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         int i;
 597
 598         if (copy_from_user(&mem, m, size))
 599                 return -EFAULT;
 600         if (mem.padding)
 601                 return -EOPNOTSUPP;
 602         if (mem.nregions > VHOST_MEMORY_MAX_NREGIONS)
 603                 return -E2BIG;
 604         newmem = kmalloc(size + mem.nregions * sizeof *m->regions, GFP_KERNEL);
 605         if (!newmem)
 606                 return -ENOMEM;
 607
 608         memcpy(newmem, &mem, size);
 609         if (copy_from_user(newmem->regions, m->regions,
 610                            mem.nregions * sizeof *m->regions)) {
 611                 kfree(newmem);
 612                 return -EFAULT;
 613         }
 614
 615         if (!memory_access_ok(d, newmem,
 616                               vhost_has_feature(d, VHOST_F_LOG_ALL))) {
 617                 kfree(newmem);
 618                 return -EFAULT;
 619         }
 620         oldmem = rcu_dereference_protected(d->memory,
 621                                            lockdep_is_held(&d->mutex));
 622         rcu_assign_pointer(d->memory, newmem);
 623
 624         /* All memory accesses are done under some VQ mutex.
 625          * So below is a faster equivalent of synchronize_rcu()
 626          */
 627         for (i = 0; i < d->nvqs; ++i) {
 628                 mutex_lock(&d->vqs[i]->mutex);
 629                 mutex_unlock(&d->vqs[i]->mutex);
 630         }
 631         kfree(oldmem);
 632         return 0;
 633 }
 634
 635 long vhost_vring_ioctl(struct vhost_dev *d, int ioctl, void __user *argp)
 636 {
 637         struct file *eventfp, *filep = NULL;
 638         bool pollstart = false, pollstop = false;
 639         struct eventfd_ctx *ctx = NULL;
 640         u32 __user *idxp = argp;
 641         struct vhost_virtqueue *vq;
 642         struct vhost_vring_state s;
 643         struct vhost_vring_file f;
 644         struct vhost_vring_addr a;
 645         u32 idx;
 646         long r;
 647
 648         r = get_user(idx, idxp);
 649         if (r < 0)
 650                 return r;
 651         if (idx >= d->nvqs)
 652                 return -ENOBUFS;
 653
 654         vq = d->vqs[idx];
 655
 656         mutex_lock(&vq->mutex);
 657
 658         switch (ioctl) {
 659         case VHOST_SET_VRING_NUM:
 660                 /* Resizing ring with an active backend?
 661                  * You don't want to do that. */
 662                 if (vq->private_data) {
 663                         r = -EBUSY;
 664                         break;
 665                 }
 666                 if (copy_from_user(&s, argp, sizeof s)) {
 667                         r = -EFAULT;
 668                         break;
 669                 }
 670                 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
 671                         r = -EINVAL;
 672                         break;
 673                 }
 674                 vq->num = s.num;
 675                 break;
 676         case VHOST_SET_VRING_BASE:
 677                 /* Moving base with an active backend?
 678                  * You don't want to do that. */
 679                 if (vq->private_data) {
 680                         r = -EBUSY;
 681                         break;
 682                 }
 683                 if (copy_from_user(&s, argp, sizeof s)) {
 684                         r = -EFAULT;
 685                         break;
 686                 }
 687                 if (s.num > 0xffff) {
 688                         r = -EINVAL;
 689                         break;
 690                 }
 691                 vq->last_avail_idx = s.num;
 692                 /* Forget the cached index value. */
 693                 vq->avail_idx = vq->last_avail_idx;
 694                 break;
 695         case VHOST_GET_VRING_BASE:
 696                 s.index = idx;
 697                 s.num = vq->last_avail_idx;
 698                 if (copy_to_user(argp, &s, sizeof s))
 699                         r = -EFAULT;
 700                 break;
 701         case VHOST_SET_VRING_ADDR:
 702                 if (copy_from_user(&a, argp, sizeof a)) {
 703                         r = -EFAULT;
 704                         break;
 705                 }
 706                 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
 707                         r = -EOPNOTSUPP;
 708                         break;
 709                 }
 710                 /* For 32bit, verify that the top 32bits of the user
 711                    data are set to zero. */
 712                 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
 713                     (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
 714                     (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
 715                         r = -EFAULT;
 716                         break;
 717                 }
 718                 if ((a.avail_user_addr & (sizeof *vq->avail->ring - 1)) ||
 719                     (a.used_user_addr & (sizeof *vq->used->ring - 1)) ||
 720                     (a.log_guest_addr & (sizeof *vq->used->ring - 1))) {
 721                         r = -EINVAL;
 722                         break;
 723                 }
 724
 725                 /* We only verify access here if backend is configured.
