Pileus Git - ~andy/linux/blob - drivers/vhost/vhost.c

   1 /* Copyright (C) 2009 Red Hat, Inc.
   2  * Copyright (C) 2006 Rusty Russell IBM Corporation
   3  *
   4  * Author: Michael S. Tsirkin <mst@redhat.com>
   5  *
   6  * Inspiration, some code, and most witty comments come from
   7  * Documentation/virtual/lguest/lguest.c, by Rusty Russell
   8  *
   9  * This work is licensed under the terms of the GNU GPL, version 2.
  10  *
  11  * Generic code for virtio server in host kernel.
  12  */
  13
  14 #include <linux/eventfd.h>
  15 #include <linux/vhost.h>
  16 #include <linux/virtio_net.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
  29 #include "vhost.h"
  30
  31 enum {
  32         VHOST_MEMORY_MAX_NREGIONS = 64,
  33         VHOST_MEMORY_F_LOG = 0x1,
  34 };
  35
  36 static unsigned vhost_zcopy_mask __read_mostly;
  37
  38 #define vhost_used_event(vq) ((u16 __user *)&vq->avail->ring[vq->num])
  39 #define vhost_avail_event(vq) ((u16 __user *)&vq->used->ring[vq->num])
  40
  41 static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
  42                             poll_table *pt)
  43 {
  44         struct vhost_poll *poll;
  45
  46         poll = container_of(pt, struct vhost_poll, table);
  47         poll->wqh = wqh;
  48         add_wait_queue(wqh, &poll->wait);
  49 }
  50
  51 static int vhost_poll_wakeup(wait_queue_t *wait, unsigned mode, int sync,
  52                              void *key)
  53 {
  54         struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
  55
  56         if (!((unsigned long)key & poll->mask))
  57                 return 0;
  58
  59         vhost_poll_queue(poll);
  60         return 0;
  61 }
  62
  63 void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
  64 {
  65         INIT_LIST_HEAD(&work->node);
  66         work->fn = fn;
  67         init_waitqueue_head(&work->done);
  68         work->flushing = 0;
  69         work->queue_seq = work->done_seq = 0;
  70 }
  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
  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
 107 /* Stop polling a file. After this function returns, it becomes safe to drop the
 108  * file reference. You must also flush afterwards. */
 109 void vhost_poll_stop(struct vhost_poll *poll)
 110 {
 111         if (poll->wqh) {
 112                 remove_wait_queue(poll->wqh, &poll->wait);
 113                 poll->wqh = NULL;
 114         }
 115 }
 116
 117 static bool vhost_work_seq_done(struct vhost_dev *dev, struct vhost_work *work,
 118                                 unsigned seq)
 119 {
 120         int left;
 121
 122         spin_lock_irq(&dev->work_lock);
 123         left = seq - work->done_seq;
 124         spin_unlock_irq(&dev->work_lock);
 125         return left <= 0;
 126 }
 127
 128 static void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work)
 129 {
 130         unsigned seq;
 131         int flushing;
 132
 133         spin_lock_irq(&dev->work_lock);
 134         seq = work->queue_seq;
 135         work->flushing++;
 136         spin_unlock_irq(&dev->work_lock);
 137         wait_event(work->done, vhost_work_seq_done(dev, work, seq));
 138         spin_lock_irq(&dev->work_lock);
 139         flushing = --work->flushing;
 140         spin_unlock_irq(&dev->work_lock);
 141         BUG_ON(flushing < 0);
 142 }
 143
 144 /* Flush any work that has been scheduled. When calling this, don't hold any
 145  * locks that are also used by the callback. */
 146 void vhost_poll_flush(struct vhost_poll *poll)
 147 {
 148         vhost_work_flush(poll->dev, &poll->work);
 149 }
 150
 151 void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work)
 152 {
 153         unsigned long flags;
 154
 155         spin_lock_irqsave(&dev->work_lock, flags);
 156         if (list_empty(&work->node)) {
 157                 list_add_tail(&work->node, &dev->work_list);
 158                 work->queue_seq++;
 159                 wake_up_process(dev->worker);
 160         }
 161         spin_unlock_irqrestore(&dev->work_lock, flags);
 162 }
 163
 164 void vhost_poll_queue(struct vhost_poll *poll)
 165 {
 166         vhost_work_queue(poll->dev, &poll->work);
 167 }
 168
 169 static void vhost_vq_reset(struct vhost_dev *dev,
 170                            struct vhost_virtqueue *vq)
 171 {
