups_struct, work);
 467         s->ret = cgroup_attach_task_all(s->owner, current);
 468 }
 469
 470 static int vhost_attach_cgroups(struct vhost_dev *dev)
 471 {
 472         struct vhost_attach_cgroups_struct attach;
 473
 474         attach.owner = current;
 475         vhost_work_init(&attach.work, vhost_attach_cgroups_work);
 476         vhost_work_queue(dev, &attach.work);
 477         vhost_work_flush(dev, &attach.work);
 478         return attach.ret;
 479 }
 480
 481 /* Caller should have device mutex */
 482 bool vhost_dev_has_owner(struct vhost_dev *dev)
 483 {
 484         return dev->mm;
 485 }
 486 EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
 487
 488 /* Caller should have device mutex */
 489 long vhost_dev_set_owner(struct vhost_dev *dev)
 490 {
 491         struct task_struct *worker;
 492         int err;
 493
 494         /* Is there an owner already? */
 495         if (vhost_dev_has_owner(dev)) {
 496                 err = -EBUSY;
 497                 goto err_mm;
 498         }
 499
 500         /* No owner, become one */
 501         dev->mm = get_task_mm(current);
 502         worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
 503         if (IS_ERR(worker)) {
 504                 err = PTR_ERR(worker);
 505                 goto err_worker;
 506         }
 507
 508         dev->worker = worker;
 509         wake_up_process(worker);        /* avoid contributing to loadavg */
 510
 511         err = vhost_attach_cgroups(dev);
 512         if (err)
 513                 goto err_cgroup;
 514
 515         err = vhost_dev_alloc_iovecs(dev);
 516         if (err)
 517                 goto err_cgroup;
 518
 519         return 0;
 520 err_cgroup:
 521         kthread_stop(worker);
 522         dev->worker = NULL;
 523 err_worker:
 524         if (dev->mm)
 525                 mmput(dev->mm);
 526         dev->mm = NULL;
 527 err_mm:
 528         return err;
 529 }
 530 EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
 531
 532 struct vhost_umem *vhost_dev_reset_owner_prepare(void)
 533 {
 534         return kvzalloc(sizeof(struct vhost_umem), GFP_KERNEL);
 535 }
 536 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
 537
 538 /* Caller should have device mutex */
 539 void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_umem *umem)
 540 {
 541         int i;
 542
 543         vhost_dev_cleanup(dev);
 544
 545         /* Restore memory to default empty mapping. */
 546         INIT_LIST_HEAD(&umem->umem_list);
 547         dev->umem = umem;
 548         /* We don't need VQ locks below since vhost_dev_cleanup makes sure
 549          * VQs aren't running.
 550          */
 551         for (i = 0; i < dev->nvqs; ++i)
 552                 dev->vqs[i]->umem = umem;
 553 }
 554 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
 555
 556 void vhost_dev_stop(struct vhost_dev *dev)
 557 {
 558         int i;
 559
 560         for (i = 0; i < dev->nvqs; ++i) {
 561                 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick) {
 562                         vhost_poll_stop(&dev->vqs[i]->poll);
 563                         vhost_poll_flush(&dev->vqs[i]->poll);
 564                 }
 565         }
 566 }
 567 EXPORT_SYMBOL_GPL(vhost_dev_stop);
 568
 569 static void vhost_umem_free(struct vhost_umem *umem,
 570                             struct vhost_umem_node *node)
 571 {
 572         vhost_umem_interval_tree_remove(node, &umem->umem_tree);
 573         list_del(&node->link);
 574         kfree(node);
 575         umem->numem--;
 576 }
 577
 578 static void vhost_umem_clean(struct vhost_umem *umem)
 579 {
 580         struct vhost_umem_node *node, *tmp;
 581
 582         if (!umem)
 583                 return;
 584
 585         list_for_each_entry_safe(node, tmp, &umem->umem_list, link)
 586                 vhost_umem_free(umem, node);
 587
 588         kvfree(umem);
 589 }
 590
 591 static void vhost_clear_msg(struct vhost_dev *dev)
 592 {
 593         struct vhost_msg_node *node, *n;
 594
 595         spin_lock(&dev->iotlb_lock);
 596
 597         list_for_each_entry_safe(node, n, &dev->read_list, node) {
 598                 list_del(&node->node);
 599                 kfree(node);
 600         }
 601
 602         list_for_each_entry_safe(node, n, &dev->pending_list, node) {
 603                 list_del(&node->node);
 604                 kfree(node);
 605         }
 606
 607         spin_unlock(&dev->iotlb_lock);
 608 }
 609
 610 void vhost_dev_cleanup(struct vhost_dev *dev)
 611 {
 612         int i;
 613
 614         for (i = 0; i < dev->nvqs; ++i) {
 615                 if (dev->vqs[i]->error_ctx)
 616                         eventfd_ctx_put(dev->vqs[i]->error_ctx);
 617                 if (dev->vqs[i]->kick)
 618                         fput(dev->vqs[i]->kick);
 619                 if (dev->vqs[i]->call_ctx)
 620                         eventfd_ctx_put(dev->vqs[i]->call_ctx);
 621                 vhost_vq_reset(dev, dev->vqs[i]);
 622         }
 623         vhost_dev_free_iovecs(dev);
 624         if (dev->log_ctx)
 625                 eventfd_ctx_put(dev->log_ctx);
 626         dev->log_ctx = NULL;
 627         /* No one will access memory at this point */
 628         vhost_umem_clean(dev->umem);
 629         dev->umem = NULL;
 630         vhost_umem_clean(dev->iotlb);
 631         dev->iotlb = NULL;
 632         vhost_clear_msg(dev);
 633         wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
 634         WARN_ON(!llist_empty(&dev->work_list));
 635         if (dev->worker) {
 636                 kthread_stop(dev->worker);
 637                 dev->worker = NULL;
 638         }
 639         if (dev->mm)
 640                 mmput(dev->mm);
 641         dev->mm = NULL;
 642 }
 643 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
 644
 645 static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
 646 {
 647         u64 a = addr / VHOST_PAGE_SIZE / 8;
 648
 649         /* Make sure 64 bit math will not overflow. */
 650         if (a > ULONG_MAX - (unsigned long)log_base ||
 651             a + (unsigned long)log_base > ULONG_MAX)
 652                 return 0;
 653
 654         return access_ok(VERIFY_WRITE, log_base + a,
 655                          (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
 656 }
 657
 658 static bool vhost_overflow(u64 uaddr, u64 size)
 659 {
 660         /* Make sure 64 bit math will not overflow. */
 661         return uaddr > ULONG_MAX || size > ULONG_MAX || uaddr > ULONG_MAX - size;
 662 }
 663
 664 /* Caller should have vq mutex and device mutex. */
 665 static int vq_memory_access_ok(void __user *log_base, struct vhost_umem *umem,
 666                                int log_all)
 667 {
 668         struct vhost_umem_node *node;
 669
 670         if (!umem)
 671                 return 0;
 672
 673         list_for_each_entry(node, &umem->umem_list, link) {
 674                 unsigned long a = node->userspace_addr;
 675
 676                 if (vhost_overflow(node->userspace_addr, node->size))
 677                         return 0;
 678
 679
 680                 if (!access_ok(VERIFY_WRITE, (void __user *)a,
 681                                     node->size))
 682                         return 0;
 683                 else if (log_all && !log_access_ok(log_base,
 684                                                    node->start,
 685                                                    node->size))
 686                         return 0;
 687         }
 688         return 1;
 689 }
 690
 691 static inline void __user *vhost_vq_meta_fetch(struct vhost_virtqueue *vq,
 692                                                u64 addr, unsigned int size,
 693                                                int type)
 694 {
 695         const struct vhost_umem_node *node = vq->meta_iotlb[type];
 696
 697         if (!node)
 698                 return NULL;
 699
 700         return (void *)(uintptr_t)(node->userspace_addr + addr - node->start);
 701 }
 702
 703 /* Can we switch to this memory table? */
 704 /* Caller should have device mutex but not vq mutex */
 705 static int memory_access_ok(struct vhost_dev *d, struct vhost_umem *umem,
 706                             int log_all)
 707 {
 708         int i;
 709
 710         for (i = 0; i < d->nvqs; ++i) {
 711                 int ok;
 712                 bool log;
 713
 714                 mutex_lock(&d->vqs[i]->mutex);
 715                 log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
 716                 /* If ring is inactive, will check when it's enabled. */
 717                 if (d->vqs[i]->private_data)
 718                         ok = vq_memory_access_ok(d->vqs[i]->log_base,
 719                                                  umem, log);
 720                 else
 721                         ok = 1;
 722                 mutex_unlock(&d->vqs[i]->mutex);
 723                 if (!ok)
 724                         return 0;
 725         }
 726         return 1;
 727 }
 728
 729 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
 730                           struct iovec iov[], int iov_size, int access);
 731
 732 static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
 733                               const void *from, unsigned size)
 734 {
 735         int ret;
 736
 737         if (!vq->iotlb)
 738                 return __copy_to_user(to, from, size);
 739         else {
 740                 /* This function should be called after iotlb
 741                  * prefetch, which means we're sure that all vq
 742                  * could be access through iotlb. So -EAGAIN should
 743                  * not happen in this case.
