drivers/vhost/vhost.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /* Copyright (C) 2009 Red Hat, Inc.
   3  * Copyright (C) 2006 Rusty Russell IBM Corporation
   4  *
   5  * Author: Michael S. Tsirkin <mst@redhat.com>
   6  *
   7  * Inspiration, some code, and most witty comments come from
   8  * Documentation/virtual/lguest/lguest.c, by Rusty Russell
   9  *
  10  * Generic code for virtio server in host kernel.
  11  */
  12
  13 #include <linux/eventfd.h>
  14 #include <linux/vhost.h>
  15 #include <linux/uio.h>
  16 #include <linux/mm.h>
  17 #include <linux/mmu_context.h>
  18 #include <linux/miscdevice.h>
  19 #include <linux/mutex.h>
  20 #include <linux/poll.h>
  21 #include <linux/file.h>
  22 #include <linux/highmem.h>
  23 #include <linux/slab.h>
  24 #include <linux/vmalloc.h>
  25 #include <linux/kthread.h>
  26 #include <linux/cgroup.h>
  27 #include <linux/module.h>
  28 #include <linux/sort.h>
  29 #include <linux/sched/mm.h>
  30 #include <linux/sched/signal.h>
  31 #include <linux/interval_tree_generic.h>
  32 #include <linux/nospec.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
 207         if (poll->wqh)
 208                 return 0;
 209
 210         mask = vfs_poll(file, &poll->table);
 211         if (mask)
 212                 vhost_poll_wakeup(&poll->wait, 0, 0, poll_to_key(mask));
 213         if (mask & EPOLLERR) {
 214                 vhost_poll_stop(poll);
 215                 return -EINVAL;
 216         }
 217
 218         return 0;
 219 }
 220 EXPORT_SYMBOL_GPL(vhost_poll_start);
 221
 222 /* Stop polling a file. After this function returns, it becomes safe to drop the
 223  * file reference. You must also flush afterwards. */
 224 void vhost_poll_stop(struct vhost_poll *poll)
 225 {
 226         if (poll->wqh) {
 227                 remove_wait_queue(poll->wqh, &poll->wait);
 228                 poll->wqh = NULL;
 229         }
 230 }
 231 EXPORT_SYMBOL_GPL(vhost_poll_stop);
 232
 233 void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work)
 234 {
 235         struct vhost_flush_struct flush;
 236
 237         if (dev->worker) {
 238                 init_completion(&flush.wait_event);
 239                 vhost_work_init(&flush.work, vhost_flush_work);
 240
 241                 vhost_work_queue(dev, &flush.work);
 242                 wait_for_completion(&flush.wait_event);
 243         }
 244 }
 245 EXPORT_SYMBOL_GPL(vhost_work_flush);
 246
 247 /* Flush any work that has been scheduled. When calling this, don't hold any
 248  * locks that are also used by the callback. */
 249 void vhost_poll_flush(struct vhost_poll *poll)
 250 {
 251         vhost_work_flush(poll->dev, &poll->work);
 252 }
 253 EXPORT_SYMBOL_GPL(vhost_poll_flush);
 254
 255 void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work)
 256 {
 257         if (!dev->worker)
 258                 return;
 259
 260         if (!test_and_set_bit(VHOST_WORK_QUEUED, &work->flags)) {
 261                 /* We can only add the work to the list after we're
 262                  * sure it was not in the list.
 263                  * test_and_set_bit() implies a memory barrier.
 264                  */
 265                 llist_add(&work->node, &dev->work_list);
 266                 wake_up_process(dev->worker);
 267         }
 268 }
 269 EXPORT_SYMBOL_GPL(vhost_work_queue);
 270
 271 /* A lockless hint for busy polling code to exit the loop */
 272 bool vhost_has_work(struct vhost_dev *dev)
 273 {
 274         return !llist_empty(&dev->work_list);
 275 }
 276 EXPORT_SYMBOL_GPL(vhost_has_work);
 277
 278 void vhost_poll_queue(struct vhost_poll *poll)
 279 {
 280         vhost_work_queue(poll->dev, &poll->work);
 281 }
 282 EXPORT_SYMBOL_GPL(vhost_poll_queue);
 283
 284 static void __vhost_vq_meta_reset(struct vhost_virtqueue *vq)
 285 {
 286         int j;
 287
 288         for (j = 0; j < VHOST_NUM_ADDRS; j++)
 289                 vq->meta_iotlb[j] = NULL;
 290 }
 291
 292 static void vhost_vq_meta_reset(struct vhost_dev *d)
 293 {
 294         int i;
 295
 296         for (i = 0; i < d->nvqs; ++i)
 297                 __vhost_vq_meta_reset(d->vqs[i]);
 298 }
 299
 300 static void vhost_vq_reset(struct vhost_dev *dev,
 301                            struct vhost_virtqueue *vq)
 302 {
 303         vq->num = 1;
 304         vq->desc = NULL;
 305         vq->avail = NULL;
 306         vq->used = NULL;
 307         vq->last_avail_idx = 0;
 308         vq->avail_idx = 0;
 309         vq->last_used_idx = 0;
 310         vq->signalled_used = 0;
 311         vq->signalled_used_valid = false;
 312         vq->used_flags = 0;
 313         vq->log_used = false;
 314         vq->log_addr = -1ull;
 315         vq->private_data = NULL;
 316         vq->acked_features = 0;
 317         vq->acked_backend_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(dev->iov_limit, sizeof(*vq->log),
 392                                         GFP_KERNEL);
 393                 vq->heads = kmalloc_array(dev->iov_limit, 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 bool vhost_exceeds_weight(struct vhost_virtqueue *vq,
 415                           int pkts, int total_len)
 416 {
 417         struct vhost_dev *dev = vq->dev;
 418
 419         if ((dev->byte_weight && total_len >= dev->byte_weight) ||
 420             pkts >= dev->weight) {
 421                 vhost_poll_queue(&vq->poll);
 422                 return true;
 423         }
 424
 425         return false;
 426 }
 427 EXPORT_SYMBOL_GPL(vhost_exceeds_weight);
 428
 429 static size_t vhost_get_avail_size(struct vhost_virtqueue *vq,
 430                                    unsigned int num)
 431 {
 432         size_t event __maybe_unused =
 433                vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
 434
 435         return sizeof(*vq->avail) +
 436                sizeof(*vq->avail->ring) * num + event;
 437 }
 438
 439 static size_t vhost_get_used_size(struct vhost_virtqueue *vq,
 440                                   unsigned int num)
 441 {
 442         size_t event __maybe_unused =
 443                vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
 444
 445         return sizeof(*vq->used) +
 446                sizeof(*vq->used->ring) * num + event;
 447 }
 448
 449 static size_t vhost_get_desc_size(struct vhost_virtqueue *vq,
 450                                   unsigned int num)
 451 {
 452         return sizeof(*vq->desc) * num;
 453 }
 454
 455 void vhost_dev_init(struct vhost_dev *dev,
 456                     struct vhost_virtqueue **vqs, int nvqs,
 457                     int iov_limit, int weight, int byte_weight)
 458 {
 459         struct vhost_virtqueue *vq;
 460         int i;
 461
 462         dev->vqs = vqs;
 463         dev->nvqs = nvqs;
 464         mutex_init(&dev->mutex);
 465         dev->log_ctx = NULL;
 466         dev->umem = NULL;
 467         dev->iotlb = NULL;
 468         dev->mm = NULL;
 469         dev->worker = NULL;
 470         dev->iov_limit = iov_limit;
 471         dev->weight = weight;
 472         dev->byte_weight = byte_weight;
 473         init_llist_head(&dev->work_list);
 474         init_waitqueue_head(&dev->wait);
 475         INIT_LIST_HEAD(&dev->read_list);
 476         INIT_LIST_HEAD(&dev->pending_list);
 477         spin_lock_init(&dev->iotlb_lock);
 478
 479
 480         for (i = 0; i < dev->nvqs; ++i) {
 481                 vq = dev->vqs[i];
 482                 vq->log = NULL;
 483                 vq->indirect = NULL;
 484                 vq->heads = NULL;
 485                 vq->dev = dev;
 486                 mutex_init(&vq->mutex);
 487                 vhost_vq_reset(dev, vq);
 488                 if (vq->handle_kick)
 489                         vhost_poll_init(&vq->poll, vq->handle_kick,
 490                                         EPOLLIN, dev);
 491         }
 492 }
 493 EXPORT_SYMBOL_GPL(vhost_dev_init);
 494
 495 /* Caller should have device mutex */
 496 long vhost_dev_check_owner(struct vhost_dev *dev)
 497 {
 498         /* Are you the owner? If not, I don't think you mean to do that */
 499         return dev->mm == current->mm ? 0 : -EPERM;
 500 }
 501 EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
 502
 503 struct vhost_attach_cgroups_struct {
 504         struct vhost_work work;
 505         struct task_struct *owner;
 506         int ret;
 507 };
 508
 509 static void vhost_attach_cgroups_work(struct vhost_work *work)
 510 {
 511         struct vhost_attach_cgroups_struct *s;
 512
 513         s = container_of(work, struct vhost_attach_cgroups_struct, work);
 514         s->ret = cgroup_attach_task_all(s->owner, current);
 515 }
 516
 517 static int vhost_attach_cgroups(struct vhost_dev *dev)
 518 {
 519         struct vhost_attach_cgroups_struct attach;
 520
 521         attach.owner = current;
 522         vhost_work_init(&attach.work, vhost_attach_cgroups_work);
 523         vhost_work_queue(dev, &attach.work);
 524         vhost_work_flush(dev, &attach.work);
 525         return attach.ret;
 526 }
 527
 528 /* Caller should have device mutex */
 529 bool vhost_dev_has_owner(struct vhost_dev *dev)
 530 {
 531         return dev->mm;
 532 }
 533 EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
 534
 535 /* Caller should have device mutex */
 536 long vhost_dev_set_owner(struct vhost_dev *dev)
 537 {
 538         struct task_struct *worker;
 539         int err;
 540
 541         /* Is there an owner already? */
 542         if (vhost_dev_has_owner(dev)) {
 543                 err = -EBUSY;
 544                 goto err_mm;
 545         }
 546
 547         /* No owner, become one */
 548         dev->mm = get_task_mm(current);
 549         worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
 550         if (IS_ERR(worker)) {
 551                 err = PTR_ERR(worker);
 552                 goto err_worker;
 553         }
 554
 555         dev->worker = worker;
 556         wake_up_process(worker);        /* avoid contributing to loadavg */
 557
 558         err = vhost_attach_cgroups(dev);
 559         if (err)
 560                 goto err_cgroup;
 561
 562         err = vhost_dev_alloc_iovecs(dev);
 563         if (err)
 564                 goto err_cgroup;
 565
 566         return 0;
 567 err_cgroup:
 568         kthread_stop(worker);
 569         dev->worker = NULL;
 570 err_worker:
 571         if (dev->mm)
 572                 mmput(dev->mm);
 573         dev->mm = NULL;
 574 err_mm:
 575         return err;
 576 }
 577 EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
 578
 579 struct vhost_umem *vhost_dev_reset_owner_prepare(void)
 580 {
 581         return kvzalloc(sizeof(struct vhost_umem), GFP_KERNEL);
 582 }
 583 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
 584
 585 /* Caller should have device mutex */
 586 void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_umem *umem)
 587 {
 588         int i;
 589
 590         vhost_dev_cleanup(dev);
 591
 592         /* Restore memory to default empty mapping. */
 593         INIT_LIST_HEAD(&umem->umem_list);
 594         dev->umem = umem;
 595         /* We don't need VQ locks below since vhost_dev_cleanup makes sure
 596          * VQs aren't running.
