drivers/vhost/vhost.c

   1 /* Copyright (C) 2009 Red Hat, Inc.
   2  * Copyright (C) 2006 Rusty Russell IBM Corporation
   3  *
   4  * Author: Michael S. Tsirkin <mst@redhat.com>
   5  *
   6  * Inspiration, some code, and most witty comments come from
   7  * Documentation/virtual/lguest/lguest.c, by Rusty Russell
   8  *
   9  * This work is licensed under the terms of the GNU GPL, version 2.
  10  *
  11  * Generic code for virtio server in host kernel.
  12  */
  13
  14 #include <linux/eventfd.h>
  15 #include <linux/vhost.h>
  16 #include <linux/uio.h>
  17 #include <linux/mm.h>
  18 #include <linux/mmu_context.h>
  19 #include <linux/miscdevice.h>
  20 #include <linux/mutex.h>
  21 #include <linux/poll.h>
  22 #include <linux/file.h>
  23 #include <linux/highmem.h>
  24 #include <linux/slab.h>
  25 #include <linux/vmalloc.h>
  26 #include <linux/kthread.h>
  27 #include <linux/cgroup.h>
  28 #include <linux/module.h>
  29 #include <linux/sort.h>
  30 #include <linux/sched/mm.h>
  31 #include <linux/sched/signal.h>
  32 #include <linux/interval_tree_generic.h>
  33
  34 #include "vhost.h"
  35
  36 static ushort max_mem_regions = 64;
  37 module_param(max_mem_regions, ushort, 0444);
  38 MODULE_PARM_DESC(max_mem_regions,
  39         "Maximum number of memory regions in memory map. (default: 64)");
  40 static int max_iotlb_entries = 2048;
  41 module_param(max_iotlb_entries, int, 0444);
  42 MODULE_PARM_DESC(max_iotlb_entries,
  43         "Maximum number of iotlb entries. (default: 2048)");
  44
  45 enum {
  46         VHOST_MEMORY_F_LOG = 0x1,
  47 };
  48
  49 #define vhost_used_event(vq) ((__virtio16 __user *)&vq->avail->ring[vq->num])
  50 #define vhost_avail_event(vq) ((__virtio16 __user *)&vq->used->ring[vq->num])
  51
  52 INTERVAL_TREE_DEFINE(struct vhost_umem_node,
  53                      rb, __u64, __subtree_last,
  54                      START, LAST, static inline, vhost_umem_interval_tree);
  55
  56 #ifdef CONFIG_VHOST_CROSS_ENDIAN_LEGACY
  57 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
  58 {
  59         vq->user_be = !virtio_legacy_is_little_endian();
  60 }
  61
  62 static void vhost_enable_cross_endian_big(struct vhost_virtqueue *vq)
  63 {
  64         vq->user_be = true;
  65 }
  66
  67 static void vhost_enable_cross_endian_little(struct vhost_virtqueue *vq)
  68 {
  69         vq->user_be = false;
  70 }
  71
  72 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
  73 {
  74         struct vhost_vring_state s;
  75
  76         if (vq->private_data)
  77                 return -EBUSY;
  78
  79         if (copy_from_user(&s, argp, sizeof(s)))
  80                 return -EFAULT;
  81
  82         if (s.num != VHOST_VRING_LITTLE_ENDIAN &&
  83             s.num != VHOST_VRING_BIG_ENDIAN)
  84                 return -EINVAL;
  85
  86         if (s.num == VHOST_VRING_BIG_ENDIAN)
  87                 vhost_enable_cross_endian_big(vq);
  88         else
  89                 vhost_enable_cross_endian_little(vq);
  90
  91         return 0;
  92 }
  93
  94 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
  95                                    int __user *argp)
  96 {
  97         struct vhost_vring_state s = {
  98                 .index = idx,
  99                 .num = vq->user_be
 100         };
 101
 102         if (copy_to_user(argp, &s, sizeof(s)))
 103                 return -EFAULT;
 104
 105         return 0;
 106 }
 107
 108 static void vhost_init_is_le(struct vhost_virtqueue *vq)
 109 {
 110         /* Note for legacy virtio: user_be is initialized at reset time
 111          * according to the host endianness. If userspace does not set an
 112          * explicit endianness, the default behavior is native endian, as
 113          * expected by legacy virtio.
 114          */
 115         vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1) || !vq->user_be;
 116 }
 117 #else
 118 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
 119 {
 120 }
 121
 122 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
 123 {
 124         return -ENOIOCTLCMD;
 125 }
 126
 127 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
 128                                    int __user *argp)
 129 {
 130         return -ENOIOCTLCMD;
 131 }
 132
 133 static void vhost_init_is_le(struct vhost_virtqueue *vq)
 134 {
 135         vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1)
 136                 || virtio_legacy_is_little_endian();
 137 }
 138 #endif /* CONFIG_VHOST_CROSS_ENDIAN_LEGACY */
 139
 140 static void vhost_reset_is_le(struct vhost_virtqueue *vq)
 141 {
 142         vhost_init_is_le(vq);
 143 }
 144
 145 struct vhost_flush_struct {
 146         struct vhost_work work;
 147         struct completion wait_event;
 148 };
 149
 150 static void vhost_flush_work(struct vhost_work *work)
 151 {
 152         struct vhost_flush_struct *s;
 153
 154         s = container_of(work, struct vhost_flush_struct, work);
 155         complete(&s->wait_event);
 156 }
 157
 158 static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
 159                             poll_table *pt)
 160 {
 161         struct vhost_poll *poll;
 162
 163         poll = container_of(pt, struct vhost_poll, table);
 164         poll->wqh = wqh;
 165         add_wait_queue(wqh, &poll->wait);
 166 }
 167
 168 static int vhost_poll_wakeup(wait_queue_entry_t *wait, unsigned mode, int sync,
 169                              void *key)
 170 {
 171         struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
 172
 173         if (!(key_to_poll(key) & poll->mask))
 174                 return 0;
 175
 176         vhost_poll_queue(poll);
 177         return 0;
 178 }
 179
 180 void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
 181 {
 182         clear_bit(VHOST_WORK_QUEUED, &work->flags);
 183         work->fn = fn;
 184 }
 185 EXPORT_SYMBOL_GPL(vhost_work_init);
 186
 187 /* Init poll structure */
 188 void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
 189                      __poll_t mask, struct vhost_dev *dev)
 190 {
 191         init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
 192         init_poll_funcptr(&poll->table, vhost_poll_func);
 193         poll->mask = mask;
 194         poll->dev = dev;
 195         poll->wqh = NULL;
 196
 197         vhost_work_init(&poll->work, fn);
 198 }
 199 EXPORT_SYMBOL_GPL(vhost_poll_init);
 200
 201 /* Start polling a file. We add ourselves to file's wait queue. The caller must
 202  * keep a reference to a file until after vhost_poll_stop is called. */
 203 int vhost_poll_start(struct vhost_poll *poll, struct file *file)
 204 {
 205         __poll_t mask;
 206         int ret = 0;
 207
 208         if (poll->wqh)
 209                 return 0;
 210
 211         mask = vfs_poll(file, &poll->table);
 212         if (mask)
 213                 vhost_poll_wakeup(&poll->wait, 0, 0, poll_to_key(mask));
 214         if (mask & EPOLLERR) {
 215                 vhost_poll_stop(poll);
 216                 ret = -EINVAL;
 217         }
 218
 219         return ret;
 220 }
 221 EXPORT_SYMBOL_GPL(vhost_poll_start);
 222
 223 /* Stop polling a file. After this function returns, it becomes safe to drop the
 224  * file reference. You must also flush afterwards. */
 225 void vhost_poll_stop(struct vhost_poll *poll)
 226 {
 227         if (poll->wqh) {
 228                 remove_wait_queue(poll->wqh, &poll->wait);
 229                 poll->wqh = NULL;
 230         }
 231 }
 232 EXPORT_SYMBOL_GPL(vhost_poll_stop);
 233
 234 void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work)
 235 {
 236         struct vhost_flush_struct flush;
 237
 238         if (dev->worker) {
 239                 init_completion(&flush.wait_event);
 240                 vhost_work_init(&flush.work, vhost_flush_work);
 241
 242                 vhost_work_queue(dev, &flush.work);
 243                 wait_for_completion(&flush.wait_event);
 244         }
 245 }
 246 EXPORT_SYMBOL_GPL(vhost_work_flush);
 247
 248 /* Flush any work that has been scheduled. When calling this, don't hold any
 249  * locks that are also used by the callback. */
 250 void vhost_poll_flush(struct vhost_poll *poll)
 251 {
 252         vhost_work_flush(poll->dev, &poll->work);
 253 }
 254 EXPORT_SYMBOL_GPL(vhost_poll_flush);
 255
 256 void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work)
 257 {
 258         if (!dev->worker)
 259                 return;
 260
 261         if (!test_and_set_bit(VHOST_WORK_QUEUED, &work->flags)) {
 262                 /* We can only add the work to the list after we're
 263                  * sure it was not in the list.
 264                  * test_and_set_bit() implies a memory barrier.
