Pileus Git - ~andy/linux/blob - 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/rcupdate.h>
  22 #include <linux/poll.h>
  23 #include <linux/file.h>
  24 #include <linux/highmem.h>
  25 #include <linux/slab.h>
  26 #include <linux/kthread.h>
  27 #include <linux/cgroup.h>
  28 #include <linux/module.h>
  29
  30 #include "vhost.h"
  31
  32 enum {
  33         VHOST_MEMORY_MAX_NREGIONS = 64,
  34         VHOST_MEMORY_F_LOG = 0x1,
  35 };
  36
  37 #define vhost_used_event(vq) ((u16 __user *)&vq->avail->ring[vq->num])
  38 #define vhost_avail_event(vq) ((u16 __user *)&vq->used->ring[vq->num])
  39
  40 static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
  41                             poll_table *pt)
  42 {
  43         struct vhost_poll *poll;
  44
  45         poll = container_of(pt, struct vhost_poll, table);
  46         poll->wqh = wqh;
  47         add_wait_queue(wqh, &poll->wait);
  48 }
  49
  50 static int vhost_poll_wakeup(wait_queue_t *wait, unsigned mode, int sync,
  51                              void *key)
  52 {
  53         struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
  54
  55         if (!((unsigned long)key & poll->mask))
  56                 return 0;
  57
  58         vhost_poll_queue(poll);
  59         return 0;
  60 }
  61
  62 void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
  63 {
  64         INIT_LIST_HEAD(&work->node);
  65         work->fn = fn;
  66         init_waitqueue_head(&work->done);
  67         work->flushing = 0;
  68         work->queue_seq = work->done_seq = 0;
  69 }
  70 EXPORT_SYMBOL_GPL(vhost_work_init);
  71
  72 /* Init poll structure */
  73 void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
  74                      unsigned long mask, struct vhost_dev *dev)
  75 {
  76         init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
  77         init_poll_funcptr(&poll->table, vhost_poll_func);
  78         poll->mask = mask;
  79         poll->dev = dev;
  80         poll->wqh = NULL;
  81
  82         vhost_work_init(&poll->work, fn);
  83 }
  84 EXPORT_SYMBOL_GPL(vhost_poll_init);
  85
  86 /* Start polling a file. We add ourselves to file's wait queue. The caller must
  87  * keep a reference to a file until after vhost_poll_stop is called. */
  88 int vhost_poll_start(struct vhost_poll *poll, struct file *file)
  89 {
  90         unsigned long mask;
  91         int ret = 0;
  92
  93         if (poll->wqh)
  94                 return 0;
  95
  96         mask = file->f_op->poll(file, &poll->table);
  97         if (mask)
  98                 vhost_poll_wakeup(&poll->wait, 0, 0, (void *)mask);
  99         if (mask & POLLERR) {
 100                 if (poll->wqh)
 101                         remove_wait_queue(poll->wqh, &poll->wait);
 102                 ret = -EINVAL;
 103         }
 104
 105         return ret;
 106 }
 107 EXPORT_SYMBOL_GPL(vhost_poll_start);
 108
 109 /* Stop polling a file. After this function returns, it becomes safe to drop the
 110  * file reference. You must also flush afterwards. */
 111 void vhost_poll_stop(struct vhost_poll *poll)
 112 {
 113         if (poll->wqh) {
 114                 remove_wait_queue(poll->wqh, &poll->wait);
 115                 poll->wqh = NULL;
 116         }
 117 }
 118 EXPORT_SYMBOL_GPL(vhost_poll_stop);
 119
 120 static bool vhost_work_seq_done(struct vhost_dev *dev, struct vhost_work *work,
 121                                 unsigned seq)
 122 {
 123         int left;
 124
 125         spin_lock_irq(&dev->work_lock);
 126         left = seq - work->done_seq;
 127         spin_unlock_irq(&dev->work_lock);
 128         return left <= 0;
 129 }
 130
 131 void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work)
 132 {
 133         unsigned seq;
 134         int flushing;
 135
 136         spin_lock_irq(&dev->work_lock);
 137         seq = work->queue_seq;
 138         work->flushing++;
 139         spin_unlock_irq(&dev->work_lock);
 140         wait_event(work->done, vhost_work_seq_done(dev, work, seq));
 141         spin_lock_irq(&dev->work_lock);
 142         flushing = --work->flushing;
 143         spin_unlock_irq(&dev->work_lock);
 144         BUG_ON(flushing < 0);
 145 }
 146 EXPORT_SYMBOL_GPL(vhost_work_flush);
 147
 148 /* Flush any work that has been scheduled. When calling this, don't hold any
 149  * locks that are also used by the callback. */
 150 void vhost_poll_flush(struct vhost_poll *poll)
 151 {
 152         vhost_work_flush(poll->dev, &poll->work);
 153 }
 154 EXPORT_SYMBOL_GPL(vhost_poll_flush);
 155
 156 void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work)
 157 {
 158         unsigned long flags;
 159
 160         spin_lock_irqsave(&dev->work_lock, flags);
 161         if (list_empty(&work->node)) {
 162                 list_add_tail(&work->node, &dev->work_list);
 163                 work->queue_seq++;
 164                 spin_unlock_irqrestore(&dev->work_lock, flags);
 165                 wake_up_process(dev->worker);
 166         } else {
 167                 spin_unlock_irqrestore(&dev->work_lock, flags);
 168         }
 169 }
 170 EXPORT_SYMBOL_GPL(vhost_work_queue);
 171
 172 void vhost_poll_queue(struct vhost_poll *poll)
 173 {
 174         vhost_work_queue(poll->dev, &poll->work);
 175 }
 176 EXPORT_SYMBOL_GPL(vhost_poll_queue);
 177
 178 static void vhost_vq_reset(struct vhost_dev *dev,
 179                            struct vhost_virtqueue *vq)
 180 {
 181         vq->num = 1;
 182         vq->desc = NULL;
 183         vq->avail = NULL;
 184         vq->used = NULL;
 185         vq->last_avail_idx = 0;
 186         vq->avail_idx = 0;
 187         vq->last_used_idx = 0;
 188         vq->signalled_used = 0;
 189         vq->signalled_used_valid = false;
 190         vq->used_flags = 0;
 191         vq->log_used = false;
 192         vq->log_addr = -1ull;
 193         vq->private_data = NULL;
 194         vq->log_base = NULL;
 195         vq->error_ctx = NULL;
 196         vq->error = NULL;
 197         vq->kick = NULL;
 198         vq->call_ctx = NULL;
 199         vq->call = NULL;
 200         vq->log_ctx = NULL;
 201 }
 202
 203 static int vhost_worker(void *data)
 204 {
 205         struct vhost_dev *dev = data;
 206         struct vhost_work *work = NULL;
 207         unsigned uninitialized_var(seq);
 208         mm_segment_t oldfs = get_fs();
 209
 210         set_fs(USER_DS);
 211         use_mm(dev->mm);
 212
 213         for (;;) {
 214                 /* mb paired w/ kthread_stop */
 215                 set_current_state(TASK_INTERRUPTIBLE);
 216
 217                 spin_lock_irq(&dev->work_lock);
 218                 if (work) {
 219                         work->done_seq = seq;
 220                         if (work->flushing)
 221                                 wake_up_all(&work->done);
 222                 }
 223
 224                 if (kthread_should_stop()) {
 225                         spin_unlock_irq(&dev->work_lock);
 226                         __set_current_state(TASK_RUNNING);
 227                         break;
 228                 }
 229                 if (!list_empty(&dev->work_list)) {
 230                         work = list_first_entry(&dev->work_list,