 726                  * If it is not, we don't as size might not have been setup.
 727                  * We will verify when backend is configured. */
 728                 if (vq->private_data) {
 729                         if (!vq_access_ok(d, vq->num,
 730                                 (void __user *)(unsigned long)a.desc_user_addr,
 731                                 (void __user *)(unsigned long)a.avail_user_addr,
 732                                 (void __user *)(unsigned long)a.used_user_addr)) {
 733                                 r = -EINVAL;
 734                                 break;
 735                         }
 736
 737                         /* Also validate log access for used ring if enabled. */
 738                         if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
 739                             !log_access_ok(vq->log_base, a.log_guest_addr,
 740                                            sizeof *vq->used +
 741                                            vq->num * sizeof *vq->used->ring)) {
 742                                 r = -EINVAL;
 743                                 break;
 744                         }
 745                 }
 746
 747                 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
 748                 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
 749                 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
 750                 vq->log_addr = a.log_guest_addr;
 751                 vq->used = (void __user *)(unsigned long)a.used_user_addr;
 752                 break;
 753         case VHOST_SET_VRING_KICK:
 754                 if (copy_from_user(&f, argp, sizeof f)) {
 755                         r = -EFAULT;
 756                         break;
 757                 }
 758                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
 759                 if (IS_ERR(eventfp)) {
 760                         r = PTR_ERR(eventfp);
 761                         break;
 762                 }
 763                 if (eventfp != vq->kick) {
 764                         pollstop = (filep = vq->kick) != NULL;
 765                         pollstart = (vq->kick = eventfp) != NULL;
 766                 } else
 767                         filep = eventfp;
 768                 break;
 769         case VHOST_SET_VRING_CALL:
 770                 if (copy_from_user(&f, argp, sizeof f)) {
 771                         r = -EFAULT;
 772                         break;
 773                 }
 774                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
 775                 if (IS_ERR(eventfp)) {
 776                         r = PTR_ERR(eventfp);
 777                         break;
 778                 }
 779                 if (eventfp != vq->call) {
 780                         filep = vq->call;
 781                         ctx = vq->call_ctx;
 782                         vq->call = eventfp;
 783                         vq->call_ctx = eventfp ?
 784                                 eventfd_ctx_fileget(eventfp) : NULL;
 785                 } else
 786                         filep = eventfp;
 787                 break;
 788         case VHOST_SET_VRING_ERR:
 789                 if (copy_from_user(&f, argp, sizeof f)) {
 790                         r = -EFAULT;
 791                         break;
 792                 }
 793                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
 794                 if (IS_ERR(eventfp)) {
 795                         r = PTR_ERR(eventfp);
 796                         break;
 797                 }
 798                 if (eventfp != vq->error) {
 799                         filep = vq->error;
 800                         vq->error = eventfp;
 801                         ctx = vq->error_ctx;
 802                         vq->error_ctx = eventfp ?
