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/interval_tree_generic.h>
  31
  32 #include "vhost.h"
  33
  34 static ushort max_mem_regions = 64;
  35 module_param(max_mem_regions, ushort, 0444);
  36 MODULE_PARM_DESC(max_mem_regions,
  37         "Maximum number of memory regions in memory map. (default: 64)");
  38 static int max_iotlb_entries = 2048;
  39 module_param(max_iotlb_entries, int, 0444);
  40 MODULE_PARM_DESC(max_iotlb_entries,
  41         "Maximum number of iotlb entries. (default: 2048)");
  42
  43 enum {
  44         VHOST_MEMORY_F_LOG = 0x1,
  45 };
  46
  47 #define vhost_used_event(vq) ((__virtio16 __user *)&vq->avail->ring[vq->num])
  48 #define vhost_avail_event(vq) ((__virtio16 __user *)&vq->used->ring[vq->num])
  49
  50 INTERVAL_TREE_DEFINE(struct vhost_umem_node,
  51                      rb, __u64, __subtree_last,
  52                      START, LAST, static inline, vhost_umem_interval_tree);
  53
  54 #ifdef CONFIG_VHOST_CROSS_ENDIAN_LEGACY
  55 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
  56 {
  57         vq->user_be = !virtio_legacy_is_little_endian();
  58 }
  59
  60 static void vhost_enable_cross_endian_big(struct vhost_virtqueue *vq)
  61 {
  62         vq->user_be = true;
  63 }
  64
  65 static void vhost_enable_cross_endian_little(struct vhost_virtqueue *vq)
  66 {
  67         vq->user_be = false;
  68 }
  69
  70 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
  71 {
  72         struct vhost_vring_state s;
  73
  74         if (vq->private_data)
  75                 return -EBUSY;
  76
  77         if (copy_from_user(&s, argp, sizeof(s)))
  78                 return -EFAULT;
  79
  80         if (s.num != VHOST_VRING_LITTLE_ENDIAN &&
  81             s.num != VHOST_VRING_BIG_ENDIAN)
  82                 return -EINVAL;
  83
  84         if (s.num == VHOST_VRING_BIG_ENDIAN)
  85                 vhost_enable_cross_endian_big(vq);
  86         else
  87                 vhost_enable_cross_endian_little(vq);
  88
  89         return 0;
  90 }
  91
  92 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
  93                                    int __user *argp)
  94 {
  95         struct vhost_vring_state s = {
  96                 .index = idx,
  97                 .num = vq->user_be
  98         };
  99
 100         if (copy_to_user(argp, &s, sizeof(s)))
 101                 return -EFAULT;
 102
 103         return 0;
 104 }
 105
 106 static void vhost_init_is_le(struct vhost_virtqueue *vq)
 107 {
 108         /* Note for legacy virtio: user_be is initialized at reset time
 109          * according to the host endianness. If userspace does not set an
 110          * explicit endianness, the default behavior is native endian, as
 111          * expected by legacy virtio.
 112          */
 113         vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1) || !vq->user_be;
 114 }
 115 #else
 116 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
 117 {
 118 }
 119
 120 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
 121 {
 122         return -ENOIOCTLCMD;
 123 }
 124
 125 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
 126                                    int __user *argp)
 127 {
 128         return -ENOIOCTLCMD;
 129 }
 130
 131 static void vhost_init_is_le(struct vhost_virtqueue *vq)
 132 {
 133         vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1)
 134                 || virtio_legacy_is_little_endian();
 135 }
 136 #endif /* CONFIG_VHOST_CROSS_ENDIAN_LEGACY */
 137
 138 static void vhost_reset_is_le(struct vhost_virtqueue *vq)
 139 {
 140         vhost_init_is_le(vq);
 141 }
 142
 143 struct vhost_flush_struct {
 144         struct vhost_work work;
 145         struct completion wait_event;
 146 };
 147
 148 static void vhost_flush_work(struct vhost_work *work)
 149 {
 150         struct vhost_flush_struct *s;
 151
 152         s = container_of(work, struct vhost_flush_struct, work);
 153         complete(&s->wait_event);
 154 }
 155
 156 static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
 157                             poll_table *pt)
 158 {
 159         struct vhost_poll *poll;
 160
 161         poll = container_of(pt, struct vhost_poll, table);
 162         poll->wqh = wqh;
 163         add_wait_queue(wqh, &poll->wait);
 164 }
 165
 166 static int vhost_poll_wakeup(wait_queue_t *wait, unsigned mode, int sync,
 167                              void *key)
 168 {
 169         struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
 170
 171         if (!((unsigned long)key & poll->mask))
 172                 return 0;
 173
 174         vhost_poll_queue(poll);
 175         return 0;
 176 }
 177
 178 void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
 179 {
 180         clear_bit(VHOST_WORK_QUEUED, &work->flags);
 181         work->fn = fn;
 182         init_waitqueue_head(&work->done);
 183 }
 184 EXPORT_SYMBOL_GPL(vhost_work_init);
 185
 186 /* Init poll structure */
 187 void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
 188                      unsigned long mask, struct vhost_dev *dev)
 189 {
 190         init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
 191         init_poll_funcptr(&poll->table, vhost_poll_func);
 192         poll->mask = mask;
 193         poll->dev = dev;
 194         poll->wqh = NULL;
 195
 196         vhost_work_init(&poll->work, fn);
 197 }
 198 EXPORT_SYMBOL_GPL(vhost_poll_init);
 199
 200 /* Start polling a file. We add ourselves to file's wait queue. The caller must
 201  * keep a reference to a file until after vhost_poll_stop is called. */
 202 int vhost_poll_start(struct vhost_poll *poll, struct file *file)
 203 {
 204         unsigned long mask;
 205         int ret = 0;
 206
 207         if (poll->wqh)
 208                 return 0;
 209
 210         mask = file->f_op->poll(file, &poll->table);
 211         if (mask)
 212                 vhost_poll_wakeup(&poll->wait, 0, 0, (void *)mask);
 213         if (mask & POLLERR) {
 214                 if (poll->wqh)
 215                         remove_wait_queue(poll->wqh, &poll->wait);
 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                 __vhost_vq_meta_reset(d->vqs[i]);
 299 }
 300
 301 static void vhost_vq_reset(struct vhost_dev *dev,
 302                            struct vhost_virtqueue *vq)
 303 {
 304         vq->num = 1;
 305         vq->desc = NULL;
 306         vq->avail = NULL;
 307         vq->used = NULL;
 308         vq->last_avail_idx = 0;
 309         vq->last_used_event = 0;
 310         vq->avail_idx = 0;
 311         vq->last_used_idx = 0;
 312         vq->signalled_used = 0;
 313         vq->signalled_used_valid = false;
 314         vq->used_flags = 0;
 315         vq->log_used = false;
 316         vq->log_addr = -1ull;
 317         vq->private_data = NULL;
 318         vq->acked_features = 0;
 319         vq->log_base = NULL;
 320         vq->error_ctx = NULL;
 321         vq->error = NULL;
 322         vq->kick = NULL;
 323         vq->call_ctx = NULL;
 324         vq->call = NULL;
 325         vq->log_ctx = NULL;
 326         vhost_reset_is_le(vq);
 327         vhost_disable_cross_endian(vq);
 328         vq->busyloop_timeout = 0;
 329         vq->umem = NULL;
 330         vq->iotlb = NULL;
 331         __vhost_vq_meta_reset(vq);
 332 }
 333
 334 static int vhost_worker(void *data)
 335 {
 336         struct vhost_dev *dev = data;
 337         struct vhost_work *work, *work_next;
 338         struct llist_node *node;
 339         mm_segment_t oldfs = get_fs();
 340
 341         set_fs(USER_DS);
 342         use_mm(dev->mm);
 343
 344         for (;;) {
 345                 /* mb paired w/ kthread_stop */
 346                 set_current_state(TASK_INTERRUPTIBLE);
 347
 348                 if (kthread_should_stop()) {
 349                         __set_current_state(TASK_RUNNING);
 350                         break;
 351                 }
 352
 353                 node = llist_del_all(&dev->work_list);
 354                 if (!node)
 355                         schedule();
 356
 357                 node = llist_reverse_order(node);
 358                 /* make sure flag is seen after deletion */
 359                 smp_wmb();
 360                 llist_for_each_entry_safe(work, work_next, node, node) {
 361                         clear_bit(VHOST_WORK_QUEUED, &work->flags);
 362                         __set_current_state(TASK_RUNNING);
 363                         work->fn(work);
 364                         if (need_resched())
 365                                 schedule();
 366                 }
 367         }
 368         unuse_mm(dev->mm);
 369         set_fs(oldfs);
 370         return 0;
 371 }
 372
 373 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
 374 {
 375         kfree(vq->indirect);
 376         vq->indirect = NULL;
 377         kfree(vq->log);
 378         vq->log = NULL;
 379         kfree(vq->heads);
 380         vq->heads = NULL;
 381 }
 382
 383 /* Helper to allocate iovec buffers for all vqs. */
 384 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
 385 {
 386         struct vhost_virtqueue *vq;
 387         int i;
 388
 389         for (i = 0; i < dev->nvqs; ++i) {
 390                 vq = dev->vqs[i];
 391                 vq->indirect = kmalloc(sizeof *vq->indirect * UIO_MAXIOV,
 392                                        GFP_KERNEL);
 393                 vq->log = kmalloc(sizeof *vq->log * UIO_MAXIOV, GFP_KERNEL);
 394                 vq->heads = kmalloc(sizeof *vq->heads * UIO_MAXIOV, GFP_KERNEL);
 395                 if (!vq->indirect || !vq->log || !vq->heads)
 396                         goto err_nomem;
 397         }
 398         return 0;
 399
 400 err_nomem:
 401         for (; i >= 0; --i)
 402                 vhost_vq_free_iovecs(dev->vqs[i]);
 403         return -ENOMEM;
 404 }
 405
 406 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
 407 {
 408         int i;
 409
 410         for (i = 0; i < dev->nvqs; ++i)
 411                 vhost_vq_free_iovecs(dev->vqs[i]);
 412 }
 413
 414 void vhost_dev_init(struct vhost_dev *dev,
 415                     struct vhost_virtqueue **vqs, int nvqs)
 416 {
 417         struct vhost_virtqueue *vq;
 418         int i;
 419
 420         dev->vqs = vqs;
 421         dev->nvqs = nvqs;
 422         mutex_init(&dev->mutex);
 423         dev->log_ctx = NULL;
 424         dev->log_file = NULL;
 425         dev->umem = NULL;
 426         dev->iotlb = NULL;
 427         dev->mm = NULL;
 428         dev->worker = NULL;
 429         init_llist_head(&dev->work_list);
 430         init_waitqueue_head(&dev->wait);
 431         INIT_LIST_HEAD(&dev->read_list);
 432         INIT_LIST_HEAD(&dev->pending_list);
 433         spin_lock_init(&dev->iotlb_lock);
 434
 435
 436         for (i = 0; i < dev->nvqs; ++i) {
 437                 vq = dev->vqs[i];
 438                 vq->log = NULL;
 439                 vq->indirect = NULL;
 440                 vq->heads = NULL;
 441                 vq->dev = dev;
 442                 mutex_init(&vq->mutex);
 443                 vhost_vq_reset(dev, vq);
 444                 if (vq->handle_kick)
