drivers/vhost/vhost.c

   1 /* Copyright (C) 2009 Red Hat, Inc.
   2  * Copyright (C) 2006 Rusty Russell IBM Corporation
   3  *
   4  * Author: Michael S. Tsirkin <mst@redhat.com>
   5  *
   6  * Inspiration, some code, and most witty comments come from
   7  * Documentation/virtual/lguest/lguest.c, by Rusty Russell
   8  *
   9  * This work is licensed under the terms of the GNU GPL, version 2.
  10  *
  11  * Generic code for virtio server in host kernel.
  12  */
  13
  14 #include <linux/eventfd.h>
  15 #include <linux/vhost.h>
  16 #include <linux/uio.h>
  17 #include <linux/mm.h>
  18 #include <linux/mmu_context.h>
  19 #include <linux/miscdevice.h>
  20 #include <linux/mutex.h>
  21 #include <linux/poll.h>
  22 #include <linux/file.h>
  23 #include <linux/highmem.h>
  24 #include <linux/slab.h>
  25 #include <linux/vmalloc.h>
  26 #include <linux/kthread.h>
  27 #include <linux/cgroup.h>
  28 #include <linux/module.h>
  29 #include <linux/sort.h>
  30 #include <linux/sched/mm.h>
  31 #include <linux/sched/signal.h>
  32 #include <linux/interval_tree_generic.h>
  33 #include <linux/nospec.h>
  34
  35 #include "vhost.h"
  36
  37 static ushort max_mem_regions = 64;
  38 module_param(max_mem_regions, ushort, 0444);
  39 MODULE_PARM_DESC(max_mem_regions,
  40         "Maximum number of memory regions in memory map. (default: 64)");
  41 static int max_iotlb_entries = 2048;
  42 module_param(max_iotlb_entries, int, 0444);
  43 MODULE_PARM_DESC(max_iotlb_entries,
  44         "Maximum number of iotlb entries. (default: 2048)");
  45
  46 enum {
  47         VHOST_MEMORY_F_LOG = 0x1,
  48 };
  49
  50 #define vhost_used_event(vq) ((__virtio16 __user *)&vq->avail->ring[vq->num])
  51 #define vhost_avail_event(vq) ((__virtio16 __user *)&vq->used->ring[vq->num])
  52
  53 INTERVAL_TREE_DEFINE(struct vhost_umem_node,
  54                      rb, __u64, __subtree_last,
  55                      START, LAST, static inline, vhost_umem_interval_tree);
  56
  57 #ifdef CONFIG_VHOST_CROSS_ENDIAN_LEGACY
  58 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
  59 {
  60         vq->user_be = !virtio_legacy_is_little_endian();
  61 }
  62
  63 static void vhost_enable_cross_endian_big(struct vhost_virtqueue *vq)
  64 {
  65         vq->user_be = true;
  66 }
  67
  68 static void vhost_enable_cross_endian_little(struct vhost_virtqueue *vq)
  69 {
  70         vq->user_be = false;
  71 }
  72
  73 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
  74 {
  75         struct vhost_vring_state s;
  76
  77         if (vq->private_data)
  78                 return -EBUSY;
  79
  80         if (copy_from_user(&s, argp, sizeof(s)))
  81                 return -EFAULT;
  82
  83         if (s.num != VHOST_VRING_LITTLE_ENDIAN &&
  84             s.num != VHOST_VRING_BIG_ENDIAN)
  85                 return -EINVAL;
  86
  87         if (s.num == VHOST_VRING_BIG_ENDIAN)
  88                 vhost_enable_cross_endian_big(vq);
  89         else
  90                 vhost_enable_cross_endian_little(vq);
  91
  92         return 0;
  93 }
  94
  95 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
  96                                    int __user *argp)
  97 {
  98         struct vhost_vring_state s = {
  99                 .index = idx,
 100                 .num = vq->user_be
 101         };
 102
 103         if (copy_to_user(argp, &s, sizeof(s)))
 104                 return -EFAULT;
 105
 106         return 0;
 107 }
 108
 109 static void vhost_init_is_le(struct vhost_virtqueue *vq)
 110 {
 111         /* Note for legacy virtio: user_be is initialized at reset time
 112          * according to the host endianness. If userspace does not set an
 113          * explicit endianness, the default behavior is native endian, as
 114          * expected by legacy virtio.
 115          */
 116         vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1) || !vq->user_be;
 117 }
 118 #else
 119 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
 120 {
 121 }
 122
 123 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
 124 {
 125         return -ENOIOCTLCMD;
 126 }
 127
 128 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
 129                                    int __user *argp)
 130 {
 131         return -ENOIOCTLCMD;
 132 }
 133
 134 static void vhost_init_is_le(struct vhost_virtqueue *vq)
 135 {
 136         vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1)
 137                 || virtio_legacy_is_little_endian();
 138 }
 139 #endif /* CONFIG_VHOST_CROSS_ENDIAN_LEGACY */
 140
 141 static void vhost_reset_is_le(struct vhost_virtqueue *vq)
 142 {
 143         vhost_init_is_le(vq);
 144 }
 145
 146 struct vhost_flush_struct {
 147         struct vhost_work work;
 148         struct completion wait_event;
 149 };
 150
 151 static void vhost_flush_work(struct vhost_work *work)
 152 {
 153         struct vhost_flush_struct *s;
 154
 155         s = container_of(work, struct vhost_flush_struct, work);
 156         complete(&s->wait_event);
 157 }
 158
 159 static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
 160                             poll_table *pt)
 161 {
 162         struct vhost_poll *poll;
 163
 164         poll = container_of(pt, struct vhost_poll, table);
 165         poll->wqh = wqh;
 166         add_wait_queue(wqh, &poll->wait);
 167 }
 168
 169 static int vhost_poll_wakeup(wait_queue_entry_t *wait, unsigned mode, int sync,
 170                              void *key)
 171 {
 172         struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
 173
 174         if (!(key_to_poll(key) & poll->mask))
 175                 return 0;
 176
 177         vhost_poll_queue(poll);
 178         return 0;
 179 }
 180
 181 void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
 182 {
 183         clear_bit(VHOST_WORK_QUEUED, &work->flags);
 184         work->fn = fn;
 185 }
 186 EXPORT_SYMBOL_GPL(vhost_work_init);
 187
 188 /* Init poll structure */
 189 void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
 190                      __poll_t mask, struct vhost_dev *dev)
 191 {
 192         init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
 193         init_poll_funcptr(&poll->table, vhost_poll_func);
 194         poll->mask = mask;
 195         poll->dev = dev;
 196         poll->wqh = NULL;
 197
 198         vhost_work_init(&poll->work, fn);
 199 }
 200 EXPORT_SYMBOL_GPL(vhost_poll_init);
 201
 202 /* Start polling a file. We add ourselves to file's wait queue. The caller must
 203  * keep a reference to a file until after vhost_poll_stop is called. */
 204 int vhost_poll_start(struct vhost_poll *poll, struct file *file)
 205 {
 206         __poll_t mask;
 207         int ret = 0;
 208
 209         if (poll->wqh)
 210                 return 0;
 211
 212         mask = vfs_poll(file, &poll->table);
 213         if (mask)
 214                 vhost_poll_wakeup(&poll->wait, 0, 0, poll_to_key(mask));
 215         if (mask & EPOLLERR) {
 216                 vhost_poll_stop(poll);
 217                 ret = -EINVAL;
 218         }
 219
 220         return ret;
 221 }
 222 EXPORT_SYMBOL_GPL(vhost_poll_start);
 223
 224 /* Stop polling a file. After this function returns, it becomes safe to drop the
 225  * file reference. You must also flush afterwards. */
 226 void vhost_poll_stop(struct vhost_poll *poll)
 227 {
 228         if (poll->wqh) {
 229                 remove_wait_queue(poll->wqh, &poll->wait);
 230                 poll->wqh = NULL;
 231         }
 232 }
 233 EXPORT_SYMBOL_GPL(vhost_poll_stop);
 234
 235 void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work)
 236 {
 237         struct vhost_flush_struct flush;
 238
 239         if (dev->worker) {
 240                 init_completion(&flush.wait_event);
 241                 vhost_work_init(&flush.work, vhost_flush_work);
 242
 243                 vhost_work_queue(dev, &flush.work);
 244                 wait_for_completion(&flush.wait_event);
 245         }
 246 }
 247 EXPORT_SYMBOL_GPL(vhost_work_flush);
 248
 249 /* Flush any work that has been scheduled. When calling this, don't hold any
 250  * locks that are also used by the callback. */
 251 void vhost_poll_flush(struct vhost_poll *poll)
 252 {
 253         vhost_work_flush(poll->dev, &poll->work);
 254 }
 255 EXPORT_SYMBOL_GPL(vhost_poll_flush);
 256
 257 void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work)
 258 {
 259         if (!dev->worker)
 260                 return;
 261
 262         if (!test_and_set_bit(VHOST_WORK_QUEUED, &work->flags)) {
 263                 /* We can only add the work to the list after we're
 264                  * sure it was not in the list.
 265                  * test_and_set_bit() implies a memory barrier.