 172         vq->num = 1;
 173         vq->desc = NULL;
 174         vq->avail = NULL;
 175         vq->used = NULL;
 176         vq->last_avail_idx = 0;
 177         vq->avail_idx = 0;
 178         vq->last_used_idx = 0;
 179         vq->signalled_used = 0;
 180         vq->signalled_used_valid = false;
 181         vq->used_flags = 0;
 182         vq->log_used = false;
 183         vq->log_addr = -1ull;
 184         vq->vhost_hlen = 0;
 185         vq->sock_hlen = 0;
 186         vq->private_data = NULL;
 187         vq->log_base = NULL;
 188         vq->error_ctx = NULL;
 189         vq->error = NULL;
 190         vq->kick = NULL;
 191         vq->call_ctx = NULL;
 192         vq->call = NULL;
 193         vq->log_ctx = NULL;
 194         vq->upend_idx = 0;
 195         vq->done_idx = 0;
 196         vq->ubufs = NULL;
 197 }
 198
 199 static int vhost_worker(void *data)
 200 {
 201         struct vhost_dev *dev = data;
 202         struct vhost_work *work = NULL;
 203         unsigned uninitialized_var(seq);
 204         mm_segment_t oldfs = get_fs();
 205
 206         set_fs(USER_DS);
 207         use_mm(dev->mm);
 208
 209         for (;;) {
 210                 /* mb paired w/ kthread_stop */
 211                 set_current_state(TASK_INTERRUPTIBLE);
 212
 213                 spin_lock_irq(&dev->work_lock);
 214                 if (work) {
 215                         work->done_seq = seq;
 216                         if (work->flushing)
 217                                 wake_up_all(&work->done);
 218                 }
 219
 220                 if (kthread_should_stop()) {
 221                         spin_unlock_irq(&dev->work_lock);
 222                         __set_current_state(TASK_RUNNING);
 223                         break;
 224                 }
 225                 if (!list_empty(&dev->work_list)) {
 226                         work = list_first_entry(&dev->work_list,
 227                                                 struct vhost_work, node);
 228                         list_del_init(&work->node);
 229                         seq = work->queue_seq;
 230                 } else
 231                         work = NULL;
 232                 spin_unlock_irq(&dev->work_lock);
 233
 234                 if (work) {
 235                         __set_current_state(TASK_RUNNING);
 236                         work->fn(work);
 237                         if (need_resched())
 238                                 schedule();
 239                 } else
 240                         schedule();
 241
 242         }
 243         unuse_mm(dev->mm);
 244         set_fs(oldfs);
 245         return 0;
 246 }
 247
 248 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
 249 {
 250         kfree(vq->indirect);
 251         vq->indirect = NULL;
 252         kfree(vq->log);
 253         vq->log = NULL;
 254         kfree(vq->heads);
 255         vq->heads = NULL;
 256         kfree(vq->ubuf_info);
 257         vq->ubuf_info = NULL;
 258 }
 259
 260 void vhost_enable_zcopy(int vq)
 261 {
 262         vhost_zcopy_mask |= 0x1 << vq;
 263 }
 264
 265 /* Helper to allocate iovec buffers for all vqs. */
 266 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
 267 {
 268         int i;
 269         bool zcopy;
 270
 271         for (i = 0; i < dev->nvqs; ++i) {
 272                 dev->vqs[i]->indirect = kmalloc(sizeof *dev->vqs[i]->indirect *
 273                                                UIO_MAXIOV, GFP_KERNEL);
 274                 dev->vqs[i]->log = kmalloc(sizeof *dev->vqs[i]->log * UIO_MAXIOV,
 275                                           GFP_KERNEL);
 276                 dev->vqs[i]->heads = kmalloc(sizeof *dev->vqs[i]->heads *
 277                                             UIO_MAXIOV, GFP_KERNEL);
 278                 zcopy = vhost_zcopy_mask & (0x1 << i);
 279                 if (zcopy)
 280                         dev->vqs[i]->ubuf_info =
 281                                 kmalloc(sizeof *dev->vqs[i]->ubuf_info *
 282                                         UIO_MAXIOV, GFP_KERNEL);
 283                 if (!dev->vqs[i]->indirect || !dev->vqs[i]->log ||
 284                         !dev->vqs[i]->heads ||
 285                         (zcopy && !dev->vqs[i]->ubuf_info))
 286                         goto err_nomem;
