kernel/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/kthread.h>
  26 #include <linux/cgroup.h>
  27 #include <linux/module.h>
  28
  29 #include "vhost.h"
  30
  31 enum {
  32         VHOST_MEMORY_MAX_NREGIONS = 64,
  33         VHOST_MEMORY_F_LOG = 0x1,
  34 };
  35
  36 #define vhost_used_event(vq) ((__virtio16 __user *)&vq->avail->ring[vq->num])
  37 #define vhost_avail_event(vq) ((__virtio16 __user *)&vq->used->ring[vq->num])
  38
  39 static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
  40                             poll_table *pt)
  41 {
  42         struct vhost_poll *poll;
  43
  44         poll = container_of(pt, struct vhost_poll, table);
  45         poll->wqh = wqh;
  46         add_wait_queue(wqh, &poll->wait);
  47 }
  48
  49 static int vhost_poll_wakeup(wait_queue_t *wait, unsigned mode, int sync,
  50                              void *key)
  51 {
  52         struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
  53
  54         if (!((unsigned long)key & poll->mask))
  55                 return 0;
  56
  57         vhost_poll_queue(poll);
  58         return 0;
  59 }
  60
  61 void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
  62 {
  63         INIT_LIST_HEAD(&work->node);
  64         work->fn = fn;
  65         init_waitqueue_head(&work->done);
  66         work->flushing = 0;
  67         work->queue_seq = work->done_seq = 0;
  68 }
  69 EXPORT_SYMBOL_GPL(vhost_work_init);
  70
  71 /* Init poll structure */
  72 void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
  73                      unsigned long mask, struct vhost_dev *dev)
  74 {
  75         init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
  76         init_poll_funcptr(&poll->table, vhost_poll_func);
  77         poll->mask = mask;
  78         poll->dev = dev;
  79         poll->wqh = NULL;
  80
  81         vhost_work_init(&poll->work, fn);
  82 }
  83 EXPORT_SYMBOL_GPL(vhost_poll_init);
  84
  85 /* Start polling a file. We add ourselves to file's wait queue. The caller must
  86  * keep a reference to a file until after vhost_poll_stop is called. */
  87 int vhost_poll_start(struct vhost_poll *poll, struct file *file)
  88 {
  89         unsigned long mask;
  90         int ret = 0;
  91
  92         if (poll->wqh)
  93                 return 0;
  94
  95         mask = file->f_op->poll(file, &poll->table);
  96         if (mask)
  97                 vhost_poll_wakeup(&poll->wait, 0, 0, (void *)mask);
  98         if (mask & POLLERR) {
  99                 if (poll->wqh)
 100                         remove_wait_queue(poll->wqh, &poll->wait);
 101                 ret = -EINVAL;
 102         }
 103
 104         return ret;
 105 }
 106 EXPORT_SYMBOL_GPL(vhost_poll_start);
 107
 108 /* Stop polling a file. After this function returns, it becomes safe to drop the
 109  * file reference. You must also flush afterwards. */
 110 void vhost_poll_stop(struct vhost_poll *poll)
 111 {
 112         if (poll->wqh) {
 113                 remove_wait_queue(poll->wqh, &poll->wait);
 114                 poll->wqh = NULL;
 115         }
 116 }
 117 EXPORT_SYMBOL_GPL(vhost_poll_stop);
 118
 119 static bool vhost_work_seq_done(struct vhost_dev *dev, struct vhost_work *work,
 120                                 unsigned seq)
 121 {
 122         int left;
 123
 124         spin_lock_irq(&dev->work_lock);
 125         left = seq - work->done_seq;
 126         spin_unlock_irq(&dev->work_lock);
 127         return left <= 0;
 128 }
 129
 130 void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work)
 131 {
 132         unsigned seq;
 133         int flushing;
 134
 135         spin_lock_irq(&dev->work_lock);
 136         seq = work->queue_seq;
 137         work->flushing++;
 138         spin_unlock_irq(&dev->work_lock);
 139         wait_event(work->done, vhost_work_seq_done(dev, work, seq));
 140         spin_lock_irq(&dev->work_lock);
 141         flushing = --work->flushing;
 142         spin_unlock_irq(&dev->work_lock);
 143         BUG_ON(flushing < 0);
 144 }
 145 EXPORT_SYMBOL_GPL(vhost_work_flush);
 146
 147 /* Flush any work that has been scheduled. When calling this, don't hold any
 148  * locks that are also used by the callback. */
 149 void vhost_poll_flush(struct vhost_poll *poll)
 150 {
 151         vhost_work_flush(poll->dev, &poll->work);
 152 }
 153 EXPORT_SYMBOL_GPL(vhost_poll_flush);
 154
 155 void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work)
 156 {
 157         unsigned long flags;
 158
 159         spin_lock_irqsave(&dev->work_lock, flags);
 160         if (list_empty(&work->node)) {
 161                 list_add_tail(&work->node, &dev->work_list);
 162                 work->queue_seq++;
 163                 spin_unlock_irqrestore(&dev->work_lock, flags);
 164                 wake_up_process(dev->worker);
 165         } else {
 166                 spin_unlock_irqrestore(&dev->work_lock, flags);
 167         }
 168 }
 169 EXPORT_SYMBOL_GPL(vhost_work_queue);
 170
 171 void vhost_poll_queue(struct vhost_poll *poll)
 172 {
 173         vhost_work_queue(poll->dev, &poll->work);
 174 }
 175 EXPORT_SYMBOL_GPL(vhost_poll_queue);
 176
 177 static void vhost_vq_reset(struct vhost_dev *dev,
 178                            struct vhost_virtqueue *vq)
 179 {
 180         vq->num = 1;
 181         vq->desc = NULL;
 182         vq->avail = NULL;
 183         vq->used = NULL;
 184         vq->last_avail_idx = 0;
 185         vq->avail_idx = 0;
 186         vq->last_used_idx = 0;
 187         vq->signalled_used = 0;
 188         vq->signalled_used_valid = false;
 189         vq->used_flags = 0;
 190         vq->log_used = false;
 191         vq->log_addr = -1ull;
 192         vq->private_data = NULL;
 193         vq->acked_features = 0;
 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         vq->memory = NULL;
 202 }
 203
 204 static int vhost_worker(void *data)
 205 {
 206         struct vhost_dev *dev = data;
 207         struct vhost_work *work = NULL;
 208         unsigned uninitialized_var(seq);
 209         mm_segment_t oldfs = get_fs();
 210
 211         set_fs(USER_DS);
 212         use_mm(dev->mm);
 213
 214         for (;;) {
 215                 /* mb paired w/ kthread_stop */
 216                 set_current_state(TASK_INTERRUPTIBLE);
 217
 218                 spin_lock_irq(&dev->work_lock);
 219                 if (work) {
 220                         work->done_seq = seq;
 221                         if (work->flushing)
 222                                 wake_up_all(&work->done);
 223                 }
 224
 225                 if (kthread_should_stop()) {
 226                         spin_unlock_irq(&dev->work_lock);
 227                         __set_current_state(TASK_RUNNING);