 445                         vhost_poll_init(&vq->poll, vq->handle_kick,
 446                                         POLLIN, dev);
 447         }
 448 }
 449 EXPORT_SYMBOL_GPL(vhost_dev_init);
 450
 451 /* Caller should have device mutex */
 452 long vhost_dev_check_owner(struct vhost_dev *dev)
 453 {
 454         /* Are you the owner? If not, I don't think you mean to do that */
 455         return dev->mm == current->mm ? 0 : -EPERM;
 456 }
 457 EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
 458
 459 struct vhost_attach_cgroups_struct {
 460         struct vhost_work work;
 461         struct task_struct *owner;
 462         int ret;
 463 };
 464
 465 static void vhost_attach_cgroups_work(struct vhost_work *work)
 466 {
 467         struct vhost_attach_cgroups_struct *s;
 468
 469         s = container_of(work, struct vhost_attach_cgroups_struct, work);
 470         s->ret = cgroup_attach_task_all(s->owner, current);
 471 }
 472
 473 static int vhost_attach_cgroups(struct vhost_dev *dev)
 474 {
 475         struct vhost_attach_cgroups_struct attach;
 476
 477         attach.owner = current;
 478         vhost_work_init(&attach.work, vhost_attach_cgroups_work);
 479         vhost_work_queue(dev, &attach.work);
 480         vhost_work_flush(dev, &attach.work);
 481         return attach.ret;
 482 }
 483
 484 /* Caller should have device mutex */
 485 bool vhost_dev_has_owner(struct vhost_dev *dev)
 486 {
 487         return dev->mm;
 488 }
 489 EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
 490
 491 /* Caller should have device mutex */
 492 long vhost_dev_set_owner(struct vhost_dev *dev)
 493 {
 494         struct task_struct *worker;
 495         int err;
 496
 497         /* Is there an owner already? */
 498         if (vhost_dev_has_owner(dev)) {
 499                 err = -EBUSY;
 500                 goto err_mm;
 501         }
 502
 503         /* No owner, become one */
 504         dev->mm = get_task_mm(current);
 505         worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
 506         if (IS_ERR(worker)) {
 507                 err = PTR_ERR(worker);
 508                 goto err_worker;
 509         }
 510
 511         dev->worker = worker;
 512         wake_up_process(worker);        /* avoid contributing to loadavg */
 513
 514         err = vhost_attach_cgroups(dev);
 515         if (err)
 516                 goto err_cgroup;
 517
 518         err = vhost_dev_alloc_iovecs(dev);
 519         if (err)
 520                 goto err_cgroup;
 521
 522         return 0;
 523 err_cgroup:
 524         kthread_stop(worker);
 525         dev->worker = NULL;
 526 err_worker:
 527         if (dev->mm)
 528                 mmput(dev->mm);
 529         dev->mm = NULL;
 530 err_mm:
 531         return err;
 532 }
 533 EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
 534
 535 static void *vhost_kvzalloc(unsigned long size)
 536 {
 537         void *n = kzalloc(size, GFP_KERNEL | __GFP_NOWARN | __GFP_REPEAT);
 538
 539         if (!n)
 540                 n = vzalloc(size);
 541         return n;
 542 }
 543
 544 struct vhost_umem *vhost_dev_reset_owner_prepare(void)
 545 {
 546         return vhost_kvzalloc(sizeof(struct vhost_umem));
 547 }
 548 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
 549
 550 /* Caller should have device mutex */
 551 void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_umem *umem)
 552 {
 553         int i;
 554
 555         vhost_dev_cleanup(dev, true);
 556
 557         /* Restore memory to default empty mapping. */
 558         INIT_LIST_HEAD(&umem->umem_list);
 559         dev->umem = umem;
 560         /* We don't need VQ locks below since vhost_dev_cleanup makes sure
 561          * VQs aren't running.
 562          */
 563         for (i = 0; i < dev->nvqs; ++i)
 564                 dev->vqs[i]->umem = umem;
 565 }
 566 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
 567
 568 void vhost_dev_stop(struct vhost_dev *dev)
 569 {
 570         int i;
 571
 572         for (i = 0; i < dev->nvqs; ++i) {
 573                 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick) {
 574                         vhost_poll_stop(&dev->vqs[i]->poll);
 575                         vhost_poll_flush(&dev->vqs[i]->poll);
 576                 }
 577         }
 578 }
 579 EXPORT_SYMBOL_GPL(vhost_dev_stop);
 580
 581 static void vhost_umem_free(struct vhost_umem *umem,
 582                             struct vhost_umem_node *node)
 583 {
 584         vhost_umem_interval_tree_remove(node, &umem->umem_tree);
 585         list_del(&node->link);
 586         kfree(node);
 587         umem->numem--;
 588 }
 589
 590 static void vhost_umem_clean(struct vhost_umem *umem)
 591 {
 592         struct vhost_umem_node *node, *tmp;
 593
 594         if (!umem)
 595                 return;
 596
 597         list_for_each_entry_safe(node, tmp, &umem->umem_list, link)
 598                 vhost_umem_free(umem, node);
 599
 600         kvfree(umem);
 601 }
 602
 603 static void vhost_clear_msg(struct vhost_dev *dev)
 604 {
 605         struct vhost_msg_node *node, *n;
 606
 607         spin_lock(&dev->iotlb_lock);
 608
 609         list_for_each_entry_safe(node, n, &dev->read_list, node) {
 610                 list_del(&node->node);
 611                 kfree(node);
 612         }
 613
 614         list_for_each_entry_safe(node, n, &dev->pending_list, node) {
 615                 list_del(&node->node);
 616                 kfree(node);
 617         }
 618
 619         spin_unlock(&dev->iotlb_lock);
 620 }
 621
 622 /* Caller should have device mutex if and only if locked is set */
 623 void vhost_dev_cleanup(struct vhost_dev *dev, bool locked)
 624 {
 625         int i;
 626
 627         for (i = 0; i < dev->nvqs; ++i) {
 628                 if (dev->vqs[i]->error_ctx)
 629                         eventfd_ctx_put(dev->vqs[i]->error_ctx);
 630                 if (dev->vqs[i]->error)
 631                         fput(dev->vqs[i]->error);
 632                 if (dev->vqs[i]->kick)
 633                         fput(dev->vqs[i]->kick);
 634                 if (dev->vqs[i]->call_ctx)
 635                         eventfd_ctx_put(dev->vqs[i]->call_ctx);
 636                 if (dev->vqs[i]->call)
 637                         fput(dev->vqs[i]->call);
 638                 vhost_vq_reset(dev, dev->vqs[i]);
 639         }
 640         vhost_dev_free_iovecs(dev);
 641         if (dev->log_ctx)
 642                 eventfd_ctx_put(dev->log_ctx);
 643         dev->log_ctx = NULL;
 644         if (dev->log_file)
 645                 fput(dev->log_file);
 646         dev->log_file = NULL;
 647         /* No one will access memory at this point */
 648         vhost_umem_clean(dev->umem);
 649         dev->umem = NULL;
 650         vhost_umem_clean(dev->iotlb);
 651         dev->iotlb = NULL;
 652         vhost_clear_msg(dev);
 653         wake_up_interruptible_poll(&dev->wait, POLLIN | POLLRDNORM);
 654         WARN_ON(!llist_empty(&dev->work_list));
 655         if (dev->worker) {
 656                 kthread_stop(dev->worker);
 657                 dev->worker = NULL;
 658         }
 659         if (dev->mm)
 660                 mmput(dev->mm);
 661         dev->mm = NULL;
 662 }
 663 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
 664
 665 static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
 666 {
 667         u64 a = addr / VHOST_PAGE_SIZE / 8;
 668
 669         /* Make sure 64 bit math will not overflow. */
 670         if (a > ULONG_MAX - (unsigned long)log_base ||
 671             a + (unsigned long)log_base > ULONG_MAX)
 672                 return 0;
 673
 674         return access_ok(VERIFY_WRITE, log_base + a,
 675                          (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
 676 }
 677
 678 static bool vhost_overflow(u64 uaddr, u64 size)
 679 {
 680         /* Make sure 64 bit math will not overflow. */
 681         return uaddr > ULONG_MAX || size > ULONG_MAX || uaddr > ULONG_MAX - size;
 682 }
 683
 684 /* Caller should have vq mutex and device mutex. */
 685 static int vq_memory_access_ok(void __user *log_base, struct vhost_umem *umem,
 686                                int log_all)
 687 {
 688         struct vhost_umem_node *node;
 689
 690         if (!umem)
 691                 return 0;
 692
 693         list_for_each_entry(node, &umem->umem_list, link) {
 694                 unsigned long a = node->userspace_addr;
 695
 696                 if (vhost_overflow(node->userspace_addr, node->size))
 697                         return 0;
 698
 699
 700                 if (!access_ok(VERIFY_WRITE, (void __user *)a,
 701                                     node->size))
 702                         return 0;
 703                 else if (log_all && !log_access_ok(log_base,
 704                                                    node->start,
 705                                                    node->size))
 706                         return 0;
 707         }
 708         return 1;
 709 }
 710
 711 static inline void __user *vhost_vq_meta_fetch(struct vhost_virtqueue *vq,
 712                                                u64 addr, unsigned int size,
 713                                                int type)
 714 {
 715         const struct vhost_umem_node *node = vq->meta_iotlb[type];
 716
 717         if (!node)
 718                 return NULL;
 719
 720         return (void *)(uintptr_t)(node->userspace_addr + addr - node->start);
 721 }
 722
 723 /* Can we switch to this memory table? */
 724 /* Caller should have device mutex but not vq mutex */
 725 static int memory_access_ok(struct vhost_dev *d, struct vhost_umem *umem,
 726                             int log_all)
 727 {
 728         int i;
 729
 730         for (i = 0; i < d->nvqs; ++i) {
 731                 int ok;
 732                 bool log;
 733
 734                 mutex_lock(&d->vqs[i]->mutex);
 735                 log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
 736                 /* If ring is inactive, will check when it's enabled. */
 737                 if (d->vqs[i]->private_data)
 738                         ok = vq_memory_access_ok(d->vqs[i]->log_base,
 739                                                  umem, log);
 740                 else
 741                         ok = 1;
 742                 mutex_unlock(&d->vqs[i]->mutex);
 743                 if (!ok)
 744                         return 0;
 745         }
 746         return 1;
 747 }
 748
 749 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
 750                           struct iovec iov[], int iov_size, int access);
 751
 752 static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
 753                               const void *from, unsigned size)
 754 {
 755         int ret;
 756
 757         if (!vq->iotlb)
 758                 return __copy_to_user(to, from, size);
 759         else {
 760                 /* This function should be called after iotlb
 761                  * prefetch, which means we're sure that all vq
 762                  * could be access through iotlb. So -EAGAIN should
 763                  * not happen in this case.