 266                  */
 267                 llist_add(&work->node, &dev->work_list);
 268                 wake_up_process(dev->worker);
 269         }
 270 }
 271 EXPORT_SYMBOL_GPL(vhost_work_queue);
 272
 273 /* A lockless hint for busy polling code to exit the loop */
 274 bool vhost_has_work(struct vhost_dev *dev)
 275 {
 276         return !llist_empty(&dev->work_list);
 277 }
 278 EXPORT_SYMBOL_GPL(vhost_has_work);
 279
 280 void vhost_poll_queue(struct vhost_poll *poll)
 281 {
 282         vhost_work_queue(poll->dev, &poll->work);
 283 }
 284 EXPORT_SYMBOL_GPL(vhost_poll_queue);
 285
 286 static void __vhost_vq_meta_reset(struct vhost_virtqueue *vq)
 287 {
 288         int j;
 289
 290         for (j = 0; j < VHOST_NUM_ADDRS; j++)
 291                 vq->meta_iotlb[j] = NULL;
 292 }
 293
 294 static void vhost_vq_meta_reset(struct vhost_dev *d)
 295 {
 296         int i;
 297
 298         for (i = 0; i < d->nvqs; ++i)
 299                 __vhost_vq_meta_reset(d->vqs[i]);
 300 }
 301
 302 static void vhost_vq_reset(struct vhost_dev *dev,
 303                            struct vhost_virtqueue *vq)
 304 {
 305         vq->num = 1;
 306         vq->desc = NULL;
 307         vq->avail = NULL;
 308         vq->used = NULL;
 309         vq->last_avail_idx = 0;
 310         vq->avail_idx = 0;
 311         vq->last_used_idx = 0;
 312         vq->signalled_used = 0;
 313         vq->signalled_used_valid = false;
 314         vq->used_flags = 0;
 315         vq->log_used = false;
 316         vq->log_addr = -1ull;
 317         vq->private_data = NULL;
 318         vq->acked_features = 0;
 319         vq->acked_backend_features = 0;
 320         vq->log_base = NULL;
 321         vq->error_ctx = NULL;
 322         vq->kick = NULL;
 323         vq->call_ctx = NULL;
 324         vq->log_ctx = NULL;
 325         vhost_disable_cross_endian(vq);
 326         vhost_reset_is_le(vq);
 327         vq->busyloop_timeout = 0;
 328         vq->umem = NULL;
 329         vq->iotlb = NULL;
 330         __vhost_vq_meta_reset(vq);
 331 }
 332
 333 static int vhost_worker(void *data)
 334 {
 335         struct vhost_dev *dev = data;
 336         struct vhost_work *work, *work_next;
 337         struct llist_node *node;
 338         mm_segment_t oldfs = get_fs();
 339
 340         set_fs(USER_DS);
 341         use_mm(dev->mm);
 342
 343         for (;;) {
 344                 /* mb paired w/ kthread_stop */
 345                 set_current_state(TASK_INTERRUPTIBLE);
 346
 347                 if (kthread_should_stop()) {
 348                         __set_current_state(TASK_RUNNING);
 349                         break;
 350                 }
 351
 352                 node = llist_del_all(&dev->work_list);
 353                 if (!node)
 354                         schedule();
 355
 356                 node = llist_reverse_order(node);
 357                 /* make sure flag is seen after deletion */
 358                 smp_wmb();
 359                 llist_for_each_entry_safe(work, work_next, node, node) {
 360                         clear_bit(VHOST_WORK_QUEUED, &work->flags);
 361                         __set_current_state(TASK_RUNNING);
 362                         work->fn(work);
 363                         if (need_resched())
 364                                 schedule();
 365                 }
 366         }
 367         unuse_mm(dev->mm);
 368         set_fs(oldfs);
 369         return 0;
 370 }
 371
 372 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
 373 {
 374         kfree(vq->indirect);
 375         vq->indirect = NULL;
 376         kfree(vq->log);
 377         vq->log = NULL;
 378         kfree(vq->heads);
 379         vq->heads = NULL;
 380 }
 381
 382 /* Helper to allocate iovec buffers for all vqs. */
 383 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
 384 {
 385         struct vhost_virtqueue *vq;
 386         int i;
 387
 388         for (i = 0; i < dev->nvqs; ++i) {
 389                 vq = dev->vqs[i];
 390                 vq->indirect = kmalloc_array(UIO_MAXIOV,
 391                                              sizeof(*vq->indirect),
 392                                              GFP_KERNEL);
 393                 vq->log = kmalloc_array(dev->iov_limit, sizeof(*vq->log),
 394                                         GFP_KERNEL);
 395                 vq->heads = kmalloc_array(dev->iov_limit, sizeof(*vq->heads),
 396                                           GFP_KERNEL);
 397                 if (!vq->indirect || !vq->log || !vq->heads)
 398                         goto err_nomem;
 399         }
 400         return 0;
 401
 402 err_nomem:
 403         for (; i >= 0; --i)
 404                 vhost_vq_free_iovecs(dev->vqs[i]);
 405         return -ENOMEM;
 406 }
 407
 408 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
 409 {
 410         int i;
 411
 412         for (i = 0; i < dev->nvqs; ++i)
 413                 vhost_vq_free_iovecs(dev->vqs[i]);
 414 }
 415
 416 bool vhost_exceeds_weight(struct vhost_virtqueue *vq,
 417                           int pkts, int total_len)
 418 {
 419         struct vhost_dev *dev = vq->dev;
 420
 421         if ((dev->byte_weight && total_len >= dev->byte_weight) ||
 422             pkts >= dev->weight) {
 423                 vhost_poll_queue(&vq->poll);
 424                 return true;
 425         }
 426
 427         return false;
 428 }
 429 EXPORT_SYMBOL_GPL(vhost_exceeds_weight);
 430
 431 void vhost_dev_init(struct vhost_dev *dev,
 432                     struct vhost_virtqueue **vqs, int nvqs,
 433                     int iov_limit, int weight, int byte_weight)
 434 {
 435         struct vhost_virtqueue *vq;
 436         int i;
 437
 438         dev->vqs = vqs;
 439         dev->nvqs = nvqs;
 440         mutex_init(&dev->mutex);
 441         dev->log_ctx = NULL;
 442         dev->umem = NULL;
 443         dev->iotlb = NULL;
 444         dev->mm = NULL;
 445         dev->worker = NULL;
 446         dev->iov_limit = iov_limit;
 447         dev->weight = weight;
 448         dev->byte_weight = byte_weight;
 449         init_llist_head(&dev->work_list);
 450         init_waitqueue_head(&dev->wait);
 451         INIT_LIST_HEAD(&dev->read_list);
 452         INIT_LIST_HEAD(&dev->pending_list);
 453         spin_lock_init(&dev->iotlb_lock);
 454
 455
 456         for (i = 0; i < dev->nvqs; ++i) {
 457                 vq = dev->vqs[i];
 458                 vq->log = NULL;
 459                 vq->indirect = NULL;
 460                 vq->heads = NULL;
 461                 vq->dev = dev;
 462                 mutex_init(&vq->mutex);
 463                 vhost_vq_reset(dev, vq);
 464                 if (vq->handle_kick)
 465                         vhost_poll_init(&vq->poll, vq->handle_kick,
 466                                         EPOLLIN, dev);
 467         }
 468 }
 469 EXPORT_SYMBOL_GPL(vhost_dev_init);
 470
 471 /* Caller should have device mutex */
 472 long vhost_dev_check_owner(struct vhost_dev *dev)
 473 {
 474         /* Are you the owner? If not, I don't think you mean to do that */
 475         return dev->mm == current->mm ? 0 : -EPERM;
 476 }
 477 EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
 478
 479 struct vhost_attach_cgroups_struct {
 480         struct vhost_work work;
 481         struct task_struct *owner;
 482         int ret;
 483 };
 484
 485 static void vhost_attach_cgroups_work(struct vhost_work *work)
 486 {
 487         struct vhost_attach_cgroups_struct *s;
 488
 489         s = container_of(work, struct vhost_attach_cgroups_struct, work);
 490         s->ret = cgroup_attach_task_all(s->owner, current);
 491 }
 492
 493 static int vhost_attach_cgroups(struct vhost_dev *dev)
 494 {
 495         struct vhost_attach_cgroups_struct attach;
 496
 497         attach.owner = current;
 498         vhost_work_init(&attach.work, vhost_attach_cgroups_work);
 499         vhost_work_queue(dev, &attach.work);
 500         vhost_work_flush(dev, &attach.work);
 501         return attach.ret;
 502 }
 503
 504 /* Caller should have device mutex */
 505 bool vhost_dev_has_owner(struct vhost_dev *dev)
 506 {
 507         return dev->mm;
 508 }
 509 EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
 510
 511 /* Caller should have device mutex */
 512 long vhost_dev_set_owner(struct vhost_dev *dev)
 513 {
 514         struct task_struct *worker;
 515         int err;
 516
 517         /* Is there an owner already? */
 518         if (vhost_dev_has_owner(dev)) {
 519                 err = -EBUSY;
 520                 goto err_mm;
 521         }
 522
 523         /* No owner, become one */
 524         dev->mm = get_task_mm(current);
 525         worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
 526         if (IS_ERR(worker)) {
 527                 err = PTR_ERR(worker);
 528                 goto err_worker;
 529         }
 530
 531         dev->worker = worker;
 532         wake_up_process(worker);        /* avoid contributing to loadavg */
 533
 534         err = vhost_attach_cgroups(dev);
 535         if (err)
 536                 goto err_cgroup;
 537
 538         err = vhost_dev_alloc_iovecs(dev);
 539         if (err)
 540                 goto err_cgroup;
 541
 542         return 0;
 543 err_cgroup:
 544         kthread_stop(worker);
 545         dev->worker = NULL;
 546 err_worker:
 547         if (dev->mm)
 548                 mmput(dev->mm);
 549         dev->mm = NULL;
 550 err_mm:
 551         return err;
 552 }
 553 EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
 554
 555 struct vhost_umem *vhost_dev_reset_owner_prepare(void)
 556 {
 557         return kvzalloc(sizeof(struct vhost_umem), GFP_KERNEL);
 558 }
 559 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
 560
 561 /* Caller should have device mutex */
 562 void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_umem *umem)
 563 {
 564         int i;
 565
 566         vhost_dev_cleanup(dev);
 567
 568         /* Restore memory to default empty mapping. */
 569         INIT_LIST_HEAD(&umem->umem_list);
 570         dev->umem = umem;
 571         /* We don't need VQ locks below since vhost_dev_cleanup makes sure
 572          * VQs aren't running.
