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