 287         }
 288         return 0;
 289
 290 err_nomem:
 291         for (; i >= 0; --i)
 292                 vhost_vq_free_iovecs(dev->vqs[i]);
 293         return -ENOMEM;
 294 }
 295
 296 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
 297 {
 298         int i;
 299
 300         for (i = 0; i < dev->nvqs; ++i)
 301                 vhost_vq_free_iovecs(dev->vqs[i]);
 302 }
 303
 304 long vhost_dev_init(struct vhost_dev *dev,
 305                     struct vhost_virtqueue **vqs, int nvqs)
 306 {
 307         int i;
 308
 309         dev->vqs = vqs;
 310         dev->nvqs = nvqs;
 311         mutex_init(&dev->mutex);
 312         dev->log_ctx = NULL;
 313         dev->log_file = NULL;
 314         dev->memory = NULL;
 315         dev->mm = NULL;
 316         spin_lock_init(&dev->work_lock);
 317         INIT_LIST_HEAD(&dev->work_list);
 318         dev->worker = NULL;
 319
 320         for (i = 0; i < dev->nvqs; ++i) {
 321                 dev->vqs[i]->log = NULL;
 322                 dev->vqs[i]->indirect = NULL;
 323                 dev->vqs[i]->heads = NULL;
 324                 dev->vqs[i]->ubuf_info = NULL;
 325                 dev->vqs[i]->dev = dev;
 326                 mutex_init(&dev->vqs[i]->mutex);
 327                 vhost_vq_reset(dev, dev->vqs[i]);
 328                 if (dev->vqs[i]->handle_kick)
 329                         vhost_poll_init(&dev->vqs[i]->poll,
 330                                         dev->vqs[i]->handle_kick, POLLIN, dev);
 331         }
 332
 333         return 0;
 334 }
 335
 336 /* Caller should have device mutex */
 337 long vhost_dev_check_owner(struct vhost_dev *dev)
 338 {
 339         /* Are you the owner? If not, I don't think you mean to do that */
 340         return dev->mm == current->mm ? 0 : -EPERM;
 341 }
 342
 343 struct vhost_attach_cgroups_struct {
 344         struct vhost_work work;
 345         struct task_struct *owner;
 346         int ret;
 347 };
 348
 349 static void vhost_attach_cgroups_work(struct vhost_work *work)
 350 {
 351         struct vhost_attach_cgroups_struct *s;
 352
 353         s = container_of(work, struct vhost_attach_cgroups_struct, work);
 354         s->ret = cgroup_attach_task_all(s->owner, current);
 355 }
 356
 357 static int vhost_attach_cgroups(struct vhost_dev *dev)
 358 {
 359         struct vhost_attach_cgroups_struct attach;
 360
 361         attach.owner = current;
 362         vhost_work_init(&attach.work, vhost_attach_cgroups_work);
 363         vhost_work_queue(dev, &attach.work);
 364         vhost_work_flush(dev, &attach.work);
 365         return attach.ret;
 366 }
 367
 368 /* Caller should have device mutex */
 369 static long vhost_dev_set_owner(struct vhost_dev *dev)
 370 {
 371         struct task_struct *worker;
 372         int err;
 373
 374         /* Is there an owner already? */
 375         if (dev->mm) {
 376                 err = -EBUSY;
 377                 goto err_mm;
 378         }
 379
 380         /* No owner, become one */
 381         dev->mm = get_task_mm(current);
 382         worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
 383         if (IS_ERR(worker)) {
 384                 err = PTR_ERR(worker);
 385                 goto err_worker;
 386         }
 387
 388         dev->worker = worker;
 389         wake_up_process(worker);        /* avoid contributing to loadavg */
 390
 391         err = vhost_attach_cgroups(dev);
 392         if (err)
 393                 goto err_cgroup;
 394
 395         err = vhost_dev_alloc_iovecs(dev);
 396         if (err)
 397                 goto err_cgroup;
 398
 399         return 0;
 400 err_cgroup:
 401         kthread_stop(worker);
 402         dev->worker = NULL;
 403 err_worker:
 404         if (dev->mm)
 405                 mmput(dev->mm);
 406         dev->mm = NULL;
 407 err_mm:
 408         return err;
 409 }
 410
 411 /* Caller should have device mutex */
 412 long vhost_dev_reset_owner(struct vhost_dev *dev)
 413 {
 414         struct vhost_memory *memory;
 415
 416         /* Restore memory to default empty mapping. */
 417         memory = kmalloc(offsetof(struct vhost_memory, regions), GFP_KERNEL);
 418         if (!memory)
 419                 return -ENOMEM;
 420
 421         vhost_dev_cleanup(dev, true);
 422
 423         memory->nregions = 0;