 803                                 eventfd_ctx_fileget(eventfp) : NULL;
 804                 } else
 805                         filep = eventfp;
 806                 break;
 807         default:
 808                 r = -ENOIOCTLCMD;
 809         }
 810
 811         if (pollstop && vq->handle_kick)
 812                 vhost_poll_stop(&vq->poll);
 813
 814         if (ctx)
 815                 eventfd_ctx_put(ctx);
 816         if (filep)
 817                 fput(filep);
 818
 819         if (pollstart && vq->handle_kick)
 820                 r = vhost_poll_start(&vq->poll, vq->kick);
 821
 822         mutex_unlock(&vq->mutex);
 823
 824         if (pollstop && vq->handle_kick)
 825                 vhost_poll_flush(&vq->poll);
 826         return r;
 827 }
 828 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
 829
 830 /* Caller must have device mutex */
 831 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
 832 {
 833         struct file *eventfp, *filep = NULL;
 834         struct eventfd_ctx *ctx = NULL;
 835         u64 p;
 836         long r;
 837         int i, fd;
 838
 839         /* If you are not the owner, you can become one */
 840         if (ioctl == VHOST_SET_OWNER) {
 841                 r = vhost_dev_set_owner(d);
 842                 goto done;
 843         }
 844
 845         /* You must be the owner to do anything else */
 846         r = vhost_dev_check_owner(d);
 847         if (r)
 848                 goto done;
 849
 850         switch (ioctl) {
 851         case VHOST_SET_MEM_TABLE:
 852                 r = vhost_set_memory(d, argp);
 853                 break;
 854         case VHOST_SET_LOG_BASE:
 855                 if (copy_from_user(&p, argp, sizeof p)) {
 856                         r = -EFAULT;
 857                         break;
 858                 }
 859                 if ((u64)(unsigned long)p != p) {
 860                         r = -EFAULT;
 861                         break;
 862                 }
 863                 for (i = 0; i < d->nvqs; ++i) {
 864                         struct vhost_virtqueue *vq;
 865                         void __user *base = (void __user *)(unsigned long)p;
 866                         vq = d->vqs[i];
 867                         mutex_lock(&vq->mutex);
 868                         /* If ring is inactive, will check when it's enabled. */
 869                         if (vq->private_data && !vq_log_access_ok(d, vq, base))
 870                                 r = -EFAULT;
 871                         else
 872                                 vq->log_base = base;
 873                         mutex_unlock(&vq->mutex);
 874                 }
 875                 break;
 876         case VHOST_SET_LOG_FD:
 877                 r = get_user(fd, (int __user *)argp);
 878                 if (r < 0)
 879                         break;
 880                 eventfp = fd == -1 ? NULL : eventfd_fget(fd);
 881                 if (IS_ERR(eventfp)) {
 882                         r = PTR_ERR(eventfp);
 883                         break;
 884                 }
 885                 if (eventfp != d->log_file) {
 886                         filep = d->log_file;
 887                         ctx = d->log_ctx;
 888                         d->log_ctx = eventfp ?
 889                                 eventfd_ctx_fileget(eventfp) : NULL;
 890                 } else
 891                         filep = eventfp;
 892                 for (i = 0; i < d->nvqs; ++i) {
 893                         mutex_lock(&d->vqs[i]->mutex);
 894                         d->vqs[i]->log_ctx = d->log_ctx;
 895                         mutex_unlock(&d->vqs[i]->mutex);
 896                 }
 897                 if (ctx)
 898                         eventfd_ctx_put(ctx);
 899                 if (filep)
 900                         fput(filep);
 901                 break;
 902         default:
 903                 r = -ENOIOCTLCMD;
 904                 break;
 905         }
 906 done:
 907         return r;
 908 }
 909 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
 910
 911 static const struct vhost_memory_region *find_region(struct vhost_memory *mem,
 912                                                      __u64 addr, __u32 len)
 913 {
 914         struct vhost_memory_region *reg;
 915         int i;
 916
 917         /* linear search is not brilliant, but we really have on the order of 6
 918          * regions in practice */
 919         for (i = 0; i < mem->nregions; ++i) {
 920                 reg = mem->regions + i;
 921                 if (reg->guest_phys_addr <= addr &&
 922                     reg->guest_phys_addr + reg->memory_size - 1 >= addr)
 923                         return reg;
 924         }
 925         return NULL;
 926 }
 927
 928 /* TODO: This is really inefficient.  We need something like get_user()
 929  * (instruction directly accesses the data, with an exception table entry
 930  * returning -EFAULT). See Documentation/x86/exception-tables.txt.