 231                                                 struct vhost_work, node);
 232                         list_del_init(&work->node);
 233                         seq = work->queue_seq;
 234                 } else
 235                         work = NULL;
 236                 spin_unlock_irq(&dev->work_lock);
 237
 238                 if (work) {
 239                         __set_current_state(TASK_RUNNING);
 240                         work->fn(work);
 241                         if (need_resched())
 242                                 schedule();
 243                 } else
 244                         schedule();
 245
 246         }
 247         unuse_mm(dev->mm);
 248         set_fs(oldfs);
 249         return 0;
 250 }
 251
 252 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
 253 {
 254         kfree(vq->indirect);
 255         vq->indirect = NULL;
 256         kfree(vq->log);
 257         vq->log = NULL;
 258         kfree(vq->heads);
 259         vq->heads = NULL;
 260 }
 261
 262 /* Helper to allocate iovec buffers for all vqs. */
 263 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
 264 {
 265         struct vhost_virtqueue *vq;
 266         int i;
 267
 268         for (i = 0; i < dev->nvqs; ++i) {
 269                 vq = dev->vqs[i];
 270                 vq->indirect = kmalloc(sizeof *vq->indirect * UIO_MAXIOV,
 271                                        GFP_KERNEL);
 272                 vq->log = kmalloc(sizeof *vq->log * UIO_MAXIOV, GFP_KERNEL);
 273                 vq->heads = kmalloc(sizeof *vq->heads * UIO_MAXIOV, GFP_KERNEL);
 274                 if (!vq->indirect || !vq->log || !vq->heads)
 275                         goto err_nomem;
 276         }
 277         return 0;
 278
 279 err_nomem:
 280         for (; i >= 0; --i)
 281                 vhost_vq_free_iovecs(dev->vqs[i]);
 282         return -ENOMEM;
 283 }
 284
 285 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
 286 {
 287         int i;
 288
 289         for (i = 0; i < dev->nvqs; ++i)
 290                 vhost_vq_free_iovecs(dev->vqs[i]);
 291 }
 292
 293 void vhost_dev_init(struct vhost_dev *dev,
 294                     struct vhost_virtqueue **vqs, int nvqs)
 295 {
 296         struct vhost_virtqueue *vq;
 297         int i;
 298
 299         dev->vqs = vqs;
 300         dev->nvqs = nvqs;
 301         mutex_init(&dev->mutex);
 302         dev->log_ctx = NULL;
 303         dev->log_file = NULL;
 304         dev->memory = NULL;
 305         dev->mm = NULL;
 306         spin_lock_init(&dev->work_lock);
 307         INIT_LIST_HEAD(&dev->work_list);
 308         dev->worker = NULL;
 309
 310         for (i = 0; i < dev->nvqs; ++i) {
 311                 vq = dev->vqs[i];
 312                 vq->log = NULL;
 313                 vq->indirect = NULL;
 314                 vq->heads = NULL;
 315                 vq->dev = dev;
 316                 mutex_init(&vq->mutex);
 317                 vhost_vq_reset(dev, vq);
 318                 if (vq->handle_kick)
 319                         vhost_poll_init(&vq->poll, vq->handle_kick,
 320                                         POLLIN, dev);
 321         }
 322 }
 323 EXPORT_SYMBOL_GPL(vhost_dev_init);
 324
 325 /* Caller should have device mutex */
 326 long vhost_dev_check_owner(struct vhost_dev *dev)
 327 {
 328         /* Are you the owner? If not, I don't think you mean to do that */
 329         return dev->mm == current->mm ? 0 : -EPERM;
 330 }
 331 EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
 332
 333 struct vhost_attach_cgroups_struct {
 334         struct vhost_work work;
 335         struct task_struct *owner;
 336         int ret;
 337 };
 338
 339 static void vhost_attach_cgroups_work(struct vhost_work *work)
 340 {
 341         struct vhost_attach_cgroups_struct *s;
 342
 343         s = container_of(work, struct vhost_attach_cgroups_struct, work);
 344         s->ret = cgroup_attach_task_all(s->owner, current);
 345 }
 346
 347 static int vhost_attach_cgroups(struct vhost_dev *dev)
 348 {
 349         struct vhost_attach_cgroups_struct attach;
 350
 351         attach.owner = current;
 352         vhost_work_init(&attach.work, vhost_attach_cgroups_work);
 353         vhost_work_queue(dev, &attach.work);
 354         vhost_work_flush(dev, &attach.work);
 355         return attach.ret;
 356 }
 357
 358 /* Caller should have device mutex */
 359 bool vhost_dev_has_owner(struct vhost_dev *dev)
 360 {
 361         return dev->mm;
 362 }
 363 EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
 364
 365 /* Caller should have device mutex */
 366 long vhost_dev_set_owner(struct vhost_dev *dev)
 367 {
 368         struct task_struct *worker;
 369         int err;
 370
 371         /* Is there an owner already? */
 372         if (vhost_dev_has_owner(dev)) {
 373                 err = -EBUSY;
 374                 goto err_mm;
 375         }
 376
 377         /* No owner, become one */
 378         dev->mm = get_task_mm(current);
 379         worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
 380         if (IS_ERR(worker)) {
 381                 err = PTR_ERR(worker);
 382                 goto err_worker;
 383         }
 384
 385         dev->worker = worker;
 386         wake_up_process(worker);        /* avoid contributing to loadavg */
 387
 388         err = vhost_attach_cgroups(dev);
 389         if (err)
 390                 goto err_cgroup;
 391
 392         err = vhost_dev_alloc_iovecs(dev);
 393         if (err)
 394                 goto err_cgroup;
 395
 396         return 0;
 397 err_cgroup:
 398         kthread_stop(worker);
 399         dev->worker = NULL;
 400 err_worker:
 401         if (dev->mm)
 402                 mmput(dev->mm);
 403         dev->mm = NULL;
 404 err_mm:
 405         return err;
 406 }
 407 EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
 408
 409 struct vhost_memory *vhost_dev_reset_owner_prepare(void)
 410 {
 411         return kmalloc(offsetof(struct vhost_memory, regions), GFP_KERNEL);
 412 }
 413 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
 414
 415 /* Caller should have device mutex */
 416 void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_memory *memory)
 417 {
 418         vhost_dev_cleanup(dev, true);
 419
 420         /* Restore memory to default empty mapping. */
 421         memory->nregions = 0;
 422         RCU_INIT_POINTER(dev->memory, memory);
 423 }
 424 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
 425
 426 void vhost_dev_stop(struct vhost_dev *dev)
 427 {
 428         int i;
 429
 430         for (i = 0; i < dev->nvqs; ++i) {
 431                 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick) {
 432                         vhost_poll_stop(&dev->vqs[i]->poll);
 433                         vhost_poll_flush(&dev->vqs[i]->poll);
 434                 }
 435         }
 436 }
 437 EXPORT_SYMBOL_GPL(vhost_dev_stop);
 438
 439 /* Caller should have device mutex if and only if locked is set */
 440 void vhost_dev_cleanup(struct vhost_dev *dev, bool locked)
 441 {
 442         int i;
 443
 444         for (i = 0; i < dev->nvqs; ++i) {
 445                 if (dev->vqs[i]->error_ctx)