 744                  */
 745                 struct iov_iter t;
 746                 void __user *uaddr = vhost_vq_meta_fetch(vq,
 747                                      (u64)(uintptr_t)to, size,
 748                                      VHOST_ADDR_DESC);
 749
 750                 if (uaddr)
 751                         return __copy_to_user(uaddr, from, size);
 752
 753                 ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
 754                                      ARRAY_SIZE(vq->iotlb_iov),
 755                                      VHOST_ACCESS_WO);
 756                 if (ret < 0)
 757                         goto out;
 758                 iov_iter_init(&t, WRITE, vq->iotlb_iov, ret, size);
 759                 ret = copy_to_iter(from, size, &t);
 760                 if (ret == size)
 761                         ret = 0;
 762         }
 763 out:
 764         return ret;
 765 }
 766
 767 static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
 768                                 void __user *from, unsigned size)
 769 {
 770         int ret;
 771
 772         if (!vq->iotlb)
 773                 return __copy_from_user(to, from, size);
 774         else {
 775                 /* This function should be called after iotlb
 776                  * prefetch, which means we're sure that vq
 777                  * could be access through iotlb. So -EAGAIN should
 778                  * not happen in this case.
 779                  */
 780                 void __user *uaddr = vhost_vq_meta_fetch(vq,
 781                                      (u64)(uintptr_t)from, size,
 782                                      VHOST_ADDR_DESC);
 783                 struct iov_iter f;
 784
 785                 if (uaddr)
 786                         return __copy_from_user(to, uaddr, size);
 787
 788                 ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
 789                                      ARRAY_SIZE(vq->iotlb_iov),
 790                                      VHOST_ACCESS_RO);
 791                 if (ret < 0) {
 792                         vq_err(vq, "IOTLB translation failure: uaddr "
 793                                "%p size 0x%llx\n", from,
 794                                (unsigned long long) size);
 795                         goto out;
 796                 }
 797                 iov_iter_init(&f, READ, vq->iotlb_iov, ret, size);
 798                 ret = copy_from_iter(to, size, &f);
 799                 if (ret == size)
 800                         ret = 0;
 801         }
 802
 803 out:
 804         return ret;
 805 }
 806
 807 static void __user *__vhost_get_user_slow(struct vhost_virtqueue *vq,
 808                                           void __user *addr, unsigned int size,
 809                                           int type)
 810 {
 811         int ret;
 812
 813         ret = translate_desc(vq, (u64)(uintptr_t)addr, size, vq->iotlb_iov,
 814                              ARRAY_SIZE(vq->iotlb_iov),
 815                              VHOST_ACCESS_RO);
 816         if (ret < 0) {
 817                 vq_err(vq, "IOTLB translation failure: uaddr "
 818                         "%p size 0x%llx\n", addr,
 819                         (unsigned long long) size);
 820                 return NULL;
 821         }
 822
 823         if (ret != 1 || vq->iotlb_iov[0].iov_len != size) {
 824                 vq_err(vq, "Non atomic userspace memory access: uaddr "
 825                         "%p size 0x%llx\n", addr,
 826                         (unsigned long long) size);
 827                 return NULL;
 828         }
 829
 830         return vq->iotlb_iov[0].iov_base;
 831 }
 832
 833 /* This function should be called after iotlb
 834  * prefetch, which means we're sure that vq
 835  * could be access through iotlb. So -EAGAIN should
 836  * not happen in this case.
 837  */
 838 static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
 839                                             void *addr, unsigned int size,
 840                                             int type)
 841 {
 842         void __user *uaddr = vhost_vq_meta_fetch(vq,
 843                              (u64)(uintptr_t)addr, size, type);
 844         if (uaddr)
 845                 return uaddr;
 846
 847         return __vhost_get_user_slow(vq, addr, size, type);
 848 }
 849
 850 #define vhost_put_user(vq, x, ptr)              \
 851 ({ \
 852         int ret = -EFAULT; \
 853         if (!vq->iotlb) { \
 854                 ret = __put_user(x, ptr); \
 855         } else { \
 856                 __typeof__(ptr) to = \
 857                         (__typeof__(ptr)) __vhost_get_user(vq, ptr,     \
 858                                           sizeof(*ptr), VHOST_ADDR_USED); \
 859                 if (to != NULL) \
 860                         ret = __put_user(x, to); \
 861                 else \
 862                         ret = -EFAULT;  \
 863         } \
 864         ret; \
 865 })
 866
 867 #define vhost_get_user(vq, x, ptr, type)                \
 868 ({ \
 869         int ret; \
 870         if (!vq->iotlb) { \
 871                 ret = __get_user(x, ptr); \
 872         } else { \
 873                 __typeof__(ptr) from = \
 874                         (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
 875                                                            sizeof(*ptr), \
 876                                                            type); \
 877                 if (from != NULL) \
 878                         ret = __get_user(x, from); \
 879                 else \
 880                         ret = -EFAULT; \
 881         } \
 882         ret; \
 883 })
 884
 885 #define vhost_get_avail(vq, x, ptr) \
 886         vhost_get_user(vq, x, ptr, VHOST_ADDR_AVAIL)
 887
 888 #define vhost_get_used(vq, x, ptr) \
 889         vhost_get_user(vq, x, ptr, VHOST_ADDR_USED)
 890
 891 static void vhost_dev_lock_vqs(struct vhost_dev *d)
 892 {
 893         int i = 0;
 894         for (i = 0; i < d->nvqs; ++i)
 895                 mutex_lock_nested(&d->vqs[i]->mutex, i);
 896 }
 897
 898 static void vhost_dev_unlock_vqs(struct vhost_dev *d)
 899 {
 900         int i = 0;
 901         for (i = 0; i < d->nvqs; ++i)
 902                 mutex_unlock(&d->vqs[i]->mutex);
 903 }
 904
 905 static int vhost_new_umem_range(struct vhost_umem *umem,
 906                                 u64 start, u64 size, u64 end,
 907                                 u64 userspace_addr, int perm)
 908 {
 909         struct vhost_umem_node *tmp, *node = kmalloc(sizeof(*node), GFP_ATOMIC);
 910
 911         if (!node)
 912                 return -ENOMEM;
 913
 914         if (umem->numem == max_iotlb_entries) {
 915                 tmp = list_first_entry(&umem->umem_list, typeof(*tmp), link);
 916                 vhost_umem_free(umem, tmp);
 917         }
 918
 919         node->start = start;
 920         node->size = size;
 921         node->last = end;
 922         node->userspace_addr = userspace_addr;
 923         node->perm = perm;
 924         INIT_LIST_HEAD(&node->link);
 925         list_add_tail(&node->link, &umem->umem_list);
 926         vhost_umem_interval_tree_insert(node, &umem->umem_tree);
 927         umem->numem++;
 928
 929         return 0;
 930 }
 931
 932 static void vhost_del_umem_range(struct vhost_umem *umem,
 933                                  u64 start, u64 end)
 934 {
 935         struct vhost_umem_node *node;
 936
 937         while ((node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
 938                                                            start, end)))
 939                 vhost_umem_free(umem, node);
 940 }
 941
 942 static void vhost_iotlb_notify_vq(struct vhost_dev *d,
 943                                   struct vhost_iotlb_msg *msg)
 944 {
 945         struct vhost_msg_node *node, *n;
 946
 947         spin_lock(&d->iotlb_lock);
 948
 949         list_for_each_entry_safe(node, n, &d->pending_list, node) {
 950                 struct vhost_iotlb_msg *vq_msg = &node->msg.iotlb;
 951                 if (msg->iova <= vq_msg->iova &&
 952                     msg->iova + msg->size - 1 > vq_msg->iova &&
 953                     vq_msg->type == VHOST_IOTLB_MISS) {