 265                  */
 266                 llist_add(&work->node, &dev->work_list);
 267                 wake_up_process(dev->worker);
 268         }
 269 }
 270 EXPORT_SYMBOL_GPL(vhost_work_queue);
 271
 272 /* A lockless hint for busy polling code to exit the loop */
 273 bool vhost_has_work(struct vhost_dev *dev)
 274 {
 275         return !llist_empty(&dev->work_list);
 276 }
 277 EXPORT_SYMBOL_GPL(vhost_has_work);
 278
 279 void vhost_poll_queue(struct vhost_poll *poll)
 280 {
 281         vhost_work_queue(poll->dev, &poll->work);
 282 }
 283 EXPORT_SYMBOL_GPL(vhost_poll_queue);
 284
 285 static void __vhost_vq_meta_reset(struct vhost_virtqueue *vq)
 286 {
 287         int j;
 288
 289         for (j = 0; j < VHOST_NUM_ADDRS; j++)
 290                 vq->meta_iotlb[j] = NULL;
 291 }
 292
 293 static void vhost_vq_meta_reset(struct vhost_dev *d)
 294 {
 295         int i;
 296
 297         for (i = 0; i < d->nvqs; ++i)
 298                 __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->avail_idx = 0;
 310         vq->last_used_idx = 0;
 311         vq->signalled_used = 0;
 312         vq->signalled_used_valid = false;
 313         vq->used_flags = 0;
 314         vq->log_used = false;
 315         vq->log_addr = -1ull;
 316         vq->private_data = NULL;
 317         vq->acked_features = 0;
 318         vq->log_base = NULL;
 319         vq->error_ctx = NULL;
 320         vq->kick = NULL;
 321         vq->call_ctx = NULL;
 322         vq->log_ctx = NULL;
 323         vhost_reset_is_le(vq);
 324         vhost_disable_cross_endian(vq);
 325         vq->busyloop_timeout = 0;
 326         vq->umem = NULL;
 327         vq->iotlb = NULL;
 328         __vhost_vq_meta_reset(vq);
 329 }
 330
 331 static int vhost_worker(void *data)
 332 {
 333         struct vhost_dev *dev = data;
 334         struct vhost_work *work, *work_next;
 335         struct llist_node *node;
 336         mm_segment_t oldfs = get_fs();
 337
 338         set_fs(USER_DS);
 339         use_mm(dev->mm);
 340
 341         for (;;) {
 342                 /* mb paired w/ kthread_stop */
 343                 set_current_state(TASK_INTERRUPTIBLE);
 344
 345                 if (kthread_should_stop()) {
 346                         __set_current_state(TASK_RUNNING);
 347                         break;
 348                 }
 349
 350                 node = llist_del_all(&dev->work_list);
 351                 if (!node)
 352                         schedule();
 353
 354                 node = llist_reverse_order(node);
 355                 /* make sure flag is seen after deletion */
 356                 smp_wmb();
 357                 llist_for_each_entry_safe(work, work_next, node, node) {
 358                         clear_bit(VHOST_WORK_QUEUED, &work->flags);
 359                         __set_current_state(TASK_RUNNING);
 360                         work->fn(work);
 361                         if (need_resched())
 362                                 schedule();
 363                 }
 364         }
 365         unuse_mm(dev->mm);
 366         set_fs(oldfs);
 367         return 0;
 368 }
 369
 370 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
 371 {
 372         kfree(vq->indirect);
 373         vq->indirect = NULL;
 374         kfree(vq->log);
 375         vq->log = NULL;
 376         kfree(vq->heads);
 377         vq->heads = NULL;
 378 }
 379
 380 /* Helper to allocate iovec buffers for all vqs. */
 381 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
 382 {
 383         struct vhost_virtqueue *vq;
 384         int i;
 385
 386         for (i = 0; i < dev->nvqs; ++i) {
 387                 vq = dev->vqs[i];
 388                 vq->indirect = kmalloc_array(UIO_MAXIOV,
 389                                              sizeof(*vq->indirect),
 390                                              GFP_KERNEL);
 391                 vq->log = kmalloc_array(UIO_MAXIOV, sizeof(*vq->log),
 392                                         GFP_KERNEL);
 393                 vq->heads = kmalloc_array(UIO_MAXIOV, sizeof(*vq->heads),
 394                                           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->umem = NULL;
 425         dev->iotlb = NULL;
 426         dev->mm = NULL;
 427         dev->worker = NULL;
 428         init_llist_head(&dev->work_list);
 429         init_waitqueue_head(&dev->wait);
 430         INIT_LIST_HEAD(&dev->read_list);
 431         INIT_LIST_HEAD(&dev->pending_list);
 432         spin_lock_init(&dev->iotlb_lock);
 433
 434
 435         for (i = 0; i < dev->nvqs; ++i) {
 436                 vq = dev->vqs[i];
 437                 vq->log = NULL;
 438                 vq->indirect = NULL;
 439                 vq->heads = NULL;
 440                 vq->dev = dev;
 441                 mutex_init(&vq->mutex);
 442                 vhost_vq_reset(dev, vq);
 443                 if (vq->handle_kick)
 444                         vhost_poll_init(&vq->poll, vq->handle_kick,
 445                                         EPOLLIN, dev);
 446         }
 447 }
 448 EXPORT_SYMBOL_GPL(vhost_dev_init);
 449
 450 /* Caller should have device mutex */
 451 long vhost_dev_check_owner(struct vhost_dev *dev)
 452 {
 453         /* Are you the owner? If not, I don't think you mean to do that */
 454         return dev->mm == current->mm ? 0 : -EPERM;
 455 }
 456 EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
 457
 458 struct vhost_attach_cgroups_struct {
 459         struct vhost_work work;
 460         struct task_struct *owner;
 461         int ret;
 462 };
 463
 464 static void vhost_attach_cgroups_work(struct vhost_work *work)
 465 {
 466         struct vhost_attach_cgroups_struct *s;
 467
 468         s = container_of(work, struct vhost_attach_cgroups_struct, work);
 469         s->ret = cgroup_attach_task_all(s->owner, current);
 470 }
 471
 472 static int vhost_attach_cgroups(struct vhost_dev *dev)
 473 {
 474         struct vhost_attach_cgroups_struct attach;
 475
 476         attach.owner = current;
 477         vhost_work_init(&attach.work, vhost_attach_cgroups_work);
 478         vhost_work_queue(dev, &attach.work);
 479         vhost_work_flush(dev, &attach.work);
 480         return attach.ret;
 481 }
 482
 483 /* Caller should have device mutex */
 484 bool vhost_dev_has_owner(struct vhost_dev *dev)
 485 {
 486         return dev->mm;
 487 }
 488 EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
 489
 490 /* Caller should have device mutex */
 491 long vhost_dev_set_owner(struct vhost_dev *dev)
 492 {
 493         struct task_struct *worker;
 494         int err;
 495
 496         /* Is there an owner already? */
 497         if (vhost_dev_has_owner(dev)) {
 498                 err = -EBUSY;
 499                 goto err_mm;
 500         }
 501
 502         /* No owner, become one */
 503         dev->mm = get_task_mm(current);
 504         worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
 505         if (IS_ERR(worker)) {
 506                 err = PTR_ERR(worker);
 507                 goto err_worker;
 508         }
 509
 510         dev->worker = worker;
 511         wake_up_process(worker);        /* avoid contributing to loadavg */
 512
 513         err = vhost_attach_cgroups(dev);
 514         if (err)
 515                 goto err_cgroup;
 516
 517         err = vhost_dev_alloc_iovecs(dev);
 518         if (err)
 519                 goto err_cgroup;
 520
 521         return 0;
 522 err_cgroup:
 523         kthread_stop(worker);
 524         dev->worker = NULL;
 525 err_worker:
 526         if (dev->mm)
 527                 mmput(dev->mm);
 528         dev->mm = NULL;
 529 err_mm:
 530         return err;
 531 }
 532 EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
 533
 534 struct vhost_umem *vhost_dev_reset_owner_prepare(void)
 535 {
 536         return kvzalloc(sizeof(struct vhost_umem), GFP_KERNEL);
 537 }
 538 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
 539
 540 /* Caller should have device mutex */
 541 void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_umem *umem)
 542 {
 543         int i;
 544
 545         vhost_dev_cleanup(dev);
 546
 547         /* Restore memory to default empty mapping. */
 548         INIT_LIST_HEAD(&umem->umem_list);
 549         dev->umem = umem;
 550         /* We don't need VQ locks below since vhost_dev_cleanup makes sure
 551          * VQs aren't running.