 597          */
 598         for (i = 0; i < dev->nvqs; ++i)
 599                 dev->vqs[i]->umem = umem;
 600 }
 601 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
 602
 603 void vhost_dev_stop(struct vhost_dev *dev)
 604 {
 605         int i;
 606
 607         for (i = 0; i < dev->nvqs; ++i) {
 608                 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick) {
 609                         vhost_poll_stop(&dev->vqs[i]->poll);
 610                         vhost_poll_flush(&dev->vqs[i]->poll);
 611                 }
 612         }
 613 }
 614 EXPORT_SYMBOL_GPL(vhost_dev_stop);
 615
 616 static void vhost_umem_free(struct vhost_umem *umem,
 617                             struct vhost_umem_node *node)
 618 {
 619         vhost_umem_interval_tree_remove(node, &umem->umem_tree);
 620         list_del(&node->link);
 621         kfree(node);
 622         umem->numem--;
 623 }
 624
 625 static void vhost_umem_clean(struct vhost_umem *umem)
 626 {
 627         struct vhost_umem_node *node, *tmp;
 628
 629         if (!umem)
 630                 return;
 631
 632         list_for_each_entry_safe(node, tmp, &umem->umem_list, link)
 633                 vhost_umem_free(umem, node);
 634
 635         kvfree(umem);
 636 }
 637
 638 static void vhost_clear_msg(struct vhost_dev *dev)
 639 {
 640         struct vhost_msg_node *node, *n;
 641
 642         spin_lock(&dev->iotlb_lock);
 643
 644         list_for_each_entry_safe(node, n, &dev->read_list, node) {
 645                 list_del(&node->node);
 646                 kfree(node);
 647         }
 648
 649         list_for_each_entry_safe(node, n, &dev->pending_list, node) {
 650                 list_del(&node->node);
 651                 kfree(node);
 652         }
 653
 654         spin_unlock(&dev->iotlb_lock);
 655 }
 656
 657 void vhost_dev_cleanup(struct vhost_dev *dev)
 658 {
 659         int i;
 660
 661         for (i = 0; i < dev->nvqs; ++i) {
 662                 if (dev->vqs[i]->error_ctx)
 663                         eventfd_ctx_put(dev->vqs[i]->error_ctx);
 664                 if (dev->vqs[i]->kick)
 665                         fput(dev->vqs[i]->kick);
 666                 if (dev->vqs[i]->call_ctx)
 667                         eventfd_ctx_put(dev->vqs[i]->call_ctx);
 668                 vhost_vq_reset(dev, dev->vqs[i]);
 669         }
 670         vhost_dev_free_iovecs(dev);
 671         if (dev->log_ctx)
 672                 eventfd_ctx_put(dev->log_ctx);
 673         dev->log_ctx = NULL;
 674         /* No one will access memory at this point */
 675         vhost_umem_clean(dev->umem);
 676         dev->umem = NULL;
 677         vhost_umem_clean(dev->iotlb);
 678         dev->iotlb = NULL;
 679         vhost_clear_msg(dev);
 680         wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
 681         WARN_ON(!llist_empty(&dev->work_list));
 682         if (dev->worker) {
 683                 kthread_stop(dev->worker);
 684                 dev->worker = NULL;
 685         }
 686         if (dev->mm)
 687                 mmput(dev->mm);
 688         dev->mm = NULL;
 689 }
 690 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
 691
 692 static bool log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
 693 {
 694         u64 a = addr / VHOST_PAGE_SIZE / 8;
 695
 696         /* Make sure 64 bit math will not overflow. */
 697         if (a > ULONG_MAX - (unsigned long)log_base ||
 698             a + (unsigned long)log_base > ULONG_MAX)
 699                 return false;
 700
 701         return access_ok(log_base + a,
 702                          (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
 703 }
 704
 705 static bool vhost_overflow(u64 uaddr, u64 size)
 706 {
 707         /* Make sure 64 bit math will not overflow. */
 708         return uaddr > ULONG_MAX || size > ULONG_MAX || uaddr > ULONG_MAX - size;
 709 }
 710
 711 /* Caller should have vq mutex and device mutex. */
 712 static bool vq_memory_access_ok(void __user *log_base, struct vhost_umem *umem,
 713                                 int log_all)
 714 {
 715         struct vhost_umem_node *node;
 716
 717         if (!umem)
 718                 return false;
 719
 720         list_for_each_entry(node, &umem->umem_list, link) {
 721                 unsigned long a = node->userspace_addr;
 722
 723                 if (vhost_overflow(node->userspace_addr, node->size))
 724                         return false;
 725
 726
 727                 if (!access_ok((void __user *)a,
 728                                     node->size))
 729                         return false;
 730                 else if (log_all && !log_access_ok(log_base,
 731                                                    node->start,
 732                                                    node->size))
 733                         return false;
 734         }
 735         return true;
 736 }
 737
 738 static inline void __user *vhost_vq_meta_fetch(struct vhost_virtqueue *vq,
 739                                                u64 addr, unsigned int size,
 740                                                int type)
 741 {
 742         const struct vhost_umem_node *node = vq->meta_iotlb[type];
 743
 744         if (!node)
 745                 return NULL;
 746
 747         return (void *)(uintptr_t)(node->userspace_addr + addr - node->start);
 748 }
 749
 750 /* Can we switch to this memory table? */
 751 /* Caller should have device mutex but not vq mutex */
 752 static bool memory_access_ok(struct vhost_dev *d, struct vhost_umem *umem,
 753                              int log_all)
 754 {
 755         int i;
 756
 757         for (i = 0; i < d->nvqs; ++i) {
 758                 bool ok;
 759                 bool log;
 760
 761                 mutex_lock(&d->vqs[i]->mutex);
 762                 log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
 763                 /* If ring is inactive, will check when it's enabled. */
 764                 if (d->vqs[i]->private_data)
 765                         ok = vq_memory_access_ok(d->vqs[i]->log_base,
 766                                                  umem, log);
 767                 else
 768                         ok = true;
 769                 mutex_unlock(&d->vqs[i]->mutex);
 770                 if (!ok)
 771                         return false;
 772         }
 773         return true;
 774 }
 775
 776 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
 777                           struct iovec iov[], int iov_size, int access);
 778
 779 static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
 780                               const void *from, unsigned size)
 781 {
 782         int ret;
 783
 784         if (!vq->iotlb)
 785                 return __copy_to_user(to, from, size);
 786         else {
 787                 /* This function should be called after iotlb
 788                  * prefetch, which means we're sure that all vq
 789                  * could be access through iotlb. So -EAGAIN should
 790                  * not happen in this case.
 791                  */
 792                 struct iov_iter t;
 793                 void __user *uaddr = vhost_vq_meta_fetch(vq,
 794                                      (u64)(uintptr_t)to, size,
 795                                      VHOST_ADDR_USED);
 796
 797                 if (uaddr)
 798                         return __copy_to_user(uaddr, from, size);
 799
 800                 ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
 801                                      ARRAY_SIZE(vq->iotlb_iov),
 802                                      VHOST_ACCESS_WO);
 803                 if (ret < 0)
 804                         goto out;
 805                 iov_iter_init(&t, WRITE, vq->iotlb_iov, ret, size);
 806                 ret = copy_to_iter(from, size, &t);
 807                 if (ret == size)
 808                         ret = 0;
 809         }
 810 out:
 811         return ret;
 812 }
 813
 814 static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
 815                                 void __user *from, unsigned size)
 816 {
 817         int ret;
 818
 819         if (!vq->iotlb)
 820                 return __copy_from_user(to, from, size);
 821         else {
 822                 /* This function should be called after iotlb
 823                  * prefetch, which means we're sure that vq
 824                  * could be access through iotlb. So -EAGAIN should
 825                  * not happen in this case.