 424         RCU_INIT_POINTER(dev->memory, memory);
 425         return 0;
 426 }
 427
 428 void vhost_dev_stop(struct vhost_dev *dev)
 429 {
 430         int i;
 431
 432         for (i = 0; i < dev->nvqs; ++i) {
 433                 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick) {
 434                         vhost_poll_stop(&dev->vqs[i]->poll);
 435                         vhost_poll_flush(&dev->vqs[i]->poll);
 436                 }
 437         }
 438 }
 439
 440 /* Caller should have device mutex if and only if locked is set */
 441 void vhost_dev_cleanup(struct vhost_dev *dev, bool locked)
 442 {
 443         int i;
 444
 445         for (i = 0; i < dev->nvqs; ++i) {
 446                 if (dev->vqs[i]->error_ctx)
 447                         eventfd_ctx_put(dev->vqs[i]->error_ctx);
 448                 if (dev->vqs[i]->error)
 449                         fput(dev->vqs[i]->error);
 450                 if (dev->vqs[i]->kick)
 451                         fput(dev->vqs[i]->kick);
 452                 if (dev->vqs[i]->call_ctx)
 453                         eventfd_ctx_put(dev->vqs[i]->call_ctx);
 454                 if (dev->vqs[i]->call)
 455                         fput(dev->vqs[i]->call);
 456                 vhost_vq_reset(dev, dev->vqs[i]);
 457         }
 458         vhost_dev_free_iovecs(dev);
 459         if (dev->log_ctx)
 460                 eventfd_ctx_put(dev->log_ctx);
 461         dev->log_ctx = NULL;
 462         if (dev->log_file)
 463                 fput(dev->log_file);
 464         dev->log_file = NULL;
 465         /* No one will access memory at this point */
 466         kfree(rcu_dereference_protected(dev->memory,
 467                                         locked ==
 468                                                 lockdep_is_held(&dev->mutex)));
 469         RCU_INIT_POINTER(dev->memory, NULL);
 470         WARN_ON(!list_empty(&dev->work_list));
 471         if (dev->worker) {
 472                 kthread_stop(dev->worker);
 473                 dev->worker = NULL;
 474         }
 475         if (dev->mm)
 476                 mmput(dev->mm);
 477         dev->mm = NULL;
 478 }
 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
 565 /* Verify access for write logging. */
 566 /* Caller should have vq mutex and device mutex */
 567 static int vq_log_access_ok(struct vhost_dev *d, struct vhost_virtqueue *vq,
 568                             void __user *log_base)
 569 {
 570         struct vhost_memory *mp;
 571         size_t s = vhost_has_feature(d, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
 572
 573         mp = rcu_dereference_protected(vq->dev->memory,
 574                                        lockdep_is_held(&vq->mutex));
 575         return vq_memory_access_ok(log_base, mp,
 576                             vhost_has_feature(vq->dev, VHOST_F_LOG_ALL)) &&
 577                 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
 578                                         sizeof *vq->used +
 579                                         vq->num * sizeof *vq->used->ring + s));
 580 }
 581
 582 /* Can we start vq? */
 583 /* Caller should have vq mutex and device mutex */
 584 int vhost_vq_access_ok(struct vhost_virtqueue *vq)
 585 {
 586         return vq_access_ok(vq->dev, vq->num, vq->desc, vq->avail, vq->used) &&
 587                 vq_log_access_ok(vq->dev, vq, vq->log_base);
 588 }
 589
 590 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
 591 {
 592         struct vhost_memory mem, *newmem, *oldmem;
 593         unsigned long size = offsetof(struct vhost_memory, regions);
 594
 595         if (copy_from_user(&mem, m, size))
 596                 return -EFAULT;
 597         if (mem.padding)
 598                 return -EOPNOTSUPP;
 599         if (mem.nregions > VHOST_MEMORY_MAX_NREGIONS)
 600                 return -E2BIG;
 601         newmem = kmalloc(size + mem.nregions * sizeof *m->regions, GFP_KERNEL);
 602         if (!newmem)
 603                 return -ENOMEM;
 604
 605         memcpy(newmem, &mem, size);
 606         if (copy_from_user(newmem->regions, m->regions,
 607                            mem.nregions * sizeof *m->regions)) {
 608                 kfree(newmem);
 609                 return -EFAULT;
 610         }
 611
 612         if (!memory_access_ok(d, newmem,