 931  */
 932 static int set_bit_to_user(int nr, void __user *addr)
 933 {
 934         unsigned long log = (unsigned long)addr;
 935         struct page *page;
 936         void *base;
 937         int bit = nr + (log % PAGE_SIZE) * 8;
 938         int r;
 939
 940         r = get_user_pages_fast(log, 1, 1, &page);
 941         if (r < 0)
 942                 return r;
 943         BUG_ON(r != 1);
 944         base = kmap_atomic(page);
 945         set_bit(bit, base);
 946         kunmap_atomic(base);
 947         set_page_dirty_lock(page);
 948         put_page(page);
 949         return 0;
 950 }
 951
 952 static int log_write(void __user *log_base,
 953                      u64 write_address, u64 write_length)
 954 {
 955         u64 write_page = write_address / VHOST_PAGE_SIZE;
 956         int r;
 957
 958         if (!write_length)
 959                 return 0;
 960         write_length += write_address % VHOST_PAGE_SIZE;
 961         for (;;) {
 962                 u64 base = (u64)(unsigned long)log_base;
 963                 u64 log = base + write_page / 8;
 964                 int bit = write_page % 8;
 965                 if ((u64)(unsigned long)log != log)
 966                         return -EFAULT;
 967                 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
 968                 if (r < 0)
 969                         return r;
 970                 if (write_length <= VHOST_PAGE_SIZE)
 971                         break;
 972                 write_length -= VHOST_PAGE_SIZE;
 973                 write_page += 1;
 974         }
 975         return r;
 976 }
 977
 978 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
 979                     unsigned int log_num, u64 len)
 980 {
 981         int i, r;
 982
 983         /* Make sure data written is seen before log. */
 984         smp_wmb();
 985         for (i = 0; i < log_num; ++i) {
 986                 u64 l = min(log[i].len, len);
 987                 r = log_write(vq->log_base, log[i].addr, l);
 988                 if (r < 0)
 989                         return r;
 990                 len -= l;
 991                 if (!len) {
 992                         if (vq->log_ctx)
 993                                 eventfd_signal(vq->log_ctx, 1);
 994                         return 0;
 995                 }
 996         }
 997         /* Length written exceeds what we have stored. This is a bug. */
 998         BUG();
 999         return 0;
1000 }
1001 EXPORT_SYMBOL_GPL(vhost_log_write);
1002
1003 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1004 {
1005         void __user *used;
1006         if (__put_user(vq->used_flags, &vq->used->flags) < 0)
1007                 return -EFAULT;
1008         if (unlikely(vq->log_used)) {
1009                 /* Make sure the flag is seen before log. */
1010                 smp_wmb();
1011                 /* Log used flag write. */
1012                 used = &vq->used->flags;
1013                 log_write(vq->log_base, vq->log_addr +
1014                           (used - (void __user *)vq->used),
1015                           sizeof vq->used->flags);
1016                 if (vq->log_ctx)
1017                         eventfd_signal(vq->log_ctx, 1);
1018         }
1019         return 0;
1020 }
1021
1022 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1023 {
1024         if (__put_user(vq->avail_idx, vhost_avail_event(vq)))
1025                 return -EFAULT;
1026         if (unlikely(vq->log_used)) {
1027                 void __user *used;
1028                 /* Make sure the event is seen before log. */
1029                 smp_wmb();
1030                 /* Log avail event write */
1031                 used = vhost_avail_event(vq);
1032                 log_write(vq->log_base, vq->log_addr +
1033                           (used - (void __user *)vq->used),
1034                           sizeof *vhost_avail_event(vq));
1035                 if (vq->log_ctx)
1036                         eventfd_signal(vq->log_ctx, 1);
1037         }
1038         return 0;
1039 }
1040
1041 int vhost_init_used(struct vhost_virtqueue *vq)
1042 {
1043         int r;
1044         if (!vq->private_data)
1045                 return 0;
1046
1047         r = vhost_update_used_flags(vq);
1048         if (r)
1049                 return r;
1050         vq->signalled_used_valid = false;
1051         return get_user(vq->last_used_idx, &vq->used->idx);
1052 }
1053 EXPORT_SYMBOL_GPL(vhost_init_used);
1054
1055 static int translate_desc(struct vhost_dev *dev, u64 addr, u32 len,
1056                           struct iovec iov[], int iov_size)
1057 {
1058         const struct vhost_memory_region *reg;
1059         struct vhost_memory *mem;
1060         struct iovec *_iov;
1061         u64 s = 0;
1062         int ret = 0;
1063
1064         rcu_read_lock();
1065
1066         mem = rcu_dereference(dev->memory);
1067         while ((u64)len > s) {
1068                 u64 size;
1069                 if (unlikely(ret >= iov_size)) {
1070                         ret = -ENOBUFS;
1071                         break;
1072                 }
1073                 reg = find_region(mem, addr, len);
1074                 if (unlikely(!reg)) {
1075                         ret = -EFAULT;
1076                         break;
1077                 }
1078                 _iov = iov + ret;
1079                 size = reg->memory_size - addr + reg->guest_phys_addr;
1080                 _iov->iov_len = min((u64)len - s, size);
1081                 _iov->iov_base = (void __user *)(unsigned long)
1082                         (reg->userspace_addr + addr - reg->guest_phys_addr);