 228                         break;
 229                 }
 230                 if (!list_empty(&dev->work_list)) {
 231                         work = list_first_entry(&dev->work_list,
 232                                                 struct vhost_work, node);
 233                         list_del_init(&work->node);
 234                         seq = work->queue_seq;
 235                 } else
 236                         work = NULL;
 237                 spin_unlock_irq(&dev->work_lock);
 238
 239                 if (work) {
 240                         __set_current_state(TASK_RUNNING);
 241                         work->fn(work);
 242                         if (need_resched())
 243                                 schedule();
 244                 } else
 245                         schedule();
 246
 247         }
 248         unuse_mm(dev->mm);
 249         set_fs(oldfs);
 250         return 0;
 251 }
 252
 253 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
 254 {
 255         kfree(vq->indirect);
 256         vq->indirect = NULL;
 257         kfree(vq->log);
 258         vq->log = NULL;
 259         kfree(vq->heads);
 260         vq->heads = NULL;
 261 }
 262
 263 /* Helper to allocate iovec buffers for all vqs. */
 264 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
 265 {
 266         struct vhost_virtqueue *vq;
 267         int i;
 268
 269         for (i = 0; i < dev->nvqs; ++i) {
 270                 vq = dev->vqs[i];
 271                 vq->indirect = kmalloc(sizeof *vq->indirect * UIO_MAXIOV,
 272                                        GFP_KERNEL);
 273                 vq->log = kmalloc(sizeof *vq->log * UIO_MAXIOV, GFP_KERNEL);
 274                 vq->heads = kmalloc(sizeof *vq->heads * UIO_MAXIOV, GFP_KERNEL);
 275                 if (!vq->indirect || !vq->log || !vq->heads)
 276                         goto err_nomem;
 277         }
 278         return 0;
 279
 280 err_nomem:
 281         for (; i >= 0; --i)
 282                 vhost_vq_free_iovecs(dev->vqs[i]);
 283         return -ENOMEM;
 284 }
 285
 286 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
 287 {
 288         int i;
 289
 290         for (i = 0; i < dev->nvqs; ++i)
 291                 vhost_vq_free_iovecs(dev->vqs[i]);
 292 }
 293
 294 void vhost_dev_init(struct vhost_dev *dev,
 295                     struct vhost_virtqueue **vqs, int nvqs)
 296 {
 297         struct vhost_virtqueue *vq;
 298         int i;
 299
 300         dev->vqs = vqs;
 301         dev->nvqs = nvqs;
 302         mutex_init(&dev->mutex);
 303         dev->log_ctx = NULL;
 304         dev->log_file = NULL;
 305         dev->memory = NULL;
 306         dev->mm = NULL;
 307         spin_lock_init(&dev->work_lock);
 308         INIT_LIST_HEAD(&dev->work_list);
 309         dev->worker = NULL;
 310
 311         for (i = 0; i < dev->nvqs; ++i) {
 312                 vq = dev->vqs[i];
 313                 vq->log = NULL;
 314                 vq->indirect = NULL;
 315                 vq->heads = NULL;
 316                 vq->dev = dev;
 317                 mutex_init(&vq->mutex);
 318                 vhost_vq_reset(dev, vq);
 319                 if (vq->handle_kick)
 320                         vhost_poll_init(&vq->poll, vq->handle_kick,
 321                                         POLLIN, dev);
 322         }
 323 }
 324 EXPORT_SYMBOL_GPL(vhost_dev_init);
 325
 326 /* Caller should have device mutex */
 327 long vhost_dev_check_owner(struct vhost_dev *dev)
 328 {
 329         /* Are you the owner? If not, I don't think you mean to do that */
 330         return dev->mm == current->mm ? 0 : -EPERM;
 331 }
 332 EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
 333
 334 struct vhost_attach_cgroups_struct {
 335         struct vhost_work work;
 336         struct task_struct *owner;
 337         int ret;
 338 };
 339
 340 static void vhost_attach_cgroups_work(struct vhost_work *work)
 341 {
 342         struct vhost_attach_cgroups_struct *s;
 343
 344         s = container_of(work, struct vhost_attach_cgroups_struct, work);
 345         s->ret = cgroup_attach_task_all(s->owner, current);
 346 }
 347
 348 static int vhost_attach_cgroups(struct vhost_dev *dev)
 349 {
 350         struct vhost_attach_cgroups_struct attach;
 351
 352         attach.owner = current;
 353         vhost_work_init(&attach.work, vhost_attach_cgroups_work);
 354         vhost_work_queue(dev, &attach.work);
 355         vhost_work_flush(dev, &attach.work);
 356         return attach.ret;
 357 }
 358
 359 /* Caller should have device mutex */
 360 bool vhost_dev_has_owner(struct vhost_dev *dev)
 361 {
 362         return dev->mm;
 363 }
 364 EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
 365
 366 /* Caller should have device mutex */
 367 long vhost_dev_set_owner(struct vhost_dev *dev)
 368 {
 369         struct task_struct *worker;
 370         int err;
 371
 372         /* Is there an owner already? */
 373         if (vhost_dev_has_owner(dev)) {
 374                 err = -EBUSY;
 375                 goto err_mm;
 376         }
 377
 378         /* No owner, become one */
 379         dev->mm = get_task_mm(current);
 380         worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
 381         if (IS_ERR(worker)) {
 382                 err = PTR_ERR(worker);
 383                 goto err_worker;
 384         }
 385
 386         dev->worker = worker;
 387         wake_up_process(worker);        /* avoid contributing to loadavg */
 388
 389         err = vhost_attach_cgroups(dev);
 390         if (err)
 391                 goto err_cgroup;
 392
 393         err = vhost_dev_alloc_iovecs(dev);
 394         if (err)
 395                 goto err_cgroup;
 396
 397         return 0;
 398 err_cgroup:
 399         kthread_stop(worker);
 400         dev->worker = NULL;
 401 err_worker:
 402         if (dev->mm)
 403                 mmput(dev->mm);
 404         dev->mm = NULL;
 405 err_mm:
 406         return err;
 407 }
 408 EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
 409
 410 struct vhost_memory *vhost_dev_reset_owner_prepare(void)
 411 {
 412         return kmalloc(offsetof(struct vhost_memory, regions), GFP_KERNEL);
 413 }
 414 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
 415
 416 /* Caller should have device mutex */
 417 void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_memory *memory)
 418 {
 419         int i;
 420
 421         vhost_dev_cleanup(dev, true);
 422
 423         /* Restore memory to default empty mapping. */
 424         memory->nregions = 0;
 425         dev->memory = memory;
 426         /* We don't need VQ locks below since vhost_dev_cleanup makes sure
 427          * VQs aren't running.