 552          */
 553         for (i = 0; i < dev->nvqs; ++i)
 554                 dev->vqs[i]->umem = umem;
 555 }
 556 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
 557
 558 void vhost_dev_stop(struct vhost_dev *dev)
 559 {
 560         int i;
 561
 562         for (i = 0; i < dev->nvqs; ++i) {
 563                 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick) {
 564                         vhost_poll_stop(&dev->vqs[i]->poll);
 565                         vhost_poll_flush(&dev->vqs[i]->poll);
 566                 }
 567         }
 568 }
 569 EXPORT_SYMBOL_GPL(vhost_dev_stop);
 570
 571 static void vhost_umem_free(struct vhost_umem *umem,
 572                             struct vhost_umem_node *node)
 573 {
 574         vhost_umem_interval_tree_remove(node, &umem->umem_tree);
 575         list_del(&node->link);
 576         kfree(node);
 577         umem->numem--;
 578 }
 579
 580 static void vhost_umem_clean(struct vhost_umem *umem)
 581 {
 582         struct vhost_umem_node *node, *tmp;
 583
 584         if (!umem)
 585                 return;
 586
 587         list_for_each_entry_safe(node, tmp, &umem->umem_list, link)
 588                 vhost_umem_free(umem, node);
 589
 590         kvfree(umem);
 591 }
 592
 593 static void vhost_clear_msg(struct vhost_dev *dev)
 594 {
 595         struct vhost_msg_node *node, *n;
 596
 597         spin_lock(&dev->iotlb_lock);
 598
 599         list_for_each_entry_safe(node, n, &dev->read_list, node) {
 600                 list_del(&node->node);
 601                 kfree(node);
 602         }
 603
 604         list_for_each_entry_safe(node, n, &dev->pending_list, node) {
 605                 list_del(&node->node);
 606                 kfree(node);
 607         }
 608
 609         spin_unlock(&dev->iotlb_lock);
 610 }
 611
 612 void vhost_dev_cleanup(struct vhost_dev *dev)
 613 {
 614         int i;
 615
 616         for (i = 0; i < dev->nvqs; ++i) {
 617                 if (dev->vqs[i]->error_ctx)
 618                         eventfd_ctx_put(dev->vqs[i]->error_ctx);
 619                 if (dev->vqs[i]->kick)
 620                         fput(dev->vqs[i]->kick);
 621                 if (dev->vqs[i]->call_ctx)
 622                         eventfd_ctx_put(dev->vqs[i]->call_ctx);
 623                 vhost_vq_reset(dev, dev->vqs[i]);
 624         }
 625         vhost_dev_free_iovecs(dev);
 626         if (dev->log_ctx)
 627                 eventfd_ctx_put(dev->log_ctx);
 628         dev->log_ctx = NULL;
 629         /* No one will access memory at this point */
 630         vhost_umem_clean(dev->umem);
 631         dev->umem = NULL;
 632         vhost_umem_clean(dev->iotlb);
 633         dev->iotlb = NULL;
 634         vhost_clear_msg(dev);
 635         wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
 636         WARN_ON(!llist_empty(&dev->work_list));
 637         if (dev->worker) {
 638                 kthread_stop(dev->worker);
 639                 dev->worker = NULL;
 640         }
 641         if (dev->mm)
 642                 mmput(dev->mm);
 643         dev->mm = NULL;
 644 }
 645 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
 646
 647 static bool log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
 648 {
 649         u64 a = addr / VHOST_PAGE_SIZE / 8;
 650
 651         /* Make sure 64 bit math will not overflow. */
 652         if (a > ULONG_MAX - (unsigned long)log_base ||
 653             a + (unsigned long)log_base > ULONG_MAX)
 654                 return false;
 655
 656         return access_ok(VERIFY_WRITE, log_base + a,
 657                          (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
 658 }
 659
 660 static bool vhost_overflow(u64 uaddr, u64 size)
 661 {
 662         /* Make sure 64 bit math will not overflow. */
 663         return uaddr > ULONG_MAX || size > ULONG_MAX || uaddr > ULONG_MAX - size;
 664 }
 665
 666 /* Caller should have vq mutex and device mutex. */
 667 static bool vq_memory_access_ok(void __user *log_base, struct vhost_umem *umem,
 668                                 int log_all)
 669 {
 670         struct vhost_umem_node *node;
 671
 672         if (!umem)
 673                 return false;
 674
 675         list_for_each_entry(node, &umem->umem_list, link) {
 676                 unsigned long a = node->userspace_addr;
 677
 678                 if (vhost_overflow(node->userspace_addr, node->size))
 679                         return false;
 680
 681
 682                 if (!access_ok(VERIFY_WRITE, (void __user *)a,
 683                                     node->size))
 684                         return false;
 685                 else if (log_all && !log_access_ok(log_base,
 686                                                    node->start,
 687                                                    node->size))
 688                         return false;
 689         }
 690         return true;
 691 }
 692
 693 static inline void __user *vhost_vq_meta_fetch(struct vhost_virtqueue *vq,
 694                                                u64 addr, unsigned int size,
 695                                                int type)
 696 {
 697         const struct vhost_umem_node *node = vq->meta_iotlb[type];
 698
 699         if (!node)
 700                 return NULL;
 701
 702         return (void *)(uintptr_t)(node->userspace_addr + addr - node->start);
 703 }
 704
 705 /* Can we switch to this memory table? */
 706 /* Caller should have device mutex but not vq mutex */
 707 static bool memory_access_ok(struct vhost_dev *d, struct vhost_umem *umem,
 708                              int log_all)
 709 {
 710         int i;
 711
 712         for (i = 0; i < d->nvqs; ++i) {
 713                 bool ok;
 714                 bool log;
 715
 716                 mutex_lock(&d->vqs[i]->mutex);
 717                 log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
 718                 /* If ring is inactive, will check when it's enabled. */
 719                 if (d->vqs[i]->private_data)
 720                         ok = vq_memory_access_ok(d->vqs[i]->log_base,
 721                                                  umem, log);
 722                 else
 723                         ok = true;
 724                 mutex_unlock(&d->vqs[i]->mutex);
 725                 if (!ok)
 726                         return false;
 727         }
 728         return true;
 729 }
 730
 731 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
 732                           struct iovec iov[], int iov_size, int access);
 733
 734 static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
 735                               const void *from, unsigned size)
 736 {
 737         int ret;
 738
 739         if (!vq->iotlb)
 740                 return __copy_to_user(to, from, size);
 741         else {
 742                 /* This function should be called after iotlb
 743                  * prefetch, which means we're sure that all vq
 744                  * could be access through iotlb. So -EAGAIN should
 745                  * not happen in this case.
 746                  */
 747                 struct iov_iter t;
 748                 void __user *uaddr = vhost_vq_meta_fetch(vq,
 749                                      (u64)(uintptr_t)to, size,
 750                                      VHOST_ADDR_USED);
 751
 752                 if (uaddr)
 753                         return __copy_to_user(uaddr, from, size);
 754
 755                 ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
 756                                      ARRAY_SIZE(vq->iotlb_iov),
 757                                      VHOST_ACCESS_WO);
 758                 if (ret < 0)
 759                         goto out;
 760                 iov_iter_init(&t, WRITE, vq->iotlb_iov, ret, size);
 761                 ret = copy_to_iter(from, size, &t);
 762                 if (ret == size)
 763                         ret = 0;
 764         }
 765 out:
 766         return ret;
 767 }
 768
 769 static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
 770                                 void __user *from, unsigned size)
 771 {
 772         int ret;
 773
 774         if (!vq->iotlb)
 775                 return __copy_from_user(to, from, size);
 776         else {
 777                 /* This function should be called after iotlb
 778                  * prefetch, which means we're sure that vq
 779                  * could be access through iotlb. So -EAGAIN should
 780                  * not happen in this case.
 781                  */
 782                 void __user *uaddr = vhost_vq_meta_fetch(vq,
 783                                      (u64)(uintptr_t)from, size,
 784                                      VHOST_ADDR_DESC);
 785                 struct iov_iter f;
 786
 787                 if (uaddr)
 788                         return __copy_from_user(to, uaddr, size);
 789
 790                 ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
 791                                      ARRAY_SIZE(vq->iotlb_iov),
 792                                      VHOST_ACCESS_RO);
 793                 if (ret < 0) {
 794                         vq_err(vq, "IOTLB translation failure: uaddr "
 795                                "%p size 0x%llx\n", from,
 796                                (unsigned long long) size);
 797                         goto out;
 798                 }
 799                 iov_iter_init(&f, READ, vq->iotlb_iov, ret, size);
 800                 ret = copy_from_iter(to, size, &f);
 801                 if (ret == size)
 802                         ret = 0;
 803         }
 804
 805 out:
 806         return ret;
 807 }
 808
 809 static void __user *__vhost_get_user_slow(struct vhost_virtqueue *vq,
 810                                           void __user *addr, unsigned int size,
 811                                           int type)
 812 {
 813         int ret;
 814
 815         ret = translate_desc(vq, (u64)(uintptr_t)addr, size, vq->iotlb_iov,
 816                              ARRAY_SIZE(vq->iotlb_iov),
 817                              VHOST_ACCESS_RO);
 818         if (ret < 0) {
 819                 vq_err(vq, "IOTLB translation failure: uaddr "
 820                         "%p size 0x%llx\n", addr,
 821                         (unsigned long long) size);
 822                 return NULL;
 823         }
 824
 825         if (ret != 1 || vq->iotlb_iov[0].iov_len != size) {
 826                 vq_err(vq, "Non atomic userspace memory access: uaddr "
 827                         "%p size 0x%llx\n", addr,
 828                         (unsigned long long) size);
 829                 return NULL;
 830         }
 831
 832         return vq->iotlb_iov[0].iov_base;
 833 }
 834
 835 /* This function should be called after iotlb
 836  * prefetch, which means we're sure that vq
 837  * could be access through iotlb. So -EAGAIN should
 838  * not happen in this case.