 613                               vhost_has_feature(d, VHOST_F_LOG_ALL))) {
 614                 kfree(newmem);
 615                 return -EFAULT;
 616         }
 617         oldmem = rcu_dereference_protected(d->memory,
 618                                            lockdep_is_held(&d->mutex));
 619         rcu_assign_pointer(d->memory, newmem);
 620         synchronize_rcu();
 621         kfree(oldmem);
 622         return 0;
 623 }
 624
 625 long vhost_vring_ioctl(struct vhost_dev *d, int ioctl, void __user *argp)
 626 {
 627         struct file *eventfp, *filep = NULL;
 628         bool pollstart = false, pollstop = false;
 629         struct eventfd_ctx *ctx = NULL;
 630         u32 __user *idxp = argp;
 631         struct vhost_virtqueue *vq;
 632         struct vhost_vring_state s;
 633         struct vhost_vring_file f;
 634         struct vhost_vring_addr a;
 635         u32 idx;
 636         long r;
 637
 638         r = get_user(idx, idxp);
 639         if (r < 0)
 640                 return r;
 641         if (idx >= d->nvqs)
 642                 return -ENOBUFS;
 643
 644         vq = d->vqs[idx];
 645
 646         mutex_lock(&vq->mutex);
 647
 648         switch (ioctl) {
 649         case VHOST_SET_VRING_NUM:
 650                 /* Resizing ring with an active backend?
 651                  * You don't want to do that. */
 652                 if (vq->private_data) {
 653                         r = -EBUSY;
 654                         break;
 655                 }
 656                 if (copy_from_user(&s, argp, sizeof s)) {
 657                         r = -EFAULT;
 658                         break;
 659                 }
 660                 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
 661                         r = -EINVAL;
 662                         break;
 663                 }
 664                 vq->num = s.num;
 665                 break;
 666         case VHOST_SET_VRING_BASE:
 667                 /* Moving base with an active backend?
 668                  * You don't want to do that. */
 669                 if (vq->private_data) {
 670                         r = -EBUSY;
 671                         break;
 672                 }
 673                 if (copy_from_user(&s, argp, sizeof s)) {
 674                         r = -EFAULT;
 675                         break;
 676                 }
 677                 if (s.num > 0xffff) {
 678                         r = -EINVAL;
 679                         break;
 680                 }
 681                 vq->last_avail_idx = s.num;
 682                 /* Forget the cached index value. */
 683                 vq->avail_idx = vq->last_avail_idx;
 684                 break;
 685         case VHOST_GET_VRING_BASE:
 686                 s.index = idx;
 687                 s.num = vq->last_avail_idx;
 688                 if (copy_to_user(argp, &s, sizeof s))
 689                         r = -EFAULT;
 690                 break;
 691         case VHOST_SET_VRING_ADDR:
 692                 if (copy_from_user(&a, argp, sizeof a)) {
 693                         r = -EFAULT;
 694                         break;
 695                 }
 696                 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
 697                         r = -EOPNOTSUPP;
 698                         break;
 699                 }
 700                 /* For 32bit, verify that the top 32bits of the user
 701                    data are set to zero. */
 702                 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
 703                     (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
 704                     (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
 705                         r = -EFAULT;
 706                         break;
 707                 }
 708                 if ((a.avail_user_addr & (sizeof *vq->avail->ring - 1)) ||
 709                     (a.used_user_addr & (sizeof *vq->used->ring - 1)) ||
 710                     (a.log_guest_addr & (sizeof *vq->used->ring - 1))) {
 711                         r = -EINVAL;
 712                         break;
 713                 }
 714
 715                 /* We only verify access here if backend is configured.
 716                  * If it is not, we don't as size might not have been setup.