 764                  */
 765                 struct iov_iter t;
 766                 void __user *uaddr = vhost_vq_meta_fetch(vq,
 767                                      (u64)(uintptr_t)to, size,
 768                                      VHOST_ADDR_DESC);
 769
 770                 if (uaddr)
 771                         return __copy_to_user(uaddr, from, size);
 772
 773                 ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
 774                                      ARRAY_SIZE(vq->iotlb_iov),
 775                                      VHOST_ACCESS_WO);
 776                 if (ret < 0)
 777                         goto out;
 778                 iov_iter_init(&t, WRITE, vq->iotlb_iov, ret, size);
 779                 ret = copy_to_iter(from, size, &t);
 780                 if (ret == size)
 781                         ret = 0;
 782         }
 783 out:
 784         return ret;
 785 }
 786
 787 static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
 788                                 void __user *from, unsigned size)
 789 {
 790         int ret;
 791
 792         if (!vq->iotlb)
 793                 return __copy_from_user(to, from, size);
 794         else {
 795                 /* This function should be called after iotlb
 796                  * prefetch, which means we're sure that vq
 797                  * could be access through iotlb. So -EAGAIN should
 798                  * not happen in this case.
 799                  */
 800                 void __user *uaddr = vhost_vq_meta_fetch(vq,
 801                                      (u64)(uintptr_t)from, size,
 802                                      VHOST_ADDR_DESC);
 803                 struct iov_iter f;
 804
 805                 if (uaddr)
 806                         return __copy_from_user(to, uaddr, size);
 807
 808                 ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
 809                                      ARRAY_SIZE(vq->iotlb_iov),
 810                                      VHOST_ACCESS_RO);
 811                 if (ret < 0) {
 812                         vq_err(vq, "IOTLB translation failure: uaddr "
 813                                "%p size 0x%llx\n", from,
 814                                (unsigned long long) size);
 815                         goto out;
 816                 }
 817                 iov_iter_init(&f, READ, vq->iotlb_iov, ret, size);
 818                 ret = copy_from_iter(to, size, &f);
 819                 if (ret == size)
 820                         ret = 0;
 821         }
 822
 823 out:
 824         return ret;
 825 }
 826
 827 static void __user *__vhost_get_user_slow(struct vhost_virtqueue *vq,
 828                                           void __user *addr, unsigned int size,
 829                                           int type)
 830 {
 831         int ret;
 832
 833         ret = translate_desc(vq, (u64)(uintptr_t)addr, size, vq->iotlb_iov,
 834                              ARRAY_SIZE(vq->iotlb_iov),
 835                              VHOST_ACCESS_RO);
 836         if (ret < 0) {
 837                 vq_err(vq, "IOTLB translation failure: uaddr "
 838                         "%p size 0x%llx\n", addr,
 839                         (unsigned long long) size);
 840                 return NULL;
 841         }
 842
 843         if (ret != 1 || vq->iotlb_iov[0].iov_len != size) {
 844                 vq_err(vq, "Non atomic userspace memory access: uaddr "
 845                         "%p size 0x%llx\n", addr,
 846                         (unsigned long long) size);
 847                 return NULL;
 848         }
 849
 850         return vq->iotlb_iov[0].iov_base;
 851 }
 852
 853 /* This function should be called after iotlb
 854  * prefetch, which means we're sure that vq
 855  * could be access through iotlb. So -EAGAIN should
 856  * not happen in this case.
 857  */
 858 static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
 859                                             void *addr, unsigned int size,
 860                                             int type)
 861 {
 862         void __user *uaddr = vhost_vq_meta_fetch(vq,
 863                              (u64)(uintptr_t)addr, size, type);
 864         if (uaddr)
 865                 return uaddr;
 866
 867         return __vhost_get_user_slow(vq, addr, size, type);
 868 }
 869
 870 #define vhost_put_user(vq, x, ptr)              \
 871 ({ \
 872         int ret = -EFAULT; \
 873         if (!vq->iotlb) { \
 874                 ret = __put_user(x, ptr); \
 875         } else { \
 876                 __typeof__(ptr) to = \
 877                         (__typeof__(ptr)) __vhost_get_user(vq, ptr,     \
 878                                           sizeof(*ptr), VHOST_ADDR_USED); \
 879                 if (to != NULL) \
 880                         ret = __put_user(x, to); \
 881                 else \
 882                         ret = -EFAULT;  \
 883         } \
 884         ret; \
 885 })
 886
 887 #define vhost_get_user(vq, x, ptr, type)                \
 888 ({ \
 889         int ret; \
 890         if (!vq->iotlb) { \
 891                 ret = __get_user(x, ptr); \
 892         } else { \
 893                 __typeof__(ptr) from = \
 894                         (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
 895                                                            sizeof(*ptr), \
 896                                                            type); \
 897                 if (from != NULL) \
 898                         ret = __get_user(x, from); \
 899                 else \
 900                         ret = -EFAULT; \
 901         } \
 902         ret; \
 903 })
 904
 905 #define vhost_get_avail(vq, x, ptr) \
 906         vhost_get_user(vq, x, ptr, VHOST_ADDR_AVAIL)
 907
 908 #define vhost_get_used(vq, x, ptr) \
 909         vhost_get_user(vq, x, ptr, VHOST_ADDR_USED)
 910
 911 static void vhost_dev_lock_vqs(struct vhost_dev *d)
 912 {
 913         int i = 0;
 914         for (i = 0; i < d->nvqs; ++i)
 915                 mutex_lock(&d->vqs[i]->mutex);
 916 }
 917
 918 static void vhost_dev_unlock_vqs(struct vhost_dev *d)
 919 {
 920         int i = 0;
 921         for (i = 0; i < d->nvqs; ++i)
 922                 mutex_unlock(&d->vqs[i]->mutex);
 923 }
 924
 925 static int vhost_new_umem_range(struct vhost_umem *umem,
 926                                 u64 start, u64 size, u64 end,
 927                                 u64 userspace_addr, int perm)
 928 {
 929         struct vhost_umem_node *tmp, *node = kmalloc(sizeof(*node), GFP_ATOMIC);
 930
 931         if (!node)
 932                 return -ENOMEM;
 933
 934         if (umem->numem == max_iotlb_entries) {
 935                 tmp = list_first_entry(&umem->umem_list, typeof(*tmp), link);
 936                 vhost_umem_free(umem, tmp);
 937         }
 938
 939         node->start = start;
 940         node->size = size;
 941         node->last = end;
 942         node->userspace_addr = userspace_addr;
 943         node->perm = perm;
 944         INIT_LIST_HEAD(&node->link);
 945         list_add_tail(&node->link, &umem->umem_list);
 946         vhost_umem_interval_tree_insert(node, &umem->umem_tree);
 947         umem->numem++;
 948
 949         return 0;
 950 }
 951
 952 static void vhost_del_umem_range(struct vhost_umem *umem,
 953                                  u64 start, u64 end)
 954 {
 955         struct vhost_umem_node *node;
 956
 957         while ((node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
 958                                                            start, end)))
 959                 vhost_umem_free(umem, node);
 960 }
 961
 962 static void vhost_iotlb_notify_vq(struct vhost_dev *d,
 963                                   struct vhost_iotlb_msg *msg)
 964 {
 965         struct vhost_msg_node *node, *n;
 966
 967         spin_lock(&d->iotlb_lock);
 968
 969         list_for_each_entry_safe(node, n, &d->pending_list, node) {
 970                 struct vhost_iotlb_msg *vq_msg = &node->msg.iotlb;
 971                 if (msg->iova <= vq_msg->iova &&
 972                     msg->iova + msg->size - 1 > vq_msg->iova &&
 973                     vq_msg->type == VHOST_IOTLB_MISS) {
 974                         vhost_poll_queue(&node->vq->poll);
 975                         list_del(&node->node);
 976                         kfree(node);
 977                 }
 978         }
 979
 980         spin_unlock(&d->iotlb_lock);
 981 }
 982
 983 static int umem_access_ok(u64 uaddr, u64 size, int access)
 984 {
 985         unsigned long a = uaddr;
 986
 987         /* Make sure 64 bit math will not overflow. */
 988         if (vhost_overflow(uaddr, size))
 989                 return -EFAULT;
 990
 991         if ((access & VHOST_ACCESS_RO) &&
 992             !access_ok(VERIFY_READ, (void __user *)a, size))
 993                 return -EFAULT;
 994         if ((access & VHOST_ACCESS_WO) &&
 995             !access_ok(VERIFY_WRITE, (void __user *)a, size))
 996                 return -EFAULT;
 997         return 0;
 998 }
 999
1000 static int vhost_process_iotlb_msg(struct vhost_dev *dev,
1001                                    struct vhost_iotlb_msg *msg)
1002 {
1003         int ret = 0;
1004
1005         vhost_dev_lock_vqs(dev);
1006         switch (msg->type) {
1007         case VHOST_IOTLB_UPDATE:
1008                 if (!dev->iotlb) {
1009                         ret = -EFAULT;
1010                         break;
1011                 }