 954                         vhost_poll_queue(&node->vq->poll);
 955                         list_del(&node->node);
 956                         kfree(node);
 957                 }
 958         }
 959
 960         spin_unlock(&d->iotlb_lock);
 961 }
 962
 963 static int umem_access_ok(u64 uaddr, u64 size, int access)
 964 {
 965         unsigned long a = uaddr;
 966
 967         /* Make sure 64 bit math will not overflow. */
 968         if (vhost_overflow(uaddr, size))
 969                 return -EFAULT;
 970
 971         if ((access & VHOST_ACCESS_RO) &&
 972             !access_ok(VERIFY_READ, (void __user *)a, size))
 973                 return -EFAULT;
 974         if ((access & VHOST_ACCESS_WO) &&
 975             !access_ok(VERIFY_WRITE, (void __user *)a, size))
 976                 return -EFAULT;
 977         return 0;
 978 }
 979
 980 static int vhost_process_iotlb_msg(struct vhost_dev *dev,
 981                                    struct vhost_iotlb_msg *msg)
 982 {
 983         int ret = 0;
 984
 985         vhost_dev_lock_vqs(dev);
 986         switch (msg->type) {
 987         case VHOST_IOTLB_UPDATE:
 988                 if (!dev->iotlb) {
 989                         ret = -EFAULT;
 990                         break;
 991                 }
 992                 if (umem_access_ok(msg->uaddr, msg->size, msg->perm)) {
 993                         ret = -EFAULT;
 994                         break;
 995                 }
 996                 vhost_vq_meta_reset(dev);
 997                 if (vhost_new_umem_range(dev->iotlb, msg->iova, msg->size,
 998                                          msg->iova + msg->size - 1,
 999                                          msg->uaddr, msg->perm)) {
1000                         ret = -ENOMEM;
1001                         break;
1002                 }
1003                 vhost_iotlb_notify_vq(dev, msg);
1004                 break;
1005         case VHOST_IOTLB_INVALIDATE:
1006                 if (!dev->iotlb) {
1007                         ret = -EFAULT;
1008                         break;
1009                 }
1010                 vhost_vq_meta_reset(dev);
1011                 vhost_del_umem_range(dev->iotlb, msg->iova,
1012                                      msg->iova + msg->size - 1);
1013                 break;
1014         default:
1015                 ret = -EINVAL;
1016                 break;
1017         }
1018
1019         vhost_dev_unlock_vqs(dev);
1020         return ret;
1021 }
1022 ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
1023                              struct iov_iter *from)
1024 {
1025         struct vhost_msg_node node;
1026         unsigned size = sizeof(struct vhost_msg);
1027         size_t ret;
1028         int err;
1029
1030         if (iov_iter_count(from) < size)
1031                 return 0;
1032         ret = copy_from_iter(&node.msg, size, from);
1033         if (ret != size)
1034                 goto done;
1035
1036         switch (node.msg.type) {
1037         case VHOST_IOTLB_MSG:
1038                 err = vhost_process_iotlb_msg(dev, &node.msg.iotlb);
1039                 if (err)
1040                         ret = err;
1041                 break;
1042         default:
1043                 ret = -EINVAL;
1044                 break;
1045         }
1046
1047 done:
1048         return ret;
1049 }
1050 EXPORT_SYMBOL(vhost_chr_write_iter);
1051
1052 __poll_t vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1053                             poll_table *wait)
1054 {
1055         __poll_t mask = 0;
1056
1057         poll_wait(file, &dev->wait, wait);
1058
1059         if (!list_empty(&dev->read_list))
1060                 mask |= EPOLLIN | EPOLLRDNORM;
1061
1062         return mask;
1063 }
1064 EXPORT_SYMBOL(vhost_chr_poll);
1065
1066 ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1067                             int noblock)
1068 {
1069         DEFINE_WAIT(wait);
1070         struct vhost_msg_node *node;
1071         ssize_t ret = 0;
1072         unsigned size = sizeof(struct vhost_msg);
1073
1074         if (iov_iter_count(to) < size)
1075                 return 0;
1076
1077         while (1) {
1078                 if (!noblock)
1079                         prepare_to_wait(&dev->wait, &wait,
1080                                         TASK_INTERRUPTIBLE);
1081
1082                 node = vhost_dequeue_msg(dev, &dev->read_list);
1083                 if (node)
1084                         break;
1085                 if (noblock) {
1086                         ret = -EAGAIN;
1087                         break;
1088                 }
1089                 if (signal_pending(current)) {
1090                         ret = -ERESTARTSYS;
1091                         break;
1092                 }
1093                 if (!dev->iotlb) {
1094                         ret = -EBADFD;
1095                         break;
1096                 }
1097
1098                 schedule();
1099         }
1100
1101         if (!noblock)
1102                 finish_wait(&dev->wait, &wait);
1103
1104         if (node) {
1105                 ret = copy_to_iter(&node->msg, size, to);
1106
1107                 if (ret != size || node->msg.type != VHOST_IOTLB_MISS) {
1108                         kfree(node);
1109                         return ret;
1110                 }
1111
1112                 vhost_enqueue_msg(dev, &dev->pending_list, node);
1113         }
1114
1115         return ret;
1116 }
1117 EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1118
1119 static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1120 {
1121         struct vhost_dev *dev = vq->dev;
1122         struct vhost_msg_node *node;
1123         struct vhost_iotlb_msg *msg;
1124
1125         node = vhost_new_msg(vq, VHOST_IOTLB_MISS);
1126         if (!node)
1127                 return -ENOMEM;
1128
1129         msg = &node->msg.iotlb;
1130         msg->type = VHOST_IOTLB_MISS;
1131         msg->iova = iova;
1132         msg->perm = access;
1133
1134         vhost_enqueue_msg(dev, &dev->read_list, node);
1135
1136         return 0;
1137 }
1138
1139 static int vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1140                         struct vring_desc __user *desc,
1141                         struct vring_avail __user *avail,
1142                         struct vring_used __user *used)
1143
1144 {
1145         size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1146
1147         return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
1148                access_ok(VERIFY_READ, avail,
1149                          sizeof *avail + num * sizeof *avail->ring + s) &&
1150                access_ok(VERIFY_WRITE, used,
1151                         sizeof *used + num * sizeof *used->ring + s);
1152 }
1153
1154 static void vhost_vq_meta_update(struct vhost_virtqueue *vq,
1155                                  const struct vhost_umem_node *node,
1156                                  int type)
1157 {
1158         int access = (type == VHOST_ADDR_USED) ?
1159                      VHOST_ACCESS_WO : VHOST_ACCESS_RO;
1160
1161         if (likely(node->perm & access))
1162                 vq->meta_iotlb[type] = node;
1163 }
1164
1165 static int iotlb_access_ok(struct vhost_virtqueue *vq,
1166                            int access, u64 addr, u64 len, int type)
1167 {
1168         const struct vhost_umem_node *node;
1169         struct vhost_umem *umem = vq->iotlb;
1170         u64 s = 0, size, orig_addr = addr, last = addr + len - 1;
1171
1172         if (vhost_vq_meta_fetch(vq, addr, len, type))
1173                 return true;
1174
1175         while (len > s) {
1176                 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1177                                                            addr,
1178                                                            last);
1179                 if (node == NULL || node->start > addr) {
1180                         vhost_iotlb_miss(vq, addr, access);
1181                         return false;
1182                 } else if (!(node->perm & access)) {
1183                         /* Report the possible access violation by
1184                          * request another translation from userspace.