 717                  * We will verify when backend is configured. */
 718                 if (vq->private_data) {
 719                         if (!vq_access_ok(d, vq->num,
 720                                 (void __user *)(unsigned long)a.desc_user_addr,
 721                                 (void __user *)(unsigned long)a.avail_user_addr,
 722                                 (void __user *)(unsigned long)a.used_user_addr)) {
 723                                 r = -EINVAL;
 724                                 break;
 725                         }
 726
 727                         /* Also validate log access for used ring if enabled. */
 728                         if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
 729                             !log_access_ok(vq->log_base, a.log_guest_addr,
 730                                            sizeof *vq->used +
 731                                            vq->num * sizeof *vq->used->ring)) {
 732                                 r = -EINVAL;
 733                                 break;
 734                         }
 735                 }
 736
 737                 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
 738                 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
 739                 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
 740                 vq->log_addr = a.log_guest_addr;
 741                 vq->used = (void __user *)(unsigned long)a.used_user_addr;
 742                 break;
 743         case VHOST_SET_VRING_KICK:
 744                 if (copy_from_user(&f, argp, sizeof f)) {
 745                         r = -EFAULT;
 746                         break;
 747                 }
 748                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
 749                 if (IS_ERR(eventfp)) {
 750                         r = PTR_ERR(eventfp);
 751                         break;
 752                 }
 753                 if (eventfp != vq->kick) {
 754                         pollstop = (filep = vq->kick) != NULL;
 755                         pollstart = (vq->kick = eventfp) != NULL;
 756                 } else
 757                         filep = eventfp;
 758                 break;
 759         case VHOST_SET_VRING_CALL:
 760                 if (copy_from_user(&f, argp, sizeof f)) {
 761                         r = -EFAULT;
 762                         break;
 763                 }
 764                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
 765                 if (IS_ERR(eventfp)) {
 766                         r = PTR_ERR(eventfp);
 767                         break;
 768                 }
 769                 if (eventfp != vq->call) {
 770                         filep = vq->call;
 771                         ctx = vq->call_ctx;
 772                         vq->call = eventfp;
 773                         vq->call_ctx = eventfp ?
 774                                 eventfd_ctx_fileget(eventfp) : NULL;
 775                 } else
 776                         filep = eventfp;
 777                 break;
 778         case VHOST_SET_VRING_ERR:
 779                 if (copy_from_user(&f, argp, sizeof f)) {
 780                         r = -EFAULT;
 781                         break;
 782                 }
 783                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
 784                 if (IS_ERR(eventfp)) {
 785                         r = PTR_ERR(eventfp);
 786                         break;
 787                 }
 788                 if (eventfp != vq->error) {
 789                         filep = vq->error;
 790                         vq->error = eventfp;
 791                         ctx = vq->error_ctx;
 792                         vq->error_ctx = eventfp ?
 793                                 eventfd_ctx_fileget(eventfp) : NULL;
 794                 } else
 795                         filep = eventfp;
 796                 break;
 797         default:
 798                 r = -ENOIOCTLCMD;
 799         }
 800
 801         if (pollstop && vq->handle_kick)
 802                 vhost_poll_stop(&vq->poll);
 803
 804         if (ctx)
 805                 eventfd_ctx_put(ctx);
 806         if (filep)
 807                 fput(filep);
 808
 809         if (pollstart && vq->handle_kick)
 810                 r = vhost_poll_start(&vq->poll, vq->kick);
 811
 812         mutex_unlock(&vq->mutex);
 813
 814         if (pollstop && vq->handle_kick)
 815                 vhost_poll_flush(&vq->poll);
 816         return r;
 817 }
 818
 819 /* Caller must have device mutex */
 820 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
 821 {
 822         struct file *eventfp, *filep = NULL;
 823         struct eventfd_ctx *ctx = NULL;
 824         u64 p;
 825         long r;
 826         int i, fd;
 827
 828         /* If you are not the owner, you can become one */
 829         if (ioctl == VHOST_SET_OWNER) {
 830                 r = vhost_dev_set_owner(d);
 831                 goto done;
 832         }
 833
 834         /* You must be the owner to do anything else */
 835         r = vhost_dev_check_owner(d);
 836         if (r)
 837                 goto done;
 838
 839         switch (ioctl) {
 840         case VHOST_SET_MEM_TABLE:
 841                 r = vhost_set_memory(d, argp);
 842                 break;
 843         case VHOST_SET_LOG_BASE:
 844                 if (copy_from_user(&p, argp, sizeof p)) {
 845                         r = -EFAULT;
 846                         break;
 847                 }
 848                 if ((u64)(unsigned long)p != p) {
 849                         r = -EFAULT;
 850                         break;
 851                 }
 852                 for (i = 0; i < d->nvqs; ++i) {
 853                         struct vhost_virtqueue *vq;
 854                         void __user *base = (void __user *)(unsigned long)p;
 855                         vq = d->vqs[i];
 856                         mutex_lock(&vq->mutex);
 857                         /* If ring is inactive, will check when it's enabled. */
 858                         if (vq->private_data && !vq_log_access_ok(d, vq, base))
 859                                 r = -EFAULT;
 860                         else
 861                                 vq->log_base = base;
 862                         mutex_unlock(&vq->mutex);
 863                 }
 864                 break;
 865         case VHOST_SET_LOG_FD:
 866                 r = get_user(fd, (int __user *)argp);
 867                 if (r < 0)
 868                         break;
 869                 eventfp = fd == -1 ? NULL : eventfd_fget(fd);
 870                 if (IS_ERR(eventfp)) {
 871                         r = PTR_ERR(eventfp);
 872                         break;
 873                 }
 874                 if (eventfp != d->log_file) {
 875                         filep = d->log_file;
 876                         ctx = d->log_ctx;
 877                         d->log_ctx = eventfp ?