1083                 s += size;
1084                 addr += size;
1085                 ++ret;
1086         }
1087
1088         rcu_read_unlock();
1089         return ret;
1090 }
1091
1092 /* Each buffer in the virtqueues is actually a chain of descriptors.  This
1093  * function returns the next descriptor in the chain,
1094  * or -1U if we're at the end. */
1095 static unsigned next_desc(struct vring_desc *desc)
1096 {
1097         unsigned int next;
1098
1099         /* If this descriptor says it doesn't chain, we're done. */
1100         if (!(desc->flags & VRING_DESC_F_NEXT))
1101                 return -1U;
1102
1103         /* Check they're not leading us off end of descriptors. */
1104         next = desc->next;
1105         /* Make sure compiler knows to grab that: we don't want it changing! */
1106         /* We will use the result as an index in an array, so most
1107          * architectures only need a compiler barrier here. */
1108         read_barrier_depends();
1109
1110         return next;
1111 }
1112
1113 static int get_indirect(struct vhost_dev *dev, struct vhost_virtqueue *vq,
1114                         struct iovec iov[], unsigned int iov_size,
1115                         unsigned int *out_num, unsigned int *in_num,
1116                         struct vhost_log *log, unsigned int *log_num,
1117                         struct vring_desc *indirect)
1118 {
1119         struct vring_desc desc;
1120         unsigned int i = 0, count, found = 0;
1121         int ret;
1122
1123         /* Sanity check */
1124         if (unlikely(indirect->len % sizeof desc)) {
1125                 vq_err(vq, "Invalid length in indirect descriptor: "
1126                        "len 0x%llx not multiple of 0x%zx\n",
1127                        (unsigned long long)indirect->len,
1128                        sizeof desc);
1129                 return -EINVAL;
1130         }
1131
1132         ret = translate_desc(dev, indirect->addr, indirect->len, vq->indirect,
1133                              UIO_MAXIOV);
1134         if (unlikely(ret < 0)) {
1135                 vq_err(vq, "Translation failure %d in indirect.\n", ret);
1136                 return ret;
1137         }
1138
1139         /* We will use the result as an address to read from, so most
1140          * architectures only need a compiler barrier here. */
1141         read_barrier_depends();
1142
1143         count = indirect->len / sizeof desc;
1144         /* Buffers are chained via a 16 bit next field, so
1145          * we can have at most 2^16 of these. */
1146         if (unlikely(count > USHRT_MAX + 1)) {
1147                 vq_err(vq, "Indirect buffer length too big: %d\n",
1148                        indirect->len);
1149                 return -E2BIG;
1150         }
1151
1152         do {
1153                 unsigned iov_count = *in_num + *out_num;
1154                 if (unlikely(++found > count)) {
1155                         vq_err(vq, "Loop detected: last one at %u "
1156                                "indirect size %u\n",
1157                                i, count);
1158                         return -EINVAL;
1159                 }
1160                 if (unlikely(memcpy_fromiovec((unsigned char *)&desc,
1161                                               vq->indirect, sizeof desc))) {
1162                         vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1163                                i, (size_t)indirect->addr + i * sizeof desc);
1164                         return -EINVAL;
1165                 }
1166                 if (unlikely(desc.flags & VRING_DESC_F_INDIRECT)) {
1167                         vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1168                                i, (size_t)indirect->addr + i * sizeof desc);
1169                         return -EINVAL;
1170                 }
1171
1172                 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
1173                                      iov_size - iov_count);
1174                 if (unlikely(ret < 0)) {
1175                         vq_err(vq, "Translation failure %d indirect idx %d\n",
1176                                ret, i);
1177                         return ret;
1178                 }
1179                 /* If this is an input descriptor, increment that count. */
1180                 if (desc.flags & VRING_DESC_F_WRITE) {
1181                         *in_num += ret;
1182                         if (unlikely(log)) {
1183                                 log[*log_num].addr = desc.addr;
1184                                 log[*log_num].len = desc.len;
1185                                 ++*log_num;
1186                         }
1187                 } else {
1188                         /* If it's an output descriptor, they're all supposed
1189                          * to come before any input descriptors. */
1190                         if (unlikely(*in_num)) {
1191                                 vq_err(vq, "Indirect descriptor "
1192                                        "has out after in: idx %d\n", i);
1193                                 return -EINVAL;
1194                         }
1195                         *out_num += ret;
1196                 }
1197         } while ((i = next_desc(&desc)) != -1);
1198         return 0;
1199 }
1200
1201 /* This looks in the virtqueue and for the first available buffer, and converts
1202  * it to an iovec for convenient access.  Since descriptors consist of some
1203  * number of output then some number of input descriptors, it's actually two
1204  * iovecs, but we pack them into one and note how many of each there were.