 573          */
 574         for (i = 0; i < dev->nvqs; ++i)
 575                 dev->vqs[i]->umem = umem;
 576 }
 577 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
 578
 579 void vhost_dev_stop(struct vhost_dev *dev)
 580 {
 581         int i;
 582
 583         for (i = 0; i < dev->nvqs; ++i) {
 584                 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick) {
 585                         vhost_poll_stop(&dev->vqs[i]->poll);
 586                         vhost_poll_flush(&dev->vqs[i]->poll);
 587                 }
 588         }
 589 }
 590 EXPORT_SYMBOL_GPL(vhost_dev_stop);
 591
 592 static void vhost_umem_free(struct vhost_umem *umem,
 593                             struct vhost_umem_node *node)
 594 {
 595         vhost_umem_interval_tree_remove(node, &umem->umem_tree);
 596         list_del(&node->link);
 597         kfree(node);
 598         umem->numem--;
 599 }
 600
 601 static void vhost_umem_clean(struct vhost_umem *umem)
 602 {
 603         struct vhost_umem_node *node, *tmp;
 604
 605         if (!umem)
 606                 return;
 607
 608         list_for_each_entry_safe(node, tmp, &umem->umem_list, link)
 609                 vhost_umem_free(umem, node);
 610
 611         kvfree(umem);
 612 }
 613
 614 static void vhost_clear_msg(struct vhost_dev *dev)
 615 {
 616         struct vhost_msg_node *node, *n;
 617
 618         spin_lock(&dev->iotlb_lock);
 619
 620         list_for_each_entry_safe(node, n, &dev->read_list, node) {
 621                 list_del(&node->node);
 622                 kfree(node);
 623         }
 624
 625         list_for_each_entry_safe(node, n, &dev->pending_list, node) {
 626                 list_del(&node->node);
 627                 kfree(node);
 628         }
 629
 630         spin_unlock(&dev->iotlb_lock);
 631 }
 632
 633 void vhost_dev_cleanup(struct vhost_dev *dev)
 634 {
 635         int i;
 636
 637         for (i = 0; i < dev->nvqs; ++i) {
 638                 if (dev->vqs[i]->error_ctx)
 639                         eventfd_ctx_put(dev->vqs[i]->error_ctx);
 640                 if (dev->vqs[i]->kick)
 641                         fput(dev->vqs[i]->kick);
 642                 if (dev->vqs[i]->call_ctx)
 643                         eventfd_ctx_put(dev->vqs[i]->call_ctx);
 644                 vhost_vq_reset(dev, dev->vqs[i]);
 645         }
 646         vhost_dev_free_iovecs(dev);
 647         if (dev->log_ctx)
 648                 eventfd_ctx_put(dev->log_ctx);
 649         dev->log_ctx = NULL;
 650         /* No one will access memory at this point */
 651         vhost_umem_clean(dev->umem);
 652         dev->umem = NULL;
 653         vhost_umem_clean(dev->iotlb);
 654         dev->iotlb = NULL;
 655         vhost_clear_msg(dev);
 656         wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
 657         WARN_ON(!llist_empty(&dev->work_list));
 658         if (dev->worker) {
 659                 kthread_stop(dev->worker);
 660                 dev->worker = NULL;
 661         }
 662         if (dev->mm)
 663                 mmput(dev->mm);
 664         dev->mm = NULL;
 665 }
 666 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
 667
 668 static bool log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
 669 {
 670         u64 a = addr / VHOST_PAGE_SIZE / 8;
 671
 672         /* Make sure 64 bit math will not overflow. */
 673         if (a > ULONG_MAX - (unsigned long)log_base ||
 674             a + (unsigned long)log_base > ULONG_MAX)
 675                 return false;
 676
 677         return access_ok(VERIFY_WRITE, log_base + a,
 678                          (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
 679 }
 680
 681 /* Make sure 64 bit math will not overflow. */
 682 static bool vhost_overflow(u64 uaddr, u64 size)
 683 {
 684         if (uaddr > ULONG_MAX || size > ULONG_MAX)
 685                 return true;
 686
 687         if (!size)
 688                 return false;
 689
 690         return uaddr > ULONG_MAX - size + 1;
 691 }
 692
 693 /* Caller should have vq mutex and device mutex. */
 694 static bool vq_memory_access_ok(void __user *log_base, struct vhost_umem *umem,
 695                                 int log_all)
 696 {
 697         struct vhost_umem_node *node;
 698
 699         if (!umem)
 700                 return false;
 701
 702         list_for_each_entry(node, &umem->umem_list, link) {
 703                 unsigned long a = node->userspace_addr;
 704
 705                 if (vhost_overflow(node->userspace_addr, node->size))
 706                         return false;
 707
 708
 709                 if (!access_ok(VERIFY_WRITE, (void __user *)a,
 710                                     node->size))
 711                         return false;
 712                 else if (log_all && !log_access_ok(log_base,
 713                                                    node->start,
 714                                                    node->size))
 715                         return false;
 716         }
 717         return true;
 718 }
 719
 720 static inline void __user *vhost_vq_meta_fetch(struct vhost_virtqueue *vq,
 721                                                u64 addr, unsigned int size,
 722                                                int type)
 723 {
 724         const struct vhost_umem_node *node = vq->meta_iotlb[type];
 725
 726         if (!node)
 727                 return NULL;
 728
 729         return (void *)(uintptr_t)(node->userspace_addr + addr - node->start);
 730 }
 731
 732 /* Can we switch to this memory table? */
 733 /* Caller should have device mutex but not vq mutex */
 734 static bool memory_access_ok(struct vhost_dev *d, struct vhost_umem *umem,
 735                              int log_all)
 736 {
 737         int i;
 738
 739         for (i = 0; i < d->nvqs; ++i) {
 740                 bool ok;
 741                 bool log;
 742
 743                 mutex_lock(&d->vqs[i]->mutex);
 744                 log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
 745                 /* If ring is inactive, will check when it's enabled. */
 746                 if (d->vqs[i]->private_data)
 747                         ok = vq_memory_access_ok(d->vqs[i]->log_base,
 748                                                  umem, log);
 749                 else
 750                         ok = true;
 751                 mutex_unlock(&d->vqs[i]->mutex);
 752                 if (!ok)
 753                         return false;
 754         }
 755         return true;
 756 }
 757
 758 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
 759                           struct iovec iov[], int iov_size, int access);
 760
 761 static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
 762                               const void *from, unsigned size)
 763 {
 764         int ret;
 765
 766         if (!vq->iotlb)
 767                 return __copy_to_user(to, from, size);
 768         else {
 769                 /* This function should be called after iotlb
 770                  * prefetch, which means we're sure that all vq
 771                  * could be access through iotlb. So -EAGAIN should
 772                  * not happen in this case.
 773                  */
 774                 struct iov_iter t;
 775                 void __user *uaddr = vhost_vq_meta_fetch(vq,
 776                                      (u64)(uintptr_t)to, size,
 777                                      VHOST_ADDR_USED);
 778
 779                 if (uaddr)
 780                         return __copy_to_user(uaddr, from, size);
 781
 782                 ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
 783                                      ARRAY_SIZE(vq->iotlb_iov),
 784                                      VHOST_ACCESS_WO);
 785                 if (ret < 0)
 786                         goto out;
 787                 iov_iter_init(&t, WRITE, vq->iotlb_iov, ret, size);
 788                 ret = copy_to_iter(from, size, &t);
 789                 if (ret == size)
 790                         ret = 0;
 791         }
 792 out:
 793         return ret;
 794 }
 795
 796 static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
 797                                 void __user *from, unsigned size)
 798 {
 799         int ret;
 800
 801         if (!vq->iotlb)
 802                 return __copy_from_user(to, from, size);
 803         else {
 804                 /* This function should be called after iotlb
 805                  * prefetch, which means we're sure that vq
 806                  * could be access through iotlb. So -EAGAIN should
 807                  * not happen in this case.
 808                  */
 809                 void __user *uaddr = vhost_vq_meta_fetch(vq,
 810                                      (u64)(uintptr_t)from, size,
 811                                      VHOST_ADDR_DESC);
 812                 struct iov_iter f;
 813
 814                 if (uaddr)
 815                         return __copy_from_user(to, uaddr, size);
 816
 817                 ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
 818                                      ARRAY_SIZE(vq->iotlb_iov),
 819                                      VHOST_ACCESS_RO);
 820                 if (ret < 0) {
 821                         vq_err(vq, "IOTLB translation failure: uaddr "
 822                                "%p size 0x%llx\n", from,
 823                                (unsigned long long) size);
 824                         goto out;
 825                 }
 826                 iov_iter_init(&f, READ, vq->iotlb_iov, ret, size);
 827                 ret = copy_from_iter(to, size, &f);
 828                 if (ret == size)
 829                         ret = 0;
 830         }
 831
 832 out:
 833         return ret;
 834 }
 835
 836 static void __user *__vhost_get_user_slow(struct vhost_virtqueue *vq,
 837                                           void __user *addr, unsigned int size,
 838                                           int type)
 839 {
 840         int ret;
 841
 842         ret = translate_desc(vq, (u64)(uintptr_t)addr, size, vq->iotlb_iov,
 843                              ARRAY_SIZE(vq->iotlb_iov),
 844                              VHOST_ACCESS_RO);
 845         if (ret < 0) {
 846                 vq_err(vq, "IOTLB translation failure: uaddr "
 847                         "%p size 0x%llx\n", addr,
 848                         (unsigned long long) size);
 849                 return NULL;
 850         }
 851
 852         if (ret != 1 || vq->iotlb_iov[0].iov_len != size) {
 853                 vq_err(vq, "Non atomic userspace memory access: uaddr "
 854                         "%p size 0x%llx\n", addr,
 855                         (unsigned long long) size);
 856                 return NULL;
 857         }
 858
 859         return vq->iotlb_iov[0].iov_base;
 860 }
 861
 862 /* This function should be called after iotlb
 863  * prefetch, which means we're sure that vq
 864  * could be access through iotlb. So -EAGAIN should
 865  * not happen in this case.