 878                                 eventfd_ctx_fileget(eventfp) : NULL;
 879                 } else
 880                         filep = eventfp;
 881                 for (i = 0; i < d->nvqs; ++i) {
 882                         mutex_lock(&d->vqs[i]->mutex);
 883                         d->vqs[i]->log_ctx = d->log_ctx;
 884                         mutex_unlock(&d->vqs[i]->mutex);
 885                 }
 886                 if (ctx)
 887                         eventfd_ctx_put(ctx);
 888                 if (filep)
 889                         fput(filep);
 890                 break;
 891         default:
 892                 r = -ENOIOCTLCMD;
 893                 break;
 894         }
 895 done:
 896         return r;
 897 }
 898
 899 static const struct vhost_memory_region *find_region(struct vhost_memory *mem,
 900                                                      __u64 addr, __u32 len)
 901 {
 902         struct vhost_memory_region *reg;
 903         int i;
 904
 905         /* linear search is not brilliant, but we really have on the order of 6
 906          * regions in practice */
 907         for (i = 0; i < mem->nregions; ++i) {
 908                 reg = mem->regions + i;
 909                 if (reg->guest_phys_addr <= addr &&
 910                     reg->guest_phys_addr + reg->memory_size - 1 >= addr)
 911                         return reg;
 912         }
 913         return NULL;
 914 }
 915
 916 /* TODO: This is really inefficient.  We need something like get_user()
 917  * (instruction directly accesses the data, with an exception table entry
 918  * returning -EFAULT). See Documentation/x86/exception-tables.txt.
 919  */
 920 static int set_bit_to_user(int nr, void __user *addr)
 921 {
 922         unsigned long log = (unsigned long)addr;
 923         struct page *page;
 924         void *base;
 925         int bit = nr + (log % PAGE_SIZE) * 8;
 926         int r;
 927
 928         r = get_user_pages_fast(log, 1, 1, &page);
 929         if (r < 0)
 930                 return r;
 931         BUG_ON(r != 1);
 932         base = kmap_atomic(page);
 933         set_bit(bit, base);
 934         kunmap_atomic(base);
 935         set_page_dirty_lock(page);
 936         put_page(page);
 937         return 0;
 938 }
 939
 940 static int log_write(void __user *log_base,
 941                      u64 write_address, u64 write_length)
 942 {
 943         u64 write_page = write_address / VHOST_PAGE_SIZE;
 944         int r;
 945
 946         if (!write_length)
 947                 return 0;
 948         write_length += write_address % VHOST_PAGE_SIZE;
 949         for (;;) {
 950                 u64 base = (u64)(unsigned long)log_base;
 951                 u64 log = base + write_page / 8;
 952                 int bit = write_page % 8;
 953                 if ((u64)(unsigned long)log != log)
 954                         return -EFAULT;
 955                 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
 956                 if (r < 0)
 957                         return r;
 958                 if (write_length <= VHOST_PAGE_SIZE)
 959                         break;
 960                 write_length -= VHOST_PAGE_SIZE;
 961                 write_page += 1;
 962         }
 963         return r;
 964 }
 965
 966 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
 967                     unsigned int log_num, u64 len)
 968 {
 969         int i, r;
 970
 971         /* Make sure data written is seen before log. */
 972         smp_wmb();
 973         for (i = 0; i < log_num; ++i) {
 974                 u64 l = min(log[i].len, len);
 975                 r = log_write(vq->log_base, log[i].addr, l);
 976                 if (r < 0)
 977                         return r;
 978                 len -= l;
 979                 if (!len) {
 980                         if (vq->log_ctx)
 981                                 eventfd_signal(vq->log_ctx, 1);
 982                         return 0;
 983                 }
 984         }
 985         /* Length written exceeds what we have stored. This is a bug. */
 986         BUG();
 987         return 0;
 988 }
 989
 990 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
 991 {
 992         void __user *used;
 993         if (__put_user(vq->used_flags, &vq->used->flags) < 0)
 994                 return -EFAULT;
 995         if (unlikely(vq->log_used)) {
 996                 /* Make sure the flag is seen before log. */
 997                 smp_wmb();