1185                          */
1186                         return false;
1187                 }
1188
1189                 size = node->size - addr + node->start;
1190
1191                 if (orig_addr == addr && size >= len)
1192                         vhost_vq_meta_update(vq, node, type);
1193
1194                 s += size;
1195                 addr += size;
1196         }
1197
1198         return true;
1199 }
1200
1201 int vq_iotlb_prefetch(struct vhost_virtqueue *vq)
1202 {
1203         size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1204         unsigned int num = vq->num;
1205
1206         if (!vq->iotlb)
1207                 return 1;
1208
1209         return iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->desc,
1210                                num * sizeof(*vq->desc), VHOST_ADDR_DESC) &&
1211                iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->avail,
1212                                sizeof *vq->avail +
1213                                num * sizeof(*vq->avail->ring) + s,
1214                                VHOST_ADDR_AVAIL) &&
1215                iotlb_access_ok(vq, VHOST_ACCESS_WO, (u64)(uintptr_t)vq->used,
1216                                sizeof *vq->used +
1217                                num * sizeof(*vq->used->ring) + s,
1218                                VHOST_ADDR_USED);
1219 }
1220 EXPORT_SYMBOL_GPL(vq_iotlb_prefetch);
1221
1222 /* Can we log writes? */
1223 /* Caller should have device mutex but not vq mutex */
1224 int vhost_log_access_ok(struct vhost_dev *dev)
1225 {
1226         return memory_access_ok(dev, dev->umem, 1);
1227 }
1228 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1229
1230 /* Verify access for write logging. */
1231 /* Caller should have vq mutex and device mutex */
1232 static int vq_log_access_ok(struct vhost_virtqueue *vq,
1233                             void __user *log_base)
1234 {
1235         size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1236
1237         return vq_memory_access_ok(log_base, vq->umem,
1238                                    vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1239                 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
1240                                         sizeof *vq->used +
1241                                         vq->num * sizeof *vq->used->ring + s));
1242 }
1243
1244 /* Can we start vq? */
1245 /* Caller should have vq mutex and device mutex */
1246 int vhost_vq_access_ok(struct vhost_virtqueue *vq)
1247 {
1248         if (vq->iotlb) {
1249                 /* When device IOTLB was used, the access validation
1250                  * will be validated during prefetching.
1251                  */
1252                 return 1;
1253         }
1254         return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used) &&
1255                 vq_log_access_ok(vq, vq->log_base);
1256 }
1257 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1258
1259 static struct vhost_umem *vhost_umem_alloc(void)
1260 {
1261         struct vhost_umem *umem = kvzalloc(sizeof(*umem), GFP_KERNEL);
1262
1263         if (!umem)
1264                 return NULL;
1265
1266         umem->umem_tree = RB_ROOT_CACHED;
1267         umem->numem = 0;
1268         INIT_LIST_HEAD(&umem->umem_list);
1269
1270         return umem;
1271 }
1272
1273 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1274 {
1275         struct vhost_memory mem, *newmem;
1276         struct vhost_memory_region *region;
1277         struct vhost_umem *newumem, *oldumem;
1278         unsigned long size = offsetof(struct vhost_memory, regions);
1279         int i;
1280
1281         if (copy_from_user(&mem, m, size))
1282                 return -EFAULT;
1283         if (mem.padding)
1284                 return -EOPNOTSUPP;
1285         if (mem.nregions > max_mem_regions)
1286                 return -E2BIG;
1287         newmem = kvzalloc(size + mem.nregions * sizeof(*m->regions), GFP_KERNEL);
1288         if (!newmem)
1289                 return -ENOMEM;
1290
1291         memcpy(newmem, &mem, size);
1292         if (copy_from_user(newmem->regions, m->regions,
1293                            mem.nregions * sizeof *m->regions)) {
1294                 kvfree(newmem);
1295                 return -EFAULT;
1296         }
1297
1298         newumem = vhost_umem_alloc();
1299         if (!newumem) {
1300                 kvfree(newmem);
1301                 return -ENOMEM;
1302         }
1303
1304         for (region = newmem->regions;
1305              region < newmem->regions + mem.nregions;
1306              region++) {
1307                 if (vhost_new_umem_range(newumem,
1308                                          region->guest_phys_addr,
1309                                          region->memory_size,
1310                                          region->guest_phys_addr +
1311                                          region->memory_size - 1,
1312                                          region->userspace_addr,
1313                                          VHOST_ACCESS_RW))
1314                         goto err;
1315         }
1316
1317         if (!memory_access_ok(d, newumem, 0))
1318                 goto err;
1319
1320         oldumem = d->umem;
1321         d->umem = newumem;
1322
1323         /* All memory accesses are done under some VQ mutex. */
1324         for (i = 0; i < d->nvqs; ++i) {
1325                 mutex_lock(&d->vqs[i]->mutex);
1326                 d->vqs[i]->umem = newumem;
1327                 mutex_unlock(&d->vqs[i]->mutex);
1328         }
1329
1330         kvfree(newmem);
1331         vhost_umem_clean(oldumem);
1332         return 0;
1333
1334 err:
1335         vhost_umem_clean(newumem);
1336         kvfree(newmem);
1337         return -EFAULT;
1338 }
1339
1340 long vhost_vring_ioctl(struct vhost_dev *d, int ioctl, void __user *argp)
1341 {
1342         struct file *eventfp, *filep = NULL;
1343         bool pollstart = false, pollstop = false;
1344         struct eventfd_ctx *ctx = NULL;
1345         u32 __user *idxp = argp;
1346         struct vhost_virtqueue *vq;
1347         struct vhost_vring_state s;
1348         struct vhost_vring_file f;
1349         struct vhost_vring_addr a;
1350         u32 idx;
1351         long r;
1352
1353         r = get_user(idx, idxp);
1354         if (r < 0)
1355                 return r;
1356         if (idx >= d->nvqs)
1357                 return -ENOBUFS;
1358
1359         vq = d->vqs[idx];
1360
1361         mutex_lock(&vq->mutex);
1362
1363         switch (ioctl) {
1364         case VHOST_SET_VRING_NUM:
1365                 /* Resizing ring with an active backend?
1366                  * You don't want to do that. */
1367                 if (vq->private_data) {
1368                         r = -EBUSY;
1369                         break;
1370                 }
1371                 if (copy_from_user(&s, argp, sizeof s)) {
1372                         r = -EFAULT;
1373                         break;
1374                 }
1375                 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
1376                         r = -EINVAL;
1377                         break;
1378                 }
1379                 vq->num = s.num;
1380                 break;
1381         case VHOST_SET_VRING_BASE:
1382                 /* Moving base with an active backend?
1383                  * You don't want to do that. */
1384                 if (vq->private_data) {
1385                         r = -EBUSY;
1386                         break;
1387                 }
1388                 if (copy_from_user(&s, argp, sizeof s)) {
1389                         r = -EFAULT;
1390                         break;
1391                 }
1392                 if (s.num > 0xffff) {
1393                         r = -EINVAL;
1394                         break;
1395                 }
1396                 vq->last_avail_idx = s.num;
1397                 /* Forget the cached index value. */
1398                 vq->avail_idx = vq->last_avail_idx;
1399                 break;
1400         case VHOST_GET_VRING_BASE:
1401                 s.index = idx;
1402                 s.num = vq->last_avail_idx;
1403                 if (copy_to_user(argp, &s, sizeof s))
1404                         r = -EFAULT;
1405                 break;
1406         case VHOST_SET_VRING_ADDR:
1407                 if (copy_from_user(&a, argp, sizeof a)) {
1408                         r = -EFAULT;
1409                         break;
1410                 }
1411                 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
1412                         r = -EOPNOTSUPP;
1413                         break;
1414                 }
1415                 /* For 32bit, verify that the top 32bits of the user
1416                    data are set to zero. */
1417                 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1418                     (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1419                     (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
1420                         r = -EFAULT;
1421                         break;
1422                 }
1423
1424                 /* Make sure it's safe to cast pointers to vring types. */
1425                 BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1426                 BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1427                 if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1428                     (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1429                     (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1))) {
1430                         r = -EINVAL;
1431                         break;
1432                 }
1433
1434                 /* We only verify access here if backend is configured.
1435                  * If it is not, we don't as size might not have been setup.