1205  *
1206  * This function returns the descriptor number found, or vq->num (which is
1207  * never a valid descriptor number) if none was found.  A negative code is
1208  * returned on error. */
1209 int vhost_get_vq_desc(struct vhost_dev *dev, struct vhost_virtqueue *vq,
1210                       struct iovec iov[], unsigned int iov_size,
1211                       unsigned int *out_num, unsigned int *in_num,
1212                       struct vhost_log *log, unsigned int *log_num)
1213 {
1214         struct vring_desc desc;
1215         unsigned int i, head, found = 0;
1216         u16 last_avail_idx;
1217         int ret;
1218
1219         /* Check it isn't doing very strange things with descriptor numbers. */
1220         last_avail_idx = vq->last_avail_idx;
1221         if (unlikely(__get_user(vq->avail_idx, &vq->avail->idx))) {
1222                 vq_err(vq, "Failed to access avail idx at %p\n",
1223                        &vq->avail->idx);
1224                 return -EFAULT;
1225         }
1226
1227         if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
1228                 vq_err(vq, "Guest moved used index from %u to %u",
1229                        last_avail_idx, vq->avail_idx);
1230                 return -EFAULT;
1231         }
1232
1233         /* If there's nothing new since last we looked, return invalid. */
1234         if (vq->avail_idx == last_avail_idx)
1235                 return vq->num;
1236
1237         /* Only get avail ring entries after they have been exposed by guest. */
1238         smp_rmb();
1239
1240         /* Grab the next descriptor number they're advertising, and increment
1241          * the index we've seen. */
1242         if (unlikely(__get_user(head,
1243                                 &vq->avail->ring[last_avail_idx % vq->num]))) {
1244                 vq_err(vq, "Failed to read head: idx %d address %p\n",
1245                        last_avail_idx,
1246                        &vq->avail->ring[last_avail_idx % vq->num]);
1247                 return -EFAULT;
1248         }
1249
1250         /* If their number is silly, that's an error. */
1251         if (unlikely(head >= vq->num)) {
1252                 vq_err(vq, "Guest says index %u > %u is available",
1253                        head, vq->num);
1254                 return -EINVAL;
1255         }
1256
1257         /* When we start there are none of either input nor output. */
1258         *out_num = *in_num = 0;
1259         if (unlikely(log))
1260                 *log_num = 0;
1261
1262         i = head;
1263         do {
1264                 unsigned iov_count = *in_num + *out_num;
1265                 if (unlikely(i >= vq->num)) {
1266                         vq_err(vq, "Desc index is %u > %u, head = %u",
1267                                i, vq->num, head);
1268                         return -EINVAL;
1269                 }
1270                 if (unlikely(++found > vq->num)) {
1271                         vq_err(vq, "Loop detected: last one at %u "
1272                                "vq size %u head %u\n",
1273                                i, vq->num, head);
1274                         return -EINVAL;
1275                 }
1276                 ret = __copy_from_user(&desc, vq->desc + i, sizeof desc);
1277                 if (unlikely(ret)) {
1278                         vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
1279                                i, vq->desc + i);
1280                         return -EFAULT;
1281                 }
1282                 if (desc.flags & VRING_DESC_F_INDIRECT) {
1283                         ret = get_indirect(dev, vq, iov, iov_size,
1284                                            out_num, in_num,
1285                                            log, log_num, &desc);
1286                         if (unlikely(ret < 0)) {
1287                                 vq_err(vq, "Failure detected "
1288                                        "in indirect descriptor at idx %d\n", i);
1289                                 return ret;
1290                         }
1291                         continue;
1292                 }
1293
1294                 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
1295                                      iov_size - iov_count);
1296                 if (unlikely(ret < 0)) {
1297                         vq_err(vq, "Translation failure %d descriptor idx %d\n",
1298                                ret, i);
1299                         return ret;
1300                 }
1301                 if (desc.flags & VRING_DESC_F_WRITE) {
1302                         /* If this is an input descriptor,
1303                          * increment that count. */
1304                         *in_num += ret;
1305                         if (unlikely(log)) {
1306                                 log[*log_num].addr = desc.addr;
1307                                 log[*log_num].len = desc.len;
1308                                 ++*log_num;
1309                         }
1310                 } else {