 446                         eventfd_ctx_put(dev->vqs[i]->error_ctx);
 447                 if (dev->vqs[i]->error)
 448                         fput(dev->vqs[i]->error);
 449                 if (dev->vqs[i]->kick)
 450                         fput(dev->vqs[i]->kick);
 451                 if (dev->vqs[i]->call_ctx)
 452                         eventfd_ctx_put(dev->vqs[i]->call_ctx);
 453                 if (dev->vqs[i]->call)
 454                         fput(dev->vqs[i]->call);
 455                 vhost_vq_reset(dev, dev->vqs[i]);
 456         }
 457         vhost_dev_free_iovecs(dev);
 458         if (dev->log_ctx)
 459                 eventfd_ctx_put(dev->log_ctx);
 460         dev->log_ctx = NULL;
 461         if (dev->log_file)
 462                 fput(dev->log_file);
 463         dev->log_file = NULL;
 464         /* No one will access memory at this point */
 465         kfree(rcu_dereference_protected(dev->memory,
 466                                         locked ==
 467                                                 lockdep_is_held(&dev->mutex)));
 468         RCU_INIT_POINTER(dev->memory, NULL);
 469         WARN_ON(!list_empty(&dev->work_list));
 470         if (dev->worker) {
 471                 kthread_stop(dev->worker);
 472                 dev->worker = NULL;
 473         }
 474         if (dev->mm)
 475                 mmput(dev->mm);
 476         dev->mm = NULL;
 477 }
 478 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
 479
 480 static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
 481 {
 482         u64 a = addr / VHOST_PAGE_SIZE / 8;
 483
 484         /* Make sure 64 bit math will not overflow. */
 485         if (a > ULONG_MAX - (unsigned long)log_base ||
 486             a + (unsigned long)log_base > ULONG_MAX)
 487                 return 0;
 488
 489         return access_ok(VERIFY_WRITE, log_base + a,
 490                          (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
 491 }
 492
 493 /* Caller should have vq mutex and device mutex. */
 494 static int vq_memory_access_ok(void __user *log_base, struct vhost_memory *mem,
 495                                int log_all)
 496 {
 497         int i;
 498
 499         if (!mem)
 500                 return 0;
 501
 502         for (i = 0; i < mem->nregions; ++i) {
 503                 struct vhost_memory_region *m = mem->regions + i;
 504                 unsigned long a = m->userspace_addr;
 505                 if (m->memory_size > ULONG_MAX)
 506                         return 0;
 507                 else if (!access_ok(VERIFY_WRITE, (void __user *)a,
 508                                     m->memory_size))
 509                         return 0;
 510                 else if (log_all && !log_access_ok(log_base,
 511                                                    m->guest_phys_addr,
 512                                                    m->memory_size))
 513                         return 0;
 514         }
 515         return 1;
 516 }
 517
 518 /* Can we switch to this memory table? */
 519 /* Caller should have device mutex but not vq mutex */
 520 static int memory_access_ok(struct vhost_dev *d, struct vhost_memory *mem,
 521                             int log_all)
 522 {
 523         int i;
 524
 525         for (i = 0; i < d->nvqs; ++i) {
 526                 int ok;
 527                 mutex_lock(&d->vqs[i]->mutex);
 528                 /* If ring is inactive, will check when it's enabled. */
 529                 if (d->vqs[i]->private_data)
 530                         ok = vq_memory_access_ok(d->vqs[i]->log_base, mem,
 531                                                  log_all);
 532                 else
 533                         ok = 1;
 534                 mutex_unlock(&d->vqs[i]->mutex);
 535                 if (!ok)
 536                         return 0;
 537         }
 538         return 1;
 539 }
 540
 541 static int vq_access_ok(struct vhost_dev *d, unsigned int num,
 542                         struct vring_desc __user *desc,
 543                         struct vring_avail __user *avail,
 544                         struct vring_used __user *used)
 545 {
 546         size_t s = vhost_has_feature(d, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
 547         return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
 548                access_ok(VERIFY_READ, avail,
 549                          sizeof *avail + num * sizeof *avail->ring + s) &&
 550                access_ok(VERIFY_WRITE, used,
 551                         sizeof *used + num * sizeof *used->ring + s);
 552 }
 553
 554 /* Can we log writes? */
 555 /* Caller should have device mutex but not vq mutex */
 556 int vhost_log_access_ok(struct vhost_dev *dev)
 557 {
 558         struct vhost_memory *mp;
 559
 560         mp = rcu_dereference_protected(dev->memory,
 561                                        lockdep_is_held(&dev->mutex));
 562         return memory_access_ok(dev, mp, 1);
 563 }
 564 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
 565
 566 /* Verify access for write logging. */
 567 /* Caller should have vq mutex and device mutex */
 568 static int vq_log_access_ok(struct vhost_dev *d, struct vhost_virtqueue *vq,
 569                             void __user *log_base)
 570 {
 571         struct vhost_memory *mp;
 572         size_t s = vhost_has_feature(d, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
 573
 574         mp = rcu_dereference_protected(vq->dev->memory,
 575                                        lockdep_is_held(&vq->mutex));
 576         return vq_memory_access_ok(log_base, mp,
 577                             vhost_has_feature(vq->dev, VHOST_F_LOG_ALL)) &&
 578                 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
 579                                         sizeof *vq->used +
 580                                         vq->num * sizeof *vq->used->ring + s));
 581 }
 582
 583 /* Can we start vq? */
 584 /* Caller should have vq mutex and device mutex */
 585 int vhost_vq_access_ok(struct vhost_virtqueue *vq)
 586 {
 587         return vq_access_ok(vq->dev, vq->num, vq->desc, vq->avail, vq->used) &&
 588                 vq_log_access_ok(vq->dev, vq, vq->log_base);
 589 }
 590 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
 591
 592 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
 593 {
 594         struct vhost_memory mem, *newmem, *oldmem;
 595         unsigned long size = offsetof(struct vhost_memory, regions);
 596
 597         if (copy_from_user(&mem, m, size))
 598                 return -EFAULT;
 599         if (mem.padding)
 600                 return -EOPNOTSUPP;
 601         if (mem.nregions > VHOST_MEMORY_MAX_NREGIONS)
 602                 return -E2BIG;
 603         newmem = kmalloc(size + mem.nregions * sizeof *m->regions, GFP_KERNEL);
 604         if (!newmem)
 605                 return -ENOMEM;
 606
 607         memcpy(newmem, &mem, size);
 608         if (copy_from_user(newmem->regions, m->regions,
 609                            mem.nregions * sizeof *m->regions)) {
 610                 kfree(newmem);
 611                 return -EFAULT;
 612         }
 613
 614         if (!memory_access_ok(d, newmem,