 839  */
 840 static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
 841                                             void *addr, unsigned int size,
 842                                             int type)
 843 {
 844         void __user *uaddr = vhost_vq_meta_fetch(vq,
 845                              (u64)(uintptr_t)addr, size, type);
 846         if (uaddr)
 847                 return uaddr;
 848
 849         return __vhost_get_user_slow(vq, addr, size, type);
 850 }
 851
 852 #define vhost_put_user(vq, x, ptr)              \
 853 ({ \
 854         int ret = -EFAULT; \
 855         if (!vq->iotlb) { \
 856                 ret = __put_user(x, ptr); \
 857         } else { \
 858                 __typeof__(ptr) to = \
 859                         (__typeof__(ptr)) __vhost_get_user(vq, ptr,     \
 860                                           sizeof(*ptr), VHOST_ADDR_USED); \
 861                 if (to != NULL) \
 862                         ret = __put_user(x, to); \
 863                 else \
 864                         ret = -EFAULT;  \
 865         } \
 866         ret; \
 867 })
 868
 869 #define vhost_get_user(vq, x, ptr, type)                \
 870 ({ \
 871         int ret; \
 872         if (!vq->iotlb) { \
 873                 ret = __get_user(x, ptr); \
 874         } else { \
 875                 __typeof__(ptr) from = \
 876                         (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
 877                                                            sizeof(*ptr), \
 878                                                            type); \
 879                 if (from != NULL) \
 880                         ret = __get_user(x, from); \
 881                 else \
 882                         ret = -EFAULT; \
 883         } \
 884         ret; \
 885 })
 886
 887 #define vhost_get_avail(vq, x, ptr) \
 888         vhost_get_user(vq, x, ptr, VHOST_ADDR_AVAIL)
 889
 890 #define vhost_get_used(vq, x, ptr) \
 891         vhost_get_user(vq, x, ptr, VHOST_ADDR_USED)
 892
 893 static void vhost_dev_lock_vqs(struct vhost_dev *d)
 894 {
 895         int i = 0;
 896         for (i = 0; i < d->nvqs; ++i)
 897                 mutex_lock_nested(&d->vqs[i]->mutex, i);
 898 }
 899
 900 static void vhost_dev_unlock_vqs(struct vhost_dev *d)
 901 {
 902         int i = 0;
 903         for (i = 0; i < d->nvqs; ++i)
 904                 mutex_unlock(&d->vqs[i]->mutex);
 905 }
 906
 907 static int vhost_new_umem_range(struct vhost_umem *umem,
 908                                 u64 start, u64 size, u64 end,
 909                                 u64 userspace_addr, int perm)
 910 {
 911         struct vhost_umem_node *tmp, *node = kmalloc(sizeof(*node), GFP_ATOMIC);
 912
 913         if (!node)
 914                 return -ENOMEM;
 915
 916         if (umem->numem == max_iotlb_entries) {
 917                 tmp = list_first_entry(&umem->umem_list, typeof(*tmp), link);
 918                 vhost_umem_free(umem, tmp);
 919         }
 920
 921         node->start = start;
 922         node->size = size;
 923         node->last = end;
 924         node->userspace_addr = userspace_addr;
 925         node->perm = perm;
 926         INIT_LIST_HEAD(&node->link);
 927         list_add_tail(&node->link, &umem->umem_list);
 928         vhost_umem_interval_tree_insert(node, &umem->umem_tree);
 929         umem->numem++;
 930
 931         return 0;
 932 }
 933
 934 static void vhost_del_umem_range(struct vhost_umem *umem,
 935                                  u64 start, u64 end)
 936 {
 937         struct vhost_umem_node *node;
 938
 939         while ((node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
 940                                                            start, end)))
 941                 vhost_umem_free(umem, node);
 942 }
 943
 944 static void vhost_iotlb_notify_vq(struct vhost_dev *d,
 945                                   struct vhost_iotlb_msg *msg)
 946 {
 947         struct vhost_msg_node *node, *n;
 948
 949         spin_lock(&d->iotlb_lock);
 950
 951         list_for_each_entry_safe(node, n, &d->pending_list, node) {
 952                 struct vhost_iotlb_msg *vq_msg = &node->msg.iotlb;
 953                 if (msg->iova <= vq_msg->iova &&
 954                     msg->iova + msg->size - 1 > vq_msg->iova &&
 955                     vq_msg->type == VHOST_IOTLB_MISS) {
 956                         vhost_poll_queue(&node->vq->poll);
 957                         list_del(&node->node);
 958                         kfree(node);
 959                 }
 960         }
 961
 962         spin_unlock(&d->iotlb_lock);
 963 }
 964
 965 static bool umem_access_ok(u64 uaddr, u64 size, int access)
 966 {
 967         unsigned long a = uaddr;
 968
 969         /* Make sure 64 bit math will not overflow. */
 970         if (vhost_overflow(uaddr, size))
 971                 return false;
 972
 973         if ((access & VHOST_ACCESS_RO) &&
 974             !access_ok(VERIFY_READ, (void __user *)a, size))
 975                 return false;
 976         if ((access & VHOST_ACCESS_WO) &&
 977             !access_ok(VERIFY_WRITE, (void __user *)a, size))
 978                 return false;
 979         return true;
 980 }
 981
 982 static int vhost_process_iotlb_msg(struct vhost_dev *dev,
 983                                    struct vhost_iotlb_msg *msg)
 984 {
 985         int ret = 0;
 986
 987         mutex_lock(&dev->mutex);
 988         vhost_dev_lock_vqs(dev);
 989         switch (msg->type) {
 990         case VHOST_IOTLB_UPDATE:
 991                 if (!dev->iotlb) {
 992                         ret = -EFAULT;
 993                         break;
 994                 }
 995                 if (!umem_access_ok(msg->uaddr, msg->size, msg->perm)) {
 996                         ret = -EFAULT;
 997                         break;
 998                 }
 999                 vhost_vq_meta_reset(dev);
1000                 if (vhost_new_umem_range(dev->iotlb, msg->iova, msg->size,
1001                                          msg->iova + msg->size - 1,
1002                                          msg->uaddr, msg->perm)) {
1003                         ret = -ENOMEM;
1004                         break;
1005                 }
1006                 vhost_iotlb_notify_vq(dev, msg);
1007                 break;
1008         case VHOST_IOTLB_INVALIDATE:
1009                 if (!dev->iotlb) {
1010                         ret = -EFAULT;
1011                         break;
1012                 }
1013                 vhost_vq_meta_reset(dev);
1014                 vhost_del_umem_range(dev->iotlb, msg->iova,
1015                                      msg->iova + msg->size - 1);
1016                 break;
1017         default:
1018                 ret = -EINVAL;
1019                 break;
1020         }
1021
1022         vhost_dev_unlock_vqs(dev);
1023         mutex_unlock(&dev->mutex);
1024
1025         return ret;
1026 }
1027 ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
1028                              struct iov_iter *from)
1029 {
1030         struct vhost_msg_node node;
1031         unsigned size = sizeof(struct vhost_msg);
1032         size_t ret;
1033         int err;
1034
1035         if (iov_iter_count(from) < size)
1036                 return 0;
1037         ret = copy_from_iter(&node.msg, size, from);
1038         if (ret != size)
1039                 goto done;
1040
1041         switch (node.msg.type) {
1042         case VHOST_IOTLB_MSG:
1043                 err = vhost_process_iotlb_msg(dev, &node.msg.iotlb);
1044                 if (err)
1045                         ret = err;
1046                 break;
1047         default:
1048                 ret = -EINVAL;
1049                 break;
1050         }
1051
1052 done:
1053         return ret;
1054 }
1055 EXPORT_SYMBOL(vhost_chr_write_iter);
1056
1057 __poll_t vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1058                             poll_table *wait)
1059 {
1060         __poll_t mask = 0;
1061
1062         poll_wait(file, &dev->wait, wait);
1063
1064         if (!list_empty(&dev->read_list))
1065                 mask |= EPOLLIN | EPOLLRDNORM;
1066
1067         return mask;
1068 }
1069 EXPORT_SYMBOL(vhost_chr_poll);
1070
1071 ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1072                             int noblock)
1073 {
1074         DEFINE_WAIT(wait);
1075         struct vhost_msg_node *node;
1076         ssize_t ret = 0;
1077         unsigned size = sizeof(struct vhost_msg);
1078
1079         if (iov_iter_count(to) < size)
1080                 return 0;
1081
1082         while (1) {
1083                 if (!noblock)
1084                         prepare_to_wait(&dev->wait, &wait,
1085                                         TASK_INTERRUPTIBLE);
1086
1087                 node = vhost_dequeue_msg(dev, &dev->read_list);
1088                 if (node)
1089                         break;
1090                 if (noblock) {
1091                         ret = -EAGAIN;
1092                         break;
1093                 }
1094                 if (signal_pending(current)) {
1095                         ret = -ERESTARTSYS;
1096                         break;
1097                 }
1098                 if (!dev->iotlb) {
1099                         ret = -EBADFD;
1100                         break;
1101                 }
1102
1103                 schedule();
1104         }
1105
1106         if (!noblock)
1107                 finish_wait(&dev->wait, &wait);
1108
1109         if (node) {
1110                 ret = copy_to_iter(&node->msg, size, to);
1111
1112                 if (ret != size || node->msg.type != VHOST_IOTLB_MISS) {
1113                         kfree(node);
1114                         return ret;
1115                 }
1116
1117                 vhost_enqueue_msg(dev, &dev->pending_list, node);
1118         }
1119
1120         return ret;
1121 }
1122 EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1123
1124 static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1125 {
1126         struct vhost_dev *dev = vq->dev;
1127         struct vhost_msg_node *node;
1128         struct vhost_iotlb_msg *msg;
1129
1130         node = vhost_new_msg(vq, VHOST_IOTLB_MISS);
1131         if (!node)
1132                 return -ENOMEM;
1133
1134         msg = &node->msg.iotlb;
1135         msg->type = VHOST_IOTLB_MISS;
1136         msg->iova = iova;
1137         msg->perm = access;
1138
1139         vhost_enqueue_msg(dev, &dev->read_list, node);
1140
1141         return 0;
1142 }
1143
1144 static bool vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1145                          struct vring_desc __user *desc,
1146                          struct vring_avail __user *avail,
1147                          struct vring_used __user *used)
1148
1149 {
1150         size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1151
1152         return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
1153                access_ok(VERIFY_READ, avail,
1154                          sizeof *avail + num * sizeof *avail->ring + s) &&
1155                access_ok(VERIFY_WRITE, used,
1156                         sizeof *used + num * sizeof *used->ring + s);
1157 }
1158
1159 static void vhost_vq_meta_update(struct vhost_virtqueue *vq,
1160                                  const struct vhost_umem_node *node,
1161                                  int type)
1162 {
1163         int access = (type == VHOST_ADDR_USED) ?
1164                      VHOST_ACCESS_WO : VHOST_ACCESS_RO;
1165
1166         if (likely(node->perm & access))
1167                 vq->meta_iotlb[type] = node;
1168 }
1169
1170 static bool iotlb_access_ok(struct vhost_virtqueue *vq,
1171                             int access, u64 addr, u64 len, int type)
1172 {
1173         const struct vhost_umem_node *node;
1174         struct vhost_umem *umem = vq->iotlb;
1175         u64 s = 0, size, orig_addr = addr, last = addr + len - 1;
1176
1177         if (vhost_vq_meta_fetch(vq, addr, len, type))
1178                 return true;
1179
1180         while (len > s) {
1181                 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1182                                                            addr,
1183                                                            last);
1184                 if (node == NULL || node->start > addr) {
1185                         vhost_iotlb_miss(vq, addr, access);
1186                         return false;
1187                 } else if (!(node->perm & access)) {
1188                         /* Report the possible access violation by
1189                          * request another translation from userspace.