1311                         /* If it's an output descriptor, they're all supposed
1312                          * to come before any input descriptors. */
1313                         if (unlikely(*in_num)) {
1314                                 vq_err(vq, "Descriptor has out after in: "
1315                                        "idx %d\n", i);
1316                                 return -EINVAL;
1317                         }
1318                         *out_num += ret;
1319                 }
1320         } while ((i = next_desc(&desc)) != -1);
1321
1322         /* On success, increment avail index. */
1323         vq->last_avail_idx++;
1324
1325         /* Assume notifications from guest are disabled at this point,
1326          * if they aren't we would need to update avail_event index. */
1327         BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
1328         return head;
1329 }
1330 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
1331
1332 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
1333 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
1334 {
1335         vq->last_avail_idx -= n;
1336 }
1337 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
1338
1339 /* After we've used one of their buffers, we tell them about it.  We'll then
1340  * want to notify the guest, using eventfd. */
1341 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
1342 {
1343         struct vring_used_elem heads = { head, len };
1344
1345         return vhost_add_used_n(vq, &heads, 1);
1346 }
1347 EXPORT_SYMBOL_GPL(vhost_add_used);
1348
1349 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
1350                             struct vring_used_elem *heads,
1351                             unsigned count)
1352 {
1353         struct vring_used_elem __user *used;
1354         u16 old, new;
1355         int start;
1356
1357         start = vq->last_used_idx % vq->num;
1358         used = vq->used->ring + start;
1359         if (count == 1) {
1360                 if (__put_user(heads[0].id, &used->id)) {
1361                         vq_err(vq, "Failed to write used id");
1362                         return -EFAULT;
1363                 }
1364                 if (__put_user(heads[0].len, &used->len)) {
1365                         vq_err(vq, "Failed to write used len");
1366                         return -EFAULT;
1367                 }
1368         } else if (__copy_to_user(used, heads, count * sizeof *used)) {
1369                 vq_err(vq, "Failed to write used");
1370                 return -EFAULT;
1371         }
1372         if (unlikely(vq->log_used)) {
1373                 /* Make sure data is seen before log. */
1374                 smp_wmb();
1375                 /* Log used ring entry write. */
1376                 log_write(vq->log_base,
1377                           vq->log_addr +
1378                            ((void __user *)used - (void __user *)vq->used),
1379                           count * sizeof *used);
1380         }
1381         old = vq->last_used_idx;
1382         new = (vq->last_used_idx += count);
1383         /* If the driver never bothers to signal in a very long while,
1384          * used index might wrap around. If that happens, invalidate
1385          * signalled_used index we stored. TODO: make sure driver
1386          * signals at least once in 2^16 and remove this. */
1387         if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
1388                 vq->signalled_used_valid = false;
1389         return 0;
1390 }
1391
1392 /* After we've used one of their buffers, we tell them about it.  We'll then
1393  * want to notify the guest, using eventfd. */
1394 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
1395                      unsigned count)
1396 {
1397         int start, n, r;
1398
1399         start = vq->last_used_idx % vq->num;
1400         n = vq->num - start;
1401         if (n < count) {
1402                 r = __vhost_add_used_n(vq, heads, n);
1403                 if (r < 0)
1404                         return r;
1405                 heads += n;
1406                 count -= n;
1407         }
1408         r = __vhost_add_used_n(vq, heads, count);
1409
1410         /* Make sure buffer is written before we update index. */
1411         smp_wmb();
1412         if (put_user(vq->last_used_idx, &vq->used->idx)) {
1413                 vq_err(vq, "Failed to increment used idx");
1414                 return -EFAULT;
1415         }
1416         if (unlikely(vq->log_used)) {
1417                 /* Log used index update. */
1418                 log_write(vq->log_base,
1419                           vq->log_addr + offsetof(struct vring_used, idx),
1420                           sizeof vq->used->idx);
1421                 if (vq->log_ctx)
1422                         eventfd_signal(vq->log_ctx, 1);
1423         }
1424         return r;
1425 }
1426 EXPORT_SYMBOL_GPL(vhost_add_used_n);
1427
1428 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1429 {
1430         __u16 old, new, event;