1436                  * We will verify when backend is configured. */
1437                 if (vq->private_data) {
1438                         if (!vq_access_ok(vq, vq->num,
1439                                 (void __user *)(unsigned long)a.desc_user_addr,
1440                                 (void __user *)(unsigned long)a.avail_user_addr,
1441                                 (void __user *)(unsigned long)a.used_user_addr)) {
1442                                 r = -EINVAL;
1443                                 break;
1444                         }
1445
1446                         /* Also validate log access for used ring if enabled. */
1447                         if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
1448                             !log_access_ok(vq->log_base, a.log_guest_addr,
1449                                            sizeof *vq->used +
1450                                            vq->num * sizeof *vq->used->ring)) {
1451                                 r = -EINVAL;
1452                                 break;
1453                         }
1454                 }
1455
1456                 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1457                 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1458                 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1459                 vq->log_addr = a.log_guest_addr;
1460                 vq->used = (void __user *)(unsigned long)a.used_user_addr;
1461                 break;
1462         case VHOST_SET_VRING_KICK:
1463                 if (copy_from_user(&f, argp, sizeof f)) {
1464                         r = -EFAULT;
1465                         break;
1466                 }
1467                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1468                 if (IS_ERR(eventfp)) {
1469                         r = PTR_ERR(eventfp);
1470                         break;
1471                 }
1472                 if (eventfp != vq->kick) {
1473                         pollstop = (filep = vq->kick) != NULL;
1474                         pollstart = (vq->kick = eventfp) != NULL;
1475                 } else
1476                         filep = eventfp;
1477                 break;
1478         case VHOST_SET_VRING_CALL:
1479                 if (copy_from_user(&f, argp, sizeof f)) {
1480                         r = -EFAULT;
1481                         break;
1482                 }
1483                 ctx = f.fd == -1 ? NULL : eventfd_ctx_fdget(f.fd);
1484                 if (IS_ERR(ctx)) {
1485                         r = PTR_ERR(ctx);
1486                         break;
1487                 }
1488                 swap(ctx, vq->call_ctx);
1489                 break;
1490         case VHOST_SET_VRING_ERR:
1491                 if (copy_from_user(&f, argp, sizeof f)) {
1492                         r = -EFAULT;
1493                         break;
1494                 }
1495                 ctx = f.fd == -1 ? NULL : eventfd_ctx_fdget(f.fd);
1496                 if (IS_ERR(ctx)) {
1497                         r = PTR_ERR(ctx);
1498                         break;
1499                 }
1500                 swap(ctx, vq->error_ctx);
1501                 break;
1502         case VHOST_SET_VRING_ENDIAN:
1503                 r = vhost_set_vring_endian(vq, argp);
1504                 break;
1505         case VHOST_GET_VRING_ENDIAN:
1506                 r = vhost_get_vring_endian(vq, idx, argp);
1507                 break;
1508         case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
1509                 if (copy_from_user(&s, argp, sizeof(s))) {
1510                         r = -EFAULT;
1511                         break;
1512                 }
1513                 vq->busyloop_timeout = s.num;
1514                 break;
1515         case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
1516                 s.index = idx;
1517                 s.num = vq->busyloop_timeout;
1518                 if (copy_to_user(argp, &s, sizeof(s)))
1519                         r = -EFAULT;
1520                 break;
1521         default:
1522                 r = -ENOIOCTLCMD;
1523         }
1524
1525         if (pollstop && vq->handle_kick)
1526                 vhost_poll_stop(&vq->poll);
1527
1528         if (!IS_ERR_OR_NULL(ctx))
1529                 eventfd_ctx_put(ctx);
1530         if (filep)
1531                 fput(filep);
1532
1533         if (pollstart && vq->handle_kick)
1534                 r = vhost_poll_start(&vq->poll, vq->kick);
1535
1536         mutex_unlock(&vq->mutex);
1537
1538         if (pollstop && vq->handle_kick)
1539                 vhost_poll_flush(&vq->poll);
1540         return r;
1541 }
1542 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
1543
1544 int vhost_init_device_iotlb(struct vhost_dev *d, bool enabled)
1545 {
1546         struct vhost_umem *niotlb, *oiotlb;
1547         int i;
1548
1549         niotlb = vhost_umem_alloc();
1550         if (!niotlb)
1551                 return -ENOMEM;
1552
1553         oiotlb = d->iotlb;
1554         d->iotlb = niotlb;
1555
1556         for (i = 0; i < d->nvqs; ++i) {
1557                 mutex_lock(&d->vqs[i]->mutex);
1558                 d->vqs[i]->iotlb = niotlb;
1559                 mutex_unlock(&d->vqs[i]->mutex);
1560         }
1561
1562         vhost_umem_clean(oiotlb);
1563
1564         return 0;
1565 }
1566 EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
1567
1568 /* Caller must have device mutex */
1569 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1570 {
1571         struct eventfd_ctx *ctx;
1572         u64 p;
1573         long r;
1574         int i, fd;
1575
1576         /* If you are not the owner, you can become one */
1577         if (ioctl == VHOST_SET_OWNER) {
1578                 r = vhost_dev_set_owner(d);
1579                 goto done;
1580         }
1581
1582         /* You must be the owner to do anything else */
1583         r = vhost_dev_check_owner(d);
1584         if (r)
1585                 goto done;
1586
1587         switch (ioctl) {
1588         case VHOST_SET_MEM_TABLE:
1589                 r = vhost_set_memory(d, argp);
1590                 break;
1591         case VHOST_SET_LOG_BASE:
1592                 if (copy_from_user(&p, argp, sizeof p)) {
1593                         r = -EFAULT;
1594                         break;
1595                 }
1596                 if ((u64)(unsigned long)p != p) {
1597                         r = -EFAULT;
1598                         break;
1599                 }
1600                 for (i = 0; i < d->nvqs; ++i) {
1601                         struct vhost_virtqueue *vq;
1602                         void __user *base = (void __user *)(unsigned long)p;
1603                         vq = d->vqs[i];
1604                         mutex_lock(&vq->mutex);
1605                         /* If ring is inactive, will check when it's enabled. */
1606                         if (vq->private_data && !vq_log_access_ok(vq, base))
1607                                 r = -EFAULT;
1608                         else
1609                                 vq->log_base = base;
1610                         mutex_unlock(&vq->mutex);
1611                 }
1612                 break;
1613         case VHOST_SET_LOG_FD:
1614                 r = get_user(fd, (int __user *)argp);
1615                 if (r < 0)
1616                         break;
1617                 ctx = fd == -1 ? NULL : eventfd_ctx_fdget(fd);
1618                 if (IS_ERR(ctx)) {
1619                         r = PTR_ERR(ctx);
1620                         break;
1621                 }
1622                 swap(ctx, d->log_ctx);
1623                 for (i = 0; i < d->nvqs; ++i) {
1624                         mutex_lock(&d->vqs[i]->mutex);
1625                         d->vqs[i]->log_ctx = d->log_ctx;
1626                         mutex_unlock(&d->vqs[i]->mutex);
1627                 }
1628                 if (ctx)
1629                         eventfd_ctx_put(ctx);
1630                 break;
1631         default:
1632                 r = -ENOIOCTLCMD;
1633                 break;
1634         }
1635 done:
1636         return r;
1637 }
1638 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1639
1640 /* TODO: This is really inefficient.  We need something like get_user()
1641  * (instruction directly accesses the data, with an exception table entry
1642  * returning -EFAULT). See Documentation/x86/exception-tables.txt.