 826                  */
 827                 void __user *uaddr = vhost_vq_meta_fetch(vq,
 828                                      (u64)(uintptr_t)from, size,
 829                                      VHOST_ADDR_DESC);
 830                 struct iov_iter f;
 831
 832                 if (uaddr)
 833                         return __copy_from_user(to, uaddr, size);
 834
 835                 ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
 836                                      ARRAY_SIZE(vq->iotlb_iov),
 837                                      VHOST_ACCESS_RO);
 838                 if (ret < 0) {
 839                         vq_err(vq, "IOTLB translation failure: uaddr "
 840                                "%p size 0x%llx\n", from,
 841                                (unsigned long long) size);
 842                         goto out;
 843                 }
 844                 iov_iter_init(&f, READ, vq->iotlb_iov, ret, size);
 845                 ret = copy_from_iter(to, size, &f);
 846                 if (ret == size)
 847                         ret = 0;
 848         }
 849
 850 out:
 851         return ret;
 852 }
 853
 854 static void __user *__vhost_get_user_slow(struct vhost_virtqueue *vq,
 855                                           void __user *addr, unsigned int size,
 856                                           int type)
 857 {
 858         int ret;
 859
 860         ret = translate_desc(vq, (u64)(uintptr_t)addr, size, vq->iotlb_iov,
 861                              ARRAY_SIZE(vq->iotlb_iov),
 862                              VHOST_ACCESS_RO);
 863         if (ret < 0) {
 864                 vq_err(vq, "IOTLB translation failure: uaddr "
 865                         "%p size 0x%llx\n", addr,
 866                         (unsigned long long) size);
 867                 return NULL;
 868         }
 869
 870         if (ret != 1 || vq->iotlb_iov[0].iov_len != size) {
 871                 vq_err(vq, "Non atomic userspace memory access: uaddr "
 872                         "%p size 0x%llx\n", addr,
 873                         (unsigned long long) size);
 874                 return NULL;
 875         }
 876
 877         return vq->iotlb_iov[0].iov_base;
 878 }
 879
 880 /* This function should be called after iotlb
 881  * prefetch, which means we're sure that vq
 882  * could be access through iotlb. So -EAGAIN should
 883  * not happen in this case.
 884  */
 885 static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
 886                                             void *addr, unsigned int size,
 887                                             int type)
 888 {
 889         void __user *uaddr = vhost_vq_meta_fetch(vq,
 890                              (u64)(uintptr_t)addr, size, type);
 891         if (uaddr)
 892                 return uaddr;
 893
 894         return __vhost_get_user_slow(vq, addr, size, type);
 895 }
 896
 897 #define vhost_put_user(vq, x, ptr)              \
 898 ({ \
 899         int ret = -EFAULT; \
 900         if (!vq->iotlb) { \
 901                 ret = __put_user(x, ptr); \
 902         } else { \
 903                 __typeof__(ptr) to = \
 904                         (__typeof__(ptr)) __vhost_get_user(vq, ptr,     \
 905                                           sizeof(*ptr), VHOST_ADDR_USED); \
 906                 if (to != NULL) \
 907                         ret = __put_user(x, to); \
 908                 else \
 909                         ret = -EFAULT;  \
 910         } \
 911         ret; \
 912 })
 913
 914 static inline int vhost_put_avail_event(struct vhost_virtqueue *vq)
 915 {
 916         return vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
 917                               vhost_avail_event(vq));
 918 }
 919
 920 static inline int vhost_put_used(struct vhost_virtqueue *vq,
 921                                  struct vring_used_elem *head, int idx,
 922                                  int count)
 923 {
 924         return vhost_copy_to_user(vq, vq->used->ring + idx, head,
 925                                   count * sizeof(*head));
 926 }
 927
 928 static inline int vhost_put_used_flags(struct vhost_virtqueue *vq)
 929
 930 {
 931         return vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
 932                               &vq->used->flags);
 933 }
 934
 935 static inline int vhost_put_used_idx(struct vhost_virtqueue *vq)
 936
 937 {
 938         return vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
 939                               &vq->used->idx);
 940 }
 941
 942 #define vhost_get_user(vq, x, ptr, type)                \
 943 ({ \
 944         int ret; \
 945         if (!vq->iotlb) { \
 946                 ret = __get_user(x, ptr); \
 947         } else { \
 948                 __typeof__(ptr) from = \
 949                         (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
 950                                                            sizeof(*ptr), \
 951                                                            type); \
 952                 if (from != NULL) \
 953                         ret = __get_user(x, from); \
 954                 else \
 955                         ret = -EFAULT; \
 956         } \
 957         ret; \
 958 })
 959
 960 #define vhost_get_avail(vq, x, ptr) \
 961         vhost_get_user(vq, x, ptr, VHOST_ADDR_AVAIL)
 962
 963 #define vhost_get_used(vq, x, ptr) \
 964         vhost_get_user(vq, x, ptr, VHOST_ADDR_USED)
 965
 966 static void vhost_dev_lock_vqs(struct vhost_dev *d)
 967 {
 968         int i = 0;
 969         for (i = 0; i < d->nvqs; ++i)
 970                 mutex_lock_nested(&d->vqs[i]->mutex, i);
 971 }
 972
 973 static void vhost_dev_unlock_vqs(struct vhost_dev *d)
 974 {
 975         int i = 0;
 976         for (i = 0; i < d->nvqs; ++i)
 977                 mutex_unlock(&d->vqs[i]->mutex);
 978 }
 979
 980 static inline int vhost_get_avail_idx(struct vhost_virtqueue *vq,
 981                                       __virtio16 *idx)
 982 {
 983         return vhost_get_avail(vq, *idx, &vq->avail->idx);
 984 }
 985
 986 static inline int vhost_get_avail_head(struct vhost_virtqueue *vq,
 987                                        __virtio16 *head, int idx)
 988 {
 989         return vhost_get_avail(vq, *head,
 990                                &vq->avail->ring[idx & (vq->num - 1)]);
 991 }
 992
 993 static inline int vhost_get_avail_flags(struct vhost_virtqueue *vq,
 994                                         __virtio16 *flags)
 995 {
 996         return vhost_get_avail(vq, *flags, &vq->avail->flags);
 997 }
 998
 999 static inline int vhost_get_used_event(struct vhost_virtqueue *vq,
1000                                        __virtio16 *event)
1001 {
1002         return vhost_get_avail(vq, *event, vhost_used_event(vq));
1003 }
1004
1005 static inline int vhost_get_used_idx(struct vhost_virtqueue *vq,
1006                                      __virtio16 *idx)
1007 {
1008         return vhost_get_used(vq, *idx, &vq->used->idx);
1009 }
1010
1011 static inline int vhost_get_desc(struct vhost_virtqueue *vq,
1012                                  struct vring_desc *desc, int idx)
1013 {
1014         return vhost_copy_from_user(vq, desc, vq->desc + idx, sizeof(*desc));
1015 }
1016
1017 static int vhost_new_umem_range(struct vhost_umem *umem,
1018                                 u64 start, u64 size, u64 end,
1019                                 u64 userspace_addr, int perm)
1020 {
1021         struct vhost_umem_node *tmp, *node;
1022
1023         if (!size)
1024                 return -EFAULT;
1025
1026         node = kmalloc(sizeof(*node), GFP_ATOMIC);
1027         if (!node)
1028                 return -ENOMEM;
1029
1030         if (umem->numem == max_iotlb_entries) {
1031                 tmp = list_first_entry(&umem->umem_list, typeof(*tmp), link);
1032                 vhost_umem_free(umem, tmp);
1033         }
1034
1035         node->start = start;
1036         node->size = size;
1037         node->last = end;
1038         node->userspace_addr = userspace_addr;
1039         node->perm = perm;
1040         INIT_LIST_HEAD(&node->link);
1041         list_add_tail(&node->link, &umem->umem_list);
1042         vhost_umem_interval_tree_insert(node, &umem->umem_tree);
1043         umem->numem++;
1044
1045         return 0;
1046 }
1047
1048 static void vhost_del_umem_range(struct vhost_umem *umem,
1049                                  u64 start, u64 end)
1050 {
1051         struct vhost_umem_node *node;
1052
1053         while ((node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1054                                                            start, end)))
1055                 vhost_umem_free(umem, node);
1056 }
1057
1058 static void vhost_iotlb_notify_vq(struct vhost_dev *d,
1059                                   struct vhost_iotlb_msg *msg)
1060 {
1061         struct vhost_msg_node *node, *n;
1062
1063         spin_lock(&d->iotlb_lock);
1064
1065         list_for_each_entry_safe(node, n, &d->pending_list, node) {
1066                 struct vhost_iotlb_msg *vq_msg = &node->msg.iotlb;
1067                 if (msg->iova <= vq_msg->iova &&
1068                     msg->iova + msg->size - 1 >= vq_msg->iova &&
1069                     vq_msg->type == VHOST_IOTLB_MISS) {
1070                         vhost_poll_queue(&node->vq->poll);
1071                         list_del(&node->node);
1072                         kfree(node);
1073                 }
1074         }
1075
1076         spin_unlock(&d->iotlb_lock);
1077 }
1078
1079 static bool umem_access_ok(u64 uaddr, u64 size, int access)
1080 {
1081         unsigned long a = uaddr;
1082
1083         /* Make sure 64 bit math will not overflow. */
1084         if (vhost_overflow(uaddr, size))
1085                 return false;
1086
1087         if ((access & VHOST_ACCESS_RO) &&
1088             !access_ok((void __user *)a, size))
1089                 return false;
1090         if ((access & VHOST_ACCESS_WO) &&
1091             !access_ok((void __user *)a, size))
1092                 return false;
1093         return true;
1094 }
1095
1096 static int vhost_process_iotlb_msg(struct vhost_dev *dev,
1097                                    struct vhost_iotlb_msg *msg)
1098 {
1099         int ret = 0;
1100
1101         mutex_lock(&dev->mutex);
1102         vhost_dev_lock_vqs(dev);
1103         switch (msg->type) {
1104         case VHOST_IOTLB_UPDATE:
1105                 if (!dev->iotlb) {
1106                         ret = -EFAULT;
1107                         break;
1108                 }
1109                 if (!umem_access_ok(msg->uaddr, msg->size, msg->perm)) {
1110                         ret = -EFAULT;
1111                         break;
1112                 }
1113                 vhost_vq_meta_reset(dev);
1114                 if (vhost_new_umem_range(dev->iotlb, msg->iova, msg->size,
1115                                          msg->iova + msg->size - 1,
1116                                          msg->uaddr, msg->perm)) {
1117                         ret = -ENOMEM;
1118                         break;
1119                 }
1120                 vhost_iotlb_notify_vq(dev, msg);
1121                 break;
1122         case VHOST_IOTLB_INVALIDATE:
1123                 if (!dev->iotlb) {
1124                         ret = -EFAULT;
1125                         break;
1126                 }
1127                 vhost_vq_meta_reset(dev);
1128                 vhost_del_umem_range(dev->iotlb, msg->iova,
1129                                      msg->iova + msg->size - 1);
1130                 break;
1131         default:
1132                 ret = -EINVAL;
1133                 break;
1134         }
1135
1136         vhost_dev_unlock_vqs(dev);
1137         mutex_unlock(&dev->mutex);
1138
1139         return ret;
1140 }
1141 ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
1142                              struct iov_iter *from)
1143 {
1144         struct vhost_iotlb_msg msg;
1145         size_t offset;
1146         int type, ret;
1147
1148         ret = copy_from_iter(&type, sizeof(type), from);
1149         if (ret != sizeof(type)) {
1150                 ret = -EINVAL;
1151                 goto done;
1152         }
1153
1154         switch (type) {
1155         case VHOST_IOTLB_MSG:
1156                 /* There maybe a hole after type for V1 message type,
1157                  * so skip it here.