 428          */
 429         for (i = 0; i < dev->nvqs; ++i)
 430                 dev->vqs[i]->memory = memory;
 431 }
 432 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
 433
 434 void vhost_dev_stop(struct vhost_dev *dev)
 435 {
 436         int i;
 437
 438         for (i = 0; i < dev->nvqs; ++i) {
 439                 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick) {
 440                         vhost_poll_stop(&dev->vqs[i]->poll);
 441                         vhost_poll_flush(&dev->vqs[i]->poll);
 442                 }
 443         }
 444 }
 445 EXPORT_SYMBOL_GPL(vhost_dev_stop);
 446
 447 /* Caller should have device mutex if and only if locked is set */
 448 void vhost_dev_cleanup(struct vhost_dev *dev, bool locked)
 449 {
 450         int i;
 451
 452         for (i = 0; i < dev->nvqs; ++i) {
 453                 if (dev->vqs[i]->error_ctx)
 454                         eventfd_ctx_put(dev->vqs[i]->error_ctx);
 455                 if (dev->vqs[i]->error)
 456                         fput(dev->vqs[i]->error);
 457                 if (dev->vqs[i]->kick)
 458                         fput(dev->vqs[i]->kick);
 459                 if (dev->vqs[i]->call_ctx)
 460                         eventfd_ctx_put(dev->vqs[i]->call_ctx);
 461                 if (dev->vqs[i]->call)
 462                         fput(dev->vqs[i]->call);
 463                 vhost_vq_reset(dev, dev->vqs[i]);
 464         }
 465         vhost_dev_free_iovecs(dev);
 466         if (dev->log_ctx)
 467                 eventfd_ctx_put(dev->log_ctx);
 468         dev->log_ctx = NULL;
 469         if (dev->log_file)
 470                 fput(dev->log_file);
 471         dev->log_file = NULL;
 472         /* No one will access memory at this point */
 473         kfree(dev->memory);
 474         dev->memory = NULL;
 475         WARN_ON(!list_empty(&dev->work_list));
 476         if (dev->worker) {
 477                 kthread_stop(dev->worker);
 478                 dev->worker = NULL;
 479         }
 480         if (dev->mm)
 481                 mmput(dev->mm);
 482         dev->mm = NULL;
 483 }
 484 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
 485
 486 static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
 487 {
 488         u64 a = addr / VHOST_PAGE_SIZE / 8;
 489
 490         /* Make sure 64 bit math will not overflow. */
 491         if (a > ULONG_MAX - (unsigned long)log_base ||
 492             a + (unsigned long)log_base > ULONG_MAX)
 493                 return 0;
 494
 495         return access_ok(VERIFY_WRITE, log_base + a,
 496                          (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
 497 }
 498
 499 /* Caller should have vq mutex and device mutex. */
 500 static int vq_memory_access_ok(void __user *log_base, struct vhost_memory *mem,
 501                                int log_all)
 502 {
 503         int i;
 504
 505         if (!mem)
 506                 return 0;
 507
 508         for (i = 0; i < mem->nregions; ++i) {
 509                 struct vhost_memory_region *m = mem->regions + i;
 510                 unsigned long a = m->userspace_addr;
 511                 if (m->memory_size > ULONG_MAX)
 512                         return 0;
 513                 else if (!access_ok(VERIFY_WRITE, (void __user *)a,
 514                                     m->memory_size))
 515                         return 0;
 516                 else if (log_all && !log_access_ok(log_base,
 517                                                    m->guest_phys_addr,
 518                                                    m->memory_size))
 519                         return 0;
 520         }
 521         return 1;
 522 }
 523
 524 /* Can we switch to this memory table? */
 525 /* Caller should have device mutex but not vq mutex */
 526 static int memory_access_ok(struct vhost_dev *d, struct vhost_memory *mem,
 527                             int log_all)
 528 {
 529         int i;
 530
 531         for (i = 0; i < d->nvqs; ++i) {
 532                 int ok;
 533                 bool log;
 534
 535                 mutex_lock(&d->vqs[i]->mutex);
 536                 log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
 537                 /* If ring is inactive, will check when it's enabled. */
 538                 if (d->vqs[i]->private_data)
 539                         ok = vq_memory_access_ok(d->vqs[i]->log_base, mem, log);
 540                 else
 541                         ok = 1;
 542                 mutex_unlock(&d->vqs[i]->mutex);
 543                 if (!ok)
 544                         return 0;
 545         }
 546         return 1;
 547 }
 548
 549 static int vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
 550                         struct vring_desc __user *desc,
 551                         struct vring_avail __user *avail,
 552                         struct vring_used __user *used)
 553 {
 554         size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
 555         return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
 556                access_ok(VERIFY_READ, avail,
 557                          sizeof *avail + num * sizeof *avail->ring + s) &&
 558                access_ok(VERIFY_WRITE, used,
 559                         sizeof *used + num * sizeof *used->ring + s);
 560 }
 561
 562 /* Can we log writes? */
 563 /* Caller should have device mutex but not vq mutex */
 564 int vhost_log_access_ok(struct vhost_dev *dev)
 565 {
 566         return memory_access_ok(dev, dev->memory, 1);
 567 }
 568 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
 569
 570 /* Verify access for write logging. */
 571 /* Caller should have vq mutex and device mutex */
 572 static int vq_log_access_ok(struct vhost_virtqueue *vq,
 573                             void __user *log_base)
 574 {
 575         size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
 576
 577         return vq_memory_access_ok(log_base, vq->memory,
 578                                    vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
 579                 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
 580                                         sizeof *vq->used +
 581                                         vq->num * sizeof *vq->used->ring + s));
 582 }
 583
 584 /* Can we start vq? */
 585 /* Caller should have vq mutex and device mutex */
 586 int vhost_vq_access_ok(struct vhost_virtqueue *vq)
 587 {
 588         return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used) &&
 589                 vq_log_access_ok(vq, vq->log_base);
 590 }
 591 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
 592
 593 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
 594 {
 595         struct vhost_memory mem, *newmem, *oldmem;
 596         unsigned long size = offsetof(struct vhost_memory, regions);
 597         int i;
 598
 599         if (copy_from_user(&mem, m, size))
 600                 return -EFAULT;
 601         if (mem.padding)
 602                 return -EOPNOTSUPP;
 603         if (mem.nregions > VHOST_MEMORY_MAX_NREGIONS)
 604                 return -E2BIG;
 605         newmem = kmalloc(size + mem.nregions * sizeof *m->regions, GFP_KERNEL);
 606         if (!newmem)
 607                 return -ENOMEM;
 608
 609         memcpy(newmem, &mem, size);
 610         if (copy_from_user(newmem->regions, m->regions,
 611                            mem.nregions * sizeof *m->regions)) {
 612                 kfree(newmem);
 613                 return -EFAULT;
 614         }
 615
 616         if (!memory_access_ok(d, newmem, 0)) {
 617                 kfree(newmem);
 618                 return -EFAULT;
 619         }
 620         oldmem = d->memory;
 621         d->memory = newmem;
 622
 623         /* All memory accesses are done under some VQ mutex. */
 624         for (i = 0; i < d->nvqs; ++i) {
 625                 mutex_lock(&d->vqs[i]->mutex);
 626                 d->vqs[i]->memory = newmem;
 627                 mutex_unlock(&d->vqs[i]->mutex);
 628         }
 629         kfree(oldmem);
 630         return 0;
 631 }
 632
 633 long vhost_vring_ioctl(struct vhost_dev *d, int ioctl, void __user *argp)
 634 {
 635         struct file *eventfp, *filep = NULL;
 636         bool pollstart = false, pollstop = false;
 637         struct eventfd_ctx *ctx = NULL;
 638         u32 __user *idxp = argp;
 639         struct vhost_virtqueue *vq;
 640         struct vhost_vring_state s;
 641         struct vhost_vring_file f;
 642         struct vhost_vring_addr a;
 643         u32 idx;
 644         long r;
 645
 646         r = get_user(idx, idxp);
 647         if (r < 0)
 648                 return r;
 649         if (idx >= d->nvqs)
 650                 return -ENOBUFS;
 651
 652         vq = d->vqs[idx];
 653
 654         mutex_lock(&vq->mutex);
 655
 656         switch (ioctl) {
 657         case VHOST_SET_VRING_NUM:
 658                 /* Resizing ring with an active backend?