 866  */
 867 static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
 868                                             void *addr, unsigned int size,
 869                                             int type)
 870 {
 871         void __user *uaddr = vhost_vq_meta_fetch(vq,
 872                              (u64)(uintptr_t)addr, size, type);
 873         if (uaddr)
 874                 return uaddr;
 875
 876         return __vhost_get_user_slow(vq, addr, size, type);
 877 }
 878
 879 #define vhost_put_user(vq, x, ptr)              \
 880 ({ \
 881         int ret = -EFAULT; \
 882         if (!vq->iotlb) { \
 883                 ret = __put_user(x, ptr); \
 884         } else { \
 885                 __typeof__(ptr) to = \
 886                         (__typeof__(ptr)) __vhost_get_user(vq, ptr,     \
 887                                           sizeof(*ptr), VHOST_ADDR_USED); \
 888                 if (to != NULL) \
 889                         ret = __put_user(x, to); \
 890                 else \
 891                         ret = -EFAULT;  \
 892         } \
 893         ret; \
 894 })
 895
 896 #define vhost_get_user(vq, x, ptr, type)                \
 897 ({ \
 898         int ret; \
 899         if (!vq->iotlb) { \
 900                 ret = __get_user(x, ptr); \
 901         } else { \
 902                 __typeof__(ptr) from = \
 903                         (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
 904                                                            sizeof(*ptr), \
 905                                                            type); \
 906                 if (from != NULL) \
 907                         ret = __get_user(x, from); \
 908                 else \
 909                         ret = -EFAULT; \
 910         } \
 911         ret; \
 912 })
 913
 914 #define vhost_get_avail(vq, x, ptr) \
 915         vhost_get_user(vq, x, ptr, VHOST_ADDR_AVAIL)
 916
 917 #define vhost_get_used(vq, x, ptr) \
 918         vhost_get_user(vq, x, ptr, VHOST_ADDR_USED)
 919
 920 static void vhost_dev_lock_vqs(struct vhost_dev *d)
 921 {
 922         int i = 0;
 923         for (i = 0; i < d->nvqs; ++i)
 924                 mutex_lock_nested(&d->vqs[i]->mutex, i);
 925 }
 926
 927 static void vhost_dev_unlock_vqs(struct vhost_dev *d)
 928 {
 929         int i = 0;
 930         for (i = 0; i < d->nvqs; ++i)
 931                 mutex_unlock(&d->vqs[i]->mutex);
 932 }
 933
 934 static int vhost_new_umem_range(struct vhost_umem *umem,
 935                                 u64 start, u64 size, u64 end,
 936                                 u64 userspace_addr, int perm)
 937 {
 938         struct vhost_umem_node *tmp, *node;
 939
 940         if (!size)
 941                 return -EFAULT;
 942
 943         node = kmalloc(sizeof(*node), GFP_ATOMIC);
 944         if (!node)
 945                 return -ENOMEM;
 946
 947         if (umem->numem == max_iotlb_entries) {
 948                 tmp = list_first_entry(&umem->umem_list, typeof(*tmp), link);
 949                 vhost_umem_free(umem, tmp);
 950         }
 951
 952         node->start = start;
 953         node->size = size;
 954         node->last = end;
 955         node->userspace_addr = userspace_addr;
 956         node->perm = perm;
 957         INIT_LIST_HEAD(&node->link);
 958         list_add_tail(&node->link, &umem->umem_list);
 959         vhost_umem_interval_tree_insert(node, &umem->umem_tree);
 960         umem->numem++;
 961
 962         return 0;
 963 }
 964
 965 static void vhost_del_umem_range(struct vhost_umem *umem,
 966                                  u64 start, u64 end)
 967 {
 968         struct vhost_umem_node *node;
 969
 970         while ((node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
 971                                                            start, end)))
 972                 vhost_umem_free(umem, node);
 973 }
 974
 975 static void vhost_iotlb_notify_vq(struct vhost_dev *d,
 976                                   struct vhost_iotlb_msg *msg)
 977 {
 978         struct vhost_msg_node *node, *n;
 979
 980         spin_lock(&d->iotlb_lock);
 981
 982         list_for_each_entry_safe(node, n, &d->pending_list, node) {
 983                 struct vhost_iotlb_msg *vq_msg = &node->msg.iotlb;
 984                 if (msg->iova <= vq_msg->iova &&
 985                     msg->iova + msg->size - 1 >= vq_msg->iova &&
 986                     vq_msg->type == VHOST_IOTLB_MISS) {
 987                         vhost_poll_queue(&node->vq->poll);
 988                         list_del(&node->node);
 989                         kfree(node);
 990                 }
 991         }
 992
 993         spin_unlock(&d->iotlb_lock);
 994 }
 995
 996 static bool umem_access_ok(u64 uaddr, u64 size, int access)
 997 {
 998         unsigned long a = uaddr;
 999
1000         /* Make sure 64 bit math will not overflow. */
1001         if (vhost_overflow(uaddr, size))
1002                 return false;
1003
1004         if ((access & VHOST_ACCESS_RO) &&
1005             !access_ok(VERIFY_READ, (void __user *)a, size))
1006                 return false;
1007         if ((access & VHOST_ACCESS_WO) &&
1008             !access_ok(VERIFY_WRITE, (void __user *)a, size))
1009                 return false;
1010         return true;
1011 }
1012
1013 static int vhost_process_iotlb_msg(struct vhost_dev *dev,
1014                                    struct vhost_iotlb_msg *msg)
1015 {
1016         int ret = 0;
1017
1018         mutex_lock(&dev->mutex);
1019         vhost_dev_lock_vqs(dev);
1020         switch (msg->type) {
1021         case VHOST_IOTLB_UPDATE:
1022                 if (!dev->iotlb) {
1023                         ret = -EFAULT;
1024                         break;
1025                 }
1026                 if (!umem_access_ok(msg->uaddr, msg->size, msg->perm)) {
1027                         ret = -EFAULT;
1028                         break;
1029                 }
1030                 vhost_vq_meta_reset(dev);
1031                 if (vhost_new_umem_range(dev->iotlb, msg->iova, msg->size,
1032                                          msg->iova + msg->size - 1,
1033                                          msg->uaddr, msg->perm)) {
1034                         ret = -ENOMEM;
1035                         break;
1036                 }
1037                 vhost_iotlb_notify_vq(dev, msg);
1038                 break;
1039         case VHOST_IOTLB_INVALIDATE:
1040                 if (!dev->iotlb) {
1041                         ret = -EFAULT;
1042                         break;
1043                 }
1044                 vhost_vq_meta_reset(dev);
1045                 vhost_del_umem_range(dev->iotlb, msg->iova,
1046                                      msg->iova + msg->size - 1);
1047                 break;
1048         default:
1049                 ret = -EINVAL;
1050                 break;
1051         }
1052
1053         vhost_dev_unlock_vqs(dev);
1054         mutex_unlock(&dev->mutex);
1055
1056         return ret;
1057 }
1058 ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
1059                              struct iov_iter *from)
1060 {
1061         struct vhost_iotlb_msg msg;
1062         size_t offset;
1063         int type, ret;
1064
1065         ret = copy_from_iter(&type, sizeof(type), from);
1066         if (ret != sizeof(type)) {
1067                 ret = -EINVAL;
1068                 goto done;
1069         }
1070
1071         switch (type) {
1072         case VHOST_IOTLB_MSG:
1073                 /* There maybe a hole after type for V1 message type,
1074                  * so skip it here.
1075                  */
1076                 offset = offsetof(struct vhost_msg, iotlb) - sizeof(int);
1077                 break;
1078         case VHOST_IOTLB_MSG_V2:
1079                 offset = sizeof(__u32);
1080                 break;
1081         default:
1082                 ret = -EINVAL;
1083                 goto done;
1084         }
1085
1086         iov_iter_advance(from, offset);
1087         ret = copy_from_iter(&msg, sizeof(msg), from);
1088         if (ret != sizeof(msg)) {
1089                 ret = -EINVAL;
1090                 goto done;
1091         }
1092         if (vhost_process_iotlb_msg(dev, &msg)) {
1093                 ret = -EFAULT;
1094                 goto done;
1095         }
1096
1097         ret = (type == VHOST_IOTLB_MSG) ? sizeof(struct vhost_msg) :
1098               sizeof(struct vhost_msg_v2);
1099 done:
1100         return ret;
1101 }
1102 EXPORT_SYMBOL(vhost_chr_write_iter);
1103
1104 __poll_t vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1105                             poll_table *wait)
1106 {
1107         __poll_t mask = 0;
1108
1109         poll_wait(file, &dev->wait, wait);
1110
1111         if (!list_empty(&dev->read_list))
1112                 mask |= EPOLLIN | EPOLLRDNORM;
1113
1114         return mask;
1115 }
1116 EXPORT_SYMBOL(vhost_chr_poll);
1117
1118 ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1119                             int noblock)
1120 {
1121         DEFINE_WAIT(wait);
1122         struct vhost_msg_node *node;
1123         ssize_t ret = 0;
1124         unsigned size = sizeof(struct vhost_msg);
1125
1126         if (iov_iter_count(to) < size)
1127                 return 0;
1128
1129         while (1) {
1130                 if (!noblock)
1131                         prepare_to_wait(&dev->wait, &wait,
1132                                         TASK_INTERRUPTIBLE);
1133
1134                 node = vhost_dequeue_msg(dev, &dev->read_list);
1135                 if (node)
1136                         break;
1137                 if (noblock) {
1138                         ret = -EAGAIN;
1139                         break;
1140                 }
1141                 if (signal_pending(current)) {
1142                         ret = -ERESTARTSYS;
1143                         break;
1144                 }
1145                 if (!dev->iotlb) {
1146                         ret = -EBADFD;
1147                         break;
1148                 }
1149
1150                 schedule();
1151         }
1152
1153         if (!noblock)
1154                 finish_wait(&dev->wait, &wait);
1155
1156         if (node) {
1157                 struct vhost_iotlb_msg *msg;
1158                 void *start = &node->msg;
1159
1160                 switch (node->msg.type) {
1161                 case VHOST_IOTLB_MSG:
1162                         size = sizeof(node->msg);
1163                         msg = &node->msg.iotlb;
1164                         break;
1165                 case VHOST_IOTLB_MSG_V2:
1166                         size = sizeof(node->msg_v2);
1167                         msg = &node->msg_v2.iotlb;
1168                         break;
1169                 default:
1170                         BUG();
1171                         break;
1172                 }
1173
1174                 ret = copy_to_iter(start, size, to);
1175                 if (ret != size || msg->type != VHOST_IOTLB_MISS) {
1176                         kfree(node);
1177                         return ret;
1178                 }
1179                 vhost_enqueue_msg(dev, &dev->pending_list, node);
1180         }
1181
1182         return ret;
1183 }
1184 EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1185
1186 static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1187 {
1188         struct vhost_dev *dev = vq->dev;
1189         struct vhost_msg_node *node;
1190         struct vhost_iotlb_msg *msg;
1191         bool v2 = vhost_backend_has_feature(vq, VHOST_BACKEND_F_IOTLB_MSG_V2);
1192
1193         node = vhost_new_msg(vq, v2 ? VHOST_IOTLB_MSG_V2 : VHOST_IOTLB_MSG);
1194         if (!node)
1195                 return -ENOMEM;
1196
1197         if (v2) {
1198                 node->msg_v2.type = VHOST_IOTLB_MSG_V2;
1199                 msg = &node->msg_v2.iotlb;
1200         } else {
1201                 msg = &node->msg.iotlb;
1202         }
1203
1204         msg->type = VHOST_IOTLB_MISS;
1205         msg->iova = iova;
1206         msg->perm = access;
1207
1208         vhost_enqueue_msg(dev, &dev->read_list, node);
1209
1210         return 0;
1211 }
1212
1213 static bool vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1214                          struct vring_desc __user *desc,
1215                          struct vring_avail __user *avail,
1216                          struct vring_used __user *used)
1217
1218 {
1219         size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1220
1221         return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
1222                access_ok(VERIFY_READ, avail,
1223                          sizeof *avail + num * sizeof *avail->ring + s) &&
1224                access_ok(VERIFY_WRITE, used,
1225                         sizeof *used + num * sizeof *used->ring + s);
1226 }
1227
1228 static void vhost_vq_meta_update(struct vhost_virtqueue *vq,
1229                                  const struct vhost_umem_node *node,
1230                                  int type)
1231 {
1232         int access = (type == VHOST_ADDR_USED) ?
1233                      VHOST_ACCESS_WO : VHOST_ACCESS_RO;
1234
1235         if (likely(node->perm & access))
1236                 vq->meta_iotlb[type] = node;
1237 }
1238
1239 static bool iotlb_access_ok(struct vhost_virtqueue *vq,
1240                             int access, u64 addr, u64 len, int type)
1241 {
1242         const struct vhost_umem_node *node;
1243         struct vhost_umem *umem = vq->iotlb;
1244         u64 s = 0, size, orig_addr = addr, last = addr + len - 1;
1245
1246         if (vhost_vq_meta_fetch(vq, addr, len, type))
1247                 return true;
1248
1249         while (len > s) {
1250                 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1251                                                            addr,
1252                                                            last);
1253                 if (node == NULL || node->start > addr) {
1254                         vhost_iotlb_miss(vq, addr, access);
1255                         return false;
1256                 } else if (!(node->perm & access)) {
1257                         /* Report the possible access violation by
1258                          * request another translation from userspace.