1190                          */
1191                         return false;
1192                 }
1193
1194                 size = node->size - addr + node->start;
1195
1196                 if (orig_addr == addr && size >= len)
1197                         vhost_vq_meta_update(vq, node, type);
1198
1199                 s += size;
1200                 addr += size;
1201         }
1202
1203         return true;
1204 }
1205
1206 int vq_iotlb_prefetch(struct vhost_virtqueue *vq)
1207 {
1208         size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1209         unsigned int num = vq->num;
1210
1211         if (!vq->iotlb)
1212                 return 1;
1213
1214         return iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->desc,
1215                                num * sizeof(*vq->desc), VHOST_ADDR_DESC) &&
1216                iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->avail,
1217                                sizeof *vq->avail +
1218                                num * sizeof(*vq->avail->ring) + s,
1219                                VHOST_ADDR_AVAIL) &&
1220                iotlb_access_ok(vq, VHOST_ACCESS_WO, (u64)(uintptr_t)vq->used,
1221                                sizeof *vq->used +
1222                                num * sizeof(*vq->used->ring) + s,
1223                                VHOST_ADDR_USED);
1224 }
1225 EXPORT_SYMBOL_GPL(vq_iotlb_prefetch);
1226
1227 /* Can we log writes? */
1228 /* Caller should have device mutex but not vq mutex */
1229 bool vhost_log_access_ok(struct vhost_dev *dev)
1230 {
1231         return memory_access_ok(dev, dev->umem, 1);
1232 }
1233 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1234
1235 /* Verify access for write logging. */
1236 /* Caller should have vq mutex and device mutex */
1237 static bool vq_log_access_ok(struct vhost_virtqueue *vq,
1238                              void __user *log_base)
1239 {
1240         size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1241
1242         return vq_memory_access_ok(log_base, vq->umem,
1243                                    vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1244                 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
1245                                         sizeof *vq->used +
1246                                         vq->num * sizeof *vq->used->ring + s));
1247 }
1248
1249 /* Can we start vq? */
1250 /* Caller should have vq mutex and device mutex */
1251 bool vhost_vq_access_ok(struct vhost_virtqueue *vq)
1252 {
1253         if (!vq_log_access_ok(vq, vq->log_base))
1254                 return false;
1255
1256         /* Access validation occurs at prefetch time with IOTLB */
1257         if (vq->iotlb)
1258                 return true;
1259
1260         return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used);
1261 }
1262 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1263
1264 static struct vhost_umem *vhost_umem_alloc(void)
1265 {
1266         struct vhost_umem *umem = kvzalloc(sizeof(*umem), GFP_KERNEL);
1267
1268         if (!umem)
1269                 return NULL;
1270
1271         umem->umem_tree = RB_ROOT_CACHED;
1272         umem->numem = 0;
1273         INIT_LIST_HEAD(&umem->umem_list);
1274
1275         return umem;
1276 }
1277
1278 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1279 {
1280         struct vhost_memory mem, *newmem;
1281         struct vhost_memory_region *region;
1282         struct vhost_umem *newumem, *oldumem;
1283         unsigned long size = offsetof(struct vhost_memory, regions);
1284         int i;
1285
1286         if (copy_from_user(&mem, m, size))
1287                 return -EFAULT;
1288         if (mem.padding)
1289                 return -EOPNOTSUPP;
1290         if (mem.nregions > max_mem_regions)
1291                 return -E2BIG;
1292         newmem = kvzalloc(struct_size(newmem, regions, mem.nregions),
1293                         GFP_KERNEL);
1294         if (!newmem)
1295                 return -ENOMEM;
1296
1297         memcpy(newmem, &mem, size);
1298         if (copy_from_user(newmem->regions, m->regions,
1299                            mem.nregions * sizeof *m->regions)) {
1300                 kvfree(newmem);
1301                 return -EFAULT;
1302         }
1303
1304         newumem = vhost_umem_alloc();
1305         if (!newumem) {
1306                 kvfree(newmem);
1307                 return -ENOMEM;
1308         }
1309
1310         for (region = newmem->regions;
1311              region < newmem->regions + mem.nregions;
1312              region++) {
1313                 if (vhost_new_umem_range(newumem,
1314                                          region->guest_phys_addr,
1315                                          region->memory_size,
1316                                          region->guest_phys_addr +
1317                                          region->memory_size - 1,
1318                                          region->userspace_addr,
1319                                          VHOST_ACCESS_RW))
1320                         goto err;
1321         }
1322
1323         if (!memory_access_ok(d, newumem, 0))
1324                 goto err;
1325
1326         oldumem = d->umem;
1327         d->umem = newumem;
1328
1329         /* All memory accesses are done under some VQ mutex. */
1330         for (i = 0; i < d->nvqs; ++i) {
1331                 mutex_lock(&d->vqs[i]->mutex);
1332                 d->vqs[i]->umem = newumem;
1333                 mutex_unlock(&d->vqs[i]->mutex);
1334         }
1335
1336         kvfree(newmem);
1337         vhost_umem_clean(oldumem);
1338         return 0;
1339
1340 err:
1341         vhost_umem_clean(newumem);
1342         kvfree(newmem);
1343         return -EFAULT;
1344 }
1345
1346 long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1347 {
1348         struct file *eventfp, *filep = NULL;
1349         bool pollstart = false, pollstop = false;
1350         struct eventfd_ctx *ctx = NULL;
1351         u32 __user *idxp = argp;
1352         struct vhost_virtqueue *vq;
1353         struct vhost_vring_state s;
1354         struct vhost_vring_file f;
1355         struct vhost_vring_addr a;
1356         u32 idx;
1357         long r;
1358
1359         r = get_user(idx, idxp);
1360         if (r < 0)
1361                 return r;
1362         if (idx >= d->nvqs)
1363                 return -ENOBUFS;
1364
1365         vq = d->vqs[idx];
1366
1367         mutex_lock(&vq->mutex);
1368
1369         switch (ioctl) {
1370         case VHOST_SET_VRING_NUM:
1371                 /* Resizing ring with an active backend?
1372                  * You don't want to do that. */
1373                 if (vq->private_data) {
1374                         r = -EBUSY;
1375                         break;
1376                 }
1377                 if (copy_from_user(&s, argp, sizeof s)) {
1378                         r = -EFAULT;
1379                         break;
1380                 }
1381                 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
1382                         r = -EINVAL;
1383                         break;
1384                 }
1385                 vq->num = s.num;
1386                 break;
1387         case VHOST_SET_VRING_BASE:
1388                 /* Moving base with an active backend?
1389                  * You don't want to do that. */
1390                 if (vq->private_data) {
1391                         r = -EBUSY;
1392                         break;
1393                 }
1394                 if (copy_from_user(&s, argp, sizeof s)) {
1395                         r = -EFAULT;
1396                         break;
1397                 }
1398                 if (s.num > 0xffff) {
1399                         r = -EINVAL;
1400                         break;
1401                 }
1402                 vq->last_avail_idx = s.num;
1403                 /* Forget the cached index value. */
1404                 vq->avail_idx = vq->last_avail_idx;
1405                 break;
1406         case VHOST_GET_VRING_BASE:
1407                 s.index = idx;
1408                 s.num = vq->last_avail_idx;
1409                 if (copy_to_user(argp, &s, sizeof s))
1410                         r = -EFAULT;
1411                 break;
1412         case VHOST_SET_VRING_ADDR:
1413                 if (copy_from_user(&a, argp, sizeof a)) {
1414                         r = -EFAULT;
1415                         break;
1416                 }
1417                 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
1418                         r = -EOPNOTSUPP;
1419                         break;
1420                 }
1421                 /* For 32bit, verify that the top 32bits of the user
1422                    data are set to zero. */
1423                 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1424                     (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1425                     (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
1426                         r = -EFAULT;
1427                         break;
1428                 }
1429
1430                 /* Make sure it's safe to cast pointers to vring types. */
1431                 BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1432                 BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1433                 if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1434                     (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1435                     (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1))) {
1436                         r = -EINVAL;
1437                         break;
1438                 }
1439
1440                 /* We only verify access here if backend is configured.
1441                  * If it is not, we don't as size might not have been setup.