1643  */
1644 static int set_bit_to_user(int nr, void __user *addr)
1645 {
1646         unsigned long log = (unsigned long)addr;
1647         struct page *page;
1648         void *base;
1649         int bit = nr + (log % PAGE_SIZE) * 8;
1650         int r;
1651
1652         r = get_user_pages_fast(log, 1, 1, &page);
1653         if (r < 0)
1654                 return r;
1655         BUG_ON(r != 1);
1656         base = kmap_atomic(page);
1657         set_bit(bit, base);
1658         kunmap_atomic(base);
1659         set_page_dirty_lock(page);
1660         put_page(page);
1661         return 0;
1662 }
1663
1664 static int log_write(void __user *log_base,
1665                      u64 write_address, u64 write_length)
1666 {
1667         u64 write_page = write_address / VHOST_PAGE_SIZE;
1668         int r;
1669
1670         if (!write_length)
1671                 return 0;
1672         write_length += write_address % VHOST_PAGE_SIZE;
1673         for (;;) {
1674                 u64 base = (u64)(unsigned long)log_base;
1675                 u64 log = base + write_page / 8;
1676                 int bit = write_page % 8;
1677                 if ((u64)(unsigned long)log != log)
1678                         return -EFAULT;
1679                 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1680                 if (r < 0)
1681                         return r;
1682                 if (write_length <= VHOST_PAGE_SIZE)
1683                         break;
1684                 write_length -= VHOST_PAGE_SIZE;
1685                 write_page += 1;
1686         }
1687         return r;
1688 }
1689
1690 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1691                     unsigned int log_num, u64 len)
1692 {
1693         int i, r;
1694
1695         /* Make sure data written is seen before log. */
1696         smp_wmb();
1697         for (i = 0; i < log_num; ++i) {
1698                 u64 l = min(log[i].len, len);
1699                 r = log_write(vq->log_base, log[i].addr, l);
1700                 if (r < 0)
1701                         return r;
1702                 len -= l;
1703                 if (!len) {
1704                         if (vq->log_ctx)
1705                                 eventfd_signal(vq->log_ctx, 1);
1706                         return 0;
1707                 }
1708         }
1709         /* Length written exceeds what we have stored. This is a bug. */
1710         BUG();
1711         return 0;
1712 }
1713 EXPORT_SYMBOL_GPL(vhost_log_write);
1714
1715 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1716 {
1717         void __user *used;
1718         if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
1719                            &vq->used->flags) < 0)
1720                 return -EFAULT;
1721         if (unlikely(vq->log_used)) {
1722                 /* Make sure the flag is seen before log. */
1723                 smp_wmb();
1724                 /* Log used flag write. */
1725                 used = &vq->used->flags;
1726                 log_write(vq->log_base, vq->log_addr +
1727                           (used - (void __user *)vq->used),
1728                           sizeof vq->used->flags);
1729                 if (vq->log_ctx)
1730                         eventfd_signal(vq->log_ctx, 1);
1731         }
1732         return 0;
1733 }
1734
1735 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1736 {
1737         if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
1738                            vhost_avail_event(vq)))
1739                 return -EFAULT;
1740         if (unlikely(vq->log_used)) {
1741                 void __user *used;
1742                 /* Make sure the event is seen before log. */
1743                 smp_wmb();
1744                 /* Log avail event write */
1745                 used = vhost_avail_event(vq);
1746                 log_write(vq->log_base, vq->log_addr +
1747                           (used - (void __user *)vq->used),
1748                           sizeof *vhost_avail_event(vq));
1749                 if (vq->log_ctx)
1750                         eventfd_signal(vq->log_ctx, 1);
1751         }
1752         return 0;
1753 }
1754
1755 int vhost_vq_init_access(struct vhost_virtqueue *vq)
1756 {
1757         __virtio16 last_used_idx;
1758         int r;
1759         bool is_le = vq->is_le;
1760
1761         if (!vq->private_data)
1762                 return 0;
1763
1764         vhost_init_is_le(vq);
1765
1766         r = vhost_update_used_flags(vq);
1767         if (r)
1768                 goto err;
1769         vq->signalled_used_valid = false;
1770         if (!vq->iotlb &&
1771             !access_ok(VERIFY_READ, &vq->used->idx, sizeof vq->used->idx)) {
1772                 r = -EFAULT;
1773                 goto err;
1774         }
1775         r = vhost_get_used(vq, last_used_idx, &vq->used->idx);
1776         if (r) {
1777                 vq_err(vq, "Can't access used idx at %p\n",
1778                        &vq->used->idx);
1779                 goto err;
1780         }
1781         vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
1782         return 0;
1783
1784 err:
1785         vq->is_le = is_le;
1786         return r;
1787 }
1788 EXPORT_SYMBOL_GPL(vhost_vq_init_access);
1789
1790 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
1791                           struct iovec iov[], int iov_size, int access)
1792 {
1793         const struct vhost_umem_node *node;
1794         struct vhost_dev *dev = vq->dev;
1795         struct vhost_umem *umem = dev->iotlb ? dev->iotlb : dev->umem;
1796         struct iovec *_iov;
1797         u64 s = 0;
1798         int ret = 0;
1799
1800         while ((u64)len > s) {
1801                 u64 size;
1802                 if (unlikely(ret >= iov_size)) {
1803                         ret = -ENOBUFS;
1804                         break;
1805                 }
1806
1807                 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1808                                                         addr, addr + len - 1);
1809                 if (node == NULL || node->start > addr) {
1810                         if (umem != dev->iotlb) {
1811                                 ret = -EFAULT;
1812                                 break;
1813                         }
1814                         ret = -EAGAIN;
1815                         break;
1816                 } else if (!(node->perm & access)) {
1817                         ret = -EPERM;
1818                         break;
1819                 }
1820
1821                 _iov = iov + ret;
1822                 size = node->size - addr + node->start;
1823                 _iov->iov_len = min((u64)len - s, size);
1824                 _iov->iov_base = (void __user *)(unsigned long)
1825                         (node->userspace_addr + addr - node->start);
1826                 s += size;
1827                 addr += size;
1828                 ++ret;
1829         }
1830
1831         if (ret == -EAGAIN)
1832                 vhost_iotlb_miss(vq, addr, access);
1833         return ret;
1834 }
1835
1836 /* Each buffer in the virtqueues is actually a chain of descriptors.  This
1837  * function returns the next descriptor in the chain,
1838  * or -1U if we're at the end. */
1839 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
1840 {
1841         unsigned int next;
1842
1843         /* If this descriptor says it doesn't chain, we're done. */
1844         if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
1845                 return -1U;
1846
1847         /* Check they're not leading us off end of descriptors. */
1848         next = vhost16_to_cpu(vq, READ_ONCE(desc->next));
1849         return next;
1850 }
1851
1852 static int get_indirect(struct vhost_virtqueue *vq,
1853                         struct iovec iov[], unsigned int iov_size,
1854                         unsigned int *out_num, unsigned int *in_num,
1855                         struct vhost_log *log, unsigned int *log_num,
1856                         struct vring_desc *indirect)
1857 {
1858         struct vring_desc desc;
1859         unsigned int i = 0, count, found = 0;
1860         u32 len = vhost32_to_cpu(vq, indirect->len);
1861         struct iov_iter from;
1862         int ret, access;
1863
1864         /* Sanity check */
1865         if (unlikely(len % sizeof desc)) {
1866                 vq_err(vq, "Invalid length in indirect descriptor: "
1867                        "len 0x%llx not multiple of 0x%zx\n",
1868                        (unsigned long long)len,
1869                        sizeof desc);
1870                 return -EINVAL;
1871         }
1872
1873         ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
1874                              UIO_MAXIOV, VHOST_ACCESS_RO);
1875         if (unlikely(ret < 0)) {
1876                 if (ret != -EAGAIN)
1877                         vq_err(vq, "Translation failure %d in indirect.\n", ret);
1878                 return ret;
1879         }
1880         iov_iter_init(&from, READ, vq->indirect, ret, len);
1881
1882         /* We will use the result as an address to read from, so most
1883          * architectures only need a compiler barrier here. */
1884         read_barrier_depends();
1885
1886         count = len / sizeof desc;
1887         /* Buffers are chained via a 16 bit next field, so
1888          * we can have at most 2^16 of these. */
1889         if (unlikely(count > USHRT_MAX + 1)) {
1890                 vq_err(vq, "Indirect buffer length too big: %d\n",
1891                        indirect->len);
1892                 return -E2BIG;
1893         }
1894
1895         do {
1896                 unsigned iov_count = *in_num + *out_num;
1897                 if (unlikely(++found > count)) {
1898                         vq_err(vq, "Loop detected: last one at %u "
1899                                "indirect size %u\n",
1900                                i, count);
1901                         return -EINVAL;
1902                 }
1903                 if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
1904                         vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1905                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1906                         return -EINVAL;
1907                 }
1908                 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
1909                         vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1910                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1911                         return -EINVAL;
1912                 }
1913
1914                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
1915                         access = VHOST_ACCESS_WO;
1916                 else
1917                         access = VHOST_ACCESS_RO;
1918
1919                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
1920                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
1921                                      iov_size - iov_count, access);
1922                 if (unlikely(ret < 0)) {
1923                         if (ret != -EAGAIN)
1924                                 vq_err(vq, "Translation failure %d indirect idx %d\n",
1925                                         ret, i);
1926                         return ret;
1927                 }
1928                 /* If this is an input descriptor, increment that count. */
1929                 if (access == VHOST_ACCESS_WO) {
1930                         *in_num += ret;
1931                         if (unlikely(log)) {
1932                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
1933                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
1934                                 ++*log_num;
1935                         }
1936                 } else {
1937                         /* If it's an output descriptor, they're all supposed
1938                          * to come before any input descriptors. */
1939                         if (unlikely(*in_num)) {
1940                                 vq_err(vq, "Indirect descriptor "
1941                                        "has out after in: idx %d\n", i);
1942                                 return -EINVAL;
1943                         }
1944                         *out_num += ret;
1945                 }
1946         } while ((i = next_desc(vq, &desc)) != -1);
1947         return 0;
1948 }
1949
1950 /* This looks in the virtqueue and for the first available buffer, and converts
1951  * it to an iovec for convenient access.  Since descriptors consist of some
1952  * number of output then some number of input descriptors, it's actually two
1953  * iovecs, but we pack them into one and note how many of each there were.