1259                          */
1260                         return false;
1261                 }
1262
1263                 size = node->size - addr + node->start;
1264
1265                 if (orig_addr == addr && size >= len)
1266                         vhost_vq_meta_update(vq, node, type);
1267
1268                 s += size;
1269                 addr += size;
1270         }
1271
1272         return true;
1273 }
1274
1275 int vq_iotlb_prefetch(struct vhost_virtqueue *vq)
1276 {
1277         size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1278         unsigned int num = vq->num;
1279
1280         if (!vq->iotlb)
1281                 return 1;
1282
1283         return iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->desc,
1284                                num * sizeof(*vq->desc), VHOST_ADDR_DESC) &&
1285                iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->avail,
1286                                sizeof *vq->avail +
1287                                num * sizeof(*vq->avail->ring) + s,
1288                                VHOST_ADDR_AVAIL) &&
1289                iotlb_access_ok(vq, VHOST_ACCESS_WO, (u64)(uintptr_t)vq->used,
1290                                sizeof *vq->used +
1291                                num * sizeof(*vq->used->ring) + s,
1292                                VHOST_ADDR_USED);
1293 }
1294 EXPORT_SYMBOL_GPL(vq_iotlb_prefetch);
1295
1296 /* Can we log writes? */
1297 /* Caller should have device mutex but not vq mutex */
1298 bool vhost_log_access_ok(struct vhost_dev *dev)
1299 {
1300         return memory_access_ok(dev, dev->umem, 1);
1301 }
1302 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1303
1304 /* Verify access for write logging. */
1305 /* Caller should have vq mutex and device mutex */
1306 static bool vq_log_access_ok(struct vhost_virtqueue *vq,
1307                              void __user *log_base)
1308 {
1309         size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1310
1311         return vq_memory_access_ok(log_base, vq->umem,
1312                                    vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1313                 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
1314                                         sizeof *vq->used +
1315                                         vq->num * sizeof *vq->used->ring + s));
1316 }
1317
1318 /* Can we start vq? */
1319 /* Caller should have vq mutex and device mutex */
1320 bool vhost_vq_access_ok(struct vhost_virtqueue *vq)
1321 {
1322         if (!vq_log_access_ok(vq, vq->log_base))
1323                 return false;
1324
1325         /* Access validation occurs at prefetch time with IOTLB */
1326         if (vq->iotlb)
1327                 return true;
1328
1329         return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used);
1330 }
1331 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1332
1333 static struct vhost_umem *vhost_umem_alloc(void)
1334 {
1335         struct vhost_umem *umem = kvzalloc(sizeof(*umem), GFP_KERNEL);
1336
1337         if (!umem)
1338                 return NULL;
1339
1340         umem->umem_tree = RB_ROOT_CACHED;
1341         umem->numem = 0;
1342         INIT_LIST_HEAD(&umem->umem_list);
1343
1344         return umem;
1345 }
1346
1347 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1348 {
1349         struct vhost_memory mem, *newmem;
1350         struct vhost_memory_region *region;
1351         struct vhost_umem *newumem, *oldumem;
1352         unsigned long size = offsetof(struct vhost_memory, regions);
1353         int i;
1354
1355         if (copy_from_user(&mem, m, size))
1356                 return -EFAULT;
1357         if (mem.padding)
1358                 return -EOPNOTSUPP;
1359         if (mem.nregions > max_mem_regions)
1360                 return -E2BIG;
1361         newmem = kvzalloc(struct_size(newmem, regions, mem.nregions),
1362                         GFP_KERNEL);
1363         if (!newmem)
1364                 return -ENOMEM;
1365
1366         memcpy(newmem, &mem, size);
1367         if (copy_from_user(newmem->regions, m->regions,
1368                            mem.nregions * sizeof *m->regions)) {
1369                 kvfree(newmem);
1370                 return -EFAULT;
1371         }
1372
1373         newumem = vhost_umem_alloc();
1374         if (!newumem) {
1375                 kvfree(newmem);
1376                 return -ENOMEM;
1377         }
1378
1379         for (region = newmem->regions;
1380              region < newmem->regions + mem.nregions;
1381              region++) {
1382                 if (vhost_new_umem_range(newumem,
1383                                          region->guest_phys_addr,
1384                                          region->memory_size,
1385                                          region->guest_phys_addr +
1386                                          region->memory_size - 1,
1387                                          region->userspace_addr,
1388                                          VHOST_ACCESS_RW))
1389                         goto err;
1390         }
1391
1392         if (!memory_access_ok(d, newumem, 0))
1393                 goto err;
1394
1395         oldumem = d->umem;
1396         d->umem = newumem;
1397
1398         /* All memory accesses are done under some VQ mutex. */
1399         for (i = 0; i < d->nvqs; ++i) {
1400                 mutex_lock(&d->vqs[i]->mutex);
1401                 d->vqs[i]->umem = newumem;
1402                 mutex_unlock(&d->vqs[i]->mutex);
1403         }
1404
1405         kvfree(newmem);
1406         vhost_umem_clean(oldumem);
1407         return 0;
1408
1409 err:
1410         vhost_umem_clean(newumem);
1411         kvfree(newmem);
1412         return -EFAULT;
1413 }
1414
1415 long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1416 {
1417         struct file *eventfp, *filep = NULL;
1418         bool pollstart = false, pollstop = false;
1419         struct eventfd_ctx *ctx = NULL;
1420         u32 __user *idxp = argp;
1421         struct vhost_virtqueue *vq;
1422         struct vhost_vring_state s;
1423         struct vhost_vring_file f;
1424         struct vhost_vring_addr a;
1425         u32 idx;
1426         long r;
1427
1428         r = get_user(idx, idxp);
1429         if (r < 0)
1430                 return r;
1431         if (idx >= d->nvqs)
1432                 return -ENOBUFS;
1433
1434         idx = array_index_nospec(idx, d->nvqs);
1435         vq = d->vqs[idx];
1436
1437         mutex_lock(&vq->mutex);
1438
1439         switch (ioctl) {
1440         case VHOST_SET_VRING_NUM:
1441                 /* Resizing ring with an active backend?
1442                  * You don't want to do that. */
1443                 if (vq->private_data) {
1444                         r = -EBUSY;
1445                         break;
1446                 }
1447                 if (copy_from_user(&s, argp, sizeof s)) {
1448                         r = -EFAULT;
1449                         break;
1450                 }
1451                 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
1452                         r = -EINVAL;
1453                         break;
1454                 }
1455                 vq->num = s.num;
1456                 break;
1457         case VHOST_SET_VRING_BASE:
1458                 /* Moving base with an active backend?
1459                  * You don't want to do that. */
1460                 if (vq->private_data) {
1461                         r = -EBUSY;
1462                         break;
1463                 }
1464                 if (copy_from_user(&s, argp, sizeof s)) {
1465                         r = -EFAULT;
1466                         break;
1467                 }
1468                 if (s.num > 0xffff) {
1469                         r = -EINVAL;
1470                         break;
1471                 }
1472                 vq->last_avail_idx = s.num;
1473                 /* Forget the cached index value. */
1474                 vq->avail_idx = vq->last_avail_idx;
1475                 break;
1476         case VHOST_GET_VRING_BASE:
1477                 s.index = idx;
1478                 s.num = vq->last_avail_idx;
1479                 if (copy_to_user(argp, &s, sizeof s))
1480                         r = -EFAULT;
1481                 break;
1482         case VHOST_SET_VRING_ADDR:
1483                 if (copy_from_user(&a, argp, sizeof a)) {
1484                         r = -EFAULT;
1485                         break;
1486                 }
1487                 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
1488                         r = -EOPNOTSUPP;
1489                         break;
1490                 }
1491                 /* For 32bit, verify that the top 32bits of the user
1492                    data are set to zero. */
1493                 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1494                     (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1495                     (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
1496                         r = -EFAULT;
1497                         break;
1498                 }
1499
1500                 /* Make sure it's safe to cast pointers to vring types. */
1501                 BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1502                 BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1503                 if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1504                     (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1505                     (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1))) {
1506                         r = -EINVAL;
1507                         break;
1508                 }
1509
1510                 /* We only verify access here if backend is configured.
1511                  * If it is not, we don't as size might not have been setup.