 659                  * You don't want to do that. */
 660                 if (vq->private_data) {
 661                         r = -EBUSY;
 662                         break;
 663                 }
 664                 if (copy_from_user(&s, argp, sizeof s)) {
 665                         r = -EFAULT;
 666                         break;
 667                 }
 668                 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
 669                         r = -EINVAL;
 670                         break;
 671                 }
 672                 vq->num = s.num;
 673                 break;
 674         case VHOST_SET_VRING_BASE:
 675                 /* Moving base with an active backend?
 676                  * You don't want to do that. */
 677                 if (vq->private_data) {
 678                         r = -EBUSY;
 679                         break;
 680                 }
 681                 if (copy_from_user(&s, argp, sizeof s)) {
 682                         r = -EFAULT;
 683                         break;
 684                 }
 685                 if (s.num > 0xffff) {
 686                         r = -EINVAL;
 687                         break;
 688                 }
 689                 vq->last_avail_idx = s.num;
 690                 /* Forget the cached index value. */
 691                 vq->avail_idx = vq->last_avail_idx;
 692                 break;
 693         case VHOST_GET_VRING_BASE:
 694                 s.index = idx;
 695                 s.num = vq->last_avail_idx;
 696                 if (copy_to_user(argp, &s, sizeof s))
 697                         r = -EFAULT;
 698                 break;
 699         case VHOST_SET_VRING_ADDR:
 700                 if (copy_from_user(&a, argp, sizeof a)) {
 701                         r = -EFAULT;
 702                         break;
 703                 }
 704                 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
 705                         r = -EOPNOTSUPP;
 706                         break;
 707                 }
 708                 /* For 32bit, verify that the top 32bits of the user
 709                    data are set to zero. */
 710                 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
 711                     (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
 712                     (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
 713                         r = -EFAULT;
 714                         break;
 715                 }
 716
 717                 /* Make sure it's safe to cast pointers to vring types. */
 718                 BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
 719                 BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
 720                 if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
 721                     (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
 722                     (a.log_guest_addr & (sizeof(u64) - 1))) {
 723                         r = -EINVAL;
 724                         break;
 725                 }
 726
 727                 /* We only verify access here if backend is configured.
 728                  * If it is not, we don't as size might not have been setup.
 729                  * We will verify when backend is configured. */
 730                 if (vq->private_data) {
 731                         if (!vq_access_ok(vq, vq->num,
 732                                 (void __user *)(unsigned long)a.desc_user_addr,
 733                                 (void __user *)(unsigned long)a.avail_user_addr,
 734                                 (void __user *)(unsigned long)a.used_user_addr)) {
 735                                 r = -EINVAL;
 736                                 break;
 737                         }
 738
 739                         /* Also validate log access for used ring if enabled. */
 740                         if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
 741                             !log_access_ok(vq->log_base, a.log_guest_addr,
 742                                            sizeof *vq->used +
 743                                            vq->num * sizeof *vq->used->ring)) {
 744                                 r = -EINVAL;
 745                                 break;
 746                         }
 747                 }
 748
 749                 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
 750                 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
 751                 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
 752                 vq->log_addr = a.log_guest_addr;
 753                 vq->used = (void __user *)(unsigned long)a.used_user_addr;
 754                 break;
 755         case VHOST_SET_VRING_KICK:
 756                 if (copy_from_user(&f, argp, sizeof f)) {
 757                         r = -EFAULT;
 758                         break;
 759                 }
 760                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
 761                 if (IS_ERR(eventfp)) {
 762                         r = PTR_ERR(eventfp);
 763                         break;
 764                 }
 765                 if (eventfp != vq->kick) {
 766                         pollstop = (filep = vq->kick) != NULL;
 767                         pollstart = (vq->kick = eventfp) != NULL;
 768                 } else
 769                         filep = eventfp;
 770                 break;
 771         case VHOST_SET_VRING_CALL:
 772                 if (copy_from_user(&f, argp, sizeof f)) {
 773                         r = -EFAULT;
 774                         break;
 775                 }
 776                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
 777                 if (IS_ERR(eventfp)) {
 778                         r = PTR_ERR(eventfp);
 779                         break;
 780                 }
 781                 if (eventfp != vq->call) {
 782                         filep = vq->call;
 783                         ctx = vq->call_ctx;
 784                         vq->call = eventfp;
 785                         vq->call_ctx = eventfp ?
 786                                 eventfd_ctx_fileget(eventfp) : NULL;
 787                 } else
 788                         filep = eventfp;
 789                 break;
 790         case VHOST_SET_VRING_ERR:
 791                 if (copy_from_user(&f, argp, sizeof f)) {
 792                         r = -EFAULT;
 793                         break;
 794                 }
 795                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
 796                 if (IS_ERR(eventfp)) {
 797                         r = PTR_ERR(eventfp);
 798                         break;
 799                 }
 800                 if (eventfp != vq->error) {
 801                         filep = vq->error;
 802                         vq->error = eventfp;
 803                         ctx = vq->error_ctx;
 804                         vq->error_ctx = eventfp ?