1158                  */
1159                 offset = offsetof(struct vhost_msg, iotlb) - sizeof(int);
1160                 break;
1161         case VHOST_IOTLB_MSG_V2:
1162                 offset = sizeof(__u32);
1163                 break;
1164         default:
1165                 ret = -EINVAL;
1166                 goto done;
1167         }
1168
1169         iov_iter_advance(from, offset);
1170         ret = copy_from_iter(&msg, sizeof(msg), from);
1171         if (ret != sizeof(msg)) {
1172                 ret = -EINVAL;
1173                 goto done;
1174         }
1175         if (vhost_process_iotlb_msg(dev, &msg)) {
1176                 ret = -EFAULT;
1177                 goto done;
1178         }
1179
1180         ret = (type == VHOST_IOTLB_MSG) ? sizeof(struct vhost_msg) :
1181               sizeof(struct vhost_msg_v2);
1182 done:
1183         return ret;
1184 }
1185 EXPORT_SYMBOL(vhost_chr_write_iter);
1186
1187 __poll_t vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1188                             poll_table *wait)
1189 {
1190         __poll_t mask = 0;
1191
1192         poll_wait(file, &dev->wait, wait);
1193
1194         if (!list_empty(&dev->read_list))
1195                 mask |= EPOLLIN | EPOLLRDNORM;
1196
1197         return mask;
1198 }
1199 EXPORT_SYMBOL(vhost_chr_poll);
1200
1201 ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1202                             int noblock)
1203 {
1204         DEFINE_WAIT(wait);
1205         struct vhost_msg_node *node;
1206         ssize_t ret = 0;
1207         unsigned size = sizeof(struct vhost_msg);
1208
1209         if (iov_iter_count(to) < size)
1210                 return 0;
1211
1212         while (1) {
1213                 if (!noblock)
1214                         prepare_to_wait(&dev->wait, &wait,
1215                                         TASK_INTERRUPTIBLE);
1216
1217                 node = vhost_dequeue_msg(dev, &dev->read_list);
1218                 if (node)
1219                         break;
1220                 if (noblock) {
1221                         ret = -EAGAIN;
1222                         break;
1223                 }
1224                 if (signal_pending(current)) {
1225                         ret = -ERESTARTSYS;
1226                         break;
1227                 }
1228                 if (!dev->iotlb) {
1229                         ret = -EBADFD;
1230                         break;
1231                 }
1232
1233                 schedule();
1234         }
1235
1236         if (!noblock)
1237                 finish_wait(&dev->wait, &wait);
1238
1239         if (node) {
1240                 struct vhost_iotlb_msg *msg;
1241                 void *start = &node->msg;
1242
1243                 switch (node->msg.type) {
1244                 case VHOST_IOTLB_MSG:
1245                         size = sizeof(node->msg);
1246                         msg = &node->msg.iotlb;
1247                         break;
1248                 case VHOST_IOTLB_MSG_V2:
1249                         size = sizeof(node->msg_v2);
1250                         msg = &node->msg_v2.iotlb;
1251                         break;
1252                 default:
1253                         BUG();
1254                         break;
1255                 }
1256
1257                 ret = copy_to_iter(start, size, to);
1258                 if (ret != size || msg->type != VHOST_IOTLB_MISS) {
1259                         kfree(node);
1260                         return ret;
1261                 }
1262                 vhost_enqueue_msg(dev, &dev->pending_list, node);
1263         }
1264
1265         return ret;
1266 }
1267 EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1268
1269 static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1270 {
1271         struct vhost_dev *dev = vq->dev;
1272         struct vhost_msg_node *node;
1273         struct vhost_iotlb_msg *msg;
1274         bool v2 = vhost_backend_has_feature(vq, VHOST_BACKEND_F_IOTLB_MSG_V2);
1275
1276         node = vhost_new_msg(vq, v2 ? VHOST_IOTLB_MSG_V2 : VHOST_IOTLB_MSG);
1277         if (!node)
1278                 return -ENOMEM;
1279
1280         if (v2) {
1281                 node->msg_v2.type = VHOST_IOTLB_MSG_V2;
1282                 msg = &node->msg_v2.iotlb;
1283         } else {
1284                 msg = &node->msg.iotlb;
1285         }
1286
1287         msg->type = VHOST_IOTLB_MISS;
1288         msg->iova = iova;
1289         msg->perm = access;
1290
1291         vhost_enqueue_msg(dev, &dev->read_list, node);
1292
1293         return 0;
1294 }
1295
1296 static bool vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1297                          struct vring_desc __user *desc,
1298                          struct vring_avail __user *avail,
1299                          struct vring_used __user *used)
1300
1301 {
1302         return access_ok(desc, vhost_get_desc_size(vq, num)) &&
1303                access_ok(avail, vhost_get_avail_size(vq, num)) &&
1304                access_ok(used, vhost_get_used_size(vq, num));
1305 }
1306
1307 static void vhost_vq_meta_update(struct vhost_virtqueue *vq,
1308                                  const struct vhost_umem_node *node,
1309                                  int type)
1310 {
1311         int access = (type == VHOST_ADDR_USED) ?
1312                      VHOST_ACCESS_WO : VHOST_ACCESS_RO;
1313
1314         if (likely(node->perm & access))
1315                 vq->meta_iotlb[type] = node;
1316 }
1317
1318 static bool iotlb_access_ok(struct vhost_virtqueue *vq,
1319                             int access, u64 addr, u64 len, int type)
1320 {
1321         const struct vhost_umem_node *node;
1322         struct vhost_umem *umem = vq->iotlb;
1323         u64 s = 0, size, orig_addr = addr, last = addr + len - 1;
1324
1325         if (vhost_vq_meta_fetch(vq, addr, len, type))
1326                 return true;
1327
1328         while (len > s) {
1329                 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1330                                                            addr,
1331                                                            last);
1332                 if (node == NULL || node->start > addr) {
1333                         vhost_iotlb_miss(vq, addr, access);
1334                         return false;
1335                 } else if (!(node->perm & access)) {
1336                         /* Report the possible access violation by
1337                          * request another translation from userspace.
1338                          */
1339                         return false;
1340                 }
1341
1342                 size = node->size - addr + node->start;
1343
1344                 if (orig_addr == addr && size >= len)
1345                         vhost_vq_meta_update(vq, node, type);
1346
1347                 s += size;
1348                 addr += size;
1349         }
1350
1351         return true;
1352 }
1353
1354 int vq_meta_prefetch(struct vhost_virtqueue *vq)
1355 {
1356         unsigned int num = vq->num;
1357
1358         if (!vq->iotlb)
1359                 return 1;
1360
1361         return iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->desc,
1362                                vhost_get_desc_size(vq, num), VHOST_ADDR_DESC) &&
1363                iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->avail,
1364                                vhost_get_avail_size(vq, num),
1365                                VHOST_ADDR_AVAIL) &&
1366                iotlb_access_ok(vq, VHOST_ACCESS_WO, (u64)(uintptr_t)vq->used,
1367                                vhost_get_used_size(vq, num), VHOST_ADDR_USED);
1368 }
1369 EXPORT_SYMBOL_GPL(vq_meta_prefetch);
1370
1371 /* Can we log writes? */
1372 /* Caller should have device mutex but not vq mutex */
1373 bool vhost_log_access_ok(struct vhost_dev *dev)
1374 {
1375         return memory_access_ok(dev, dev->umem, 1);
1376 }
1377 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1378
1379 /* Verify access for write logging. */
1380 /* Caller should have vq mutex and device mutex */
1381 static bool vq_log_access_ok(struct vhost_virtqueue *vq,
1382                              void __user *log_base)
1383 {
1384         return vq_memory_access_ok(log_base, vq->umem,
1385                                    vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1386                 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
1387                                   vhost_get_used_size(vq, vq->num)));
1388 }
1389
1390 /* Can we start vq? */
1391 /* Caller should have vq mutex and device mutex */
1392 bool vhost_vq_access_ok(struct vhost_virtqueue *vq)
1393 {
1394         if (!vq_log_access_ok(vq, vq->log_base))
1395                 return false;
1396
1397         /* Access validation occurs at prefetch time with IOTLB */
1398         if (vq->iotlb)
1399                 return true;
1400
1401         return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used);
1402 }
1403 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1404
1405 static struct vhost_umem *vhost_umem_alloc(void)
1406 {
1407         struct vhost_umem *umem = kvzalloc(sizeof(*umem), GFP_KERNEL);
1408
1409         if (!umem)
1410                 return NULL;
1411
1412         umem->umem_tree = RB_ROOT_CACHED;
1413         umem->numem = 0;
1414         INIT_LIST_HEAD(&umem->umem_list);
1415
1416         return umem;
1417 }
1418
1419 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1420 {
1421         struct vhost_memory mem, *newmem;
1422         struct vhost_memory_region *region;
1423         struct vhost_umem *newumem, *oldumem;
1424         unsigned long size = offsetof(struct vhost_memory, regions);
1425         int i;
1426
1427         if (copy_from_user(&mem, m, size))
1428                 return -EFAULT;
1429         if (mem.padding)
1430                 return -EOPNOTSUPP;
1431         if (mem.nregions > max_mem_regions)
1432                 return -E2BIG;
1433         newmem = kvzalloc(struct_size(newmem, regions, mem.nregions),
1434                         GFP_KERNEL);
1435         if (!newmem)
1436                 return -ENOMEM;
1437
1438         memcpy(newmem, &mem, size);
1439         if (copy_from_user(newmem->regions, m->regions,
1440                            mem.nregions * sizeof *m->regions)) {
1441                 kvfree(newmem);
1442                 return -EFAULT;
1443         }
1444
1445         newumem = vhost_umem_alloc();
1446         if (!newumem) {
1447                 kvfree(newmem);
1448                 return -ENOMEM;
1449         }
1450
1451         for (region = newmem->regions;
1452              region < newmem->regions + mem.nregions;
1453              region++) {
1454                 if (vhost_new_umem_range(newumem,
1455                                          region->guest_phys_addr,
1456                                          region->memory_size,
1457                                          region->guest_phys_addr +
1458                                          region->memory_size - 1,
1459                                          region->userspace_addr,
1460                                          VHOST_ACCESS_RW))
1461                         goto err;
1462         }
1463
1464         if (!memory_access_ok(d, newumem, 0))
1465                 goto err;
1466
1467         oldumem = d->umem;
1468         d->umem = newumem;
1469
1470         /* All memory accesses are done under some VQ mutex. */
1471         for (i = 0; i < d->nvqs; ++i) {
1472                 mutex_lock(&d->vqs[i]->mutex);
1473                 d->vqs[i]->umem = newumem;
1474                 mutex_unlock(&d->vqs[i]->mutex);
1475         }
1476
1477         kvfree(newmem);
1478         vhost_umem_clean(oldumem);
1479         return 0;
1480
1481 err:
1482         vhost_umem_clean(newumem);
1483         kvfree(newmem);
1484         return -EFAULT;
1485 }
1486
1487 static long vhost_vring_set_num(struct vhost_dev *d,
1488                                 struct vhost_virtqueue *vq,
1489                                 void __user *argp)
1490 {
1491         struct vhost_vring_state s;
1492
1493         /* Resizing ring with an active backend?