1512                  * We will verify when backend is configured. */
1513                 if (vq->private_data) {
1514                         if (!vq_access_ok(vq, vq->num,
1515                                 (void __user *)(unsigned long)a.desc_user_addr,
1516                                 (void __user *)(unsigned long)a.avail_user_addr,
1517                                 (void __user *)(unsigned long)a.used_user_addr)) {
1518                                 r = -EINVAL;
1519                                 break;
1520                         }
1521
1522                         /* Also validate log access for used ring if enabled. */
1523                         if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
1524                             !log_access_ok(vq->log_base, a.log_guest_addr,
1525                                            sizeof *vq->used +
1526                                            vq->num * sizeof *vq->used->ring)) {
1527                                 r = -EINVAL;
1528                                 break;
1529                         }
1530                 }
1531
1532                 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1533                 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1534                 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1535                 vq->log_addr = a.log_guest_addr;
1536                 vq->used = (void __user *)(unsigned long)a.used_user_addr;
1537                 break;
1538         case VHOST_SET_VRING_KICK:
1539                 if (copy_from_user(&f, argp, sizeof f)) {
1540                         r = -EFAULT;
1541                         break;
1542                 }
1543                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1544                 if (IS_ERR(eventfp)) {
1545                         r = PTR_ERR(eventfp);
1546                         break;
1547                 }
1548                 if (eventfp != vq->kick) {
1549                         pollstop = (filep = vq->kick) != NULL;
1550                         pollstart = (vq->kick = eventfp) != NULL;
1551                 } else
1552                         filep = eventfp;
1553                 break;
1554         case VHOST_SET_VRING_CALL:
1555                 if (copy_from_user(&f, argp, sizeof f)) {
1556                         r = -EFAULT;
1557                         break;
1558                 }
1559                 ctx = f.fd == -1 ? NULL : eventfd_ctx_fdget(f.fd);
1560                 if (IS_ERR(ctx)) {
1561                         r = PTR_ERR(ctx);
1562                         break;
1563                 }
1564                 swap(ctx, vq->call_ctx);
1565                 break;
1566         case VHOST_SET_VRING_ERR:
1567                 if (copy_from_user(&f, argp, sizeof f)) {
1568                         r = -EFAULT;
1569                         break;
1570                 }
1571                 ctx = f.fd == -1 ? NULL : eventfd_ctx_fdget(f.fd);
1572                 if (IS_ERR(ctx)) {
1573                         r = PTR_ERR(ctx);
1574                         break;
1575                 }
1576                 swap(ctx, vq->error_ctx);
1577                 break;
1578         case VHOST_SET_VRING_ENDIAN:
1579                 r = vhost_set_vring_endian(vq, argp);
1580                 break;
1581         case VHOST_GET_VRING_ENDIAN:
1582                 r = vhost_get_vring_endian(vq, idx, argp);
1583                 break;
1584         case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
1585                 if (copy_from_user(&s, argp, sizeof(s))) {
1586                         r = -EFAULT;
1587                         break;
1588                 }
1589                 vq->busyloop_timeout = s.num;
1590                 break;
1591         case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
1592                 s.index = idx;
1593                 s.num = vq->busyloop_timeout;
1594                 if (copy_to_user(argp, &s, sizeof(s)))
1595                         r = -EFAULT;
1596                 break;
1597         default:
1598                 r = -ENOIOCTLCMD;
1599         }
1600
1601         if (pollstop && vq->handle_kick)
1602                 vhost_poll_stop(&vq->poll);
1603
1604         if (!IS_ERR_OR_NULL(ctx))
1605                 eventfd_ctx_put(ctx);
1606         if (filep)
1607                 fput(filep);
1608
1609         if (pollstart && vq->handle_kick)
1610                 r = vhost_poll_start(&vq->poll, vq->kick);
1611
1612         mutex_unlock(&vq->mutex);
1613
1614         if (pollstop && vq->handle_kick)
1615                 vhost_poll_flush(&vq->poll);
1616         return r;
1617 }
1618 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
1619
1620 int vhost_init_device_iotlb(struct vhost_dev *d, bool enabled)
1621 {
1622         struct vhost_umem *niotlb, *oiotlb;
1623         int i;
1624
1625         niotlb = vhost_umem_alloc();
1626         if (!niotlb)
1627                 return -ENOMEM;
1628
1629         oiotlb = d->iotlb;
1630         d->iotlb = niotlb;
1631
1632         for (i = 0; i < d->nvqs; ++i) {
1633                 struct vhost_virtqueue *vq = d->vqs[i];
1634
1635                 mutex_lock(&vq->mutex);
1636                 vq->iotlb = niotlb;
1637                 __vhost_vq_meta_reset(vq);
1638                 mutex_unlock(&vq->mutex);
1639         }
1640
1641         vhost_umem_clean(oiotlb);
1642
1643         return 0;
1644 }
1645 EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
1646
1647 /* Caller must have device mutex */
1648 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1649 {
1650         struct eventfd_ctx *ctx;
1651         u64 p;
1652         long r;
1653         int i, fd;
1654
1655         /* If you are not the owner, you can become one */
1656         if (ioctl == VHOST_SET_OWNER) {
1657                 r = vhost_dev_set_owner(d);
1658                 goto done;
1659         }
1660
1661         /* You must be the owner to do anything else */
1662         r = vhost_dev_check_owner(d);
1663         if (r)
1664                 goto done;
1665
1666         switch (ioctl) {
1667         case VHOST_SET_MEM_TABLE:
1668                 r = vhost_set_memory(d, argp);
1669                 break;
1670         case VHOST_SET_LOG_BASE:
1671                 if (copy_from_user(&p, argp, sizeof p)) {
1672                         r = -EFAULT;
1673                         break;
1674                 }
1675                 if ((u64)(unsigned long)p != p) {
1676                         r = -EFAULT;
1677                         break;
1678                 }
1679                 for (i = 0; i < d->nvqs; ++i) {
1680                         struct vhost_virtqueue *vq;
1681                         void __user *base = (void __user *)(unsigned long)p;
1682                         vq = d->vqs[i];
1683                         mutex_lock(&vq->mutex);
1684                         /* If ring is inactive, will check when it's enabled. */
1685                         if (vq->private_data && !vq_log_access_ok(vq, base))
1686                                 r = -EFAULT;
1687                         else
1688                                 vq->log_base = base;
1689                         mutex_unlock(&vq->mutex);
1690                 }
1691                 break;
1692         case VHOST_SET_LOG_FD:
1693                 r = get_user(fd, (int __user *)argp);
1694                 if (r < 0)
1695                         break;
1696                 ctx = fd == -1 ? NULL : eventfd_ctx_fdget(fd);
1697                 if (IS_ERR(ctx)) {
1698                         r = PTR_ERR(ctx);
1699                         break;
1700                 }
1701                 swap(ctx, d->log_ctx);
1702                 for (i = 0; i < d->nvqs; ++i) {
1703                         mutex_lock(&d->vqs[i]->mutex);
1704                         d->vqs[i]->log_ctx = d->log_ctx;
1705                         mutex_unlock(&d->vqs[i]->mutex);
1706                 }
1707                 if (ctx)
1708                         eventfd_ctx_put(ctx);
1709                 break;
1710         default:
1711                 r = -ENOIOCTLCMD;
1712                 break;
1713         }
1714 done:
1715         return r;
1716 }
1717 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1718
1719 /* TODO: This is really inefficient.  We need something like get_user()
1720  * (instruction directly accesses the data, with an exception table entry
1721  * returning -EFAULT). See Documentation/x86/exception-tables.txt.
1722  */
1723 static int set_bit_to_user(int nr, void __user *addr)
1724 {
1725         unsigned long log = (unsigned long)addr;
1726         struct page *page;
1727         void *base;
1728         int bit = nr + (log % PAGE_SIZE) * 8;
1729         int r;
1730
1731         r = get_user_pages_fast(log, 1, 1, &page);
1732         if (r < 0)
1733                 return r;
1734         BUG_ON(r != 1);
1735         base = kmap_atomic(page);
1736         set_bit(bit, base);
1737         kunmap_atomic(base);
1738         set_page_dirty_lock(page);
1739         put_page(page);
1740         return 0;
1741 }
1742
1743 static int log_write(void __user *log_base,
1744                      u64 write_address, u64 write_length)
1745 {
1746         u64 write_page = write_address / VHOST_PAGE_SIZE;
1747         int r;
1748
1749         if (!write_length)
1750                 return 0;
1751         write_length += write_address % VHOST_PAGE_SIZE;
1752         for (;;) {
1753                 u64 base = (u64)(unsigned long)log_base;
1754                 u64 log = base + write_page / 8;
1755                 int bit = write_page % 8;
1756                 if ((u64)(unsigned long)log != log)
1757                         return -EFAULT;
1758                 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1759                 if (r < 0)
1760                         return r;
1761                 if (write_length <= VHOST_PAGE_SIZE)
1762                         break;
1763                 write_length -= VHOST_PAGE_SIZE;
1764                 write_page += 1;
1765         }
1766         return r;
1767 }
1768
1769 static int log_write_hva(struct vhost_virtqueue *vq, u64 hva, u64 len)
1770 {
1771         struct vhost_umem *umem = vq->umem;
1772         struct vhost_umem_node *u;
1773         u64 start, end, l, min;
1774         int r;
1775         bool hit = false;
1776
1777         while (len) {
1778                 min = len;
1779                 /* More than one GPAs can be mapped into a single HVA. So
1780                  * iterate all possible umems here to be safe.
1781                  */
1782                 list_for_each_entry(u, &umem->umem_list, link) {
1783                         if (u->userspace_addr > hva - 1 + len ||
1784                             u->userspace_addr - 1 + u->size < hva)
1785                                 continue;
1786                         start = max(u->userspace_addr, hva);
1787                         end = min(u->userspace_addr - 1 + u->size,
1788                                   hva - 1 + len);
1789                         l = end - start + 1;
1790                         r = log_write(vq->log_base,
1791                                       u->start + start - u->userspace_addr,
1792                                       l);
1793                         if (r < 0)
1794                                 return r;
1795                         hit = true;
1796                         min = min(l, min);
1797                 }
1798
1799                 if (!hit)
1800                         return -EFAULT;
1801
1802                 len -= min;
1803                 hva += min;
1804         }
1805
1806         return 0;
1807 }
1808
1809 static int log_used(struct vhost_virtqueue *vq, u64 used_offset, u64 len)
1810 {
1811         struct iovec iov[64];
1812         int i, ret;
1813
1814         if (!vq->iotlb)
1815                 return log_write(vq->log_base, vq->log_addr + used_offset, len);
1816
1817         ret = translate_desc(vq, (uintptr_t)vq->used + used_offset,
1818                              len, iov, 64, VHOST_ACCESS_WO);
1819         if (ret < 0)
1820                 return ret;
1821
1822         for (i = 0; i < ret; i++) {
1823                 ret = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1824                                     iov[i].iov_len);
1825                 if (ret)
1826                         return ret;
1827         }
1828
1829         return 0;
1830 }
1831
1832 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1833                     unsigned int log_num, u64 len, struct iovec *iov, int count)
1834 {
1835         int i, r;
1836
1837         /* Make sure data written is seen before log. */
1838         smp_wmb();
1839
1840         if (vq->iotlb) {
1841                 for (i = 0; i < count; i++) {
1842                         r = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1843                                           iov[i].iov_len);