 805                                 eventfd_ctx_fileget(eventfp) : NULL;
 806                 } else
 807                         filep = eventfp;
 808                 break;
 809         default:
 810                 r = -ENOIOCTLCMD;
 811         }
 812
 813         if (pollstop && vq->handle_kick)
 814                 vhost_poll_stop(&vq->poll);
 815
 816         if (ctx)
 817                 eventfd_ctx_put(ctx);
 818         if (filep)
 819                 fput(filep);
 820
 821         if (pollstart && vq->handle_kick)
 822                 r = vhost_poll_start(&vq->poll, vq->kick);
 823
 824         mutex_unlock(&vq->mutex);
 825
 826         if (pollstop && vq->handle_kick)
 827                 vhost_poll_flush(&vq->poll);
 828         return r;
 829 }
 830 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
 831
 832 /* Caller must have device mutex */
 833 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
 834 {
 835         struct file *eventfp, *filep = NULL;
 836         struct eventfd_ctx *ctx = NULL;
 837         u64 p;
 838         long r;
 839         int i, fd;
 840
 841         /* If you are not the owner, you can become one */
 842         if (ioctl == VHOST_SET_OWNER) {
 843                 r = vhost_dev_set_owner(d);
 844                 goto done;
 845         }
 846
 847         /* You must be the owner to do anything else */
 848         r = vhost_dev_check_owner(d);
 849         if (r)
 850                 goto done;
 851
 852         switch (ioctl) {
 853         case VHOST_SET_MEM_TABLE:
 854                 r = vhost_set_memory(d, argp);
 855                 break;
 856         case VHOST_SET_LOG_BASE:
 857                 if (copy_from_user(&p, argp, sizeof p)) {
 858                         r = -EFAULT;
 859                         break;
 860                 }
 861                 if ((u64)(unsigned long)p != p) {
 862                         r = -EFAULT;
 863                         break;
 864                 }
 865                 for (i = 0; i < d->nvqs; ++i) {
 866                         struct vhost_virtqueue *vq;
 867                         void __user *base = (void __user *)(unsigned long)p;
 868                         vq = d->vqs[i];
 869                         mutex_lock(&vq->mutex);
 870                         /* If ring is inactive, will check when it's enabled. */
 871                         if (vq->private_data && !vq_log_access_ok(vq, base))
 872                                 r = -EFAULT;
 873                         else
 874                                 vq->log_base = base;
 875                         mutex_unlock(&vq->mutex);
 876                 }
 877                 break;
 878         case VHOST_SET_LOG_FD:
 879                 r = get_user(fd, (int __user *)argp);
 880                 if (r < 0)
 881                         break;
 882                 eventfp = fd == -1 ? NULL : eventfd_fget(fd);
 883                 if (IS_ERR(eventfp)) {
 884                         r = PTR_ERR(eventfp);
 885                         break;
 886                 }
 887                 if (eventfp != d->log_file) {
 888                         filep = d->log_file;
 889                         ctx = d->log_ctx;
 890                         d->log_ctx = eventfp ?
 891                                 eventfd_ctx_fileget(eventfp) : NULL;
 892                 } else
 893                         filep = eventfp;
 894                 for (i = 0; i < d->nvqs; ++i) {
 895                         mutex_lock(&d->vqs[i]->mutex);
 896                         d->vqs[i]->log_ctx = d->log_ctx;
 897                         mutex_unlock(&d->vqs[i]->mutex);
 898                 }
 899                 if (ctx)
 900                         eventfd_ctx_put(ctx);
 901                 if (filep)
 902                         fput(filep);
 903                 break;
 904         default:
 905                 r = -ENOIOCTLCMD;
 906                 break;
 907         }
 908 done:
 909         return r;
 910 }
 911 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
 912
 913 static const struct vhost_memory_region *find_region(struct vhost_memory *mem,
 914                                                      __u64 addr, __u32 len)
 915 {
 916         struct vhost_memory_region *reg;
 917         int i;
 918
 919         /* linear search is not brilliant, but we really have on the order of 6
 920          * regions in practice */
 921         for (i = 0; i < mem->nregions; ++i) {
 922                 reg = mem->regions + i;
 923                 if (reg->guest_phys_addr <= addr &&
 924                     reg->guest_phys_addr + reg->memory_size - 1 >= addr)
 925                         return reg;
 926         }
 927         return NULL;
 928 }
 929
 930 /* TODO: This is really inefficient.  We need something like get_user()
 931  * (instruction directly accesses the data, with an exception table entry
 932  * returning -EFAULT). See Documentation/x86/exception-tables.txt.
 933  */
 934 static int set_bit_to_user(int nr, void __user *addr)
 935 {
 936         unsigned long log = (unsigned long)addr;
 937         struct page *page;
 938         void *base;
 939         int bit = nr + (log % PAGE_SIZE) * 8;
 940         int r;
 941
 942         r = get_user_pages_fast(log, 1, 1, &page);
 943         if (r < 0)
 944                 return r;
 945         BUG_ON(r != 1);
 946         base = kmap_atomic(page);
 947         set_bit(bit, base);
 948         kunmap_atomic(base);
 949         set_page_dirty_lock(page);
 950         put_page(page);
 951         return 0;
 952 }
 953
 954 static int log_write(void __user *log_base,
 955                      u64 write_address, u64 write_length)
 956 {
 957         u64 write_page = write_address / VHOST_PAGE_SIZE;
 958         int r;
 959
 960         if (!write_length)
 961                 return 0;
 962         write_length += write_address % VHOST_PAGE_SIZE;
 963         for (;;) {
 964                 u64 base = (u64)(unsigned long)log_base;
 965                 u64 log = base + write_page / 8;
 966                 int bit = write_page % 8;
 967                 if ((u64)(unsigned long)log != log)
 968                         return -EFAULT;
 969                 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
 970                 if (r < 0)
 971                         return r;
 972                 if (write_length <= VHOST_PAGE_SIZE)
 973                         break;
 974                 write_length -= VHOST_PAGE_SIZE;
 975                 write_page += 1;
 976         }
 977         return r;
 978 }
 979
 980 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
 981                     unsigned int log_num, u64 len)
 982 {
 983         int i, r;
 984
 985         /* Make sure data written is seen before log. */
 986         smp_wmb();
 987         for (i = 0; i < log_num; ++i) {
 988                 u64 l = min(log[i].len, len);
 989                 r = log_write(vq->log_base, log[i].addr, l);
 990                 if (r < 0)
 991                         return r;
 992                 len -= l;
 993                 if (!len) {
 994                         if (vq->log_ctx)
 995                                 eventfd_signal(vq->log_ctx, 1);
 996                         return 0;
 997                 }
 998         }
 999         /* Length written exceeds what we have stored. This is a bug. */
1000         BUG();
1001         return 0;
1002 }
1003 EXPORT_SYMBOL_GPL(vhost_log_write);
1004
1005 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1006 {
1007         void __user *used;
1008         if (__put_user(cpu_to_vhost16(vq, vq->used_flags), &vq->used->flags) < 0)
1009                 return -EFAULT;
1010         if (unlikely(vq->log_used)) {
1011                 /* Make sure the flag is seen before log. */
1012                 smp_wmb();
1013                 /* Log used flag write. */
1014                 used = &vq->used->flags;
1015                 log_write(vq->log_base, vq->log_addr +
1016                           (used - (void __user *)vq->used),
1017                           sizeof vq->used->flags);
1018                 if (vq->log_ctx)
1019                         eventfd_signal(vq->log_ctx, 1);
1020         }
1021         return 0;
1022 }
1023
1024 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1025 {
