drivers/vhost/vhost.c

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