1012                 if (umem_access_ok(msg->uaddr, msg->size, msg->perm)) {
1013                         ret = -EFAULT;
1014                         break;
1015                 }
1016                 vhost_vq_meta_reset(dev);
1017                 if (vhost_new_umem_range(dev->iotlb, msg->iova, msg->size,
1018                                          msg->iova + msg->size - 1,
1019                                          msg->uaddr, msg->perm)) {
1020                         ret = -ENOMEM;
1021                         break;
1022                 }
1023                 vhost_iotlb_notify_vq(dev, msg);
1024                 break;
1025         case VHOST_IOTLB_INVALIDATE:
1026                 vhost_vq_meta_reset(dev);
1027                 vhost_del_umem_range(dev->iotlb, msg->iova,
1028                                      msg->iova + msg->size - 1);
1029                 break;
1030         default:
1031                 ret = -EINVAL;
1032                 break;
1033         }
1034
1035         vhost_dev_unlock_vqs(dev);
1036         return ret;
1037 }
1038 ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
1039                              struct iov_iter *from)
1040 {
1041         struct vhost_msg_node node;
1042         unsigned size = sizeof(struct vhost_msg);
1043         size_t ret;
1044         int err;
1045
1046         if (iov_iter_count(from) < size)
1047                 return 0;
1048         ret = copy_from_iter(&node.msg, size, from);
1049         if (ret != size)
1050                 goto done;
1051
1052         switch (node.msg.type) {
1053         case VHOST_IOTLB_MSG:
1054                 err = vhost_process_iotlb_msg(dev, &node.msg.iotlb);
1055                 if (err)
1056                         ret = err;
1057                 break;
1058         default:
1059                 ret = -EINVAL;
1060                 break;
1061         }
1062
1063 done:
1064         return ret;
1065 }
1066 EXPORT_SYMBOL(vhost_chr_write_iter);
1067
1068 unsigned int vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1069                             poll_table *wait)
1070 {
1071         unsigned int mask = 0;
1072
1073         poll_wait(file, &dev->wait, wait);
1074
1075         if (!list_empty(&dev->read_list))
1076                 mask |= POLLIN | POLLRDNORM;
1077
1078         return mask;
1079 }
1080 EXPORT_SYMBOL(vhost_chr_poll);
1081
1082 ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1083                             int noblock)
1084 {
1085         DEFINE_WAIT(wait);
1086         struct vhost_msg_node *node;
1087         ssize_t ret = 0;
1088         unsigned size = sizeof(struct vhost_msg);
1089
1090         if (iov_iter_count(to) < size)
1091                 return 0;
1092
1093         while (1) {
1094                 if (!noblock)
1095                         prepare_to_wait(&dev->wait, &wait,
1096                                         TASK_INTERRUPTIBLE);
1097
1098                 node = vhost_dequeue_msg(dev, &dev->read_list);
1099                 if (node)
1100                         break;
1101                 if (noblock) {
1102                         ret = -EAGAIN;
1103                         break;
1104                 }
1105                 if (signal_pending(current)) {
1106                         ret = -ERESTARTSYS;
1107                         break;
1108                 }
1109                 if (!dev->iotlb) {
1110                         ret = -EBADFD;
1111                         break;
1112                 }
1113
1114                 schedule();
1115         }
1116
1117         if (!noblock)
1118                 finish_wait(&dev->wait, &wait);
1119
1120         if (node) {
1121                 ret = copy_to_iter(&node->msg, size, to);
1122
1123                 if (ret != size || node->msg.type != VHOST_IOTLB_MISS) {
1124                         kfree(node);
1125                         return ret;
1126                 }
1127
1128                 vhost_enqueue_msg(dev, &dev->pending_list, node);
1129         }
1130
1131         return ret;
1132 }
1133 EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1134
1135 static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1136 {
1137         struct vhost_dev *dev = vq->dev;
1138         struct vhost_msg_node *node;
1139         struct vhost_iotlb_msg *msg;
1140
1141         node = vhost_new_msg(vq, VHOST_IOTLB_MISS);
1142         if (!node)
1143                 return -ENOMEM;
1144
1145         msg = &node->msg.iotlb;
1146         msg->type = VHOST_IOTLB_MISS;
1147         msg->iova = iova;
1148         msg->perm = access;
1149
1150         vhost_enqueue_msg(dev, &dev->read_list, node);
1151
1152         return 0;
1153 }
1154
1155 static int vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1156                         struct vring_desc __user *desc,
1157                         struct vring_avail __user *avail,
1158                         struct vring_used __user *used)
1159
1160 {
1161         size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1162
1163         return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
1164                access_ok(VERIFY_READ, avail,
1165                          sizeof *avail + num * sizeof *avail->ring + s) &&
1166                access_ok(VERIFY_WRITE, used,
1167                         sizeof *used + num * sizeof *used->ring + s);
1168 }
1169
1170 static void vhost_vq_meta_update(struct vhost_virtqueue *vq,
1171                                  const struct vhost_umem_node *node,
1172                                  int type)
1173 {
1174         int access = (type == VHOST_ADDR_USED) ?
1175                      VHOST_ACCESS_WO : VHOST_ACCESS_RO;
1176
1177         if (likely(node->perm & access))
1178                 vq->meta_iotlb[type] = node;
1179 }
1180
1181 static int iotlb_access_ok(struct vhost_virtqueue *vq,
1182                            int access, u64 addr, u64 len, int type)
1183 {
1184         const struct vhost_umem_node *node;
1185         struct vhost_umem *umem = vq->iotlb;
1186         u64 s = 0, size, orig_addr = addr;
1187
1188         if (vhost_vq_meta_fetch(vq, addr, len, type))
1189                 return true;
1190
1191         while (len > s) {
1192                 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1193                                                            addr,
1194                                                            addr + len - 1);
1195                 if (node == NULL || node->start > addr) {
1196                         vhost_iotlb_miss(vq, addr, access);
1197                         return false;
1198                 } else if (!(node->perm & access)) {
1199                         /* Report the possible access violation by
1200                          * request another translation from userspace.
1201                          */
1202                         return false;
1203                 }
1204
1205                 size = node->size - addr + node->start;
1206
1207                 if (orig_addr == addr && size >= len)
1208                         vhost_vq_meta_update(vq, node, type);
1209
1210                 s += size;
1211                 addr += size;
1212         }
1213
1214         return true;
1215 }
1216
1217 int vq_iotlb_prefetch(struct vhost_virtqueue *vq)
1218 {
1219         size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1220         unsigned int num = vq->num;
1221
1222         if (!vq->iotlb)
1223                 return 1;
1224
1225         return iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->desc,
1226                                num * sizeof(*vq->desc), VHOST_ADDR_DESC) &&
1227                iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->avail,
1228                                sizeof *vq->avail +
1229                                num * sizeof(*vq->avail->ring) + s,
1230                                VHOST_ADDR_AVAIL) &&
1231                iotlb_access_ok(vq, VHOST_ACCESS_WO, (u64)(uintptr_t)vq->used,
1232                                sizeof *vq->used +
1233                                num * sizeof(*vq->used->ring) + s,
1234                                VHOST_ADDR_USED);
1235 }
1236 EXPORT_SYMBOL_GPL(vq_iotlb_prefetch);
1237
1238 /* Can we log writes? */
1239 /* Caller should have device mutex but not vq mutex */
1240 int vhost_log_access_ok(struct vhost_dev *dev)
1241 {
1242         return memory_access_ok(dev, dev->umem, 1);
1243 }
1244 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1245
1246 /* Verify access for write logging. */
1247 /* Caller should have vq mutex and device mutex */
1248 static int vq_log_access_ok(struct vhost_virtqueue *vq,
1249                             void __user *log_base)
1250 {
1251         size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1252
1253         return vq_memory_access_ok(log_base, vq->umem,
1254                                    vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1255                 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
1256                                         sizeof *vq->used +
1257                                         vq->num * sizeof *vq->used->ring + s));
1258 }
1259
1260 /* Can we start vq? */
1261 /* Caller should have vq mutex and device mutex */
1262 int vhost_vq_access_ok(struct vhost_virtqueue *vq)
1263 {
1264         if (vq->iotlb) {
1265                 /* When device IOTLB was used, the access validation
1266                  * will be validated during prefetching.