1026         if (__put_user(cpu_to_vhost16(vq, vq->avail_idx), vhost_avail_event(vq)))
1027                 return -EFAULT;
1028         if (unlikely(vq->log_used)) {
1029                 void __user *used;
1030                 /* Make sure the event is seen before log. */
1031                 smp_wmb();
1032                 /* Log avail event write */
1033                 used = vhost_avail_event(vq);
1034                 log_write(vq->log_base, vq->log_addr +
1035                           (used - (void __user *)vq->used),
1036                           sizeof *vhost_avail_event(vq));
1037                 if (vq->log_ctx)
1038                         eventfd_signal(vq->log_ctx, 1);
1039         }
1040         return 0;
1041 }
1042
1043 int vhost_init_used(struct vhost_virtqueue *vq)
1044 {
1045         __virtio16 last_used_idx;
1046         int r;
1047         if (!vq->private_data)
1048                 return 0;
1049
1050         r = vhost_update_used_flags(vq);
1051         if (r)
1052                 return r;
1053         vq->signalled_used_valid = false;
1054         if (!access_ok(VERIFY_READ, &vq->used->idx, sizeof vq->used->idx))
1055                 return -EFAULT;
1056         r = __get_user(last_used_idx, &vq->used->idx);
1057         if (r)
1058                 return r;
1059         vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
1060         return 0;
1061 }
1062 EXPORT_SYMBOL_GPL(vhost_init_used);
1063
1064 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
1065                           struct iovec iov[], int iov_size)
1066 {
1067         const struct vhost_memory_region *reg;
1068         struct vhost_memory *mem;
1069         struct iovec *_iov;
1070         u64 s = 0;
1071         int ret = 0;
1072
1073         mem = vq->memory;
1074         while ((u64)len > s) {
1075                 u64 size;
1076                 if (unlikely(ret >= iov_size)) {
1077                         ret = -ENOBUFS;
1078                         break;
1079                 }
1080                 reg = find_region(mem, addr, len);
1081                 if (unlikely(!reg)) {
1082                         ret = -EFAULT;
1083                         break;
1084                 }
1085                 _iov = iov + ret;
1086                 size = reg->memory_size - addr + reg->guest_phys_addr;
1087                 _iov->iov_len = min((u64)len - s, size);
1088                 _iov->iov_base = (void __user *)(unsigned long)
1089                         (reg->userspace_addr + addr - reg->guest_phys_addr);
1090                 s += size;
1091                 addr += size;
1092                 ++ret;
1093         }
1094
1095         return ret;
1096 }
1097
1098 /* Each buffer in the virtqueues is actually a chain of descriptors.  This
1099  * function returns the next descriptor in the chain,
1100  * or -1U if we're at the end. */
1101 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
1102 {
1103         unsigned int next;
1104
1105         /* If this descriptor says it doesn't chain, we're done. */
1106         if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
1107                 return -1U;
1108
1109         /* Check they're not leading us off end of descriptors. */
1110         next = vhost16_to_cpu(vq, desc->next);
1111         /* Make sure compiler knows to grab that: we don't want it changing! */
1112         /* We will use the result as an index in an array, so most
1113          * architectures only need a compiler barrier here. */
1114         read_barrier_depends();
1115
1116         return next;
1117 }
1118
1119 static int get_indirect(struct vhost_virtqueue *vq,
1120                         struct iovec iov[], unsigned int iov_size,
1121                         unsigned int *out_num, unsigned int *in_num,
1122                         struct vhost_log *log, unsigned int *log_num,
1123                         struct vring_desc *indirect)
1124 {
1125         struct vring_desc desc;
1126         unsigned int i = 0, count, found = 0;
1127         u32 len = vhost32_to_cpu(vq, indirect->len);
1128         struct iov_iter from;
1129         int ret;
1130
1131         /* Sanity check */
1132         if (unlikely(len % sizeof desc)) {
1133                 vq_err(vq, "Invalid length in indirect descriptor: "
1134                        "len 0x%llx not multiple of 0x%zx\n",
1135                        (unsigned long long)len,
1136                        sizeof desc);
1137                 return -EINVAL;
1138         }
1139
1140         ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
1141                              UIO_MAXIOV);
1142         if (unlikely(ret < 0)) {
1143                 vq_err(vq, "Translation failure %d in indirect.\n", ret);
1144                 return ret;
1145         }
1146         iov_iter_init(&from, READ, vq->indirect, ret, len);
1147
1148         /* We will use the result as an address to read from, so most
1149          * architectures only need a compiler barrier here. */
1150         read_barrier_depends();
1151
1152         count = len / sizeof desc;
1153         /* Buffers are chained via a 16 bit next field, so
1154          * we can have at most 2^16 of these. */
1155         if (unlikely(count > USHRT_MAX + 1)) {
1156                 vq_err(vq, "Indirect buffer length too big: %d\n",
1157                        indirect->len);
1158                 return -E2BIG;
1159         }
1160
1161         do {
1162                 unsigned iov_count = *in_num + *out_num;
1163                 if (unlikely(++found > count)) {
1164                         vq_err(vq, "Loop detected: last one at %u "
1165                                "indirect size %u\n",
1166                                i, count);
1167                         return -EINVAL;
1168                 }
1169                 if (unlikely(copy_from_iter(&desc, sizeof(desc), &from) !=
1170                              sizeof(desc))) {
1171                         vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1172                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1173                         return -EINVAL;
1174                 }
1175                 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
1176                         vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1177                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1178                         return -EINVAL;
1179                 }
1180
1181                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
1182                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
1183                                      iov_size - iov_count);
1184                 if (unlikely(ret < 0)) {
1185                         vq_err(vq, "Translation failure %d indirect idx %d\n",
1186                                ret, i);
1187                         return ret;
1188                 }
1189                 /* If this is an input descriptor, increment that count. */
1190                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE)) {
1191                         *in_num += ret;
1192                         if (unlikely(log)) {
1193                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
1194                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
1195                                 ++*log_num;
1196                         }
1197                 } else {
1198                         /* If it's an output descriptor, they're all supposed
1199                          * to come before any input descriptors. */
1200                         if (unlikely(*in_num)) {
1201                                 vq_err(vq, "Indirect descriptor "
1202                                        "has out after in: idx %d\n", i);
1203                                 return -EINVAL;
1204                         }
1205                         *out_num += ret;
1206                 }
1207         } while ((i = next_desc(vq, &desc)) != -1);
1208         return 0;
1209 }
1210
1211 /* This looks in the virtqueue and for the first available buffer, and converts
1212  * it to an iovec for convenient access.  Since descriptors consist of some
1213  * number of output then some number of input descriptors, it's actually two
1214  * iovecs, but we pack them into one and note how many of each there were.
