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