1494          * You don't want to do that. */
1495         if (vq->private_data)
1496                 return -EBUSY;
1497
1498         if (copy_from_user(&s, argp, sizeof s))
1499                 return -EFAULT;
1500
1501         if (!s.num || s.num > 0xffff || (s.num & (s.num - 1)))
1502                 return -EINVAL;
1503         vq->num = s.num;
1504
1505         return 0;
1506 }
1507
1508 static long vhost_vring_set_addr(struct vhost_dev *d,
1509                                  struct vhost_virtqueue *vq,
1510                                  void __user *argp)
1511 {
1512         struct vhost_vring_addr a;
1513
1514         if (copy_from_user(&a, argp, sizeof a))
1515                 return -EFAULT;
1516         if (a.flags & ~(0x1 << VHOST_VRING_F_LOG))
1517                 return -EOPNOTSUPP;
1518
1519         /* For 32bit, verify that the top 32bits of the user
1520            data are set to zero. */
1521         if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1522             (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1523             (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr)
1524                 return -EFAULT;
1525
1526         /* Make sure it's safe to cast pointers to vring types. */
1527         BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1528         BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1529         if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1530             (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1531             (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1)))
1532                 return -EINVAL;
1533
1534         /* We only verify access here if backend is configured.
1535          * If it is not, we don't as size might not have been setup.
1536          * We will verify when backend is configured. */
1537         if (vq->private_data) {
1538                 if (!vq_access_ok(vq, vq->num,
1539                         (void __user *)(unsigned long)a.desc_user_addr,
1540                         (void __user *)(unsigned long)a.avail_user_addr,
1541                         (void __user *)(unsigned long)a.used_user_addr))
1542                         return -EINVAL;
1543
1544                 /* Also validate log access for used ring if enabled. */
1545                 if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
1546                         !log_access_ok(vq->log_base, a.log_guest_addr,
1547                                 sizeof *vq->used +
1548                                 vq->num * sizeof *vq->used->ring))
1549                         return -EINVAL;
1550         }
1551
1552         vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1553         vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1554         vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1555         vq->log_addr = a.log_guest_addr;
1556         vq->used = (void __user *)(unsigned long)a.used_user_addr;
1557
1558         return 0;
1559 }
1560
1561 static long vhost_vring_set_num_addr(struct vhost_dev *d,
1562                                      struct vhost_virtqueue *vq,
1563                                      unsigned int ioctl,
1564                                      void __user *argp)
1565 {
1566         long r;
1567
1568         mutex_lock(&vq->mutex);
1569
1570         switch (ioctl) {
1571         case VHOST_SET_VRING_NUM:
1572                 r = vhost_vring_set_num(d, vq, argp);
1573                 break;
1574         case VHOST_SET_VRING_ADDR:
1575                 r = vhost_vring_set_addr(d, vq, argp);
1576                 break;
1577         default:
1578                 BUG();
1579         }
1580
1581         mutex_unlock(&vq->mutex);
1582
1583         return r;
1584 }
1585 long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1586 {
1587         struct file *eventfp, *filep = NULL;
1588         bool pollstart = false, pollstop = false;
1589         struct eventfd_ctx *ctx = NULL;
1590         u32 __user *idxp = argp;
1591         struct vhost_virtqueue *vq;
1592         struct vhost_vring_state s;
1593         struct vhost_vring_file f;
1594         u32 idx;
1595         long r;
1596
1597         r = get_user(idx, idxp);
1598         if (r < 0)
1599                 return r;
1600         if (idx >= d->nvqs)
1601                 return -ENOBUFS;
1602
1603         idx = array_index_nospec(idx, d->nvqs);
1604         vq = d->vqs[idx];
1605
1606         if (ioctl == VHOST_SET_VRING_NUM ||
1607             ioctl == VHOST_SET_VRING_ADDR) {
1608                 return vhost_vring_set_num_addr(d, vq, ioctl, argp);
1609         }
1610
1611         mutex_lock(&vq->mutex);
1612
1613         switch (ioctl) {
1614         case VHOST_SET_VRING_BASE:
1615                 /* Moving base with an active backend?
1616                  * You don't want to do that. */
1617                 if (vq->private_data) {
1618                         r = -EBUSY;
1619                         break;
1620                 }
1621                 if (copy_from_user(&s, argp, sizeof s)) {
1622                         r = -EFAULT;
1623                         break;
1624                 }
1625                 if (s.num > 0xffff) {
1626                         r = -EINVAL;
1627                         break;
1628                 }
1629                 vq->last_avail_idx = s.num;
1630                 /* Forget the cached index value. */
1631                 vq->avail_idx = vq->last_avail_idx;
1632                 break;
1633         case VHOST_GET_VRING_BASE:
1634                 s.index = idx;
1635                 s.num = vq->last_avail_idx;
1636                 if (copy_to_user(argp, &s, sizeof s))
1637                         r = -EFAULT;
1638                 break;
1639         case VHOST_SET_VRING_KICK:
1640                 if (copy_from_user(&f, argp, sizeof f)) {
1641                         r = -EFAULT;
1642                         break;
1643                 }
1644                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1645                 if (IS_ERR(eventfp)) {
1646                         r = PTR_ERR(eventfp);
1647                         break;
1648                 }
1649                 if (eventfp != vq->kick) {
1650                         pollstop = (filep = vq->kick) != NULL;
1651                         pollstart = (vq->kick = eventfp) != NULL;
1652                 } else
1653                         filep = eventfp;
1654                 break;
1655         case VHOST_SET_VRING_CALL:
1656                 if (copy_from_user(&f, argp, sizeof f)) {
1657                         r = -EFAULT;
1658                         break;
1659                 }
1660                 ctx = f.fd == -1 ? NULL : eventfd_ctx_fdget(f.fd);
1661                 if (IS_ERR(ctx)) {
1662                         r = PTR_ERR(ctx);
1663                         break;
1664                 }
1665                 swap(ctx, vq->call_ctx);
1666                 break;
1667         case VHOST_SET_VRING_ERR:
1668                 if (copy_from_user(&f, argp, sizeof f)) {
1669                         r = -EFAULT;
1670                         break;
1671                 }
1672                 ctx = f.fd == -1 ? NULL : eventfd_ctx_fdget(f.fd);
1673                 if (IS_ERR(ctx)) {
1674                         r = PTR_ERR(ctx);
1675                         break;
1676                 }
1677                 swap(ctx, vq->error_ctx);
1678                 break;
1679         case VHOST_SET_VRING_ENDIAN:
1680                 r = vhost_set_vring_endian(vq, argp);
1681                 break;
1682         case VHOST_GET_VRING_ENDIAN:
1683                 r = vhost_get_vring_endian(vq, idx, argp);
1684                 break;
1685         case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
1686                 if (copy_from_user(&s, argp, sizeof(s))) {
1687                         r = -EFAULT;
1688                         break;
1689                 }
1690                 vq->busyloop_timeout = s.num;
1691                 break;
1692         case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
1693                 s.index = idx;
1694                 s.num = vq->busyloop_timeout;
1695                 if (copy_to_user(argp, &s, sizeof(s)))
1696                         r = -EFAULT;
1697                 break;
1698         default:
1699                 r = -ENOIOCTLCMD;
1700         }
1701
1702         if (pollstop && vq->handle_kick)
1703                 vhost_poll_stop(&vq->poll);
1704
1705         if (!IS_ERR_OR_NULL(ctx))
1706                 eventfd_ctx_put(ctx);
1707         if (filep)
1708                 fput(filep);
1709
1710         if (pollstart && vq->handle_kick)
1711                 r = vhost_poll_start(&vq->poll, vq->kick);
1712
1713         mutex_unlock(&vq->mutex);
1714
1715         if (pollstop && vq->handle_kick)
1716                 vhost_poll_flush(&vq->poll);
1717         return r;
1718 }
1719 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
1720
1721 int vhost_init_device_iotlb(struct vhost_dev *d, bool enabled)
1722 {
1723         struct vhost_umem *niotlb, *oiotlb;
1724         int i;
1725
1726         niotlb = vhost_umem_alloc();
1727         if (!niotlb)
1728                 return -ENOMEM;
1729
1730         oiotlb = d->iotlb;
1731         d->iotlb = niotlb;
1732
1733         for (i = 0; i < d->nvqs; ++i) {
1734                 struct vhost_virtqueue *vq = d->vqs[i];
1735
1736                 mutex_lock(&vq->mutex);
1737                 vq->iotlb = niotlb;
1738                 __vhost_vq_meta_reset(vq);
1739                 mutex_unlock(&vq->mutex);
1740         }
1741
1742         vhost_umem_clean(oiotlb);
1743
1744         return 0;
1745 }
1746 EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
1747
1748 /* Caller must have device mutex */
1749 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1750 {
1751         struct eventfd_ctx *ctx;
1752         u64 p;
1753         long r;
1754         int i, fd;