 615                               vhost_has_feature(d, VHOST_F_LOG_ALL))) {
 616                 kfree(newmem);
 617                 return -EFAULT;
 618         }
 619         oldmem = rcu_dereference_protected(d->memory,
 620                                            lockdep_is_held(&d->mutex));
 621         rcu_assign_pointer(d->memory, newmem);
 622         synchronize_rcu();
 623         kfree(oldmem);
 624         return 0;
 625 }
 626
 627 long vhost_vring_ioctl(struct vhost_dev *d, int ioctl, void __user *argp)
 628 {
 629         struct file *eventfp, *filep = NULL;
 630         bool pollstart = false, pollstop = false;
 631         struct eventfd_ctx *ctx = NULL;
 632         u32 __user *idxp = argp;
 633         struct vhost_virtqueue *vq;
 634         struct vhost_vring_state s;
 635         struct vhost_vring_file f;
 636         struct vhost_vring_addr a;
 637         u32 idx;
 638         long r;
 639
 640         r = get_user(idx, idxp);
 641         if (r < 0)
 642                 return r;
 643         if (idx >= d->nvqs)
 644                 return -ENOBUFS;
 645
 646         vq = d->vqs[idx];
 647
 648         mutex_lock(&vq->mutex);
 649
 650         switch (ioctl) {
 651         case VHOST_SET_VRING_NUM:
 652                 /* Resizing ring with an active backend?
 653                  * You don't want to do that. */
 654                 if (vq->private_data) {
 655                         r = -EBUSY;
 656                         break;
 657                 }
 658                 if (copy_from_user(&s, argp, sizeof s)) {
 659                         r = -EFAULT;
 660                         break;
 661                 }
 662                 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
 663                         r = -EINVAL;
 664                         break;
 665                 }
 666                 vq->num = s.num;
 667                 break;
 668         case VHOST_SET_VRING_BASE:
 669                 /* Moving base with an active backend?
 670                  * You don't want to do that. */
 671                 if (vq->private_data) {
 672                         r = -EBUSY;
 673                         break;
 674                 }
 675                 if (copy_from_user(&s, argp, sizeof s)) {
 676                         r = -EFAULT;
 677                         break;
 678                 }
 679                 if (s.num > 0xffff) {
 680                         r = -EINVAL;
 681                         break;
 682                 }
 683                 vq->last_avail_idx = s.num;
 684                 /* Forget the cached index value. */
 685                 vq->avail_idx = vq->last_avail_idx;
 686                 break;
 687         case VHOST_GET_VRING_BASE:
 688                 s.index = idx;
 689                 s.num = vq->last_avail_idx;
 690                 if (copy_to_user(argp, &s, sizeof s))
 691                         r = -EFAULT;
 692                 break;
 693         case VHOST_SET_VRING_ADDR:
 694                 if (copy_from_user(&a, argp, sizeof a)) {
 695                         r = -EFAULT;
 696                         break;
 697                 }
 698                 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
 699                         r = -EOPNOTSUPP;
 700                         break;
 701                 }
 702                 /* For 32bit, verify that the top 32bits of the user
 703                    data are set to zero. */
 704                 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
 705                     (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
 706                     (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
 707                         r = -EFAULT;
 708                         break;
 709                 }
 710                 if ((a.avail_user_addr & (sizeof *vq->avail->ring - 1)) ||
 711                     (a.used_user_addr & (sizeof *vq->used->ring - 1)) ||
 712                     (a.log_guest_addr & (sizeof *vq->used->ring - 1))) {
 713                         r = -EINVAL;
 714                         break;
 715                 }
 716
 717                 /* We only verify access here if backend is configured.
 718                  * If it is not, we don't as size might not have been setup.
 719                  * We will verify when backend is configured. */
 720                 if (vq->private_data) {
 721                         if (!vq_access_ok(d, vq->num,
 722                                 (void __user *)(unsigned long)a.desc_user_addr,
 723                                 (void __user *)(unsigned long)a.avail_user_addr,
 724                                 (void __user *)(unsigned long)a.used_user_addr)) {
 725                                 r = -EINVAL;
 726                                 break;
 727                         }
 728
 729                         /* Also validate log access for used ring if enabled. */
 730                         if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
 731                             !log_access_ok(vq->log_base, a.log_guest_addr,
 732                                            sizeof *vq->used +
 733                                            vq->num * sizeof *vq->used->ring)) {
 734                                 r = -EINVAL;
 735                                 break;
 736                         }
 737                 }
 738
 739                 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
 740                 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
 741                 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
 742                 vq->log_addr = a.log_guest_addr;
 743                 vq->used = (void __user *)(unsigned long)a.used_user_addr;
 744                 break;
 745         case VHOST_SET_VRING_KICK:
 746                 if (copy_from_user(&f, argp, sizeof f)) {
 747                         r = -EFAULT;
 748                         break;
 749                 }
 750                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
 751                 if (IS_ERR(eventfp)) {
 752                         r = PTR_ERR(eventfp);
 753                         break;
 754                 }
 755                 if (eventfp != vq->kick) {
 756                         pollstop = (filep = vq->kick) != NULL;
 757                         pollstart = (vq->kick = eventfp) != NULL;
 758                 } else
 759                         filep = eventfp;
 760                 break;
 761         case VHOST_SET_VRING_CALL:
 762                 if (copy_from_user(&f, argp, sizeof f)) {
 763                         r = -EFAULT;
 764                         break;
 765                 }
 766                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
 767                 if (IS_ERR(eventfp)) {
 768                         r = PTR_ERR(eventfp);
 769                         break;
 770                 }
 771                 if (eventfp != vq->call) {
 772                         filep = vq->call;
 773                         ctx = vq->call_ctx;
 774                         vq->call = eventfp;
 775                         vq->call_ctx = eventfp ?