1267                  */
1268                 return 1;
1269         }
1270         return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used) &&
1271                 vq_log_access_ok(vq, vq->log_base);
1272 }
1273 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1274
1275 static struct vhost_umem *vhost_umem_alloc(void)
1276 {
1277         struct vhost_umem *umem = vhost_kvzalloc(sizeof(*umem));
1278
1279         if (!umem)
1280                 return NULL;
1281
1282         umem->umem_tree = RB_ROOT;
1283         umem->numem = 0;
1284         INIT_LIST_HEAD(&umem->umem_list);
1285
1286         return umem;
1287 }
1288
1289 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1290 {
1291         struct vhost_memory mem, *newmem;
1292         struct vhost_memory_region *region;
1293         struct vhost_umem *newumem, *oldumem;
1294         unsigned long size = offsetof(struct vhost_memory, regions);
1295         int i;
1296
1297         if (copy_from_user(&mem, m, size))
1298                 return -EFAULT;
1299         if (mem.padding)
1300                 return -EOPNOTSUPP;
1301         if (mem.nregions > max_mem_regions)
1302                 return -E2BIG;
1303         newmem = vhost_kvzalloc(size + mem.nregions * sizeof(*m->regions));
1304         if (!newmem)
1305                 return -ENOMEM;
1306
1307         memcpy(newmem, &mem, size);
1308         if (copy_from_user(newmem->regions, m->regions,
1309                            mem.nregions * sizeof *m->regions)) {
1310                 kvfree(newmem);
1311                 return -EFAULT;
1312         }
1313
1314         newumem = vhost_umem_alloc();
1315         if (!newumem) {
1316                 kvfree(newmem);
1317                 return -ENOMEM;
1318         }
1319
1320         for (region = newmem->regions;
1321              region < newmem->regions + mem.nregions;
1322              region++) {
1323                 if (vhost_new_umem_range(newumem,
1324                                          region->guest_phys_addr,
1325                                          region->memory_size,
1326                                          region->guest_phys_addr +
1327                                          region->memory_size - 1,
1328                                          region->userspace_addr,
1329                                          VHOST_ACCESS_RW))
1330                         goto err;
1331         }
1332
1333         if (!memory_access_ok(d, newumem, 0))
1334                 goto err;
1335
1336         oldumem = d->umem;
1337         d->umem = newumem;
1338
1339         /* All memory accesses are done under some VQ mutex. */
1340         for (i = 0; i < d->nvqs; ++i) {
1341                 mutex_lock(&d->vqs[i]->mutex);
1342                 d->vqs[i]->umem = newumem;
1343                 mutex_unlock(&d->vqs[i]->mutex);
1344         }
1345
1346         kvfree(newmem);
1347         vhost_umem_clean(oldumem);
1348         return 0;
1349
1350 err:
1351         vhost_umem_clean(newumem);
1352         kvfree(newmem);
1353         return -EFAULT;
1354 }
1355
1356 long vhost_vring_ioctl(struct vhost_dev *d, int ioctl, void __user *argp)
1357 {
1358         struct file *eventfp, *filep = NULL;
1359         bool pollstart = false, pollstop = false;
1360         struct eventfd_ctx *ctx = NULL;
1361         u32 __user *idxp = argp;
1362         struct vhost_virtqueue *vq;
1363         struct vhost_vring_state s;
1364         struct vhost_vring_file f;
1365         struct vhost_vring_addr a;
1366         u32 idx;
1367         long r;
1368
1369         r = get_user(idx, idxp);
1370         if (r < 0)
1371                 return r;
1372         if (idx >= d->nvqs)
1373                 return -ENOBUFS;
1374
1375         vq = d->vqs[idx];
1376
1377         mutex_lock(&vq->mutex);
1378
1379         switch (ioctl) {
1380         case VHOST_SET_VRING_NUM:
1381                 /* Resizing ring with an active backend?
1382                  * You don't want to do that. */
1383                 if (vq->private_data) {
1384                         r = -EBUSY;
1385                         break;
1386                 }
1387                 if (copy_from_user(&s, argp, sizeof s)) {
1388                         r = -EFAULT;
1389                         break;
1390                 }
1391                 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
1392                         r = -EINVAL;
1393                         break;
1394                 }
1395                 vq->num = s.num;
1396                 break;
1397         case VHOST_SET_VRING_BASE:
1398                 /* Moving base with an active backend?
1399                  * You don't want to do that. */
1400                 if (vq->private_data) {
1401                         r = -EBUSY;
1402                         break;
1403                 }
1404                 if (copy_from_user(&s, argp, sizeof s)) {
1405                         r = -EFAULT;
1406                         break;
1407                 }
1408                 if (s.num > 0xffff) {
1409                         r = -EINVAL;
1410                         break;
1411                 }
1412                 vq->last_avail_idx = vq->last_used_event = s.num;
1413                 /* Forget the cached index value. */
1414                 vq->avail_idx = vq->last_avail_idx;
1415                 break;
1416         case VHOST_GET_VRING_BASE:
1417                 s.index = idx;
1418                 s.num = vq->last_avail_idx;
1419                 if (copy_to_user(argp, &s, sizeof s))
1420                         r = -EFAULT;
1421                 break;
1422         case VHOST_SET_VRING_ADDR:
1423                 if (copy_from_user(&a, argp, sizeof a)) {
1424                         r = -EFAULT;
1425                         break;
1426                 }
1427                 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
1428                         r = -EOPNOTSUPP;
1429                         break;
1430                 }
1431                 /* For 32bit, verify that the top 32bits of the user
1432                    data are set to zero. */
1433                 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1434                     (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1435                     (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
1436                         r = -EFAULT;
1437                         break;
1438                 }
1439
1440                 /* Make sure it's safe to cast pointers to vring types. */
1441                 BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1442                 BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1443                 if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1444                     (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1445                     (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1))) {
1446                         r = -EINVAL;
1447                         break;
1448                 }
1449
1450                 /* We only verify access here if backend is configured.
1451                  * If it is not, we don't as size might not have been setup.
1452                  * We will verify when backend is configured. */
1453                 if (vq->private_data) {
1454                         if (!vq_access_ok(vq, vq->num,
1455                                 (void __user *)(unsigned long)a.desc_user_addr,
1456                                 (void __user *)(unsigned long)a.avail_user_addr,
1457                                 (void __user *)(unsigned long)a.used_user_addr)) {
1458                                 r = -EINVAL;
1459                                 break;
1460                         }
1461
1462                         /* Also validate log access for used ring if enabled. */
1463                         if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
1464                             !log_access_ok(vq->log_base, a.log_guest_addr,
1465                                            sizeof *vq->used +
1466                                            vq->num * sizeof *vq->used->ring)) {
1467                                 r = -EINVAL;
1468                                 break;
1469                         }
1470                 }
1471
1472                 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1473                 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1474                 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1475                 vq->log_addr = a.log_guest_addr;
1476                 vq->used = (void __user *)(unsigned long)a.used_user_addr;
1477                 break;
1478         case VHOST_SET_VRING_KICK:
1479                 if (copy_from_user(&f, argp, sizeof f)) {
1480                         r = -EFAULT;
1481                         break;
1482                 }
1483                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1484                 if (IS_ERR(eventfp)) {
1485                         r = PTR_ERR(eventfp);
1486                         break;
1487                 }
1488                 if (eventfp != vq->kick) {
1489                         pollstop = (filep = vq->kick) != NULL;
1490                         pollstart = (vq->kick = eventfp) != NULL;
1491                 } else
1492                         filep = eventfp;
1493                 break;
1494         case VHOST_SET_VRING_CALL:
1495                 if (copy_from_user(&f, argp, sizeof f)) {
1496                         r = -EFAULT;
1497                         break;
1498                 }
1499                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1500                 if (IS_ERR(eventfp)) {
1501                         r = PTR_ERR(eventfp);
1502                         break;
1503                 }
1504                 if (eventfp != vq->call) {
1505                         filep = vq->call;
1506                         ctx = vq->call_ctx;
1507                         vq->call = eventfp;
1508                         vq->call_ctx = eventfp ?
1509                                 eventfd_ctx_fileget(eventfp) : NULL;
1510                 } else
1511                         filep = eventfp;
1512                 break;
1513         case VHOST_SET_VRING_ERR:
1514                 if (copy_from_user(&f, argp, sizeof f)) {
1515                         r = -EFAULT;
1516                         break;
1517                 }
1518                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1519                 if (IS_ERR(eventfp)) {
1520                         r = PTR_ERR(eventfp);
1521                         break;
1522                 }
1523                 if (eventfp != vq->error) {
1524                         filep = vq->error;
1525                         vq->error = eventfp;
1526                         ctx = vq->error_ctx;
1527                         vq->error_ctx = eventfp ?