1755
1756         /* If you are not the owner, you can become one */
1757         if (ioctl == VHOST_SET_OWNER) {
1758                 r = vhost_dev_set_owner(d);
1759                 goto done;
1760         }
1761
1762         /* You must be the owner to do anything else */
1763         r = vhost_dev_check_owner(d);
1764         if (r)
1765                 goto done;
1766
1767         switch (ioctl) {
1768         case VHOST_SET_MEM_TABLE:
1769                 r = vhost_set_memory(d, argp);
1770                 break;
1771         case VHOST_SET_LOG_BASE:
1772                 if (copy_from_user(&p, argp, sizeof p)) {
1773                         r = -EFAULT;
1774                         break;
1775                 }
1776                 if ((u64)(unsigned long)p != p) {
1777                         r = -EFAULT;
1778                         break;
1779                 }
1780                 for (i = 0; i < d->nvqs; ++i) {
1781                         struct vhost_virtqueue *vq;
1782                         void __user *base = (void __user *)(unsigned long)p;
1783                         vq = d->vqs[i];
1784                         mutex_lock(&vq->mutex);
1785                         /* If ring is inactive, will check when it's enabled. */
1786                         if (vq->private_data && !vq_log_access_ok(vq, base))
1787                                 r = -EFAULT;
1788                         else
1789                                 vq->log_base = base;
1790                         mutex_unlock(&vq->mutex);
1791                 }
1792                 break;
1793         case VHOST_SET_LOG_FD:
1794                 r = get_user(fd, (int __user *)argp);
1795                 if (r < 0)
1796                         break;
1797                 ctx = fd == -1 ? NULL : eventfd_ctx_fdget(fd);
1798                 if (IS_ERR(ctx)) {
1799                         r = PTR_ERR(ctx);
1800                         break;
1801                 }
1802                 swap(ctx, d->log_ctx);
1803                 for (i = 0; i < d->nvqs; ++i) {
1804                         mutex_lock(&d->vqs[i]->mutex);
1805                         d->vqs[i]->log_ctx = d->log_ctx;
1806                         mutex_unlock(&d->vqs[i]->mutex);
1807                 }
1808                 if (ctx)
1809                         eventfd_ctx_put(ctx);
1810                 break;
1811         default:
1812                 r = -ENOIOCTLCMD;
1813                 break;
1814         }
1815 done:
1816         return r;
1817 }
1818 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1819
1820 /* TODO: This is really inefficient.  We need something like get_user()
1821  * (instruction directly accesses the data, with an exception table entry
1822  * returning -EFAULT). See Documentation/x86/exception-tables.rst.
1823  */
1824 static int set_bit_to_user(int nr, void __user *addr)
1825 {
1826         unsigned long log = (unsigned long)addr;
1827         struct page *page;
1828         void *base;
1829         int bit = nr + (log % PAGE_SIZE) * 8;
1830         int r;
1831
1832         r = get_user_pages_fast(log, 1, FOLL_WRITE, &page);
1833         if (r < 0)
1834                 return r;
1835         BUG_ON(r != 1);
1836         base = kmap_atomic(page);
1837         set_bit(bit, base);
1838         kunmap_atomic(base);
1839         set_page_dirty_lock(page);
1840         put_page(page);
1841         return 0;
1842 }
1843
1844 static int log_write(void __user *log_base,
1845                      u64 write_address, u64 write_length)
1846 {
1847         u64 write_page = write_address / VHOST_PAGE_SIZE;
1848         int r;
1849
1850         if (!write_length)
1851                 return 0;
1852         write_length += write_address % VHOST_PAGE_SIZE;
1853         for (;;) {
1854                 u64 base = (u64)(unsigned long)log_base;
1855                 u64 log = base + write_page / 8;
1856                 int bit = write_page % 8;
1857                 if ((u64)(unsigned long)log != log)
1858                         return -EFAULT;
1859                 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1860                 if (r < 0)
1861                         return r;
1862                 if (write_length <= VHOST_PAGE_SIZE)
1863                         break;
1864                 write_length -= VHOST_PAGE_SIZE;
1865                 write_page += 1;
1866         }
1867         return r;
1868 }
1869
1870 static int log_write_hva(struct vhost_virtqueue *vq, u64 hva, u64 len)
1871 {
1872         struct vhost_umem *umem = vq->umem;
1873         struct vhost_umem_node *u;
1874         u64 start, end, l, min;
1875         int r;
1876         bool hit = false;
1877
1878         while (len) {
1879                 min = len;
1880                 /* More than one GPAs can be mapped into a single HVA. So
1881                  * iterate all possible umems here to be safe.
1882                  */
1883                 list_for_each_entry(u, &umem->umem_list, link) {
1884                         if (u->userspace_addr > hva - 1 + len ||
1885                             u->userspace_addr - 1 + u->size < hva)
1886                                 continue;
1887                         start = max(u->userspace_addr, hva);
1888                         end = min(u->userspace_addr - 1 + u->size,
1889                                   hva - 1 + len);
1890                         l = end - start + 1;
1891                         r = log_write(vq->log_base,
1892                                       u->start + start - u->userspace_addr,
1893                                       l);
1894                         if (r < 0)
1895                                 return r;
1896                         hit = true;
1897                         min = min(l, min);
1898                 }
1899
1900                 if (!hit)
1901                         return -EFAULT;
1902
1903                 len -= min;
1904                 hva += min;
1905         }
1906
1907         return 0;
1908 }
1909
1910 static int log_used(struct vhost_virtqueue *vq, u64 used_offset, u64 len)
1911 {
1912         struct iovec iov[64];
1913         int i, ret;
1914
1915         if (!vq->iotlb)
1916                 return log_write(vq->log_base, vq->log_addr + used_offset, len);
1917
1918         ret = translate_desc(vq, (uintptr_t)vq->used + used_offset,
1919                              len, iov, 64, VHOST_ACCESS_WO);
1920         if (ret < 0)
1921                 return ret;
1922
1923         for (i = 0; i < ret; i++) {
1924                 ret = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1925                                     iov[i].iov_len);
1926                 if (ret)
1927                         return ret;
1928         }
1929
1930         return 0;
1931 }
1932
1933 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1934                     unsigned int log_num, u64 len, struct iovec *iov, int count)
1935 {
1936         int i, r;
1937
1938         /* Make sure data written is seen before log. */
1939         smp_wmb();
1940
1941         if (vq->iotlb) {
1942                 for (i = 0; i < count; i++) {
1943                         r = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1944                                           iov[i].iov_len);
1945                         if (r < 0)
1946                                 return r;
1947                 }
1948                 return 0;
1949         }
1950
1951         for (i = 0; i < log_num; ++i) {
1952                 u64 l = min(log[i].len, len);
1953                 r = log_write(vq->log_base, log[i].addr, l);
1954                 if (r < 0)
1955                         return r;
1956                 len -= l;
1957                 if (!len) {
1958                         if (vq->log_ctx)
1959                                 eventfd_signal(vq->log_ctx, 1);
1960                         return 0;
1961                 }
1962         }
1963         /* Length written exceeds what we have stored. This is a bug. */
1964         BUG();
1965         return 0;
1966 }
1967 EXPORT_SYMBOL_GPL(vhost_log_write);
1968
1969 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1970 {
1971         void __user *used;
1972         if (vhost_put_used_flags(vq))
1973                 return -EFAULT;
1974         if (unlikely(vq->log_used)) {
1975                 /* Make sure the flag is seen before log. */
1976                 smp_wmb();
1977                 /* Log used flag write. */
1978                 used = &vq->used->flags;
1979                 log_used(vq, (used - (void __user *)vq->used),
1980                          sizeof vq->used->flags);
1981                 if (vq->log_ctx)
1982                         eventfd_signal(vq->log_ctx, 1);
1983         }
1984         return 0;
1985 }
1986
1987 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1988 {
1989         if (vhost_put_avail_event(vq))
1990                 return -EFAULT;
1991         if (unlikely(vq->log_used)) {
1992                 void __user *used;
1993                 /* Make sure the event is seen before log. */
1994                 smp_wmb();
1995                 /* Log avail event write */
1996                 used = vhost_avail_event(vq);
1997                 log_used(vq, (used - (void __user *)vq->used),
1998                          sizeof *vhost_avail_event(vq));
1999                 if (vq->log_ctx)
2000                         eventfd_signal(vq->log_ctx, 1);
2001         }
2002         return 0;
2003 }
2004
2005 int vhost_vq_init_access(struct vhost_virtqueue *vq)
2006 {
2007         __virtio16 last_used_idx;
2008         int r;
2009         bool is_le = vq->is_le;
2010
2011         if (!vq->private_data)
2012                 return 0;
2013
2014         vhost_init_is_le(vq);
2015
2016         r = vhost_update_used_flags(vq);
2017         if (r)
2018                 goto err;
2019         vq->signalled_used_valid = false;
2020         if (!vq->iotlb &&
