drivers/vhost/vhost.c

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