1442                  * We will verify when backend is configured. */
1443                 if (vq->private_data) {
1444                         if (!vq_access_ok(vq, vq->num,
1445                                 (void __user *)(unsigned long)a.desc_user_addr,
1446                                 (void __user *)(unsigned long)a.avail_user_addr,
1447                                 (void __user *)(unsigned long)a.used_user_addr)) {
1448                                 r = -EINVAL;
1449                                 break;
1450                         }
1451
1452                         /* Also validate log access for used ring if enabled. */
1453                         if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
1454                             !log_access_ok(vq->log_base, a.log_guest_addr,
1455                                            sizeof *vq->used +
1456                                            vq->num * sizeof *vq->used->ring)) {
1457                                 r = -EINVAL;
1458                                 break;
1459                         }
1460                 }
1461
1462                 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1463                 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1464                 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1465                 vq->log_addr = a.log_guest_addr;
1466                 vq->used = (void __user *)(unsigned long)a.used_user_addr;
1467                 break;
1468         case VHOST_SET_VRING_KICK:
1469                 if (copy_from_user(&f, argp, sizeof f)) {
1470                         r = -EFAULT;
1471                         break;
1472                 }
1473                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1474                 if (IS_ERR(eventfp)) {
1475                         r = PTR_ERR(eventfp);
1476                         break;
1477                 }
1478                 if (eventfp != vq->kick) {
1479                         pollstop = (filep = vq->kick) != NULL;
1480                         pollstart = (vq->kick = eventfp) != NULL;
1481                 } else
1482                         filep = eventfp;
1483                 break;
1484         case VHOST_SET_VRING_CALL:
1485                 if (copy_from_user(&f, argp, sizeof f)) {
1486                         r = -EFAULT;
1487                         break;
1488                 }
1489                 ctx = f.fd == -1 ? NULL : eventfd_ctx_fdget(f.fd);
1490                 if (IS_ERR(ctx)) {
1491                         r = PTR_ERR(ctx);
1492                         break;
1493                 }
1494                 swap(ctx, vq->call_ctx);
1495                 break;
1496         case VHOST_SET_VRING_ERR:
1497                 if (copy_from_user(&f, argp, sizeof f)) {
1498                         r = -EFAULT;
1499                         break;
1500                 }
1501                 ctx = f.fd == -1 ? NULL : eventfd_ctx_fdget(f.fd);
1502                 if (IS_ERR(ctx)) {
1503                         r = PTR_ERR(ctx);
1504                         break;
1505                 }
1506                 swap(ctx, vq->error_ctx);
1507                 break;
1508         case VHOST_SET_VRING_ENDIAN:
1509                 r = vhost_set_vring_endian(vq, argp);
1510                 break;
1511         case VHOST_GET_VRING_ENDIAN:
1512                 r = vhost_get_vring_endian(vq, idx, argp);
1513                 break;
1514         case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
1515                 if (copy_from_user(&s, argp, sizeof(s))) {
1516                         r = -EFAULT;
1517                         break;
1518                 }
1519                 vq->busyloop_timeout = s.num;
1520                 break;
1521         case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
1522                 s.index = idx;
1523                 s.num = vq->busyloop_timeout;
1524                 if (copy_to_user(argp, &s, sizeof(s)))
1525                         r = -EFAULT;
1526                 break;
1527         default:
1528                 r = -ENOIOCTLCMD;
1529         }
1530
1531         if (pollstop && vq->handle_kick)
1532                 vhost_poll_stop(&vq->poll);
1533
1534         if (!IS_ERR_OR_NULL(ctx))
1535                 eventfd_ctx_put(ctx);
1536         if (filep)
1537                 fput(filep);
1538
1539         if (pollstart && vq->handle_kick)
1540                 r = vhost_poll_start(&vq->poll, vq->kick);
1541
1542         mutex_unlock(&vq->mutex);
1543
1544         if (pollstop && vq->handle_kick)
1545                 vhost_poll_flush(&vq->poll);
1546         return r;
1547 }
1548 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
1549
1550 int vhost_init_device_iotlb(struct vhost_dev *d, bool enabled)
1551 {
1552         struct vhost_umem *niotlb, *oiotlb;
1553         int i;
1554
1555         niotlb = vhost_umem_alloc();
1556         if (!niotlb)
1557                 return -ENOMEM;
1558
1559         oiotlb = d->iotlb;
1560         d->iotlb = niotlb;
1561
1562         for (i = 0; i < d->nvqs; ++i) {
1563                 mutex_lock(&d->vqs[i]->mutex);
1564                 d->vqs[i]->iotlb = niotlb;
1565                 mutex_unlock(&d->vqs[i]->mutex);
1566         }
1567
1568         vhost_umem_clean(oiotlb);
1569
1570         return 0;
1571 }
1572 EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
1573
1574 /* Caller must have device mutex */
1575 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1576 {
1577         struct eventfd_ctx *ctx;
1578         u64 p;
1579         long r;
1580         int i, fd;
1581
1582         /* If you are not the owner, you can become one */
1583         if (ioctl == VHOST_SET_OWNER) {
1584                 r = vhost_dev_set_owner(d);
1585                 goto done;
1586         }
1587
1588         /* You must be the owner to do anything else */
1589         r = vhost_dev_check_owner(d);
1590         if (r)
1591                 goto done;
1592
1593         switch (ioctl) {
1594         case VHOST_SET_MEM_TABLE:
1595                 r = vhost_set_memory(d, argp);
1596                 break;
1597         case VHOST_SET_LOG_BASE:
1598                 if (copy_from_user(&p, argp, sizeof p)) {
1599                         r = -EFAULT;
1600                         break;
1601                 }
1602                 if ((u64)(unsigned long)p != p) {
1603                         r = -EFAULT;
1604                         break;
1605                 }
1606                 for (i = 0; i < d->nvqs; ++i) {
1607                         struct vhost_virtqueue *vq;
1608                         void __user *base = (void __user *)(unsigned long)p;
1609                         vq = d->vqs[i];
1610                         mutex_lock(&vq->mutex);
1611                         /* If ring is inactive, will check when it's enabled. */
1612                         if (vq->private_data && !vq_log_access_ok(vq, base))
1613                                 r = -EFAULT;
1614                         else
1615                                 vq->log_base = base;
1616                         mutex_unlock(&vq->mutex);
1617                 }
1618                 break;
1619         case VHOST_SET_LOG_FD:
1620                 r = get_user(fd, (int __user *)argp);
1621                 if (r < 0)
1622                         break;
1623                 ctx = fd == -1 ? NULL : eventfd_ctx_fdget(fd);
1624                 if (IS_ERR(ctx)) {
1625                         r = PTR_ERR(ctx);
1626                         break;
1627                 }
1628                 swap(ctx, d->log_ctx);
1629                 for (i = 0; i < d->nvqs; ++i) {
1630                         mutex_lock(&d->vqs[i]->mutex);
1631                         d->vqs[i]->log_ctx = d->log_ctx;
1632                         mutex_unlock(&d->vqs[i]->mutex);
1633                 }
1634                 if (ctx)
1635                         eventfd_ctx_put(ctx);
1636                 break;
1637         default:
1638                 r = -ENOIOCTLCMD;
1639                 break;
1640         }
1641 done:
1642         return r;
1643 }
1644 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1645
1646 /* TODO: This is really inefficient.  We need something like get_user()
1647  * (instruction directly accesses the data, with an exception table entry
1648  * returning -EFAULT). See Documentation/x86/exception-tables.txt.
1649  */
1650 static int set_bit_to_user(int nr, void __user *addr)
1651 {
1652         unsigned long log = (unsigned long)addr;
1653         struct page *page;
1654         void *base;
1655         int bit = nr + (log % PAGE_SIZE) * 8;
1656         int r;
1657
1658         r = get_user_pages_fast(log, 1, 1, &page);
1659         if (r < 0)
1660                 return r;
1661         BUG_ON(r != 1);
1662         base = kmap_atomic(page);
1663         set_bit(bit, base);
1664         kunmap_atomic(base);
1665         set_page_dirty_lock(page);
1666         put_page(page);
1667         return 0;
1668 }
1669
1670 static int log_write(void __user *log_base,
1671                      u64 write_address, u64 write_length)
1672 {
1673         u64 write_page = write_address / VHOST_PAGE_SIZE;
1674         int r;
1675
1676         if (!write_length)
1677                 return 0;
1678         write_length += write_address % VHOST_PAGE_SIZE;
1679         for (;;) {
1680                 u64 base = (u64)(unsigned long)log_base;
1681                 u64 log = base + write_page / 8;
1682                 int bit = write_page % 8;
1683                 if ((u64)(unsigned long)log != log)
1684                         return -EFAULT;
1685                 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1686                 if (r < 0)
1687                         return r;
1688                 if (write_length <= VHOST_PAGE_SIZE)
1689                         break;
1690                 write_length -= VHOST_PAGE_SIZE;
1691                 write_page += 1;
1692         }
1693         return r;
1694 }
1695
1696 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1697                     unsigned int log_num, u64 len)
1698 {
1699         int i, r;
1700
1701         /* Make sure data written is seen before log. */
1702         smp_wmb();
1703         for (i = 0; i < log_num; ++i) {
1704                 u64 l = min(log[i].len, len);
1705                 r = log_write(vq->log_base, log[i].addr, l);
1706                 if (r < 0)
1707                         return r;
1708                 len -= l;
1709                 if (!len) {
1710                         if (vq->log_ctx)
1711                                 eventfd_signal(vq->log_ctx, 1);
1712                         return 0;
1713                 }
1714         }
1715         /* Length written exceeds what we have stored. This is a bug. */
1716         BUG();
1717         return 0;
1718 }
1719 EXPORT_SYMBOL_GPL(vhost_log_write);
1720
1721 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1722 {
1723         void __user *used;
1724         if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
1725                            &vq->used->flags) < 0)
1726                 return -EFAULT;
1727         if (unlikely(vq->log_used)) {
1728                 /* Make sure the flag is seen before log. */
1729                 smp_wmb();
1730                 /* Log used flag write. */
1731                 used = &vq->used->flags;
1732                 log_write(vq->log_base, vq->log_addr +
1733                           (used - (void __user *)vq->used),
1734                           sizeof vq->used->flags);
1735                 if (vq->log_ctx)
1736                         eventfd_signal(vq->log_ctx, 1);
1737         }
1738         return 0;
1739 }
1740
1741 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1742 {
1743         if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
1744                            vhost_avail_event(vq)))
1745                 return -EFAULT;
1746         if (unlikely(vq->log_used)) {
1747                 void __user *used;
1748                 /* Make sure the event is seen before log. */
1749                 smp_wmb();
1750                 /* Log avail event write */
1751                 used = vhost_avail_event(vq);
1752                 log_write(vq->log_base, vq->log_addr +
1753                           (used - (void __user *)vq->used),
1754                           sizeof *vhost_avail_event(vq));
1755                 if (vq->log_ctx)
1756                         eventfd_signal(vq->log_ctx, 1);
1757         }
1758         return 0;
1759 }
1760
1761 int vhost_vq_init_access(struct vhost_virtqueue *vq)
1762 {
1763         __virtio16 last_used_idx;
1764         int r;
1765         bool is_le = vq->is_le;
1766
1767         if (!vq->private_data)
1768                 return 0;
1769
1770         vhost_init_is_le(vq);
1771
1772         r = vhost_update_used_flags(vq);
1773         if (r)
1774                 goto err;
1775         vq->signalled_used_valid = false;
1776         if (!vq->iotlb &&
1777             !access_ok(VERIFY_READ, &vq->used->idx, sizeof vq->used->idx)) {
1778                 r = -EFAULT;
1779                 goto err;
1780         }