1431         bool v;
1432         /* Flush out used index updates. This is paired
1433          * with the barrier that the Guest executes when enabling
1434          * interrupts. */
1435         smp_mb();
1436
1437         if (vhost_has_feature(dev, VIRTIO_F_NOTIFY_ON_EMPTY) &&
1438             unlikely(vq->avail_idx == vq->last_avail_idx))
1439                 return true;
1440
1441         if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1442                 __u16 flags;
1443                 if (__get_user(flags, &vq->avail->flags)) {
1444                         vq_err(vq, "Failed to get flags");
1445                         return true;
1446                 }
1447                 return !(flags & VRING_AVAIL_F_NO_INTERRUPT);
1448         }
1449         old = vq->signalled_used;
1450         v = vq->signalled_used_valid;
1451         new = vq->signalled_used = vq->last_used_idx;
1452         vq->signalled_used_valid = true;
1453
1454         if (unlikely(!v))
1455                 return true;
1456
1457         if (get_user(event, vhost_used_event(vq))) {
1458                 vq_err(vq, "Failed to get used event idx");
1459                 return true;
1460         }
1461         return vring_need_event(event, new, old);
1462 }
1463
1464 /* This actually signals the guest, using eventfd. */
1465 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1466 {
1467         /* Signal the Guest tell them we used something up. */
1468         if (vq->call_ctx && vhost_notify(dev, vq))
1469                 eventfd_signal(vq->call_ctx, 1);
1470 }
1471 EXPORT_SYMBOL_GPL(vhost_signal);
1472
1473 /* And here's the combo meal deal.  Supersize me! */
1474 void vhost_add_used_and_signal(struct vhost_dev *dev,
1475                                struct vhost_virtqueue *vq,
1476                                unsigned int head, int len)
1477 {
1478         vhost_add_used(vq, head, len);
1479         vhost_signal(dev, vq);
1480 }
1481 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
1482
1483 /* multi-buffer version of vhost_add_used_and_signal */
1484 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
1485                                  struct vhost_virtqueue *vq,
1486                                  struct vring_used_elem *heads, unsigned count)
1487 {
1488         vhost_add_used_n(vq, heads, count);
1489         vhost_signal(dev, vq);
1490 }
1491 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
1492
1493 /* OK, now we need to know about added descriptors. */
1494 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1495 {
1496         u16 avail_idx;
1497         int r;
1498
1499         if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
1500                 return false;
1501         vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
1502         if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1503                 r = vhost_update_used_flags(vq);
1504                 if (r) {
1505                         vq_err(vq, "Failed to enable notification at %p: %d\n",
1506                                &vq->used->flags, r);
1507                         return false;
1508                 }
1509         } else {
1510                 r = vhost_update_avail_event(vq, vq->avail_idx);
1511                 if (r) {
1512                         vq_err(vq, "Failed to update avail event index at %p: %d\n",
1513                                vhost_avail_event(vq), r);
1514                         return false;
1515                 }
1516         }
1517         /* They could have slipped one in as we were doing that: make
1518          * sure it's written, then check again. */
1519         smp_mb();
1520         r = __get_user(avail_idx, &vq->avail->idx);
1521         if (r) {
1522                 vq_err(vq, "Failed to check avail idx at %p: %d\n",
1523                        &vq->avail->idx, r);
1524                 return false;
1525         }
1526
1527         return avail_idx != vq->avail_idx;
1528 }
1529 EXPORT_SYMBOL_GPL(vhost_enable_notify);
1530
1531 /* We don't need to be notified again. */
1532 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1533 {
1534         int r;
1535
1536         if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
1537                 return;
1538         vq->used_flags |= VRING_USED_F_NO_NOTIFY;
1539         if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1540                 r = vhost_update_used_flags(vq);
1541                 if (r)
1542                         vq_err(vq, "Failed to enable notification at %p: %d\n",
1543                                &vq->used->flags, r);
1544         }
1545 }
1546 EXPORT_SYMBOL_GPL(vhost_disable_notify);
1547
1548 static int __init vhost_init(void)
1549 {
1550         return 0;
1551 }
1552
1553 static void __exit vhost_exit(void)
1554 {
1555 }
1556
1557 module_init(vhost_init);
1558 module_exit(vhost_exit);
1559
1560 MODULE_VERSION("0.0.1");
1561 MODULE_LICENSE("GPL v2");
1562 MODULE_AUTHOR("Michael S. Tsirkin");
1563 MODULE_DESCRIPTION("Host kernel accelerator for virtio");