1528                                 eventfd_ctx_fileget(eventfp) : NULL;
1529                 } else
1530                         filep = eventfp;
1531                 break;
1532         case VHOST_SET_VRING_ENDIAN:
1533                 r = vhost_set_vring_endian(vq, argp);
1534                 break;
1535         case VHOST_GET_VRING_ENDIAN:
1536                 r = vhost_get_vring_endian(vq, idx, argp);
1537                 break;
1538         case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
1539                 if (copy_from_user(&s, argp, sizeof(s))) {
1540                         r = -EFAULT;
1541                         break;
1542                 }
1543                 vq->busyloop_timeout = s.num;
1544                 break;
1545         case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
1546                 s.index = idx;
1547                 s.num = vq->busyloop_timeout;
1548                 if (copy_to_user(argp, &s, sizeof(s)))
1549                         r = -EFAULT;
1550                 break;
1551         default:
1552                 r = -ENOIOCTLCMD;
1553         }
1554
1555         if (pollstop && vq->handle_kick)
1556                 vhost_poll_stop(&vq->poll);
1557
1558         if (ctx)
1559                 eventfd_ctx_put(ctx);
1560         if (filep)
1561                 fput(filep);
1562
1563         if (pollstart && vq->handle_kick)
1564                 r = vhost_poll_start(&vq->poll, vq->kick);
1565
1566         mutex_unlock(&vq->mutex);
1567
1568         if (pollstop && vq->handle_kick)
1569                 vhost_poll_flush(&vq->poll);
1570         return r;
1571 }
1572 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
1573
1574 int vhost_init_device_iotlb(struct vhost_dev *d, bool enabled)
1575 {
1576         struct vhost_umem *niotlb, *oiotlb;
1577         int i;
1578
1579         niotlb = vhost_umem_alloc();
1580         if (!niotlb)
1581                 return -ENOMEM;
1582
1583         oiotlb = d->iotlb;
1584         d->iotlb = niotlb;
1585
1586         for (i = 0; i < d->nvqs; ++i) {
1587                 mutex_lock(&d->vqs[i]->mutex);
1588                 d->vqs[i]->iotlb = niotlb;
1589                 mutex_unlock(&d->vqs[i]->mutex);
1590         }
1591
1592         vhost_umem_clean(oiotlb);
1593
1594         return 0;
1595 }
1596 EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
1597
1598 /* Caller must have device mutex */
1599 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1600 {
1601         struct file *eventfp, *filep = NULL;
1602         struct eventfd_ctx *ctx = NULL;
1603         u64 p;
1604         long r;
1605         int i, fd;
1606
1607         /* If you are not the owner, you can become one */
1608         if (ioctl == VHOST_SET_OWNER) {
1609                 r = vhost_dev_set_owner(d);
1610                 goto done;
1611         }
1612
1613         /* You must be the owner to do anything else */
1614         r = vhost_dev_check_owner(d);
1615         if (r)
1616                 goto done;
1617
1618         switch (ioctl) {
1619         case VHOST_SET_MEM_TABLE:
1620                 r = vhost_set_memory(d, argp);
1621                 break;
1622         case VHOST_SET_LOG_BASE:
1623                 if (copy_from_user(&p, argp, sizeof p)) {
1624                         r = -EFAULT;
1625                         break;
1626                 }
1627                 if ((u64)(unsigned long)p != p) {
1628                         r = -EFAULT;
1629                         break;
1630                 }
1631                 for (i = 0; i < d->nvqs; ++i) {
1632                         struct vhost_virtqueue *vq;
1633                         void __user *base = (void __user *)(unsigned long)p;
1634                         vq = d->vqs[i];
1635                         mutex_lock(&vq->mutex);
1636                         /* If ring is inactive, will check when it's enabled. */
1637                         if (vq->private_data && !vq_log_access_ok(vq, base))
1638                                 r = -EFAULT;
1639                         else
1640                                 vq->log_base = base;
1641                         mutex_unlock(&vq->mutex);
1642                 }
1643                 break;
1644         case VHOST_SET_LOG_FD:
1645                 r = get_user(fd, (int __user *)argp);
1646                 if (r < 0)
1647                         break;
1648                 eventfp = fd == -1 ? NULL : eventfd_fget(fd);
1649                 if (IS_ERR(eventfp)) {
1650                         r = PTR_ERR(eventfp);
1651                         break;
1652                 }
1653                 if (eventfp != d->log_file) {
1654                         filep = d->log_file;
1655                         d->log_file = eventfp;
1656                         ctx = d->log_ctx;
1657                         d->log_ctx = eventfp ?
1658                                 eventfd_ctx_fileget(eventfp) : NULL;
1659                 } else
1660                         filep = eventfp;
1661                 for (i = 0; i < d->nvqs; ++i) {
1662                         mutex_lock(&d->vqs[i]->mutex);
1663                         d->vqs[i]->log_ctx = d->log_ctx;
1664                         mutex_unlock(&d->vqs[i]->mutex);
1665                 }
1666                 if (ctx)
1667                         eventfd_ctx_put(ctx);
1668                 if (filep)
1669                         fput(filep);
1670                 break;
1671         default:
1672                 r = -ENOIOCTLCMD;
1673                 break;
1674         }
1675 done:
1676         return r;
1677 }
1678 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1679
1680 /* TODO: This is really inefficient.  We need something like get_user()
1681  * (instruction directly accesses the data, with an exception table entry
1682  * returning -EFAULT). See Documentation/x86/exception-tables.txt.
1683  */
1684 static int set_bit_to_user(int nr, void __user *addr)
1685 {
1686         unsigned long log = (unsigned long)addr;
1687         struct page *page;
1688         void *base;
1689         int bit = nr + (log % PAGE_SIZE) * 8;
1690         int r;
1691
1692         r = get_user_pages_fast(log, 1, 1, &page);
1693         if (r < 0)
1694                 return r;
1695         BUG_ON(r != 1);
1696         base = kmap_atomic(page);
1697         set_bit(bit, base);
1698         kunmap_atomic(base);
1699         set_page_dirty_lock(page);
1700         put_page(page);
1701         return 0;
1702 }
1703
1704 static int log_write(void __user *log_base,
1705                      u64 write_address, u64 write_length)
1706 {
1707         u64 write_page = write_address / VHOST_PAGE_SIZE;
1708         int r;
1709
1710         if (!write_length)
1711                 return 0;
1712         write_length += write_address % VHOST_PAGE_SIZE;
1713         for (;;) {
1714                 u64 base = (u64)(unsigned long)log_base;
1715                 u64 log = base + write_page / 8;
1716                 int bit = write_page % 8;
1717                 if ((u64)(unsigned long)log != log)
1718                         return -EFAULT;
1719                 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1720                 if (r < 0)
1721                         return r;
1722                 if (write_length <= VHOST_PAGE_SIZE)
1723                         break;
1724                 write_length -= VHOST_PAGE_SIZE;
1725                 write_page += 1;
1726         }
1727         return r;
1728 }
1729
1730 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1731                     unsigned int log_num, u64 len)
1732 {
1733         int i, r;
1734
1735         /* Make sure data written is seen before log. */
1736         smp_wmb();
1737         for (i = 0; i < log_num; ++i) {
1738                 u64 l = min(log[i].len, len);
1739                 r = log_write(vq->log_base, log[i].addr, l);
1740                 if (r < 0)
1741                         return r;
1742                 len -= l;
1743                 if (!len) {
1744                         if (vq->log_ctx)
1745                                 eventfd_signal(vq->log_ctx, 1);
1746                         return 0;
1747                 }
1748         }
1749         /* Length written exceeds what we have stored. This is a bug. */
1750         BUG();
1751         return 0;
1752 }
1753 EXPORT_SYMBOL_GPL(vhost_log_write);
1754
1755 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1756 {
1757         void __user *used;
1758         if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
1759                            &vq->used->flags) < 0)
1760                 return -EFAULT;
1761         if (unlikely(vq->log_used)) {
1762                 /* Make sure the flag is seen before log. */
1763                 smp_wmb();
1764                 /* Log used flag write. */
1765                 used = &vq->used->flags;
1766                 log_write(vq->log_base, vq->log_addr +
1767                           (used - (void __user *)vq->used),
1768                           sizeof vq->used->flags);
1769                 if (vq->log_ctx)
1770                         eventfd_signal(vq->log_ctx, 1);
1771         }
1772         return 0;
1773 }
1774
1775 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1776 {
1777         if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
1778                            vhost_avail_event(vq)))
1779                 return -EFAULT;
1780         if (unlikely(vq->log_used)) {
1781                 void __user *used;
1782                 /* Make sure the event is seen before log. */
1783                 smp_wmb();
1784                 /* Log avail event write */
1785                 used = vhost_avail_event(vq);
1786                 log_write(vq->log_base, vq->log_addr +
1787                           (used - (void __user *)vq->used),
1788                           sizeof *vhost_avail_event(vq));
1789                 if (vq->log_ctx)
1790                         eventfd_signal(vq->log_ctx, 1);
1791         }
1792         return 0;
1793 }
1794
1795 int vhost_vq_init_access(struct vhost_virtqueue *vq)
1796 {
1797         __virtio16 last_used_idx;
1798         int r;
1799         bool is_le = vq->is_le;
1800
1801         if (!vq->private_data)
1802                 return 0;
1803
1804         vhost_init_is_le(vq);
1805
1806         r = vhost_update_used_flags(vq);
1807         if (r)
1808                 goto err;
1809         vq->signalled_used_valid = false;
1810         if (!vq->iotlb &&
1811             !access_ok(VERIFY_READ, &vq->used->idx, sizeof vq->used->idx)) {
1812                 r = -EFAULT;
1813                 goto err;
1814         }
1815         r = vhost_get_used(vq, last_used_idx, &vq->used->idx);
1816         if (r) {
1817                 vq_err(vq, "Can't access used idx at %p\n",
1818                        &vq->used->idx);
1819                 goto err;
1820         }
1821         vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
1822         return 0;
1823
1824 err:
1825         vq->is_le = is_le;
1826         return r;
1827 }
1828 EXPORT_SYMBOL_GPL(vhost_vq_init_access);
1829
1830 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
1831                           struct iovec iov[], int iov_size, int access)
1832 {
1833         const struct vhost_umem_node *node;
1834         struct vhost_dev *dev = vq->dev;
1835         struct vhost_umem *umem = dev->iotlb ? dev->iotlb : dev->umem;
1836         struct iovec *_iov;
1837         u64 s = 0;
1838         int ret = 0;
1839
1840         while ((u64)len > s) {
1841                 u64 size;
1842                 if (unlikely(ret >= iov_size)) {
1843                         ret = -ENOBUFS;
1844                         break;
1845                 }
1846
1847                 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1848                                                         addr, addr + len - 1);
1849                 if (node == NULL || node->start > addr) {
1850                         if (umem != dev->iotlb) {
1851                                 ret = -EFAULT;
1852                                 break;
1853                         }
1854                         ret = -EAGAIN;
1855                         break;
1856                 } else if (!(node->perm & access)) {
1857                         ret = -EPERM;
1858                         break;
1859                 }
1860
1861                 _iov = iov + ret;
1862                 size = node->size - addr + node->start;
1863                 _iov->iov_len = min((u64)len - s, size);
1864                 _iov->iov_base = (void __user *)(unsigned long)