1781         r = vhost_get_used(vq, last_used_idx, &vq->used->idx);
1782         if (r) {
1783                 vq_err(vq, "Can't access used idx at %p\n",
1784                        &vq->used->idx);
1785                 goto err;
1786         }
1787         vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
1788         return 0;
1789
1790 err:
1791         vq->is_le = is_le;
1792         return r;
1793 }
1794 EXPORT_SYMBOL_GPL(vhost_vq_init_access);
1795
1796 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
1797                           struct iovec iov[], int iov_size, int access)
1798 {
1799         const struct vhost_umem_node *node;
1800         struct vhost_dev *dev = vq->dev;
1801         struct vhost_umem *umem = dev->iotlb ? dev->iotlb : dev->umem;
1802         struct iovec *_iov;
1803         u64 s = 0;
1804         int ret = 0;
1805
1806         while ((u64)len > s) {
1807                 u64 size;
1808                 if (unlikely(ret >= iov_size)) {
1809                         ret = -ENOBUFS;
1810                         break;
1811                 }
1812
1813                 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1814                                                         addr, addr + len - 1);
1815                 if (node == NULL || node->start > addr) {
1816                         if (umem != dev->iotlb) {
1817                                 ret = -EFAULT;
1818                                 break;
1819                         }
1820                         ret = -EAGAIN;
1821                         break;
1822                 } else if (!(node->perm & access)) {
1823                         ret = -EPERM;
1824                         break;
1825                 }
1826
1827                 _iov = iov + ret;
1828                 size = node->size - addr + node->start;
1829                 _iov->iov_len = min((u64)len - s, size);
1830                 _iov->iov_base = (void __user *)(unsigned long)
1831                         (node->userspace_addr + addr - node->start);
1832                 s += size;
1833                 addr += size;
1834                 ++ret;
1835         }
1836
1837         if (ret == -EAGAIN)
1838                 vhost_iotlb_miss(vq, addr, access);
1839         return ret;
1840 }
1841
1842 /* Each buffer in the virtqueues is actually a chain of descriptors.  This
1843  * function returns the next descriptor in the chain,
1844  * or -1U if we're at the end. */
1845 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
1846 {
1847         unsigned int next;
1848
1849         /* If this descriptor says it doesn't chain, we're done. */
1850         if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
1851                 return -1U;
1852
1853         /* Check they're not leading us off end of descriptors. */
1854         next = vhost16_to_cpu(vq, READ_ONCE(desc->next));
1855         return next;
1856 }
1857
1858 static int get_indirect(struct vhost_virtqueue *vq,
1859                         struct iovec iov[], unsigned int iov_size,
1860                         unsigned int *out_num, unsigned int *in_num,
1861                         struct vhost_log *log, unsigned int *log_num,
1862                         struct vring_desc *indirect)
1863 {
1864         struct vring_desc desc;
1865         unsigned int i = 0, count, found = 0;
1866         u32 len = vhost32_to_cpu(vq, indirect->len);
1867         struct iov_iter from;
1868         int ret, access;
1869
1870         /* Sanity check */
1871         if (unlikely(len % sizeof desc)) {
1872                 vq_err(vq, "Invalid length in indirect descriptor: "
1873                        "len 0x%llx not multiple of 0x%zx\n",
1874                        (unsigned long long)len,
1875                        sizeof desc);
1876                 return -EINVAL;
1877         }
1878
1879         ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
1880                              UIO_MAXIOV, VHOST_ACCESS_RO);
1881         if (unlikely(ret < 0)) {
1882                 if (ret != -EAGAIN)
1883                         vq_err(vq, "Translation failure %d in indirect.\n", ret);
1884                 return ret;
1885         }
1886         iov_iter_init(&from, READ, vq->indirect, ret, len);
1887
1888         /* We will use the result as an address to read from, so most
1889          * architectures only need a compiler barrier here. */
1890         read_barrier_depends();
1891
1892         count = len / sizeof desc;
1893         /* Buffers are chained via a 16 bit next field, so
1894          * we can have at most 2^16 of these. */
1895         if (unlikely(count > USHRT_MAX + 1)) {
1896                 vq_err(vq, "Indirect buffer length too big: %d\n",
1897                        indirect->len);
1898                 return -E2BIG;
1899         }
1900
1901         do {
1902                 unsigned iov_count = *in_num + *out_num;
1903                 if (unlikely(++found > count)) {
1904                         vq_err(vq, "Loop detected: last one at %u "
1905                                "indirect size %u\n",
1906                                i, count);
1907                         return -EINVAL;
1908                 }
1909                 if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
1910                         vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1911                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1912                         return -EINVAL;
1913                 }
1914                 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
1915                         vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1916                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1917                         return -EINVAL;
1918                 }
1919
1920                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
1921                         access = VHOST_ACCESS_WO;
1922                 else
1923                         access = VHOST_ACCESS_RO;
1924
1925                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
1926                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
1927                                      iov_size - iov_count, access);
1928                 if (unlikely(ret < 0)) {
1929                         if (ret != -EAGAIN)
1930                                 vq_err(vq, "Translation failure %d indirect idx %d\n",
1931                                         ret, i);
1932                         return ret;
1933                 }
1934                 /* If this is an input descriptor, increment that count. */
1935                 if (access == VHOST_ACCESS_WO) {
1936                         *in_num += ret;
1937                         if (unlikely(log)) {
1938                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
1939                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
1940                                 ++*log_num;
1941                         }
1942                 } else {
1943                         /* If it's an output descriptor, they're all supposed
1944                          * to come before any input descriptors. */
1945                         if (unlikely(*in_num)) {
1946                                 vq_err(vq, "Indirect descriptor "
1947                                        "has out after in: idx %d\n", i);
1948                                 return -EINVAL;
1949                         }
1950                         *out_num += ret;
1951                 }
1952         } while ((i = next_desc(vq, &desc)) != -1);
1953         return 0;
1954 }
1955
1956 /* This looks in the virtqueue and for the first available buffer, and converts
1957  * it to an iovec for convenient access.  Since descriptors consist of some
1958  * number of output then some number of input descriptors, it's actually two
1959  * iovecs, but we pack them into one and note how many of each there were.
1960  *
1961  * This function returns the descriptor number found, or vq->num (which is
1962  * never a valid descriptor number) if none was found.  A negative code is
1963  * returned on error. */
1964 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
1965                       struct iovec iov[], unsigned int iov_size,
1966                       unsigned int *out_num, unsigned int *in_num,
1967                       struct vhost_log *log, unsigned int *log_num)
1968 {
1969         struct vring_desc desc;
1970         unsigned int i, head, found = 0;
1971         u16 last_avail_idx;
1972         __virtio16 avail_idx;
1973         __virtio16 ring_head;
1974         int ret, access;
1975
1976         /* Check it isn't doing very strange things with descriptor numbers. */
1977         last_avail_idx = vq->last_avail_idx;
1978
1979         if (vq->avail_idx == vq->last_avail_idx) {
1980                 if (unlikely(vhost_get_avail(vq, avail_idx, &vq->avail->idx))) {
1981                         vq_err(vq, "Failed to access avail idx at %p\n",
1982                                 &vq->avail->idx);
1983                         return -EFAULT;
1984                 }
1985                 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
1986
1987                 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
1988                         vq_err(vq, "Guest moved used index from %u to %u",
1989                                 last_avail_idx, vq->avail_idx);
1990                         return -EFAULT;
1991                 }
1992
1993                 /* If there's nothing new since last we looked, return
1994                  * invalid.
1995                  */
1996                 if (vq->avail_idx == last_avail_idx)
1997                         return vq->num;
1998
1999                 /* Only get avail ring entries after they have been
2000                  * exposed by guest.
2001                  */
2002                 smp_rmb();
2003         }
2004
2005         /* Grab the next descriptor number they're advertising, and increment
2006          * the index we've seen. */
2007         if (unlikely(vhost_get_avail(vq, ring_head,
2008                      &vq->avail->ring[last_avail_idx & (vq->num - 1)]))) {
2009                 vq_err(vq, "Failed to read head: idx %d address %p\n",
2010                        last_avail_idx,
2011                        &vq->avail->ring[last_avail_idx % vq->num]);
2012                 return -EFAULT;
2013         }
2014
2015         head = vhost16_to_cpu(vq, ring_head);
2016
2017         /* If their number is silly, that's an error. */
2018         if (unlikely(head >= vq->num)) {
2019                 vq_err(vq, "Guest says index %u > %u is available",
2020                        head, vq->num);
2021                 return -EINVAL;
2022         }
2023
2024         /* When we start there are none of either input nor output. */
2025         *out_num = *in_num = 0;
2026         if (unlikely(log))
2027                 *log_num = 0;
2028
2029         i = head;
2030         do {
2031                 unsigned iov_count = *in_num + *out_num;
2032                 if (unlikely(i >= vq->num)) {
2033                         vq_err(vq, "Desc index is %u > %u, head = %u",
2034                                i, vq->num, head);
2035                         return -EINVAL;
2036                 }
2037                 if (unlikely(++found > vq->num)) {
2038                         vq_err(vq, "Loop detected: last one at %u "
2039                                "vq size %u head %u\n",
2040                                i, vq->num, head);
2041                         return -EINVAL;
2042                 }
2043                 ret = vhost_copy_from_user(vq, &desc, vq->desc + i,
2044                                            sizeof desc);
2045                 if (unlikely(ret)) {
2046                         vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
2047                                i, vq->desc + i);
2048                         return -EFAULT;
2049                 }
2050                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2051                         ret = get_indirect(vq, iov, iov_size,
2052                                            out_num, in_num,
2053                                            log, log_num, &desc);
2054                         if (unlikely(ret < 0)) {
2055                                 if (ret != -EAGAIN)
2056                                         vq_err(vq, "Failure detected "
2057                                                 "in indirect descriptor at idx %d\n", i);
2058                                 return ret;
2059                         }
2060                         continue;
2061                 }
2062
2063                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2064                         access = VHOST_ACCESS_WO;
2065                 else
2066                         access = VHOST_ACCESS_RO;
2067                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2068                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
2069                                      iov_size - iov_count, access);
2070                 if (unlikely(ret < 0)) {
2071                         if (ret != -EAGAIN)