1954  *
1955  * This function returns the descriptor number found, or vq->num (which is
1956  * never a valid descriptor number) if none was found.  A negative code is
1957  * returned on error. */
1958 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
1959                       struct iovec iov[], unsigned int iov_size,
1960                       unsigned int *out_num, unsigned int *in_num,
1961                       struct vhost_log *log, unsigned int *log_num)
1962 {
1963         struct vring_desc desc;
1964         unsigned int i, head, found = 0;
1965         u16 last_avail_idx;
1966         __virtio16 avail_idx;
1967         __virtio16 ring_head;
1968         int ret, access;
1969
1970         /* Check it isn't doing very strange things with descriptor numbers. */
1971         last_avail_idx = vq->last_avail_idx;
1972
1973         if (vq->avail_idx == vq->last_avail_idx) {
1974                 if (unlikely(vhost_get_avail(vq, avail_idx, &vq->avail->idx))) {
1975                         vq_err(vq, "Failed to access avail idx at %p\n",
1976                                 &vq->avail->idx);
1977                         return -EFAULT;
1978                 }
1979                 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
1980
1981                 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
1982                         vq_err(vq, "Guest moved used index from %u to %u",
1983                                 last_avail_idx, vq->avail_idx);
1984                         return -EFAULT;
1985                 }
1986
1987                 /* If there's nothing new since last we looked, return
1988                  * invalid.
1989                  */
1990                 if (vq->avail_idx == last_avail_idx)
1991                         return vq->num;
1992
1993                 /* Only get avail ring entries after they have been
1994                  * exposed by guest.
1995                  */
1996                 smp_rmb();
1997         }
1998
1999         /* Grab the next descriptor number they're advertising, and increment
2000          * the index we've seen. */
2001         if (unlikely(vhost_get_avail(vq, ring_head,
2002                      &vq->avail->ring[last_avail_idx & (vq->num - 1)]))) {
2003                 vq_err(vq, "Failed to read head: idx %d address %p\n",
2004                        last_avail_idx,
2005                        &vq->avail->ring[last_avail_idx % vq->num]);
2006                 return -EFAULT;
2007         }
2008
2009         head = vhost16_to_cpu(vq, ring_head);
2010
2011         /* If their number is silly, that's an error. */
2012         if (unlikely(head >= vq->num)) {
2013                 vq_err(vq, "Guest says index %u > %u is available",
2014                        head, vq->num);
2015                 return -EINVAL;
2016         }
2017
2018         /* When we start there are none of either input nor output. */
2019         *out_num = *in_num = 0;
2020         if (unlikely(log))
2021                 *log_num = 0;
2022
2023         i = head;
2024         do {
2025                 unsigned iov_count = *in_num + *out_num;
2026                 if (unlikely(i >= vq->num)) {
2027                         vq_err(vq, "Desc index is %u > %u, head = %u",
2028                                i, vq->num, head);
2029                         return -EINVAL;
2030                 }
2031                 if (unlikely(++found > vq->num)) {
2032                         vq_err(vq, "Loop detected: last one at %u "
2033                                "vq size %u head %u\n",
2034                                i, vq->num, head);
2035                         return -EINVAL;
2036                 }
2037                 ret = vhost_copy_from_user(vq, &desc, vq->desc + i,
2038                                            sizeof desc);
2039                 if (unlikely(ret)) {
2040                         vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
2041                                i, vq->desc + i);
2042                         return -EFAULT;
2043                 }
2044                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2045                         ret = get_indirect(vq, iov, iov_size,
2046                                            out_num, in_num,
2047                                            log, log_num, &desc);
2048                         if (unlikely(ret < 0)) {
2049                                 if (ret != -EAGAIN)
2050                                         vq_err(vq, "Failure detected "
2051                                                 "in indirect descriptor at idx %d\n", i);
2052                                 return ret;
2053                         }
2054                         continue;
2055                 }
2056
2057                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2058                         access = VHOST_ACCESS_WO;
2059                 else
2060                         access = VHOST_ACCESS_RO;
2061                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2062                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
2063                                      iov_size - iov_count, access);
2064                 if (unlikely(ret < 0)) {
2065                         if (ret != -EAGAIN)
2066                                 vq_err(vq, "Translation failure %d descriptor idx %d\n",
2067                                         ret, i);
2068                         return ret;
2069                 }
2070                 if (access == VHOST_ACCESS_WO) {
2071                         /* If this is an input descriptor,
2072                          * increment that count. */
2073                         *in_num += ret;
2074                         if (unlikely(log)) {
2075                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2076                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2077                                 ++*log_num;
2078                         }
2079                 } else {
2080                         /* If it's an output descriptor, they're all supposed
2081                          * to come before any input descriptors. */
2082                         if (unlikely(*in_num)) {
2083                                 vq_err(vq, "Descriptor has out after in: "
2084                                        "idx %d\n", i);
2085                                 return -EINVAL;
2086                         }
2087                         *out_num += ret;
2088                 }
2089         } while ((i = next_desc(vq, &desc)) != -1);
2090
2091         /* On success, increment avail index. */
2092         vq->last_avail_idx++;
2093
2094         /* Assume notifications from guest are disabled at this point,
2095          * if they aren't we would need to update avail_event index. */
2096         BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2097         return head;
2098 }
2099 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2100
2101 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2102 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2103 {
2104         vq->last_avail_idx -= n;
2105 }
2106 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2107
2108 /* After we've used one of their buffers, we tell them about it.  We'll then
2109  * want to notify the guest, using eventfd. */
2110 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2111 {
2112         struct vring_used_elem heads = {
2113                 cpu_to_vhost32(vq, head),
2114                 cpu_to_vhost32(vq, len)
2115         };
2116
2117         return vhost_add_used_n(vq, &heads, 1);
2118 }
2119 EXPORT_SYMBOL_GPL(vhost_add_used);
2120
2121 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2122                             struct vring_used_elem *heads,
2123                             unsigned count)
2124 {
2125         struct vring_used_elem __user *used;
2126         u16 old, new;
2127         int start;
2128
2129         start = vq->last_used_idx & (vq->num - 1);
2130         used = vq->used->ring + start;
2131         if (count == 1) {
2132                 if (vhost_put_user(vq, heads[0].id, &used->id)) {
2133                         vq_err(vq, "Failed to write used id");
2134                         return -EFAULT;
2135                 }
2136                 if (vhost_put_user(vq, heads[0].len, &used->len)) {
2137                         vq_err(vq, "Failed to write used len");
2138                         return -EFAULT;
2139                 }
2140         } else if (vhost_copy_to_user(vq, used, heads, count * sizeof *used)) {
2141                 vq_err(vq, "Failed to write used");
2142                 return -EFAULT;
2143         }
2144         if (unlikely(vq->log_used)) {
2145                 /* Make sure data is seen before log. */
2146                 smp_wmb();
2147                 /* Log used ring entry write. */
2148                 log_write(vq->log_base,
2149                           vq->log_addr +
2150                            ((void __user *)used - (void __user *)vq->used),
2151                           count * sizeof *used);
2152         }
2153         old = vq->last_used_idx;