1844                         if (r < 0)
1845                                 return r;
1846                 }
1847                 return 0;
1848         }
1849
1850         for (i = 0; i < log_num; ++i) {
1851                 u64 l = min(log[i].len, len);
1852                 r = log_write(vq->log_base, log[i].addr, l);
1853                 if (r < 0)
1854                         return r;
1855                 len -= l;
1856                 if (!len) {
1857                         if (vq->log_ctx)
1858                                 eventfd_signal(vq->log_ctx, 1);
1859                         return 0;
1860                 }
1861         }
1862         /* Length written exceeds what we have stored. This is a bug. */
1863         BUG();
1864         return 0;
1865 }
1866 EXPORT_SYMBOL_GPL(vhost_log_write);
1867
1868 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1869 {
1870         void __user *used;
1871         if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
1872                            &vq->used->flags) < 0)
1873                 return -EFAULT;
1874         if (unlikely(vq->log_used)) {
1875                 /* Make sure the flag is seen before log. */
1876                 smp_wmb();
1877                 /* Log used flag write. */
1878                 used = &vq->used->flags;
1879                 log_used(vq, (used - (void __user *)vq->used),
1880                          sizeof vq->used->flags);
1881                 if (vq->log_ctx)
1882                         eventfd_signal(vq->log_ctx, 1);
1883         }
1884         return 0;
1885 }
1886
1887 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1888 {
1889         if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
1890                            vhost_avail_event(vq)))
1891                 return -EFAULT;
1892         if (unlikely(vq->log_used)) {
1893                 void __user *used;
1894                 /* Make sure the event is seen before log. */
1895                 smp_wmb();
1896                 /* Log avail event write */
1897                 used = vhost_avail_event(vq);
1898                 log_used(vq, (used - (void __user *)vq->used),
1899                          sizeof *vhost_avail_event(vq));
1900                 if (vq->log_ctx)
1901                         eventfd_signal(vq->log_ctx, 1);
1902         }
1903         return 0;
1904 }
1905
1906 int vhost_vq_init_access(struct vhost_virtqueue *vq)
1907 {
1908         __virtio16 last_used_idx;
1909         int r;
1910         bool is_le = vq->is_le;
1911
1912         if (!vq->private_data)
1913                 return 0;
1914
1915         vhost_init_is_le(vq);
1916
1917         r = vhost_update_used_flags(vq);
1918         if (r)
1919                 goto err;
1920         vq->signalled_used_valid = false;
1921         if (!vq->iotlb &&
1922             !access_ok(VERIFY_READ, &vq->used->idx, sizeof vq->used->idx)) {
1923                 r = -EFAULT;
1924                 goto err;
1925         }
1926         r = vhost_get_used(vq, last_used_idx, &vq->used->idx);
1927         if (r) {
1928                 vq_err(vq, "Can't access used idx at %p\n",
1929                        &vq->used->idx);
1930                 goto err;
1931         }
1932         vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
1933         return 0;
1934
1935 err:
1936         vq->is_le = is_le;
1937         return r;
1938 }
1939 EXPORT_SYMBOL_GPL(vhost_vq_init_access);
1940
1941 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
1942                           struct iovec iov[], int iov_size, int access)
1943 {
1944         const struct vhost_umem_node *node;
1945         struct vhost_dev *dev = vq->dev;
1946         struct vhost_umem *umem = dev->iotlb ? dev->iotlb : dev->umem;
1947         struct iovec *_iov;
1948         u64 s = 0;
1949         int ret = 0;
1950
1951         while ((u64)len > s) {
1952                 u64 size;
1953                 if (unlikely(ret >= iov_size)) {
1954                         ret = -ENOBUFS;
1955                         break;
1956                 }
1957
1958                 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1959                                                         addr, addr + len - 1);
1960                 if (node == NULL || node->start > addr) {
1961                         if (umem != dev->iotlb) {
1962                                 ret = -EFAULT;
1963                                 break;
1964                         }
1965                         ret = -EAGAIN;
1966                         break;
1967                 } else if (!(node->perm & access)) {
1968                         ret = -EPERM;
1969                         break;
1970                 }
1971
1972                 _iov = iov + ret;
1973                 size = node->size - addr + node->start;
1974                 _iov->iov_len = min((u64)len - s, size);
1975                 _iov->iov_base = (void __user *)(unsigned long)
1976                         (node->userspace_addr + addr - node->start);
1977                 s += size;
1978                 addr += size;
1979                 ++ret;
1980         }
1981
1982         if (ret == -EAGAIN)
1983                 vhost_iotlb_miss(vq, addr, access);
1984         return ret;
1985 }
1986
1987 /* Each buffer in the virtqueues is actually a chain of descriptors.  This
1988  * function returns the next descriptor in the chain,
1989  * or -1U if we're at the end. */
1990 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
1991 {
1992         unsigned int next;
1993
1994         /* If this descriptor says it doesn't chain, we're done. */
1995         if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
1996                 return -1U;
1997
1998         /* Check they're not leading us off end of descriptors. */
1999         next = vhost16_to_cpu(vq, READ_ONCE(desc->next));
2000         return next;
2001 }
2002
2003 static int get_indirect(struct vhost_virtqueue *vq,
2004                         struct iovec iov[], unsigned int iov_size,
2005                         unsigned int *out_num, unsigned int *in_num,
2006                         struct vhost_log *log, unsigned int *log_num,
2007                         struct vring_desc *indirect)
2008 {
2009         struct vring_desc desc;
2010         unsigned int i = 0, count, found = 0;
2011         u32 len = vhost32_to_cpu(vq, indirect->len);
2012         struct iov_iter from;
2013         int ret, access;
2014
2015         /* Sanity check */
2016         if (unlikely(len % sizeof desc)) {
2017                 vq_err(vq, "Invalid length in indirect descriptor: "
2018                        "len 0x%llx not multiple of 0x%zx\n",
2019                        (unsigned long long)len,
2020                        sizeof desc);
2021                 return -EINVAL;
2022         }
2023
2024         ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
2025                              UIO_MAXIOV, VHOST_ACCESS_RO);
2026         if (unlikely(ret < 0)) {
2027                 if (ret != -EAGAIN)
2028                         vq_err(vq, "Translation failure %d in indirect.\n", ret);
2029                 return ret;
2030         }
2031         iov_iter_init(&from, READ, vq->indirect, ret, len);
2032
2033         /* We will use the result as an address to read from, so most
2034          * architectures only need a compiler barrier here. */
2035         read_barrier_depends();
2036
2037         count = len / sizeof desc;
2038         /* Buffers are chained via a 16 bit next field, so
2039          * we can have at most 2^16 of these. */
2040         if (unlikely(count > USHRT_MAX + 1)) {
2041                 vq_err(vq, "Indirect buffer length too big: %d\n",
2042                        indirect->len);
2043                 return -E2BIG;
2044         }
2045
2046         do {
2047                 unsigned iov_count = *in_num + *out_num;
2048                 if (unlikely(++found > count)) {
2049                         vq_err(vq, "Loop detected: last one at %u "
2050                                "indirect size %u\n",
2051                                i, count);
2052                         return -EINVAL;
2053                 }
2054                 if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
2055                         vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
2056                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2057                         return -EINVAL;
2058                 }
2059                 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
2060                         vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
2061                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2062                         return -EINVAL;
2063                 }
2064
2065                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2066                         access = VHOST_ACCESS_WO;
2067                 else
2068                         access = VHOST_ACCESS_RO;
2069
2070                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2071                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
2072                                      iov_size - iov_count, access);
2073                 if (unlikely(ret < 0)) {
2074                         if (ret != -EAGAIN)
2075                                 vq_err(vq, "Translation failure %d indirect idx %d\n",
2076                                         ret, i);
2077                         return ret;
2078                 }
2079                 /* If this is an input descriptor, increment that count. */
2080                 if (access == VHOST_ACCESS_WO) {
2081                         *in_num += ret;
2082                         if (unlikely(log && ret)) {
2083                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2084                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2085                                 ++*log_num;
2086                         }
2087                 } else {
2088                         /* If it's an output descriptor, they're all supposed
2089                          * to come before any input descriptors. */
2090                         if (unlikely(*in_num)) {
2091                                 vq_err(vq, "Indirect descriptor "
2092                                        "has out after in: idx %d\n", i);
2093                                 return -EINVAL;
2094                         }
2095                         *out_num += ret;
2096                 }
2097         } while ((i = next_desc(vq, &desc)) != -1);
2098         return 0;
2099 }
2100
2101 /* This looks in the virtqueue and for the first available buffer, and converts
2102  * it to an iovec for convenient access.  Since descriptors consist of some
2103  * number of output then some number of input descriptors, it's actually two
2104  * iovecs, but we pack them into one and note how many of each there were.
2105  *
2106  * This function returns the descriptor number found, or vq->num (which is
2107  * never a valid descriptor number) if none was found.  A negative code is
2108  * returned on error. */
2109 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
2110                       struct iovec iov[], unsigned int iov_size,
2111                       unsigned int *out_num, unsigned int *in_num,
2112                       struct vhost_log *log, unsigned int *log_num)
2113 {
2114         struct vring_desc desc;
2115         unsigned int i, head, found = 0;
2116         u16 last_avail_idx;
2117         __virtio16 avail_idx;
2118         __virtio16 ring_head;
2119         int ret, access;
2120
2121         /* Check it isn't doing very strange things with descriptor numbers. */
2122         last_avail_idx = vq->last_avail_idx;
2123
2124         if (vq->avail_idx == vq->last_avail_idx) {
2125                 if (unlikely(vhost_get_avail(vq, avail_idx, &vq->avail->idx))) {
2126                         vq_err(vq, "Failed to access avail idx at %p\n",
2127                                 &vq->avail->idx);
2128                         return -EFAULT;
2129                 }
2130                 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2131
2132                 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
2133                         vq_err(vq, "Guest moved used index from %u to %u",
2134                                 last_avail_idx, vq->avail_idx);
2135                         return -EFAULT;
2136                 }
2137
2138                 /* If there's nothing new since last we looked, return
2139                  * invalid.
2140                  */
2141                 if (vq->avail_idx == last_avail_idx)
2142                         return vq->num;
2143
2144                 /* Only get avail ring entries after they have been
2145                  * exposed by guest.