1865                         (node->userspace_addr + addr - node->start);
1866                 s += size;
1867                 addr += size;
1868                 ++ret;
1869         }
1870
1871         if (ret == -EAGAIN)
1872                 vhost_iotlb_miss(vq, addr, access);
1873         return ret;
1874 }
1875
1876 /* Each buffer in the virtqueues is actually a chain of descriptors.  This
1877  * function returns the next descriptor in the chain,
1878  * or -1U if we're at the end. */
1879 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
1880 {
1881         unsigned int next;
1882
1883         /* If this descriptor says it doesn't chain, we're done. */
1884         if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
1885                 return -1U;
1886
1887         /* Check they're not leading us off end of descriptors. */
1888         next = vhost16_to_cpu(vq, desc->next);
1889         /* Make sure compiler knows to grab that: we don't want it changing! */
1890         /* We will use the result as an index in an array, so most
1891          * architectures only need a compiler barrier here. */
1892         read_barrier_depends();
1893
1894         return next;
1895 }
1896
1897 static int get_indirect(struct vhost_virtqueue *vq,
1898                         struct iovec iov[], unsigned int iov_size,
1899                         unsigned int *out_num, unsigned int *in_num,
1900                         struct vhost_log *log, unsigned int *log_num,
1901                         struct vring_desc *indirect)
1902 {
1903         struct vring_desc desc;
1904         unsigned int i = 0, count, found = 0;
1905         u32 len = vhost32_to_cpu(vq, indirect->len);
1906         struct iov_iter from;
1907         int ret, access;
1908
1909         /* Sanity check */
1910         if (unlikely(len % sizeof desc)) {
1911                 vq_err(vq, "Invalid length in indirect descriptor: "
1912                        "len 0x%llx not multiple of 0x%zx\n",
1913                        (unsigned long long)len,
1914                        sizeof desc);
1915                 return -EINVAL;
1916         }
1917
1918         ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
1919                              UIO_MAXIOV, VHOST_ACCESS_RO);
1920         if (unlikely(ret < 0)) {
1921                 if (ret != -EAGAIN)
1922                         vq_err(vq, "Translation failure %d in indirect.\n", ret);
1923                 return ret;
1924         }
1925         iov_iter_init(&from, READ, vq->indirect, ret, len);
1926
1927         /* We will use the result as an address to read from, so most
1928          * architectures only need a compiler barrier here. */
1929         read_barrier_depends();
1930
1931         count = len / sizeof desc;
1932         /* Buffers are chained via a 16 bit next field, so
1933          * we can have at most 2^16 of these. */
1934         if (unlikely(count > USHRT_MAX + 1)) {
1935                 vq_err(vq, "Indirect buffer length too big: %d\n",
1936                        indirect->len);
1937                 return -E2BIG;
1938         }
1939
1940         do {
1941                 unsigned iov_count = *in_num + *out_num;
1942                 if (unlikely(++found > count)) {
1943                         vq_err(vq, "Loop detected: last one at %u "
1944                                "indirect size %u\n",
1945                                i, count);
1946                         return -EINVAL;
1947                 }
1948                 if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
1949                         vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1950                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1951                         return -EINVAL;
1952                 }
1953                 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
1954                         vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1955                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1956                         return -EINVAL;
1957                 }
1958
1959                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
1960                         access = VHOST_ACCESS_WO;
1961                 else
1962                         access = VHOST_ACCESS_RO;
1963
1964                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
1965                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
1966                                      iov_size - iov_count, access);
1967                 if (unlikely(ret < 0)) {
1968                         if (ret != -EAGAIN)
1969                                 vq_err(vq, "Translation failure %d indirect idx %d\n",
1970                                         ret, i);
1971                         return ret;
1972                 }
1973                 /* If this is an input descriptor, increment that count. */
1974                 if (access == VHOST_ACCESS_WO) {
1975                         *in_num += ret;
1976                         if (unlikely(log)) {
1977                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
1978                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
1979                                 ++*log_num;
1980                         }
1981                 } else {
1982                         /* If it's an output descriptor, they're all supposed
1983                          * to come before any input descriptors. */
1984                         if (unlikely(*in_num)) {
1985                                 vq_err(vq, "Indirect descriptor "
1986                                        "has out after in: idx %d\n", i);
1987                                 return -EINVAL;
1988                         }
1989                         *out_num += ret;
1990                 }
1991         } while ((i = next_desc(vq, &desc)) != -1);
1992         return 0;
1993 }
1994
1995 /* This looks in the virtqueue and for the first available buffer, and converts
1996  * it to an iovec for convenient access.  Since descriptors consist of some
1997  * number of output then some number of input descriptors, it's actually two
1998  * iovecs, but we pack them into one and note how many of each there were.
1999  *
2000  * This function returns the descriptor number found, or vq->num (which is
2001  * never a valid descriptor number) if none was found.  A negative code is
2002  * returned on error. */
2003 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
2004                       struct iovec iov[], unsigned int iov_size,
2005                       unsigned int *out_num, unsigned int *in_num,
2006                       struct vhost_log *log, unsigned int *log_num)
2007 {
2008         struct vring_desc desc;
2009         unsigned int i, head, found = 0;
2010         u16 last_avail_idx;
2011         __virtio16 avail_idx;
2012         __virtio16 ring_head;
2013         int ret, access;
2014
2015         /* Check it isn't doing very strange things with descriptor numbers. */
2016         last_avail_idx = vq->last_avail_idx;
2017
2018         if (vq->avail_idx == vq->last_avail_idx) {
2019                 if (unlikely(vhost_get_avail(vq, avail_idx, &vq->avail->idx))) {
2020                         vq_err(vq, "Failed to access avail idx at %p\n",
2021                                 &vq->avail->idx);
2022                         return -EFAULT;
2023                 }
2024                 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2025
2026                 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
2027                         vq_err(vq, "Guest moved used index from %u to %u",
2028                                 last_avail_idx, vq->avail_idx);
2029                         return -EFAULT;
2030                 }
2031
2032                 /* If there's nothing new since last we looked, return
2033                  * invalid.
2034                  */
2035                 if (vq->avail_idx == last_avail_idx)
2036                         return vq->num;
2037
2038                 /* Only get avail ring entries after they have been
2039                  * exposed by guest.
2040                  */
2041                 smp_rmb();
2042         }
2043
2044         /* Grab the next descriptor number they're advertising, and increment
2045          * the index we've seen. */
2046         if (unlikely(vhost_get_avail(vq, ring_head,
2047                      &vq->avail->ring[last_avail_idx & (vq->num - 1)]))) {
2048                 vq_err(vq, "Failed to read head: idx %d address %p\n",
2049                        last_avail_idx,
2050                        &vq->avail->ring[last_avail_idx % vq->num]);
2051                 return -EFAULT;
2052         }
2053
2054         head = vhost16_to_cpu(vq, ring_head);
2055
2056         /* If their number is silly, that's an error. */
2057         if (unlikely(head >= vq->num)) {
2058                 vq_err(vq, "Guest says index %u > %u is available",
2059                        head, vq->num);
2060                 return -EINVAL;
2061         }
2062
2063         /* When we start there are none of either input nor output. */
2064         *out_num = *in_num = 0;
2065         if (unlikely(log))
2066                 *log_num = 0;
2067
2068         i = head;
2069         do {
2070                 unsigned iov_count = *in_num + *out_num;
2071                 if (unlikely(i >= vq->num)) {
2072                         vq_err(vq, "Desc index is %u > %u, head = %u",
2073                                i, vq->num, head);
2074                         return -EINVAL;
2075                 }
2076                 if (unlikely(++found > vq->num)) {
2077                         vq_err(vq, "Loop detected: last one at %u "
2078                                "vq size %u head %u\n",
2079                                i, vq->num, head);
2080                         return -EINVAL;
2081                 }
2082                 ret = vhost_copy_from_user(vq, &desc, vq->desc + i,
2083                                            sizeof desc);
2084                 if (unlikely(ret)) {
2085                         vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
2086                                i, vq->desc + i);
2087                         return -EFAULT;
2088                 }
2089                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2090                         ret = get_indirect(vq, iov, iov_size,
2091                                            out_num, in_num,
2092                                            log, log_num, &desc);
2093                         if (unlikely(ret < 0)) {
2094                                 if (ret != -EAGAIN)
2095                                         vq_err(vq, "Failure detected "
2096                                                 "in indirect descriptor at idx %d\n", i);
2097                                 return ret;
2098                         }
2099                         continue;
2100                 }
2101
2102                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2103                         access = VHOST_ACCESS_WO;
2104                 else
2105                         access = VHOST_ACCESS_RO;
2106                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2107                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
2108                                      iov_size - iov_count, access);
2109                 if (unlikely(ret < 0)) {
2110                         if (ret != -EAGAIN)
2111                                 vq_err(vq, "Translation failure %d descriptor idx %d\n",
2112                                         ret, i);
2113                         return ret;
2114                 }
2115                 if (access == VHOST_ACCESS_WO) {
2116                         /* If this is an input descriptor,
2117                          * increment that count. */
2118                         *in_num += ret;
2119                         if (unlikely(log)) {
2120                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2121                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2122                                 ++*log_num;
2123                         }
2124                 } else {
2125                         /* If it's an output descriptor, they're all supposed
2126                          * to come before any input descriptors. */
2127                         if (unlikely(*in_num)) {
2128                                 vq_err(vq, "Descriptor has out after in: "
2129                                        "idx %d\n", i);
2130                                 return -EINVAL;