 776                                 eventfd_ctx_fileget(eventfp) : NULL;
 777                 } else
 778                         filep = eventfp;
 779                 break;
 780         case VHOST_SET_VRING_ERR:
 781                 if (copy_from_user(&f, argp, sizeof f)) {
 782                         r = -EFAULT;
 783                         break;
 784                 }
 785                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
 786                 if (IS_ERR(eventfp)) {
 787                         r = PTR_ERR(eventfp);
 788                         break;
 789                 }
 790                 if (eventfp != vq->error) {
 791                         filep = vq->error;
 792                         vq->error = eventfp;
 793                         ctx = vq->error_ctx;
 794                         vq->error_ctx = eventfp ?
 795                                 eventfd_ctx_fileget(eventfp) : NULL;
 796                 } else
 797                         filep = eventfp;
 798                 break;
 799         default:
 800                 r = -ENOIOCTLCMD;
 801         }
 802
 803         if (pollstop && vq->handle_kick)
 804                 vhost_poll_stop(&vq->poll);
 805
 806         if (ctx)
 807                 eventfd_ctx_put(ctx);
 808         if (filep)
 809                 fput(filep);
 810
 811         if (pollstart && vq->handle_kick)
 812                 r = vhost_poll_start(&vq->poll, vq->kick);
 813
 814         mutex_unlock(&vq->mutex);
 815
 816         if (pollstop && vq->handle_kick)
 817                 vhost_poll_flush(&vq->poll);
 818         return r;
 819 }
 820 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
 821
 822 /* Caller must have device mutex */
 823 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
 824 {
 825         struct file *eventfp, *filep = NULL;
 826         struct eventfd_ctx *ctx = NULL;
 827         u64 p;
 828         long r;
 829         int i, fd;
 830
 831         /* If you are not the owner, you can become one */
 832         if (ioctl == VHOST_SET_OWNER) {
 833                 r = vhost_dev_set_owner(d);
 834                 goto done;
 835         }
 836
 837         /* You must be the owner to do anything else */
 838         r = vhost_dev_check_owner(d);
 839         if (r)
 840                 goto done;
 841
 842         switch (ioctl) {
 843         case VHOST_SET_MEM_TABLE:
 844                 r = vhost_set_memory(d, argp);
 845                 break;
 846         case VHOST_SET_LOG_BASE:
 847                 if (copy_from_user(&p, argp, sizeof p)) {
 848                         r = -EFAULT;
 849                         break;
 850                 }
 851                 if ((u64)(unsigned long)p != p) {
 852                         r = -EFAULT;
 853                         break;
 854                 }
 855                 for (i = 0; i < d->nvqs; ++i) {
 856                         struct vhost_virtqueue *vq;
 857                         void __user *base = (void __user *)(unsigned long)p;
 858                         vq = d->vqs[i];
 859                         mutex_lock(&vq->mutex);
 860                         /* If ring is inactive, will check when it's enabled. */
 861                         if (vq->private_data && !vq_log_access_ok(d, vq, base))
 862                                 r = -EFAULT;
 863                         else
 864                                 vq->log_base = base;
 865                         mutex_unlock(&vq->mutex);
 866                 }
 867                 break;
 868         case VHOST_SET_LOG_FD:
 869                 r = get_user(fd, (int __user *)argp);
 870                 if (r < 0)
 871                         break;
 872                 eventfp = fd == -1 ? NULL : eventfd_fget(fd);
 873                 if (IS_ERR(eventfp)) {
 874                         r = PTR_ERR(eventfp);
 875                         break;
 876                 }
 877                 if (eventfp != d->log_file) {
 878                         filep = d->log_file;
 879                         ctx = d->log_ctx;
 880                         d->log_ctx = eventfp ?
 881                                 eventfd_ctx_fileget(eventfp) : NULL;
 882                 } else
 883                         filep = eventfp;
 884                 for (i = 0; i < d->nvqs; ++i) {
 885                         mutex_lock(&d->vqs[i]->mutex);
 886                         d->vqs[i]->log_ctx = d->log_ctx;
 887                         mutex_unlock(&d->vqs[i]->mutex);
 888                 }
 889                 if (ctx)
 890                         eventfd_ctx_put(ctx);
 891                 if (filep)
 892                         fput(filep);
 893                 break;
 894         default:
 895                 r = -ENOIOCTLCMD;
 896                 break;
 897         }
 898 done:
 899         return r;
 900 }
 901 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
 902
 903 static const struct vhost_memory_region *find_region(struct vhost_memory *mem,
 904                                                      __u64 addr, __u32 len)
 905 {
 906         struct vhost_memory_region *reg;
 907         int i;
 908
 909         /* linear search is not brilliant, but we really have on the order of 6
 910          * regions in practice */
 911         for (i = 0; i < mem->nregions; ++i) {
 912                 reg = mem->regions + i;
 913                 if (reg->guest_phys_addr <= addr &&
 914                     reg->guest_phys_addr + reg->memory_size - 1 >= addr)
 915                         return reg;
 916         }
 917         return NULL;
 918 }
 919
 920 /* TODO: This is really inefficient.  We need something like get_user()
 921  * (instruction directly accesses the data, with an exception table entry
 922  * returning -EFAULT). See Documentation/x86/exception-tables.txt.