2154         new = (vq->last_used_idx += count);
2155         /* If the driver never bothers to signal in a very long while,
2156          * used index might wrap around. If that happens, invalidate
2157          * signalled_used index we stored. TODO: make sure driver
2158          * signals at least once in 2^16 and remove this. */
2159         if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2160                 vq->signalled_used_valid = false;
2161         return 0;
2162 }
2163
2164 /* After we've used one of their buffers, we tell them about it.  We'll then
2165  * want to notify the guest, using eventfd. */
2166 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2167                      unsigned count)
2168 {
2169         int start, n, r;
2170
2171         start = vq->last_used_idx & (vq->num - 1);
2172         n = vq->num - start;
2173         if (n < count) {
2174                 r = __vhost_add_used_n(vq, heads, n);
2175                 if (r < 0)
2176                         return r;
2177                 heads += n;
2178                 count -= n;
2179         }
2180         r = __vhost_add_used_n(vq, heads, count);
2181
2182         /* Make sure buffer is written before we update index. */
2183         smp_wmb();
2184         if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
2185                            &vq->used->idx)) {
2186                 vq_err(vq, "Failed to increment used idx");
2187                 return -EFAULT;
2188         }
2189         if (unlikely(vq->log_used)) {
2190                 /* Log used index update. */
2191                 log_write(vq->log_base,
2192                           vq->log_addr + offsetof(struct vring_used, idx),
2193                           sizeof vq->used->idx);
2194                 if (vq->log_ctx)
2195                         eventfd_signal(vq->log_ctx, 1);
2196         }
2197         return r;
2198 }
2199 EXPORT_SYMBOL_GPL(vhost_add_used_n);
2200
2201 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2202 {
2203         __u16 old, new;
2204         __virtio16 event;
2205         bool v;
2206         /* Flush out used index updates. This is paired
2207          * with the barrier that the Guest executes when enabling
2208          * interrupts. */
2209         smp_mb();
2210
2211         if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2212             unlikely(vq->avail_idx == vq->last_avail_idx))
2213                 return true;
2214
2215         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2216                 __virtio16 flags;
2217                 if (vhost_get_avail(vq, flags, &vq->avail->flags)) {
2218                         vq_err(vq, "Failed to get flags");
2219                         return true;
2220                 }
2221                 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2222         }
2223         old = vq->signalled_used;
2224         v = vq->signalled_used_valid;
2225         new = vq->signalled_used = vq->last_used_idx;
2226         vq->signalled_used_valid = true;
2227
2228         if (unlikely(!v))
2229                 return true;
2230
2231         if (vhost_get_avail(vq, event, vhost_used_event(vq))) {
2232                 vq_err(vq, "Failed to get used event idx");
2233                 return true;
2234         }
2235         return vring_need_event(vhost16_to_cpu(vq, event), new, old);
2236 }
2237
2238 /* This actually signals the guest, using eventfd. */
2239 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2240 {
2241         /* Signal the Guest tell them we used something up. */
2242         if (vq->call_ctx && vhost_notify(dev, vq))
2243                 eventfd_signal(vq->call_ctx, 1);
2244 }
2245 EXPORT_SYMBOL_GPL(vhost_signal);
2246
2247 /* And here's the combo meal deal.  Supersize me! */
2248 void vhost_add_used_and_signal(struct vhost_dev *dev,
2249                                struct vhost_virtqueue *vq,
2250                                unsigned int head, int len)
2251 {
2252         vhost_add_used(vq, head, len);
2253         vhost_signal(dev, vq);
2254 }
2255 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2256
2257 /* multi-buffer version of vhost_add_used_and_signal */
2258 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2259                                  struct vhost_virtqueue *vq,
2260                                  struct vring_used_elem *heads, unsigned count)
2261 {
2262         vhost_add_used_n(vq, heads, count);
2263         vhost_signal(dev, vq);
2264 }
2265 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2266
2267 /* return true if we're sure that avaiable ring is empty */
2268 bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2269 {
2270         __virtio16 avail_idx;
2271         int r;
2272
2273         if (vq->avail_idx != vq->last_avail_idx)
2274                 return false;
2275
2276         r = vhost_get_avail(vq, avail_idx, &vq->avail->idx);
2277         if (unlikely(r))
2278                 return false;
2279         vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2280
2281         return vq->avail_idx == vq->last_avail_idx;
2282 }
2283 EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2284
2285 /* OK, now we need to know about added descriptors. */
2286 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2287 {
2288         __virtio16 avail_idx;
2289         int r;
2290
2291         if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2292                 return false;
2293         vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2294         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2295                 r = vhost_update_used_flags(vq);
2296                 if (r) {
2297                         vq_err(vq, "Failed to enable notification at %p: %d\n",
2298                                &vq->used->flags, r);
2299                         return false;
2300                 }
2301         } else {
2302                 r = vhost_update_avail_event(vq, vq->avail_idx);
2303                 if (r) {
2304                         vq_err(vq, "Failed to update avail event index at %p: %d\n",
2305                                vhost_avail_event(vq), r);
2306                         return false;
2307                 }
2308         }
2309         /* They could have slipped one in as we were doing that: make
2310          * sure it's written, then check again. */
2311         smp_mb();
2312         r = vhost_get_avail(vq, avail_idx, &vq->avail->idx);
2313         if (r) {
2314                 vq_err(vq, "Failed to check avail idx at %p: %d\n",
2315                        &vq->avail->idx, r);
2316                 return false;
2317         }
2318
2319         return vhost16_to_cpu(vq, avail_idx) != vq->avail_idx;
2320 }
2321 EXPORT_SYMBOL_GPL(vhost_enable_notify);
2322
2323 /* We don't need to be notified again. */
2324 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2325 {
2326         int r;
2327
2328         if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2329                 return;
2330         vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2331         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2332                 r = vhost_update_used_flags(vq);
2333                 if (r)
2334                         vq_err(vq, "Failed to enable notification at %p: %d\n",
2335                                &vq->used->flags, r);
2336         }
2337 }
2338 EXPORT_SYMBOL_GPL(vhost_disable_notify);
2339
2340 /* Create a new message. */
2341 struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2342 {
2343         struct vhost_msg_node *node = kmalloc(sizeof *node, GFP_KERNEL);
2344         if (!node)
2345                 return NULL;
2346         node->vq = vq;
2347         node->msg.type = type;
2348         return node;
2349 }
2350 EXPORT_SYMBOL_GPL(vhost_new_msg);
2351
2352 void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2353                        struct vhost_msg_node *node)
2354 {
2355         spin_lock(&dev->iotlb_lock);
2356         list_add_tail(&node->node, head);
2357         spin_unlock(&dev->iotlb_lock);
2358
2359         wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
2360 }
2361 EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2362
2363 struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2364                                          struct list_head *head)
2365 {
2366         struct vhost_msg_node *node = NULL;
2367
2368         spin_lock(&dev->iotlb_lock);
2369         if (!list_empty(head)) {
2370                 node = list_first_entry(head, struct vhost_msg_node,
2371                                         node);
2372                 list_del(&node->node);
2373         }
2374         spin_unlock(&dev->iotlb_lock);
2375
2376         return node;
2377 }
2378 EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2379
2380
2381 static int __init vhost_init(void)
2382 {
2383         return 0;
2384 }
2385
2386 static void __exit vhost_exit(void)
2387 {
2388 }
2389
2390 module_init(vhost_init);
2391 module_exit(vhost_exit);
2392
2393 MODULE_VERSION("0.0.1");
2394 MODULE_LICENSE("GPL v2");
2395 MODULE_AUTHOR("Michael S. Tsirkin");
2396 MODULE_DESCRIPTION("Host kernel accelerator for virtio");