2021             !access_ok(&vq->used->idx, sizeof vq->used->idx)) {
2022                 r = -EFAULT;
2023                 goto err;
2024         }
2025         r = vhost_get_used_idx(vq, &last_used_idx);
2026         if (r) {
2027                 vq_err(vq, "Can't access used idx at %p\n",
2028                        &vq->used->idx);
2029                 goto err;
2030         }
2031         vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
2032         return 0;
2033
2034 err:
2035         vq->is_le = is_le;
2036         return r;
2037 }
2038 EXPORT_SYMBOL_GPL(vhost_vq_init_access);
2039
2040 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
2041                           struct iovec iov[], int iov_size, int access)
2042 {
2043         const struct vhost_umem_node *node;
2044         struct vhost_dev *dev = vq->dev;
2045         struct vhost_umem *umem = dev->iotlb ? dev->iotlb : dev->umem;
2046         struct iovec *_iov;
2047         u64 s = 0;
2048         int ret = 0;
2049
2050         while ((u64)len > s) {
2051                 u64 size;
2052                 if (unlikely(ret >= iov_size)) {
2053                         ret = -ENOBUFS;
2054                         break;
2055                 }
2056
2057                 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
2058                                                         addr, addr + len - 1);
2059                 if (node == NULL || node->start > addr) {
2060                         if (umem != dev->iotlb) {
2061                                 ret = -EFAULT;
2062                                 break;
2063                         }
2064                         ret = -EAGAIN;
2065                         break;
2066                 } else if (!(node->perm & access)) {
2067                         ret = -EPERM;
2068                         break;
2069                 }
2070
2071                 _iov = iov + ret;
2072                 size = node->size - addr + node->start;
2073                 _iov->iov_len = min((u64)len - s, size);
2074                 _iov->iov_base = (void __user *)(unsigned long)
2075                         (node->userspace_addr + addr - node->start);
2076                 s += size;
2077                 addr += size;
2078                 ++ret;
2079         }
2080
2081         if (ret == -EAGAIN)
2082                 vhost_iotlb_miss(vq, addr, access);
2083         return ret;
2084 }
2085
2086 /* Each buffer in the virtqueues is actually a chain of descriptors.  This
2087  * function returns the next descriptor in the chain,
2088  * or -1U if we're at the end. */
2089 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
2090 {
2091         unsigned int next;
2092
2093         /* If this descriptor says it doesn't chain, we're done. */
2094         if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
2095                 return -1U;
2096
2097         /* Check they're not leading us off end of descriptors. */
2098         next = vhost16_to_cpu(vq, READ_ONCE(desc->next));
2099         return next;
2100 }
2101
2102 static int get_indirect(struct vhost_virtqueue *vq,
2103                         struct iovec iov[], unsigned int iov_size,
2104                         unsigned int *out_num, unsigned int *in_num,
2105                         struct vhost_log *log, unsigned int *log_num,
2106                         struct vring_desc *indirect)
2107 {
2108         struct vring_desc desc;
2109         unsigned int i = 0, count, found = 0;
2110         u32 len = vhost32_to_cpu(vq, indirect->len);
2111         struct iov_iter from;
2112         int ret, access;
2113
2114         /* Sanity check */
2115         if (unlikely(len % sizeof desc)) {
2116                 vq_err(vq, "Invalid length in indirect descriptor: "
2117                        "len 0x%llx not multiple of 0x%zx\n",
2118                        (unsigned long long)len,
2119                        sizeof desc);
2120                 return -EINVAL;
2121         }
2122
2123         ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
2124                              UIO_MAXIOV, VHOST_ACCESS_RO);
2125         if (unlikely(ret < 0)) {
2126                 if (ret != -EAGAIN)
2127                         vq_err(vq, "Translation failure %d in indirect.\n", ret);
2128                 return ret;
2129         }
2130         iov_iter_init(&from, READ, vq->indirect, ret, len);
2131
2132         /* We will use the result as an address to read from, so most
2133          * architectures only need a compiler barrier here. */
2134         read_barrier_depends();
2135
2136         count = len / sizeof desc;
2137         /* Buffers are chained via a 16 bit next field, so
2138          * we can have at most 2^16 of these. */
2139         if (unlikely(count > USHRT_MAX + 1)) {
2140                 vq_err(vq, "Indirect buffer length too big: %d\n",
2141                        indirect->len);
2142                 return -E2BIG;
2143         }
2144
2145         do {
2146                 unsigned iov_count = *in_num + *out_num;
2147                 if (unlikely(++found > count)) {
2148                         vq_err(vq, "Loop detected: last one at %u "
2149                                "indirect size %u\n",
2150                                i, count);
2151                         return -EINVAL;
2152                 }
2153                 if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
2154                         vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
2155                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2156                         return -EINVAL;
2157                 }
2158                 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
2159                         vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
2160                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2161                         return -EINVAL;
2162                 }
2163
2164                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2165                         access = VHOST_ACCESS_WO;
2166                 else
2167                         access = VHOST_ACCESS_RO;
2168
2169                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2170                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
2171                                      iov_size - iov_count, access);
2172                 if (unlikely(ret < 0)) {
2173                         if (ret != -EAGAIN)
2174                                 vq_err(vq, "Translation failure %d indirect idx %d\n",
2175                                         ret, i);
2176                         return ret;
2177                 }
2178                 /* If this is an input descriptor, increment that count. */
2179                 if (access == VHOST_ACCESS_WO) {
2180                         *in_num += ret;
2181                         if (unlikely(log && ret)) {
2182                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2183                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2184                                 ++*log_num;
2185                         }
2186                 } else {
2187                         /* If it's an output descriptor, they're all supposed
2188                          * to come before any input descriptors. */
2189                         if (unlikely(*in_num)) {
2190                                 vq_err(vq, "Indirect descriptor "
2191                                        "has out after in: idx %d\n", i);
2192                                 return -EINVAL;
2193                         }
2194                         *out_num += ret;
2195                 }
2196         } while ((i = next_desc(vq, &desc)) != -1);
2197         return 0;
2198 }
2199
2200 /* This looks in the virtqueue and for the first available buffer, and converts
2201  * it to an iovec for convenient access.  Since descriptors consist of some
2202  * number of output then some number of input descriptors, it's actually two
2203  * iovecs, but we pack them into one and note how many of each there were.
2204  *
2205  * This function returns the descriptor number found, or vq->num (which is
2206  * never a valid descriptor number) if none was found.  A negative code is
2207  * returned on error. */
2208 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
2209                       struct iovec iov[], unsigned int iov_size,
2210                       unsigned int *out_num, unsigned int *in_num,
2211                       struct vhost_log *log, unsigned int *log_num)
2212 {
2213         struct vring_desc desc;
2214         unsigned int i, head, found = 0;
2215         u16 last_avail_idx;
2216         __virtio16 avail_idx;
2217         __virtio16 ring_head;
2218         int ret, access;
2219
2220         /* Check it isn't doing very strange things with descriptor numbers. */
2221         last_avail_idx = vq->last_avail_idx;
2222
2223         if (vq->avail_idx == vq->last_avail_idx) {
2224                 if (unlikely(vhost_get_avail_idx(vq, &avail_idx))) {
2225                         vq_err(vq, "Failed to access avail idx at %p\n",
2226                                 &vq->avail->idx);
2227                         return -EFAULT;
2228                 }
2229                 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2230
2231                 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
2232                         vq_err(vq, "Guest moved used index from %u to %u",
2233                                 last_avail_idx, vq->avail_idx);
2234                         return -EFAULT;
2235                 }
2236
2237                 /* If there's nothing new since last we looked, return
2238                  * invalid.
2239                  */
2240                 if (vq->avail_idx == last_avail_idx)
2241                         return vq->num;
2242
2243                 /* Only get avail ring entries after they have been
2244                  * exposed by guest.