2072                                 vq_err(vq, "Translation failure %d descriptor idx %d\n",
2073                                         ret, i);
2074                         return ret;
2075                 }
2076                 if (access == VHOST_ACCESS_WO) {
2077                         /* If this is an input descriptor,
2078                          * increment that count. */
2079                         *in_num += ret;
2080                         if (unlikely(log)) {
2081                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2082                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2083                                 ++*log_num;
2084                         }
2085                 } else {
2086                         /* If it's an output descriptor, they're all supposed
2087                          * to come before any input descriptors. */
2088                         if (unlikely(*in_num)) {
2089                                 vq_err(vq, "Descriptor has out after in: "
2090                                        "idx %d\n", i);
2091                                 return -EINVAL;
2092                         }
2093                         *out_num += ret;
2094                 }
2095         } while ((i = next_desc(vq, &desc)) != -1);
2096
2097         /* On success, increment avail index. */
2098         vq->last_avail_idx++;
2099
2100         /* Assume notifications from guest are disabled at this point,
2101          * if they aren't we would need to update avail_event index. */
2102         BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2103         return head;
2104 }
2105 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2106
2107 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2108 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2109 {
2110         vq->last_avail_idx -= n;
2111 }
2112 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2113
2114 /* After we've used one of their buffers, we tell them about it.  We'll then
2115  * want to notify the guest, using eventfd. */
2116 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2117 {
2118         struct vring_used_elem heads = {
2119                 cpu_to_vhost32(vq, head),
2120                 cpu_to_vhost32(vq, len)
2121         };
2122
2123         return vhost_add_used_n(vq, &heads, 1);
2124 }
2125 EXPORT_SYMBOL_GPL(vhost_add_used);
2126
2127 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2128                             struct vring_used_elem *heads,
2129                             unsigned count)
2130 {
2131         struct vring_used_elem __user *used;
2132         u16 old, new;
2133         int start;
2134
2135         start = vq->last_used_idx & (vq->num - 1);
2136         used = vq->used->ring + start;
2137         if (count == 1) {
2138                 if (vhost_put_user(vq, heads[0].id, &used->id)) {
2139                         vq_err(vq, "Failed to write used id");
2140                         return -EFAULT;
2141                 }
2142                 if (vhost_put_user(vq, heads[0].len, &used->len)) {
2143                         vq_err(vq, "Failed to write used len");
2144                         return -EFAULT;
2145                 }
2146         } else if (vhost_copy_to_user(vq, used, heads, count * sizeof *used)) {
2147                 vq_err(vq, "Failed to write used");
2148                 return -EFAULT;
2149         }
2150         if (unlikely(vq->log_used)) {
2151                 /* Make sure data is seen before log. */
2152                 smp_wmb();
2153                 /* Log used ring entry write. */
2154                 log_write(vq->log_base,
2155                           vq->log_addr +
2156                            ((void __user *)used - (void __user *)vq->used),
2157                           count * sizeof *used);
2158         }
2159         old = vq->last_used_idx;
2160         new = (vq->last_used_idx += count);
2161         /* If the driver never bothers to signal in a very long while,
2162          * used index might wrap around. If that happens, invalidate
2163          * signalled_used index we stored. TODO: make sure driver
2164          * signals at least once in 2^16 and remove this. */
2165         if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2166                 vq->signalled_used_valid = false;
2167         return 0;
2168 }
2169
2170 /* After we've used one of their buffers, we tell them about it.  We'll then
2171  * want to notify the guest, using eventfd. */
2172 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2173                      unsigned count)
2174 {
2175         int start, n, r;
2176
2177         start = vq->last_used_idx & (vq->num - 1);
2178         n = vq->num - start;
2179         if (n < count) {
2180                 r = __vhost_add_used_n(vq, heads, n);
2181                 if (r < 0)
2182                         return r;
2183                 heads += n;
2184                 count -= n;
2185         }
2186         r = __vhost_add_used_n(vq, heads, count);
2187
2188         /* Make sure buffer is written before we update index. */
2189         smp_wmb();
2190         if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
2191                            &vq->used->idx)) {
2192                 vq_err(vq, "Failed to increment used idx");
2193                 return -EFAULT;
2194         }
2195         if (unlikely(vq->log_used)) {
2196                 /* Log used index update. */
2197                 log_write(vq->log_base,
2198                           vq->log_addr + offsetof(struct vring_used, idx),
2199                           sizeof vq->used->idx);
2200                 if (vq->log_ctx)
2201                         eventfd_signal(vq->log_ctx, 1);
2202         }
2203         return r;
2204 }
2205 EXPORT_SYMBOL_GPL(vhost_add_used_n);
2206
2207 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2208 {
2209         __u16 old, new;
2210         __virtio16 event;
2211         bool v;
2212         /* Flush out used index updates. This is paired
2213          * with the barrier that the Guest executes when enabling
2214          * interrupts. */
2215         smp_mb();
2216
2217         if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2218             unlikely(vq->avail_idx == vq->last_avail_idx))
2219                 return true;
2220
2221         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2222                 __virtio16 flags;
2223                 if (vhost_get_avail(vq, flags, &vq->avail->flags)) {
2224                         vq_err(vq, "Failed to get flags");
2225                         return true;
2226                 }
2227                 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2228         }
2229         old = vq->signalled_used;
2230         v = vq->signalled_used_valid;
2231         new = vq->signalled_used = vq->last_used_idx;
2232         vq->signalled_used_valid = true;
2233
2234         if (unlikely(!v))
2235                 return true;
2236
2237         if (vhost_get_avail(vq, event, vhost_used_event(vq))) {
2238                 vq_err(vq, "Failed to get used event idx");
2239                 return true;
2240         }
2241         return vring_need_event(vhost16_to_cpu(vq, event), new, old);
2242 }
2243
2244 /* This actually signals the guest, using eventfd. */
2245 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2246 {
2247         /* Signal the Guest tell them we used something up. */
2248         if (vq->call_ctx && vhost_notify(dev, vq))
2249                 eventfd_signal(vq->call_ctx, 1);
2250 }
2251 EXPORT_SYMBOL_GPL(vhost_signal);
2252
2253 /* And here's the combo meal deal.  Supersize me! */
2254 void vhost_add_used_and_signal(struct vhost_dev *dev,
2255                                struct vhost_virtqueue *vq,
2256                                unsigned int head, int len)
2257 {
2258         vhost_add_used(vq, head, len);
2259         vhost_signal(dev, vq);
2260 }
2261 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2262
2263 /* multi-buffer version of vhost_add_used_and_signal */
2264 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2265                                  struct vhost_virtqueue *vq,
2266                                  struct vring_used_elem *heads, unsigned count)
2267 {
2268         vhost_add_used_n(vq, heads, count);
2269         vhost_signal(dev, vq);
2270 }
2271 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2272
2273 /* return true if we're sure that avaiable ring is empty */
2274 bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2275 {
2276         __virtio16 avail_idx;
2277         int r;
2278
2279         if (vq->avail_idx != vq->last_avail_idx)
2280                 return false;
2281
2282         r = vhost_get_avail(vq, avail_idx, &vq->avail->idx);
2283         if (unlikely(r))
2284                 return false;
2285         vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2286
2287         return vq->avail_idx == vq->last_avail_idx;
2288 }
2289 EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2290
2291 /* OK, now we need to know about added descriptors. */
2292 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2293 {
2294         __virtio16 avail_idx;
2295         int r;
2296
2297         if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2298                 return false;
2299         vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2300         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2301                 r = vhost_update_used_flags(vq);
2302                 if (r) {
2303                         vq_err(vq, "Failed to enable notification at %p: %d\n",
2304                                &vq->used->flags, r);
2305                         return false;
2306                 }
2307         } else {
2308                 r = vhost_update_avail_event(vq, vq->avail_idx);
2309                 if (r) {
2310                         vq_err(vq, "Failed to update avail event index at %p: %d\n",
2311                                vhost_avail_event(vq), r);
2312                         return false;
2313                 }
2314         }
2315         /* They could have slipped one in as we were doing that: make
2316          * sure it's written, then check again. */
2317         smp_mb();
2318         r = vhost_get_avail(vq, avail_idx, &vq->avail->idx);
2319         if (r) {
2320                 vq_err(vq, "Failed to check avail idx at %p: %d\n",
2321                        &vq->avail->idx, r);
2322                 return false;
2323         }
2324
2325         return vhost16_to_cpu(vq, avail_idx) != vq->avail_idx;
2326 }
2327 EXPORT_SYMBOL_GPL(vhost_enable_notify);
2328
2329 /* We don't need to be notified again. */
2330 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2331 {
2332         int r;
2333
2334         if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2335                 return;
2336         vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2337         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2338                 r = vhost_update_used_flags(vq);
2339                 if (r)
2340                         vq_err(vq, "Failed to enable notification at %p: %d\n",
2341                                &vq->used->flags, r);
2342         }
2343 }
2344 EXPORT_SYMBOL_GPL(vhost_disable_notify);
2345
2346 /* Create a new message. */
2347 struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2348 {
2349         struct vhost_msg_node *node = kmalloc(sizeof *node, GFP_KERNEL);
2350         if (!node)
2351                 return NULL;
2352
2353         /* Make sure all padding within the structure is initialized. */
2354         memset(&node->msg, 0, sizeof node->msg);
2355         node->vq = vq;
2356         node->msg.type = type;
2357         return node;
2358 }
2359 EXPORT_SYMBOL_GPL(vhost_new_msg);
2360
2361 void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2362                        struct vhost_msg_node *node)
2363 {
2364         spin_lock(&dev->iotlb_lock);
2365         list_add_tail(&node->node, head);
2366         spin_unlock(&dev->iotlb_lock);
2367
2368         wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
2369 }
2370 EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2371
2372 struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2373                                          struct list_head *head)
2374 {
2375         struct vhost_msg_node *node = NULL;
2376
2377         spin_lock(&dev->iotlb_lock);
2378         if (!list_empty(head)) {
2379                 node = list_first_entry(head, struct vhost_msg_node,
2380                                         node);
2381                 list_del(&node->node);
2382         }
2383         spin_unlock(&dev->iotlb_lock);
2384
2385         return node;
2386 }
2387 EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2388
2389
2390 static int __init vhost_init(void)
2391 {
2392         return 0;
2393 }
2394
2395 static void __exit vhost_exit(void)
2396 {
2397 }
2398
2399 module_init(vhost_init);
2400 module_exit(vhost_exit);
2401
2402 MODULE_VERSION("0.0.1");
2403 MODULE_LICENSE("GPL v2");
2404 MODULE_AUTHOR("Michael S. Tsirkin");
2405 MODULE_DESCRIPTION("Host kernel accelerator for virtio");