 923  */
 924 static int set_bit_to_user(int nr, void __user *addr)
 925 {
 926         unsigned long log = (unsigned long)addr;
 927         struct page *page;
 928         void *base;
 929         int bit = nr + (log % PAGE_SIZE) * 8;
 930         int r;
 931
 932         r = get_user_pages_fast(log, 1, 1, &page);
 933         if (r < 0)
 934                 return r;
 935         BUG_ON(r != 1);
 936         base = kmap_atomic(page);
 937         set_bit(bit, base);
 938         kunmap_atomic(base);
 939         set_page_dirty_lock(page);
 940         put_page(page);
 941         return 0;
 942 }
 943
 944 static int log_write(void __user *log_base,
 945                      u64 write_address, u64 write_length)
 946 {
 947         u64 write_page = write_address / VHOST_PAGE_SIZE;
 948         int r;
 949
 950         if (!write_length)
 951                 return 0;
 952         write_length += write_address % VHOST_PAGE_SIZE;
 953         for (;;) {
 954                 u64 base = (u64)(unsigned long)log_base;
 955                 u64 log = base + write_page / 8;
 956                 int bit = write_page % 8;
 957                 if ((u64)(unsigned long)log != log)
 958                         return -EFAULT;
 959                 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
 960                 if (r < 0)
 961                         return r;
 962                 if (write_length <= VHOST_PAGE_SIZE)
 963                         break;
 964                 write_length -= VHOST_PAGE_SIZE;
 965                 write_page += 1;
 966         }
 967         return r;
 968 }
 969
 970 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
 971                     unsigned int log_num, u64 len)
 972 {
 973         int i, r;
 974
 975         /* Make sure data written is seen before log. */
 976         smp_wmb();
 977         for (i = 0; i < log_num; ++i) {
 978                 u64 l = min(log[i].len, len);
 979                 r = log_write(vq->log_base, log[i].addr, l);
 980                 if (r < 0)
 981                         return r;
 982                 len -= l;
 983                 if (!len) {
 984                         if (vq->log_ctx)
 985                                 eventfd_signal(vq->log_ctx, 1);
 986                         return 0;
 987                 }
 988         }
 989         /* Length written exceeds what we have stored. This is a bug. */
 990         BUG();
 991         return 0;
 992 }
 993 EXPORT_SYMBOL_GPL(vhost_log_write);
 994
 995 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
 996 {
 997         void __user *used;
 998         if (__put_user(vq->used_flags, &vq->used->flags) < 0)
 999                 return -EFAULT;
1000         if (unlikely(vq->log_used)) {
1001                 /* Make sure the flag is seen before log. */
1002                 smp_wmb();
1003                 /* Log used flag write. */
1004                 used = &vq->used->flags;
1005                 log_write(vq->log_base, vq->log_addr +
1006                           (used - (void __user *)vq->used),
1007                           sizeof vq->used->flags);
1008                 if (vq->log_ctx)
1009                         eventfd_signal(vq->log_ctx, 1);
1010         }
1011         return 0;
1012 }
1013
1014 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1015 {
1016         if (__put_user(vq->avail_idx, vhost_avail_event(vq)))
1017                 return -EFAULT;
1018         if (unlikely(vq->log_used)) {
1019                 void __user *used;
1020                 /* Make sure the event is seen before log. */
1021                 smp_wmb();
1022                 /* Log avail event write */
1023                 used = vhost_avail_event(vq);
1024                 log_write(vq->log_base, vq->log_addr +
1025                           (used - (void __user *)vq->used),
1026                           sizeof *vhost_avail_event(vq));
1027                 if (vq->log_ctx)
1028                         eventfd_signal(vq->log_ctx, 1);
1029         }
1030         return 0;
1031 }
1032
1033 int vhost_init_used(struct vhost_virtqueue *vq)
1034 {
1035         int r;
1036         if (!vq->private_data)
1037                 return 0;
1038
1039         r = vhost_update_used_flags(vq);
1040         if (r)
1041                 return r;
1042         vq->signalled_used_valid = false;
1043         return get_user(vq->last_used_idx, &vq->used->idx);
1044 }
1045 EXPORT_SYMBOL_GPL(vhost_init_used);
1046
1047 static int translate_desc(struct vhost_dev *dev, u64 addr, u32 len,
1048                           struct iovec iov[], int iov_size)
1049 {
1050         const struct vhost_memory_region *reg;
1051         struct vhost_memory *mem;
1052         struct iovec *_iov;
1053         u64 s = 0;
1054         int ret = 0;
1055
1056         rcu_read_lock();
1057
1058         mem = rcu_dereference(dev->memory);
1059         while ((u64)len > s) {
1060                 u64 size;
1061                 if (unlikely(ret >= iov_size)) {
1062                         ret = -ENOBUFS;
1063                         break;
1064                 }
1065                 reg = find_region(mem, addr, len);
1066                 if (unlikely(!reg)) {
1067                         ret = -EFAULT;
1068                         break;
1069                 }
1070                 _iov = iov + ret;
1071                 size = reg->memory_size - addr + reg->guest_phys_addr;
1072                 _iov->iov_len = min((u64)len - s, size);
1073                 _iov->iov_base = (void __user *)(unsigned long)
1074                         (reg->userspace_addr + addr - reg->guest_phys_addr);
1075                 s += size;
1076                 addr += size;
1077                 ++ret;
1078         }
1079
1080         rcu_read_unlock();
1081         return ret;
1082 }
1083
1084 /* Each buffer in the virtqueues is actually a chain of descriptors.  This
1085  * function returns the next descriptor in the chain,
1086  * or -1U if we're at the end. */
1087 static unsigned next_desc(struct vring_desc *desc)
1088 {
1089         unsigned int next;
1090
1091         /* If this descriptor says it doesn't chain, we're done. */
1092         if (!(desc->flags & VRING_DESC_F_NEXT))
1093                 return -1U;
1094
1095         /* Check they're not leading us off end of descriptors. */
1096         next = desc->next;
1097         /* Make sure compiler knows to grab that: we don't want it changing! */
1098         /* We will use the result as an index in an array, so most
1099          * architectures only need a compiler barrier here. */
1100         read_barrier_depends();
1101
1102         return next;
1103 }
1104
1105 static int get_indirect(struct vhost_dev *dev, struct vhost_virtqueue *vq,
1106                         struct iovec iov[], unsigned int iov_size,
1107                         unsigned int *out_num, unsigned int *in_num,
1108                         struct vhost_log *log, unsigned int *log_num,
1109                         struct vring_desc *indirect)
1110 {
1111         struct vring_desc desc;
1112         unsigned int i = 0, count, found = 0;
1113         int ret;
1114
1115         /* Sanity check */
1116         if (unlikely(indirect->len % sizeof desc)) {
1117                 vq_err(vq, "Invalid length in indirect descriptor: "
1118                        "len 0x%llx not multiple of 0x%zx\n",
1119                        (unsigned long long)indirect->len,
1120                        sizeof desc);
1121                 return -EINVAL;
1122         }
1123
1124         ret = translate_desc(dev, indirect->addr, indirect->len, vq->indirect,
1125                              UIO_MAXIOV);
1126         if (unlikely(ret < 0)) {
1127                 vq_err(vq, "Translation failure %d in indirect.\n", ret);
1128                 return ret;
1129         }
1130
1131         /* We will use the result as an address to read from, so most
1132          * architectures only need a compiler barrier here. */
1133         read_barrier_depends();
1134
1135         count = indirect->len / sizeof desc;
1136         /* Buffers are chained via a 16 bit next field, so
1137          * we can have at most 2^16 of these. */
1138         if (unlikely(count > USHRT_MAX + 1)) {
1139                 vq_err(vq, "Indirect buffer length too big: %d\n",
1140                        indirect->len);
1141                 return -E2BIG;
1142         }
1143
1144         do {
1145                 unsigned iov_count = *in_num + *out_num;
1146                 if (unlikely(++found > count)) {
1147                         vq_err(vq, "Loop detected: last one at %u "
1148                                "indirect size %u\n",
1149                                i, count);
1150                         return -EINVAL;
1151                 }
1152                 if (unlikely(memcpy_fromiovec((unsigned char *)&desc,
1153                                               vq->indirect, sizeof desc))) {
1154                         vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1155                                i, (size_t)indirect->addr + i * sizeof desc);
1156                         return -EINVAL;
1157                 }
1158                 if (unlikely(desc.flags & VRING_DESC_F_INDIRECT)) {
1159                         vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1160                                i, (size_t)indirect->addr + i * sizeof desc);
1161                         return -EINVAL;
1162                 }
1163
1164                 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
1165                                      iov_size - iov_count);
1166                 if (unlikely(ret < 0)) {
1167                         vq_err(vq, "Translation failure %d indirect idx %d\n",
1168                                ret, i);
1169                         return ret;
1170                 }
1171                 /* If this is an input descriptor, increment that count. */
1172                 if (desc.flags & VRING_DESC_F_WRITE) {
1173                         *in_num += ret;
1174                         if (unlikely(log)) {
1175                                 log[*log_num].addr = desc.addr;
1176                                 log[*log_num].len = desc.len;
1177                                 ++*log_num;
1178                         }
1179                 } else {
1180                         /* If it's an output descriptor, they're all supposed
1181                          * to come before any input descriptors. */
1182                         if (unlikely(*in_num)) {
1183                                 vq_err(vq, "Indirect descriptor "
1184                                        "has out after in: idx %d\n", i);
1185                                 return -EINVAL;
1186                         }
1187                         *out_num += ret;
1188                 }
1189         } while ((i = next_desc(&desc)) != -1);
1190         return 0;
1191 }
1192
1193 /* This looks in the virtqueue and for the first available buffer, and converts
1194  * it to an iovec for convenient access.  Since descriptors consist of some
1195  * number of output then some number of input descriptors, it's actually two
1196  * iovecs, but we pack them into one and note how many of each there were.