2245                  */
2246                 smp_rmb();
2247         }
2248
2249         /* Grab the next descriptor number they're advertising, and increment
2250          * the index we've seen. */
2251         if (unlikely(vhost_get_avail_head(vq, &ring_head, last_avail_idx))) {
2252                 vq_err(vq, "Failed to read head: idx %d address %p\n",
2253                        last_avail_idx,
2254                        &vq->avail->ring[last_avail_idx % vq->num]);
2255                 return -EFAULT;
2256         }
2257
2258         head = vhost16_to_cpu(vq, ring_head);
2259
2260         /* If their number is silly, that's an error. */
2261         if (unlikely(head >= vq->num)) {
2262                 vq_err(vq, "Guest says index %u > %u is available",
2263                        head, vq->num);
2264                 return -EINVAL;
2265         }
2266
2267         /* When we start there are none of either input nor output. */
2268         *out_num = *in_num = 0;
2269         if (unlikely(log))
2270                 *log_num = 0;
2271
2272         i = head;
2273         do {
2274                 unsigned iov_count = *in_num + *out_num;
2275                 if (unlikely(i >= vq->num)) {
2276                         vq_err(vq, "Desc index is %u > %u, head = %u",
2277                                i, vq->num, head);
2278                         return -EINVAL;
2279                 }
2280                 if (unlikely(++found > vq->num)) {
2281                         vq_err(vq, "Loop detected: last one at %u "
2282                                "vq size %u head %u\n",
2283                                i, vq->num, head);
2284                         return -EINVAL;
2285                 }
2286                 ret = vhost_get_desc(vq, &desc, i);
2287                 if (unlikely(ret)) {
2288                         vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
2289                                i, vq->desc + i);
2290                         return -EFAULT;
2291                 }
2292                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2293                         ret = get_indirect(vq, iov, iov_size,
2294                                            out_num, in_num,
2295                                            log, log_num, &desc);
2296                         if (unlikely(ret < 0)) {
2297                                 if (ret != -EAGAIN)
2298                                         vq_err(vq, "Failure detected "
2299                                                 "in indirect descriptor at idx %d\n", i);
2300                                 return ret;
2301                         }
2302                         continue;
2303                 }
2304
2305                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2306                         access = VHOST_ACCESS_WO;
2307                 else
2308                         access = VHOST_ACCESS_RO;
2309                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2310                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
2311                                      iov_size - iov_count, access);
2312                 if (unlikely(ret < 0)) {
2313                         if (ret != -EAGAIN)
2314                                 vq_err(vq, "Translation failure %d descriptor idx %d\n",
2315                                         ret, i);
2316                         return ret;
2317                 }
2318                 if (access == VHOST_ACCESS_WO) {
2319                         /* If this is an input descriptor,
2320                          * increment that count. */
2321                         *in_num += ret;
2322                         if (unlikely(log && ret)) {
2323                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2324                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2325                                 ++*log_num;
2326                         }
2327                 } else {
2328                         /* If it's an output descriptor, they're all supposed
2329                          * to come before any input descriptors. */
2330                         if (unlikely(*in_num)) {
2331                                 vq_err(vq, "Descriptor has out after in: "
2332                                        "idx %d\n", i);
2333                                 return -EINVAL;
2334                         }
2335                         *out_num += ret;
2336                 }
2337         } while ((i = next_desc(vq, &desc)) != -1);
2338
2339         /* On success, increment avail index. */
2340         vq->last_avail_idx++;
2341
2342         /* Assume notifications from guest are disabled at this point,
2343          * if they aren't we would need to update avail_event index. */
2344         BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2345         return head;
2346 }
2347 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2348
2349 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2350 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2351 {
2352         vq->last_avail_idx -= n;
2353 }
2354 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2355
2356 /* After we've used one of their buffers, we tell them about it.  We'll then
2357  * want to notify the guest, using eventfd. */
2358 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2359 {
2360         struct vring_used_elem heads = {
2361                 cpu_to_vhost32(vq, head),
2362                 cpu_to_vhost32(vq, len)
2363         };
2364
2365         return vhost_add_used_n(vq, &heads, 1);
2366 }
2367 EXPORT_SYMBOL_GPL(vhost_add_used);
2368
2369 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2370                             struct vring_used_elem *heads,
2371                             unsigned count)
2372 {
2373         struct vring_used_elem __user *used;
2374         u16 old, new;
2375         int start;
2376
2377         start = vq->last_used_idx & (vq->num - 1);
2378         used = vq->used->ring + start;
2379         if (vhost_put_used(vq, heads, start, count)) {
2380                 vq_err(vq, "Failed to write used");
2381                 return -EFAULT;
2382         }
2383         if (unlikely(vq->log_used)) {
2384                 /* Make sure data is seen before log. */
2385                 smp_wmb();
2386                 /* Log used ring entry write. */
2387                 log_used(vq, ((void __user *)used - (void __user *)vq->used),
2388                          count * sizeof *used);
2389         }
2390         old = vq->last_used_idx;
2391         new = (vq->last_used_idx += count);
2392         /* If the driver never bothers to signal in a very long while,
2393          * used index might wrap around. If that happens, invalidate
2394          * signalled_used index we stored. TODO: make sure driver
2395          * signals at least once in 2^16 and remove this. */
2396         if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2397                 vq->signalled_used_valid = false;
2398         return 0;
2399 }
2400
2401 /* After we've used one of their buffers, we tell them about it.  We'll then
2402  * want to notify the guest, using eventfd. */
2403 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2404                      unsigned count)
2405 {
2406         int start, n, r;
2407
2408         start = vq->last_used_idx & (vq->num - 1);
2409         n = vq->num - start;
2410         if (n < count) {
2411                 r = __vhost_add_used_n(vq, heads, n);
2412                 if (r < 0)
2413                         return r;
2414                 heads += n;
2415                 count -= n;
2416         }
2417         r = __vhost_add_used_n(vq, heads, count);
2418
2419         /* Make sure buffer is written before we update index. */
2420         smp_wmb();
2421         if (vhost_put_used_idx(vq)) {
2422                 vq_err(vq, "Failed to increment used idx");
2423                 return -EFAULT;
2424         }
2425         if (unlikely(vq->log_used)) {
2426                 /* Make sure used idx is seen before log. */
2427                 smp_wmb();
2428                 /* Log used index update. */
2429                 log_used(vq, offsetof(struct vring_used, idx),
2430                          sizeof vq->used->idx);
2431                 if (vq->log_ctx)
2432                         eventfd_signal(vq->log_ctx, 1);
2433         }
2434         return r;
2435 }
2436 EXPORT_SYMBOL_GPL(vhost_add_used_n);
2437
2438 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2439 {
2440         __u16 old, new;
2441         __virtio16 event;
2442         bool v;
2443         /* Flush out used index updates. This is paired
2444          * with the barrier that the Guest executes when enabling
2445          * interrupts. */
2446         smp_mb();
2447
2448         if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2449             unlikely(vq->avail_idx == vq->last_avail_idx))
2450                 return true;
2451
2452         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2453                 __virtio16 flags;
2454                 if (vhost_get_avail_flags(vq, &flags)) {
2455                         vq_err(vq, "Failed to get flags");
2456                         return true;
2457                 }
2458                 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2459         }
2460         old = vq->signalled_used;
2461         v = vq->signalled_used_valid;
2462         new = vq->signalled_used = vq->last_used_idx;
2463         vq->signalled_used_valid = true;
2464
2465         if (unlikely(!v))
2466                 return true;
2467
2468         if (vhost_get_used_event(vq, &event)) {
2469                 vq_err(vq, "Failed to get used event idx");
2470                 return true;
2471         }
2472         return vring_need_event(vhost16_to_cpu(vq, event), new, old);
2473 }
2474
2475 /* This actually signals the guest, using eventfd. */
2476 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2477 {
2478         /* Signal the Guest tell them we used something up. */
2479         if (vq->call_ctx && vhost_notify(dev, vq))
2480                 eventfd_signal(vq->call_ctx, 1);
2481 }
2482 EXPORT_SYMBOL_GPL(vhost_signal);
2483
2484 /* And here's the combo meal deal.  Supersize me! */
2485 void vhost_add_used_and_signal(struct vhost_dev *dev,
2486                                struct vhost_virtqueue *vq,
2487                                unsigned int head, int len)
2488 {
2489         vhost_add_used(vq, head, len);
2490         vhost_signal(dev, vq);
2491 }
2492 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2493
2494 /* multi-buffer version of vhost_add_used_and_signal */
2495 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2496                                  struct vhost_virtqueue *vq,
2497                                  struct vring_used_elem *heads, unsigned count)
2498 {
2499         vhost_add_used_n(vq, heads, count);
2500         vhost_signal(dev, vq);
2501 }
2502 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2503
2504 /* return true if we're sure that avaiable ring is empty */
2505 bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2506 {
2507         __virtio16 avail_idx;
2508         int r;
2509
2510         if (vq->avail_idx != vq->last_avail_idx)
2511                 return false;
2512
2513         r = vhost_get_avail_idx(vq, &avail_idx);
2514         if (unlikely(r))
2515                 return false;
2516         vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2517
2518         return vq->avail_idx == vq->last_avail_idx;
2519 }
2520 EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2521
2522 /* OK, now we need to know about added descriptors. */
2523 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2524 {
2525         __virtio16 avail_idx;
2526         int r;
2527
2528         if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2529                 return false;
2530         vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2531         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2532                 r = vhost_update_used_flags(vq);
2533                 if (r) {
2534                         vq_err(vq, "Failed to enable notification at %p: %d\n",
2535                                &vq->used->flags, r);
2536                         return false;
2537                 }
2538         } else {
2539                 r = vhost_update_avail_event(vq, vq->avail_idx);
2540                 if (r) {
2541                         vq_err(vq, "Failed to update avail event index at %p: %d\n",
2542                                vhost_avail_event(vq), r);
2543                         return false;
2544                 }
2545         }
2546         /* They could have slipped one in as we were doing that: make
2547          * sure it's written, then check again. */
2548         smp_mb();
2549         r = vhost_get_avail_idx(vq, &avail_idx);
2550         if (r) {
2551                 vq_err(vq, "Failed to check avail idx at %p: %d\n",
2552                        &vq->avail->idx, r);
2553                 return false;
2554         }
2555
2556         return vhost16_to_cpu(vq, avail_idx) != vq->avail_idx;
2557 }
2558 EXPORT_SYMBOL_GPL(vhost_enable_notify);
2559
2560 /* We don't need to be notified again. */
2561 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2562 {
2563         int r;
2564
2565         if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2566                 return;
2567         vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2568         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2569                 r = vhost_update_used_flags(vq);
2570                 if (r)
2571                         vq_err(vq, "Failed to enable notification at %p: %d\n",
2572                                &vq->used->flags, r);
2573         }
2574 }
2575 EXPORT_SYMBOL_GPL(vhost_disable_notify);
2576
2577 /* Create a new message. */
2578 struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2579 {
2580         struct vhost_msg_node *node = kmalloc(sizeof *node, GFP_KERNEL);
2581         if (!node)
2582                 return NULL;
2583
2584         /* Make sure all padding within the structure is initialized. */
2585         memset(&node->msg, 0, sizeof node->msg);
2586         node->vq = vq;
2587         node->msg.type = type;
2588         return node;
2589 }
2590 EXPORT_SYMBOL_GPL(vhost_new_msg);
2591
2592 void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2593                        struct vhost_msg_node *node)
2594 {
2595         spin_lock(&dev->iotlb_lock);
2596         list_add_tail(&node->node, head);
2597         spin_unlock(&dev->iotlb_lock);
2598
2599         wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
2600 }
2601 EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2602
2603 struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2604                                          struct list_head *head)
2605 {
2606         struct vhost_msg_node *node = NULL;
2607
2608         spin_lock(&dev->iotlb_lock);
2609         if (!list_empty(head)) {
2610                 node = list_first_entry(head, struct vhost_msg_node,
2611                                         node);
2612                 list_del(&node->node);
2613         }
2614         spin_unlock(&dev->iotlb_lock);
2615
2616         return node;
2617 }
2618 EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2619
2620
2621 static int __init vhost_init(void)
2622 {
2623         return 0;
2624 }
2625
2626 static void __exit vhost_exit(void)
2627 {
2628 }
2629
2630 module_init(vhost_init);
2631 module_exit(vhost_exit);
2632
2633 MODULE_VERSION("0.0.1");
2634 MODULE_LICENSE("GPL v2");
2635 MODULE_AUTHOR("Michael S. Tsirkin");
2636 MODULE_DESCRIPTION("Host kernel accelerator for virtio");