1215  *
1216  * This function returns the descriptor number found, or vq->num (which is
1217  * never a valid descriptor number) if none was found.  A negative code is
1218  * returned on error. */
1219 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
1220                       struct iovec iov[], unsigned int iov_size,
1221                       unsigned int *out_num, unsigned int *in_num,
1222                       struct vhost_log *log, unsigned int *log_num)
1223 {
1224         struct vring_desc desc;
1225         unsigned int i, head, found = 0;
1226         u16 last_avail_idx;
1227         __virtio16 avail_idx;
1228         __virtio16 ring_head;
1229         int ret;
1230
1231         /* Check it isn't doing very strange things with descriptor numbers. */
1232         last_avail_idx = vq->last_avail_idx;
1233         if (unlikely(__get_user(avail_idx, &vq->avail->idx))) {
1234                 vq_err(vq, "Failed to access avail idx at %p\n",
1235                        &vq->avail->idx);
1236                 return -EFAULT;
1237         }
1238         vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
1239
1240         if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
1241                 vq_err(vq, "Guest moved used index from %u to %u",
1242                        last_avail_idx, vq->avail_idx);
1243                 return -EFAULT;
1244         }
1245
1246         /* If there's nothing new since last we looked, return invalid. */
1247         if (vq->avail_idx == last_avail_idx)
1248                 return vq->num;
1249
1250         /* Only get avail ring entries after they have been exposed by guest. */
1251         smp_rmb();
1252
1253         /* Grab the next descriptor number they're advertising, and increment
1254          * the index we've seen. */
1255         if (unlikely(__get_user(ring_head,
1256                                 &vq->avail->ring[last_avail_idx % vq->num]))) {
1257                 vq_err(vq, "Failed to read head: idx %d address %p\n",
1258                        last_avail_idx,
1259                        &vq->avail->ring[last_avail_idx % vq->num]);
1260                 return -EFAULT;
1261         }
1262
1263         head = vhost16_to_cpu(vq, ring_head);
1264
1265         /* If their number is silly, that's an error. */
1266         if (unlikely(head >= vq->num)) {
1267                 vq_err(vq, "Guest says index %u > %u is available",
1268                        head, vq->num);
1269                 return -EINVAL;
1270         }
1271
1272         /* When we start there are none of either input nor output. */
1273         *out_num = *in_num = 0;
1274         if (unlikely(log))
1275                 *log_num = 0;
1276
1277         i = head;
1278         do {
1279                 unsigned iov_count = *in_num + *out_num;
1280                 if (unlikely(i >= vq->num)) {
1281                         vq_err(vq, "Desc index is %u > %u, head = %u",
1282                                i, vq->num, head);
1283                         return -EINVAL;
1284                 }
1285                 if (unlikely(++found > vq->num)) {
1286                         vq_err(vq, "Loop detected: last one at %u "
1287                                "vq size %u head %u\n",
1288                                i, vq->num, head);
1289                         return -EINVAL;
1290                 }
1291                 ret = __copy_from_user(&desc, vq->desc + i, sizeof desc);
1292                 if (unlikely(ret)) {
1293                         vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
1294                                i, vq->desc + i);
1295                         return -EFAULT;
1296                 }
1297                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
1298                         ret = get_indirect(vq, iov, iov_size,
1299                                            out_num, in_num,
1300                                            log, log_num, &desc);
1301                         if (unlikely(ret < 0)) {
1302                                 vq_err(vq, "Failure detected "
1303                                        "in indirect descriptor at idx %d\n", i);
1304                                 return ret;
1305                         }
1306                         continue;
1307                 }
1308
1309                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
1310                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
1311                                      iov_size - iov_count);
1312                 if (unlikely(ret < 0)) {
1313                         vq_err(vq, "Translation failure %d descriptor idx %d\n",
1314                                ret, i);
1315                         return ret;
1316                 }
1317                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE)) {
1318                         /* If this is an input descriptor,
1319                          * increment that count. */
1320                         *in_num += ret;
1321                         if (unlikely(log)) {
1322                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
1323                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
1324                                 ++*log_num;
1325                         }
1326                 } else {
1327                         /* If it's an output descriptor, they're all supposed
1328                          * to come before any input descriptors. */
1329                         if (unlikely(*in_num)) {
1330                                 vq_err(vq, "Descriptor has out after in: "
1331                                        "idx %d\n", i);
1332                                 return -EINVAL;
1333                         }
1334                         *out_num += ret;
1335                 }
1336         } while ((i = next_desc(vq, &desc)) != -1);
1337
1338         /* On success, increment avail index. */
1339         vq->last_avail_idx++;
1340
1341         /* Assume notifications from guest are disabled at this point,
1342          * if they aren't we would need to update avail_event index. */
1343         BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
1344         return head;
1345 }
1346 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
1347
1348 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
1349 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
1350 {
1351         vq->last_avail_idx -= n;
1352 }
1353 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
1354
1355 /* After we've used one of their buffers, we tell them about it.  We'll then
1356  * want to notify the guest, using eventfd. */
1357 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
1358 {
1359         struct vring_used_elem heads = {
1360                 cpu_to_vhost32(vq, head),
1361                 cpu_to_vhost32(vq, len)
1362         };
1363
1364         return vhost_add_used_n(vq, &heads, 1);
1365 }
1366 EXPORT_SYMBOL_GPL(vhost_add_used);
1367
1368 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
1369                             struct vring_used_elem *heads,
1370                             unsigned count)
1371 {
1372         struct vring_used_elem __user *used;
1373         u16 old, new;
1374         int start;
1375
1376         start = vq->last_used_idx % vq->num;
1377         used = vq->used->ring + start;
1378         if (count == 1) {
1379                 if (__put_user(heads[0].id, &used->id)) {
1380                         vq_err(vq, "Failed to write used id");
1381                         return -EFAULT;
1382                 }
1383                 if (__put_user(heads[0].len, &used->len)) {
1384                         vq_err(vq, "Failed to write used len");
1385                         return -EFAULT;
1386                 }