1197  *
1198  * This function returns the descriptor number found, or vq->num (which is
1199  * never a valid descriptor number) if none was found.  A negative code is
1200  * returned on error. */
1201 int vhost_get_vq_desc(struct vhost_dev *dev, struct vhost_virtqueue *vq,
1202                       struct iovec iov[], unsigned int iov_size,
1203                       unsigned int *out_num, unsigned int *in_num,
1204                       struct vhost_log *log, unsigned int *log_num)
1205 {
1206         struct vring_desc desc;
1207         unsigned int i, head, found = 0;
1208         u16 last_avail_idx;
1209         int ret;
1210
1211         /* Check it isn't doing very strange things with descriptor numbers. */
1212         last_avail_idx = vq->last_avail_idx;
1213         if (unlikely(__get_user(vq->avail_idx, &vq->avail->idx))) {
1214                 vq_err(vq, "Failed to access avail idx at %p\n",
1215                        &vq->avail->idx);
1216                 return -EFAULT;
1217         }
1218
1219         if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
1220                 vq_err(vq, "Guest moved used index from %u to %u",
1221                        last_avail_idx, vq->avail_idx);
1222                 return -EFAULT;
1223         }
1224
1225         /* If there's nothing new since last we looked, return invalid. */
1226         if (vq->avail_idx == last_avail_idx)
1227                 return vq->num;
1228
1229         /* Only get avail ring entries after they have been exposed by guest. */
1230         smp_rmb();
1231
1232         /* Grab the next descriptor number they're advertising, and increment
1233          * the index we've seen. */
1234         if (unlikely(__get_user(head,
1235                                 &vq->avail->ring[last_avail_idx % vq->num]))) {
1236                 vq_err(vq, "Failed to read head: idx %d address %p\n",
1237                        last_avail_idx,
1238                        &vq->avail->ring[last_avail_idx % vq->num]);
1239                 return -EFAULT;
1240         }
1241
1242         /* If their number is silly, that's an error. */
1243         if (unlikely(head >= vq->num)) {
1244                 vq_err(vq, "Guest says index %u > %u is available",
1245                        head, vq->num);
1246                 return -EINVAL;
1247         }
1248
1249         /* When we start there are none of either input nor output. */
1250         *out_num = *in_num = 0;
1251         if (unlikely(log))
1252                 *log_num = 0;
1253
1254         i = head;
1255         do {
1256                 unsigned iov_count = *in_num + *out_num;
1257                 if (unlikely(i >= vq->num)) {
1258                         vq_err(vq, "Desc index is %u > %u, head = %u",
1259                                i, vq->num, head);
1260                         return -EINVAL;
1261                 }
1262                 if (unlikely(++found > vq->num)) {
1263                         vq_err(vq, "Loop detected: last one at %u "
1264                                "vq size %u head %u\n",
1265                                i, vq->num, head);
1266                         return -EINVAL;
1267                 }
1268                 ret = __copy_from_user(&desc, vq->desc + i, sizeof desc);
1269                 if (unlikely(ret)) {
1270                         vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
1271                                i, vq->desc + i);
1272                         return -EFAULT;
1273                 }
1274                 if (desc.flags & VRING_DESC_F_INDIRECT) {
1275                         ret = get_indirect(dev, vq, iov, iov_size,
1276                                            out_num, in_num,
1277                                            log, log_num, &desc);
1278                         if (unlikely(ret < 0)) {
1279                                 vq_err(vq, "Failure detected "
1280                                        "in indirect descriptor at idx %d\n", i);
1281                                 return ret;
1282                         }
1283                         continue;
1284                 }
1285
1286                 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
1287                                      iov_size - iov_count);
1288                 if (unlikely(ret < 0)) {
1289                         vq_err(vq, "Translation failure %d descriptor idx %d\n",
1290                                ret, i);
1291                         return ret;
1292                 }
1293                 if (desc.flags & VRING_DESC_F_WRITE) {
1294                         /* If this is an input descriptor,
1295                          * increment that count. */
1296                         *in_num += ret;
1297                         if (unlikely(log)) {
1298                                 log[*log_num].addr = desc.addr;
1299                                 log[*log_num].len = desc.len;
1300                                 ++*log_num;
1301                         }
1302                 } else {
1303                         /* If it's an output descriptor, they're all supposed
1304                          * to come before any input descriptors. */
1305                         if (unlikely(*in_num)) {
1306                                 vq_err(vq, "Descriptor has out after in: "
1307                                        "idx %d\n", i);
1308                                 return -EINVAL;
1309                         }
1310                         *out_num += ret;
1311                 }
1312         } while ((i = next_desc(&desc)) != -1);
1313
1314         /* On success, increment avail index. */
1315         vq->last_avail_idx++;
1316
1317         /* Assume notifications from guest are disabled at this point,
1318          * if they aren't we would need to update avail_event index. */
1319         BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
1320         return head;
1321 }
1322 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
1323
1324 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
1325 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
1326 {
1327         vq->last_avail_idx -= n;
1328 }
1329 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
1330
1331 /* After we've used one of their buffers, we tell them about it.  We'll then
1332  * want to notify the guest, using eventfd. */
1333 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
1334 {
1335         struct vring_used_elem heads = { head, len };
1336
1337         return vhost_add_used_n(vq, &heads, 1);
1338 }
1339 EXPORT_SYMBOL_GPL(vhost_add_used);
1340
1341 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
1342                             struct vring_used_elem *heads,
1343                             unsigned count)
1344 {
1345         struct vring_used_elem __user *used;
1346         u16 old, new;
1347         int start;
1348
1349         start = vq->last_used_idx % vq->num;
1350         used = vq->used->ring + start;
1351         if (count == 1) {
1352                 if (__put_user(heads[0].id, &used->id)) {
1353                         vq_err(vq, "Failed to write used id");
1354                         return -EFAULT;
1355                 }
1356                 if (__put_user(heads[0].len, &used->len)) {
1357                         vq_err(vq, "Failed to write used len");
1358                         return -EFAULT;
1359                 }
1360         } else if (__copy_to_user(used, heads, count * sizeof *used)) {
1361                 vq_err(vq, "Failed to write used");