2146                  */
2147                 smp_rmb();
2148         }
2149
2150         /* Grab the next descriptor number they're advertising, and increment
2151          * the index we've seen. */
2152         if (unlikely(vhost_get_avail(vq, ring_head,
2153                      &vq->avail->ring[last_avail_idx & (vq->num - 1)]))) {
2154                 vq_err(vq, "Failed to read head: idx %d address %p\n",
2155                        last_avail_idx,
2156                        &vq->avail->ring[last_avail_idx % vq->num]);
2157                 return -EFAULT;
2158         }
2159
2160         head = vhost16_to_cpu(vq, ring_head);
2161
2162         /* If their number is silly, that's an error. */
2163         if (unlikely(head >= vq->num)) {
2164                 vq_err(vq, "Guest says index %u > %u is available",
2165                        head, vq->num);
2166                 return -EINVAL;
2167         }
2168
2169         /* When we start there are none of either input nor output. */
2170         *out_num = *in_num = 0;
2171         if (unlikely(log))
2172                 *log_num = 0;
2173
2174         i = head;
2175         do {
2176                 unsigned iov_count = *in_num + *out_num;
2177                 if (unlikely(i >= vq->num)) {
2178                         vq_err(vq, "Desc index is %u > %u, head = %u",
2179                                i, vq->num, head);
2180                         return -EINVAL;
2181                 }
2182                 if (unlikely(++found > vq->num)) {
2183                         vq_err(vq, "Loop detected: last one at %u "
2184                                "vq size %u head %u\n",
2185                                i, vq->num, head);
2186                         return -EINVAL;
2187                 }
2188                 ret = vhost_copy_from_user(vq, &desc, vq->desc + i,
2189                                            sizeof desc);
2190                 if (unlikely(ret)) {
2191                         vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
2192                                i, vq->desc + i);
2193                         return -EFAULT;
2194                 }
2195                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2196                         ret = get_indirect(vq, iov, iov_size,
2197                                            out_num, in_num,
2198                                            log, log_num, &desc);
2199                         if (unlikely(ret < 0)) {
2200                                 if (ret != -EAGAIN)
2201                                         vq_err(vq, "Failure detected "
2202                                                 "in indirect descriptor at idx %d\n", i);
2203                                 return ret;
2204                         }
2205                         continue;
2206                 }
2207
2208                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2209                         access = VHOST_ACCESS_WO;
2210                 else
2211                         access = VHOST_ACCESS_RO;
2212                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2213                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
2214                                      iov_size - iov_count, access);
2215                 if (unlikely(ret < 0)) {
2216                         if (ret != -EAGAIN)
2217                                 vq_err(vq, "Translation failure %d descriptor idx %d\n",
2218                                         ret, i);
2219                         return ret;
2220                 }
2221                 if (access == VHOST_ACCESS_WO) {
2222                         /* If this is an input descriptor,
2223                          * increment that count. */
2224                         *in_num += ret;
2225                         if (unlikely(log && ret)) {
2226                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2227                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2228                                 ++*log_num;
2229                         }
2230                 } else {
2231                         /* If it's an output descriptor, they're all supposed
2232                          * to come before any input descriptors. */
2233                         if (unlikely(*in_num)) {
2234                                 vq_err(vq, "Descriptor has out after in: "
2235                                        "idx %d\n", i);
2236                                 return -EINVAL;
2237                         }
2238                         *out_num += ret;
2239                 }
2240         } while ((i = next_desc(vq, &desc)) != -1);
2241
2242         /* On success, increment avail index. */
2243         vq->last_avail_idx++;
2244
2245         /* Assume notifications from guest are disabled at this point,
2246          * if they aren't we would need to update avail_event index. */
2247         BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2248         return head;
2249 }
2250 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2251
2252 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2253 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2254 {
2255         vq->last_avail_idx -= n;
2256 }
2257 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2258
2259 /* After we've used one of their buffers, we tell them about it.  We'll then
2260  * want to notify the guest, using eventfd. */
2261 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2262 {
2263         struct vring_used_elem heads = {
2264                 cpu_to_vhost32(vq, head),
2265                 cpu_to_vhost32(vq, len)
2266         };
2267
2268         return vhost_add_used_n(vq, &heads, 1);
2269 }
2270 EXPORT_SYMBOL_GPL(vhost_add_used);
2271
2272 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2273                             struct vring_used_elem *heads,
2274                             unsigned count)
2275 {
2276         struct vring_used_elem __user *used;
2277         u16 old, new;
2278         int start;
2279
2280         start = vq->last_used_idx & (vq->num - 1);
2281         used = vq->used->ring + start;
2282         if (count == 1) {
2283                 if (vhost_put_user(vq, heads[0].id, &used->id)) {
2284                         vq_err(vq, "Failed to write used id");
2285                         return -EFAULT;
2286                 }
2287                 if (vhost_put_user(vq, heads[0].len, &used->len)) {
2288                         vq_err(vq, "Failed to write used len");
2289                         return -EFAULT;
2290                 }
2291         } else if (vhost_copy_to_user(vq, used, heads, count * sizeof *used)) {
2292                 vq_err(vq, "Failed to write used");
2293                 return -EFAULT;
2294         }
2295         if (unlikely(vq->log_used)) {
2296                 /* Make sure data is seen before log. */
2297                 smp_wmb();
2298                 /* Log used ring entry write. */
2299                 log_used(vq, ((void __user *)used - (void __user *)vq->used),
2300                          count * sizeof *used);
2301         }
2302         old = vq->last_used_idx;
2303         new = (vq->last_used_idx += count);
2304         /* If the driver never bothers to signal in a very long while,
2305          * used index might wrap around. If that happens, invalidate
2306          * signalled_used index we stored. TODO: make sure driver
2307          * signals at least once in 2^16 and remove this. */
2308         if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2309                 vq->signalled_used_valid = false;
2310         return 0;
2311 }
2312
2313 /* After we've used one of their buffers, we tell them about it.  We'll then
2314  * want to notify the guest, using eventfd. */
2315 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2316                      unsigned count)
2317 {
2318         int start, n, r;
2319
2320         start = vq->last_used_idx & (vq->num - 1);
2321         n = vq->num - start;
2322         if (n < count) {
2323                 r = __vhost_add_used_n(vq, heads, n);
2324                 if (r < 0)
2325                         return r;
2326                 heads += n;
2327                 count -= n;
2328         }
2329         r = __vhost_add_used_n(vq, heads, count);
2330
2331         /* Make sure buffer is written before we update index. */
2332         smp_wmb();
2333         if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
2334                            &vq->used->idx)) {
2335                 vq_err(vq, "Failed to increment used idx");
2336                 return -EFAULT;
2337         }
2338         if (unlikely(vq->log_used)) {
2339                 /* Make sure used idx is seen before log. */
2340                 smp_wmb();
2341                 /* Log used index update. */
2342                 log_used(vq, offsetof(struct vring_used, idx),
2343                          sizeof vq->used->idx);
2344                 if (vq->log_ctx)
2345                         eventfd_signal(vq->log_ctx, 1);
2346         }
2347         return r;
2348 }
2349 EXPORT_SYMBOL_GPL(vhost_add_used_n);
2350
2351 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2352 {
2353         __u16 old, new;
2354         __virtio16 event;
2355         bool v;
2356         /* Flush out used index updates. This is paired
2357          * with the barrier that the Guest executes when enabling
2358          * interrupts. */
2359         smp_mb();
2360
2361         if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2362             unlikely(vq->avail_idx == vq->last_avail_idx))
2363                 return true;
2364
2365         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2366                 __virtio16 flags;
2367                 if (vhost_get_avail(vq, flags, &vq->avail->flags)) {
2368                         vq_err(vq, "Failed to get flags");
2369                         return true;
2370                 }
2371                 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2372         }
2373         old = vq->signalled_used;
2374         v = vq->signalled_used_valid;
2375         new = vq->signalled_used = vq->last_used_idx;
2376         vq->signalled_used_valid = true;
2377
2378         if (unlikely(!v))
2379                 return true;
2380
2381         if (vhost_get_avail(vq, event, vhost_used_event(vq))) {
2382                 vq_err(vq, "Failed to get used event idx");
2383                 return true;
2384         }
2385         return vring_need_event(vhost16_to_cpu(vq, event), new, old);
2386 }
2387
2388 /* This actually signals the guest, using eventfd. */
2389 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2390 {
2391         /* Signal the Guest tell them we used something up. */
2392         if (vq->call_ctx && vhost_notify(dev, vq))
2393                 eventfd_signal(vq->call_ctx, 1);
2394 }
2395 EXPORT_SYMBOL_GPL(vhost_signal);
2396
2397 /* And here's the combo meal deal.  Supersize me! */
2398 void vhost_add_used_and_signal(struct vhost_dev *dev,
2399                                struct vhost_virtqueue *vq,
2400                                unsigned int head, int len)
2401 {
2402         vhost_add_used(vq, head, len);
2403         vhost_signal(dev, vq);
2404 }
2405 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2406
2407 /* multi-buffer version of vhost_add_used_and_signal */
2408 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2409                                  struct vhost_virtqueue *vq,
2410                                  struct vring_used_elem *heads, unsigned count)
2411 {
2412         vhost_add_used_n(vq, heads, count);
2413         vhost_signal(dev, vq);
2414 }
2415 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2416
2417 /* return true if we're sure that avaiable ring is empty */
2418 bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2419 {
2420         __virtio16 avail_idx;
2421         int r;
2422
2423         if (vq->avail_idx != vq->last_avail_idx)
2424                 return false;
2425
2426         r = vhost_get_avail(vq, avail_idx, &vq->avail->idx);
2427         if (unlikely(r))
2428                 return false;
2429         vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2430
2431         return vq->avail_idx == vq->last_avail_idx;
2432 }
2433 EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2434
2435 /* OK, now we need to know about added descriptors. */
2436 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2437 {
2438         __virtio16 avail_idx;
2439         int r;
2440
2441         if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2442                 return false;
2443         vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2444         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2445                 r = vhost_update_used_flags(vq);
2446                 if (r) {
2447                         vq_err(vq, "Failed to enable notification at %p: %d\n",
2448                                &vq->used->flags, r);
2449                         return false;
2450                 }
2451         } else {
2452                 r = vhost_update_avail_event(vq, vq->avail_idx);
2453                 if (r) {
2454                         vq_err(vq, "Failed to update avail event index at %p: %d\n",
2455                                vhost_avail_event(vq), r);
2456                         return false;
2457                 }
2458         }
2459         /* They could have slipped one in as we were doing that: make
2460          * sure it's written, then check again. */
2461         smp_mb();
2462         r = vhost_get_avail(vq, avail_idx, &vq->avail->idx);
2463         if (r) {
2464                 vq_err(vq, "Failed to check avail idx at %p: %d\n",
2465                        &vq->avail->idx, r);
2466                 return false;
2467         }
2468
2469         return vhost16_to_cpu(vq, avail_idx) != vq->avail_idx;
2470 }
2471 EXPORT_SYMBOL_GPL(vhost_enable_notify);
2472
2473 /* We don't need to be notified again. */
2474 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2475 {
2476         int r;
2477
2478         if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2479                 return;
2480         vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2481         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2482                 r = vhost_update_used_flags(vq);
2483                 if (r)
2484                         vq_err(vq, "Failed to enable notification at %p: %d\n",
2485                                &vq->used->flags, r);
2486         }
2487 }
2488 EXPORT_SYMBOL_GPL(vhost_disable_notify);
2489
2490 /* Create a new message. */
2491 struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2492 {
2493         struct vhost_msg_node *node = kmalloc(sizeof *node, GFP_KERNEL);
2494         if (!node)
2495                 return NULL;
2496
2497         /* Make sure all padding within the structure is initialized. */
2498         memset(&node->msg, 0, sizeof node->msg);
2499         node->vq = vq;
2500         node->msg.type = type;
2501         return node;
2502 }
2503 EXPORT_SYMBOL_GPL(vhost_new_msg);
2504
2505 void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2506                        struct vhost_msg_node *node)
2507 {
2508         spin_lock(&dev->iotlb_lock);
2509         list_add_tail(&node->node, head);
2510         spin_unlock(&dev->iotlb_lock);
2511
2512         wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
2513 }
2514 EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2515
2516 struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2517                                          struct list_head *head)
2518 {
2519         struct vhost_msg_node *node = NULL;
2520
2521         spin_lock(&dev->iotlb_lock);
2522         if (!list_empty(head)) {
2523                 node = list_first_entry(head, struct vhost_msg_node,
2524                                         node);
2525                 list_del(&node->node);
2526         }
2527         spin_unlock(&dev->iotlb_lock);
2528
2529         return node;
2530 }
2531 EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2532
2533
2534 static int __init vhost_init(void)
2535 {
2536         return 0;
2537 }
2538
2539 static void __exit vhost_exit(void)
2540 {
2541 }
2542
2543 module_init(vhost_init);
2544 module_exit(vhost_exit);
2545
2546 MODULE_VERSION("0.0.1");
2547 MODULE_LICENSE("GPL v2");
2548 MODULE_AUTHOR("Michael S. Tsirkin");
2549 MODULE_DESCRIPTION("Host kernel accelerator for virtio");