1387         } else if (__copy_to_user(used, heads, count * sizeof *used)) {
1388                 vq_err(vq, "Failed to write used");
1389                 return -EFAULT;
1390         }
1391         if (unlikely(vq->log_used)) {
1392                 /* Make sure data is seen before log. */
1393                 smp_wmb();
1394                 /* Log used ring entry write. */
1395                 log_write(vq->log_base,
1396                           vq->log_addr +
1397                            ((void __user *)used - (void __user *)vq->used),
1398                           count * sizeof *used);
1399         }
1400         old = vq->last_used_idx;
1401         new = (vq->last_used_idx += count);
1402         /* If the driver never bothers to signal in a very long while,
1403          * used index might wrap around. If that happens, invalidate
1404          * signalled_used index we stored. TODO: make sure driver
1405          * signals at least once in 2^16 and remove this. */
1406         if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
1407                 vq->signalled_used_valid = false;
1408         return 0;
1409 }
1410
1411 /* After we've used one of their buffers, we tell them about it.  We'll then
1412  * want to notify the guest, using eventfd. */
1413 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
1414                      unsigned count)
1415 {
1416         int start, n, r;
1417
1418         start = vq->last_used_idx % vq->num;
1419         n = vq->num - start;
1420         if (n < count) {
1421                 r = __vhost_add_used_n(vq, heads, n);
1422                 if (r < 0)
1423                         return r;
1424                 heads += n;
1425                 count -= n;
1426         }
1427         r = __vhost_add_used_n(vq, heads, count);
1428
1429         /* Make sure buffer is written before we update index. */
1430         smp_wmb();
1431         if (__put_user(cpu_to_vhost16(vq, vq->last_used_idx), &vq->used->idx)) {
1432                 vq_err(vq, "Failed to increment used idx");
1433                 return -EFAULT;
1434         }
1435         if (unlikely(vq->log_used)) {
1436                 /* Log used index update. */
1437                 log_write(vq->log_base,
1438                           vq->log_addr + offsetof(struct vring_used, idx),
1439                           sizeof vq->used->idx);
1440                 if (vq->log_ctx)
1441                         eventfd_signal(vq->log_ctx, 1);
1442         }
1443         return r;
1444 }
1445 EXPORT_SYMBOL_GPL(vhost_add_used_n);
1446
1447 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1448 {
1449         __u16 old, new;
1450         __virtio16 event;
1451         bool v;
1452         /* Flush out used index updates. This is paired
1453          * with the barrier that the Guest executes when enabling
1454          * interrupts. */
1455         smp_mb();
1456
1457         if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
1458             unlikely(vq->avail_idx == vq->last_avail_idx))
1459                 return true;
1460
1461         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
1462                 __virtio16 flags;
1463                 if (__get_user(flags, &vq->avail->flags)) {
1464                         vq_err(vq, "Failed to get flags");
1465                         return true;
1466                 }
1467                 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
1468         }
1469         old = vq->signalled_used;
1470         v = vq->signalled_used_valid;
1471         new = vq->signalled_used = vq->last_used_idx;
1472         vq->signalled_used_valid = true;
1473
1474         if (unlikely(!v))
1475                 return true;
1476
1477         if (__get_user(event, vhost_used_event(vq))) {
1478                 vq_err(vq, "Failed to get used event idx");
1479                 return true;
1480         }
1481         return vring_need_event(vhost16_to_cpu(vq, event), new, old);
1482 }
1483
1484 /* This actually signals the guest, using eventfd. */
1485 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1486 {
1487         /* Signal the Guest tell them we used something up. */
1488         if (vq->call_ctx && vhost_notify(dev, vq))
1489                 eventfd_signal(vq->call_ctx, 1);
1490 }
1491 EXPORT_SYMBOL_GPL(vhost_signal);
1492
1493 /* And here's the combo meal deal.  Supersize me! */
1494 void vhost_add_used_and_signal(struct vhost_dev *dev,
1495                                struct vhost_virtqueue *vq,
1496                                unsigned int head, int len)
1497 {
1498         vhost_add_used(vq, head, len);
1499         vhost_signal(dev, vq);
1500 }
1501 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
1502
1503 /* multi-buffer version of vhost_add_used_and_signal */
1504 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
1505                                  struct vhost_virtqueue *vq,
1506                                  struct vring_used_elem *heads, unsigned count)
1507 {
1508         vhost_add_used_n(vq, heads, count);
1509         vhost_signal(dev, vq);
1510 }
1511 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
1512
1513 /* OK, now we need to know about added descriptors. */
1514 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1515 {
1516         __virtio16 avail_idx;
1517         int r;
1518
1519         if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
1520                 return false;
1521         vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
1522         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
1523                 r = vhost_update_used_flags(vq);
1524                 if (r) {
1525                         vq_err(vq, "Failed to enable notification at %p: %d\n",
1526                                &vq->used->flags, r);
1527                         return false;
1528                 }
1529         } else {
1530                 r = vhost_update_avail_event(vq, vq->avail_idx);
1531                 if (r) {
1532                         vq_err(vq, "Failed to update avail event index at %p: %d\n",
1533                                vhost_avail_event(vq), r);
1534                         return false;
1535                 }
1536         }
1537         /* They could have slipped one in as we were doing that: make
1538          * sure it's written, then check again. */
1539         smp_mb();
1540         r = __get_user(avail_idx, &vq->avail->idx);
1541         if (r) {
1542                 vq_err(vq, "Failed to check avail idx at %p: %d\n",
1543                        &vq->avail->idx, r);
1544                 return false;
1545         }
1546
1547         return vhost16_to_cpu(vq, avail_idx) != vq->avail_idx;
1548 }
1549 EXPORT_SYMBOL_GPL(vhost_enable_notify);
1550
1551 /* We don't need to be notified again. */
1552 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1553 {
1554         int r;
1555
1556         if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
1557                 return;
1558         vq->used_flags |= VRING_USED_F_NO_NOTIFY;
1559         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
1560                 r = vhost_update_used_flags(vq);
1561                 if (r)
1562                         vq_err(vq, "Failed to enable notification at %p: %d\n",
1563                                &vq->used->flags, r);
1564         }
1565 }
1566 EXPORT_SYMBOL_GPL(vhost_disable_notify);
1567
1568 static int __init vhost_init(void)
1569 {
1570         return 0;
1571 }
1572
1573 static void __exit vhost_exit(void)
1574 {
1575 }
1576
1577 module_init(vhost_init);
1578 module_exit(vhost_exit);
1579
1580 MODULE_VERSION("0.0.1");
1581 MODULE_LICENSE("GPL v2");
1582 MODULE_AUTHOR("Michael S. Tsirkin");
1583 MODULE_DESCRIPTION("Host kernel accelerator for virtio");