1362                 return -EFAULT;
1363         }
1364         if (unlikely(vq->log_used)) {
1365                 /* Make sure data is seen before log. */
1366                 smp_wmb();
1367                 /* Log used ring entry write. */
1368                 log_write(vq->log_base,
1369                           vq->log_addr +
1370                            ((void __user *)used - (void __user *)vq->used),
1371                           count * sizeof *used);
1372         }
1373         old = vq->last_used_idx;
1374         new = (vq->last_used_idx += count);
1375         /* If the driver never bothers to signal in a very long while,
1376          * used index might wrap around. If that happens, invalidate
1377          * signalled_used index we stored. TODO: make sure driver
1378          * signals at least once in 2^16 and remove this. */
1379         if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
1380                 vq->signalled_used_valid = false;
1381         return 0;
1382 }
1383
1384 /* After we've used one of their buffers, we tell them about it.  We'll then
1385  * want to notify the guest, using eventfd. */
1386 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
1387                      unsigned count)
1388 {
1389         int start, n, r;
1390
1391         start = vq->last_used_idx % vq->num;
1392         n = vq->num - start;
1393         if (n < count) {
1394                 r = __vhost_add_used_n(vq, heads, n);
1395                 if (r < 0)
1396                         return r;
1397                 heads += n;
1398                 count -= n;
1399         }
1400         r = __vhost_add_used_n(vq, heads, count);
1401
1402         /* Make sure buffer is written before we update index. */
1403         smp_wmb();
1404         if (put_user(vq->last_used_idx, &vq->used->idx)) {
1405                 vq_err(vq, "Failed to increment used idx");
1406                 return -EFAULT;
1407         }
1408         if (unlikely(vq->log_used)) {
1409                 /* Log used index update. */
1410                 log_write(vq->log_base,
1411                           vq->log_addr + offsetof(struct vring_used, idx),
1412                           sizeof vq->used->idx);
1413                 if (vq->log_ctx)
1414                         eventfd_signal(vq->log_ctx, 1);
1415         }
1416         return r;
1417 }
1418 EXPORT_SYMBOL_GPL(vhost_add_used_n);
1419
1420 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1421 {
1422         __u16 old, new, event;
1423         bool v;
1424         /* Flush out used index updates. This is paired
1425          * with the barrier that the Guest executes when enabling
1426          * interrupts. */
1427         smp_mb();
1428
1429         if (vhost_has_feature(dev, VIRTIO_F_NOTIFY_ON_EMPTY) &&
1430             unlikely(vq->avail_idx == vq->last_avail_idx))
1431                 return true;
1432
1433         if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1434                 __u16 flags;
1435                 if (__get_user(flags, &vq->avail->flags)) {
1436                         vq_err(vq, "Failed to get flags");
1437                         return true;
1438                 }
1439                 return !(flags & VRING_AVAIL_F_NO_INTERRUPT);
1440         }
1441         old = vq->signalled_used;
1442         v = vq->signalled_used_valid;
1443         new = vq->signalled_used = vq->last_used_idx;
1444         vq->signalled_used_valid = true;
1445
1446         if (unlikely(!v))
1447                 return true;
1448
1449         if (get_user(event, vhost_used_event(vq))) {
1450                 vq_err(vq, "Failed to get used event idx");
1451                 return true;
1452         }
1453         return vring_need_event(event, new, old);
1454 }
1455
1456 /* This actually signals the guest, using eventfd. */
1457 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1458 {
1459         /* Signal the Guest tell them we used something up. */
1460         if (vq->call_ctx && vhost_notify(dev, vq))
1461                 eventfd_signal(vq->call_ctx, 1);
1462 }
1463 EXPORT_SYMBOL_GPL(vhost_signal);
1464
1465 /* And here's the combo meal deal.  Supersize me! */
1466 void vhost_add_used_and_signal(struct vhost_dev *dev,
1467                                struct vhost_virtqueue *vq,
1468                                unsigned int head, int len)
1469 {
1470         vhost_add_used(vq, head, len);
1471         vhost_signal(dev, vq);
1472 }
1473 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
1474
1475 /* multi-buffer version of vhost_add_used_and_signal */
1476 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
1477                                  struct vhost_virtqueue *vq,
1478                                  struct vring_used_elem *heads, unsigned count)
1479 {
1480         vhost_add_used_n(vq, heads, count);
1481         vhost_signal(dev, vq);
1482 }
1483 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
1484
1485 /* OK, now we need to know about added descriptors. */
1486 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1487 {
1488         u16 avail_idx;
1489         int r;
1490
1491         if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
1492                 return false;
1493         vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
1494         if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1495                 r = vhost_update_used_flags(vq);
1496                 if (r) {
1497                         vq_err(vq, "Failed to enable notification at %p: %d\n",
1498                                &vq->used->flags, r);
1499                         return false;
1500                 }
1501         } else {
1502                 r = vhost_update_avail_event(vq, vq->avail_idx);
1503                 if (r) {
1504                         vq_err(vq, "Failed to update avail event index at %p: %d\n",
1505                                vhost_avail_event(vq), r);
1506                         return false;
1507                 }
1508         }
1509         /* They could have slipped one in as we were doing that: make
1510          * sure it's written, then check again. */
1511         smp_mb();
1512         r = __get_user(avail_idx, &vq->avail->idx);
1513         if (r) {
1514                 vq_err(vq, "Failed to check avail idx at %p: %d\n",
1515                        &vq->avail->idx, r);
1516                 return false;
1517         }
1518
1519         return avail_idx != vq->avail_idx;
1520 }
1521 EXPORT_SYMBOL_GPL(vhost_enable_notify);
1522
1523 /* We don't need to be notified again. */
1524 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1525 {
1526         int r;
1527
1528         if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
1529                 return;
1530         vq->used_flags |= VRING_USED_F_NO_NOTIFY;
1531         if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1532                 r = vhost_update_used_flags(vq);
1533                 if (r)
1534                         vq_err(vq, "Failed to enable notification at %p: %d\n",
1535                                &vq->used->flags, r);
1536         }
1537 }
1538 EXPORT_SYMBOL_GPL(vhost_disable_notify);
1539
1540 static int __init vhost_init(void)
1541 {
1542         return 0;
1543 }
1544
1545 static void __exit vhost_exit(void)
1546 {
1547 }
1548
1549 module_init(vhost_init);
1550 module_exit(vhost_exit);
1551
1552 MODULE_VERSION("0.0.1");
1553 MODULE_LICENSE("GPL v2");
1554 MODULE_AUTHOR("Michael S. Tsirkin");
1555 MODULE_DESCRIPTION("Host kernel accelerator for virtio");