2131                         }
2132                         *out_num += ret;
2133                 }
2134         } while ((i = next_desc(vq, &desc)) != -1);
2135
2136         /* On success, increment avail index. */
2137         vq->last_avail_idx++;
2138
2139         /* Assume notifications from guest are disabled at this point,
2140          * if they aren't we would need to update avail_event index. */
2141         BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2142         return head;
2143 }
2144 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2145
2146 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2147 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2148 {
2149         vq->last_avail_idx -= n;
2150 }
2151 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2152
2153 /* After we've used one of their buffers, we tell them about it.  We'll then
2154  * want to notify the guest, using eventfd. */
2155 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2156 {
2157         struct vring_used_elem heads = {
2158                 cpu_to_vhost32(vq, head),
2159                 cpu_to_vhost32(vq, len)
2160         };
2161
2162         return vhost_add_used_n(vq, &heads, 1);
2163 }
2164 EXPORT_SYMBOL_GPL(vhost_add_used);
2165
2166 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2167                             struct vring_used_elem *heads,
2168                             unsigned count)
2169 {
2170         struct vring_used_elem __user *used;
2171         u16 old, new;
2172         int start;
2173
2174         start = vq->last_used_idx & (vq->num - 1);
2175         used = vq->used->ring + start;
2176         if (count == 1) {
2177                 if (vhost_put_user(vq, heads[0].id, &used->id)) {
2178                         vq_err(vq, "Failed to write used id");
2179                         return -EFAULT;
2180                 }
2181                 if (vhost_put_user(vq, heads[0].len, &used->len)) {
2182                         vq_err(vq, "Failed to write used len");
2183                         return -EFAULT;
2184                 }
2185         } else if (vhost_copy_to_user(vq, used, heads, count * sizeof *used)) {
2186                 vq_err(vq, "Failed to write used");
2187                 return -EFAULT;
2188         }
2189         if (unlikely(vq->log_used)) {
2190                 /* Make sure data is seen before log. */
2191                 smp_wmb();
2192                 /* Log used ring entry write. */
2193                 log_write(vq->log_base,
2194                           vq->log_addr +
2195                            ((void __user *)used - (void __user *)vq->used),
2196                           count * sizeof *used);
2197         }
2198         old = vq->last_used_idx;
2199         new = (vq->last_used_idx += count);
2200         /* If the driver never bothers to signal in a very long while,
2201          * used index might wrap around. If that happens, invalidate
2202          * signalled_used index we stored. TODO: make sure driver
2203          * signals at least once in 2^16 and remove this. */
2204         if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2205                 vq->signalled_used_valid = false;
2206         return 0;
2207 }
2208
2209 /* After we've used one of their buffers, we tell them about it.  We'll then
2210  * want to notify the guest, using eventfd. */
2211 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2212                      unsigned count)
2213 {
2214         int start, n, r;
2215
2216         start = vq->last_used_idx & (vq->num - 1);
2217         n = vq->num - start;
2218         if (n < count) {
2219                 r = __vhost_add_used_n(vq, heads, n);
2220                 if (r < 0)
2221                         return r;
2222                 heads += n;
2223                 count -= n;
2224         }
2225         r = __vhost_add_used_n(vq, heads, count);
2226
2227         /* Make sure buffer is written before we update index. */
2228         smp_wmb();
2229         if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
2230                            &vq->used->idx)) {
2231                 vq_err(vq, "Failed to increment used idx");
2232                 return -EFAULT;
2233         }
2234         if (unlikely(vq->log_used)) {
2235                 /* Log used index update. */
2236                 log_write(vq->log_base,
2237                           vq->log_addr + offsetof(struct vring_used, idx),
2238                           sizeof vq->used->idx);
2239                 if (vq->log_ctx)
2240                         eventfd_signal(vq->log_ctx, 1);
2241         }
2242         return r;
2243 }
2244 EXPORT_SYMBOL_GPL(vhost_add_used_n);
2245
2246 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2247 {
2248         __u16 old, new;
2249         __virtio16 event;
2250         bool v;
2251
2252         if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2253             unlikely(vq->avail_idx == vq->last_avail_idx))
2254                 return true;
2255
2256         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2257                 __virtio16 flags;
2258                 /* Flush out used index updates. This is paired
2259                  * with the barrier that the Guest executes when enabling
2260                  * interrupts. */
2261                 smp_mb();
2262                 if (vhost_get_avail(vq, flags, &vq->avail->flags)) {
2263                         vq_err(vq, "Failed to get flags");
2264                         return true;
2265                 }
2266                 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2267         }
2268         old = vq->signalled_used;
2269         v = vq->signalled_used_valid;
2270         new = vq->signalled_used = vq->last_used_idx;
2271         vq->signalled_used_valid = true;
2272
2273         if (unlikely(!v))
2274                 return true;
2275
2276         /* We're sure if the following conditions are met, there's no
2277          * need to notify guest:
2278          * 1) cached used event is ahead of new
2279          * 2) old to new updating does not cross cached used event. */
2280         if (vring_need_event(vq->last_used_event, new + vq->num, new) &&
2281             !vring_need_event(vq->last_used_event, new, old))
2282                 return false;
2283
2284         /* Flush out used index updates. This is paired
2285          * with the barrier that the Guest executes when enabling
2286          * interrupts. */
2287         smp_mb();
2288
2289         if (vhost_get_avail(vq, event, vhost_used_event(vq))) {
2290                 vq_err(vq, "Failed to get used event idx");
2291                 return true;
2292         }
2293         vq->last_used_event = vhost16_to_cpu(vq, event);
2294
2295         return vring_need_event(vq->last_used_event, new, old);
2296 }
2297
2298 /* This actually signals the guest, using eventfd. */
2299 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2300 {
2301         /* Signal the Guest tell them we used something up. */
2302         if (vq->call_ctx && vhost_notify(dev, vq))
2303                 eventfd_signal(vq->call_ctx, 1);
2304 }
2305 EXPORT_SYMBOL_GPL(vhost_signal);
2306
2307 /* And here's the combo meal deal.  Supersize me! */
2308 void vhost_add_used_and_signal(struct vhost_dev *dev,
2309                                struct vhost_virtqueue *vq,
2310                                unsigned int head, int len)
2311 {
2312         vhost_add_used(vq, head, len);
2313         vhost_signal(dev, vq);
2314 }
2315 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2316
2317 /* multi-buffer version of vhost_add_used_and_signal */
2318 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2319                                  struct vhost_virtqueue *vq,
2320                                  struct vring_used_elem *heads, unsigned count)
2321 {
2322         vhost_add_used_n(vq, heads, count);
2323         vhost_signal(dev, vq);
2324 }
2325 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2326
2327 /* return true if we're sure that avaiable ring is empty */
2328 bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2329 {
2330         __virtio16 avail_idx;
2331         int r;
2332
2333         if (vq->avail_idx != vq->last_avail_idx)
2334                 return false;
2335
2336         r = vhost_get_avail(vq, avail_idx, &vq->avail->idx);
2337         if (unlikely(r))
2338                 return false;
2339         vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2340
2341         return vq->avail_idx == vq->last_avail_idx;
2342 }
2343 EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2344
2345 /* OK, now we need to know about added descriptors. */
2346 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2347 {
2348         __virtio16 avail_idx;
2349         int r;
2350
2351         if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2352                 return false;
2353         vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2354         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2355                 r = vhost_update_used_flags(vq);
2356                 if (r) {
2357                         vq_err(vq, "Failed to enable notification at %p: %d\n",
2358                                &vq->used->flags, r);
2359                         return false;
2360                 }
2361         } else {
2362                 r = vhost_update_avail_event(vq, vq->avail_idx);
2363                 if (r) {
2364                         vq_err(vq, "Failed to update avail event index at %p: %d\n",
2365                                vhost_avail_event(vq), r);
2366                         return false;
2367                 }
2368         }
2369         /* They could have slipped one in as we were doing that: make
2370          * sure it's written, then check again. */
2371         smp_mb();
2372         r = vhost_get_avail(vq, avail_idx, &vq->avail->idx);
2373         if (r) {
2374                 vq_err(vq, "Failed to check avail idx at %p: %d\n",
2375                        &vq->avail->idx, r);
2376                 return false;
2377         }
2378
2379         return vhost16_to_cpu(vq, avail_idx) != vq->avail_idx;
2380 }
2381 EXPORT_SYMBOL_GPL(vhost_enable_notify);
2382
2383 /* We don't need to be notified again. */
2384 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2385 {
2386         int r;
2387
2388         if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2389                 return;
2390         vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2391         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2392                 r = vhost_update_used_flags(vq);
2393                 if (r)
2394                         vq_err(vq, "Failed to enable notification at %p: %d\n",
2395                                &vq->used->flags, r);
2396         }
2397 }
2398 EXPORT_SYMBOL_GPL(vhost_disable_notify);
2399
2400 /* Create a new message. */
2401 struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2402 {
2403         struct vhost_msg_node *node = kmalloc(sizeof *node, GFP_KERNEL);
2404         if (!node)
2405                 return NULL;
2406         node->vq = vq;
2407         node->msg.type = type;
2408         return node;
2409 }
2410 EXPORT_SYMBOL_GPL(vhost_new_msg);
2411
2412 void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2413                        struct vhost_msg_node *node)
2414 {
2415         spin_lock(&dev->iotlb_lock);
2416         list_add_tail(&node->node, head);
2417         spin_unlock(&dev->iotlb_lock);
2418
2419         wake_up_interruptible_poll(&dev->wait, POLLIN | POLLRDNORM);
2420 }
2421 EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2422
2423 struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2424                                          struct list_head *head)
2425 {
2426         struct vhost_msg_node *node = NULL;
2427
2428         spin_lock(&dev->iotlb_lock);
2429         if (!list_empty(head)) {
2430                 node = list_first_entry(head, struct vhost_msg_node,
2431                                         node);
2432                 list_del(&node->node);
2433         }
2434         spin_unlock(&dev->iotlb_lock);
2435
2436         return node;
2437 }
2438 EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2439
2440
2441 static int __init vhost_init(void)
2442 {
2443         return 0;
2444 }
2445
2446 static void __exit vhost_exit(void)
2447 {
2448 }
2449
2450 module_init(vhost_init);
2451 module_exit(vhost_exit);
2452
2453 MODULE_VERSION("0.0.1");
2454 MODULE_LICENSE("GPL v2");
2455 MODULE_AUTHOR("Michael S. Tsirkin");
2456 MODULE_DESCRIPTION("Host kernel accelerator for virtio");