 998                 /* Log used flag write. */
 999                 used = &vq->used->flags;
1000                 log_write(vq->log_base, vq->log_addr +
1001                           (used - (void __user *)vq->used),
1002                           sizeof vq->used->flags);
1003                 if (vq->log_ctx)
1004                         eventfd_signal(vq->log_ctx, 1);
1005         }
1006         return 0;
1007 }
1008
1009 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1010 {
1011         if (__put_user(vq->avail_idx, vhost_avail_event(vq)))
1012                 return -EFAULT;
1013         if (unlikely(vq->log_used)) {
1014                 void __user *used;
1015                 /* Make sure the event is seen before log. */
1016                 smp_wmb();
1017                 /* Log avail event write */
1018                 used = vhost_avail_event(vq);
1019                 log_write(vq->log_base, vq->log_addr +
1020                           (used - (void __user *)vq->used),
1021                           sizeof *vhost_avail_event(vq));
1022                 if (vq->log_ctx)
1023                         eventfd_signal(vq->log_ctx, 1);
1024         }
1025         return 0;
1026 }
1027
1028 int vhost_init_used(struct vhost_virtqueue *vq)
1029 {
1030         int r;
1031         if (!vq->private_data)
1032                 return 0;
1033
1034         r = vhost_update_used_flags(vq);
1035         if (r)
1036                 return r;
1037         vq->signalled_used_valid = false;
1038         return get_user(vq->last_used_idx, &vq->used->idx);
1039 }
1040
1041 static int translate_desc(struct vhost_dev *dev, u64 addr, u32 len,
1042                           struct iovec iov[], int iov_size)
1043 {
1044         const struct vhost_memory_region *reg;
1045         struct vhost_memory *mem;
1046         struct iovec *_iov;
1047         u64 s = 0;
1048         int ret = 0;
1049
1050         rcu_read_lock();
1051
1052         mem = rcu_dereference(dev->memory);
1053         while ((u64)len > s) {
1054                 u64 size;
1055                 if (unlikely(ret >= iov_size)) {
1056                         ret = -ENOBUFS;
1057                         break;
1058                 }
1059                 reg = find_region(mem, addr, len);
1060                 if (unlikely(!reg)) {
1061                         ret = -EFAULT;
1062                         break;
1063                 }
1064                 _iov = iov + ret;
1065                 size = reg->memory_size - addr + reg->guest_phys_addr;
1066                 _iov->iov_len = min((u64)len - s, size);
1067                 _iov->iov_base = (void __user *)(unsigned long)
1068                         (reg->userspace_addr + addr - reg->guest_phys_addr);
1069                 s += size;
1070                 addr += size;
1071                 ++ret;
1072         }
1073
1074         rcu_read_unlock();
1075         return ret;
1076 }
1077
1078 /* Each buffer in the virtqueues is actually a chain of descriptors.  This
1079  * function returns the next descriptor in the chain,
1080  * or -1U if we're at the end. */
1081 static unsigned next_desc(struct vring_desc *desc)
1082 {
1083         unsigned int next;
1084
1085         /* If this descriptor says it doesn't chain, we're done. */
1086         if (!(desc->flags & VRING_DESC_F_NEXT))
1087                 return -1U;
1088
1089         /* Check they're not leading us off end of descriptors. */
1090         next = desc->next;
1091         /* Make sure compiler knows to grab that: we don't want it changing! */
1092         /* We will use the result as an index in an array, so most
1093          * architectures only need a compiler barrier here. */
1094         read_barrier_depends();
1095
1096         return next;
1097 }
1098
1099 static int get_indirect(struct vhost_dev *dev, struct vhost_virtqueue *vq,
1100                         struct iovec iov[], unsigned int iov_size,
1101                         unsigned int *out_num, unsigned int *in_num,
1102                         struct vhost_log *log, unsigned int *log_num,
1103                         struct vring_desc *indirect)
1104 {
1105         struct vring_desc desc;
1106         unsigned int i = 0, count, found = 0;
1107         int ret;
1108
1109         /* Sanity check */
1110         if (unlikely(indirect->len % sizeof desc)) {
1111                 vq_err(vq, "Invalid length in indirect descriptor: "
1112                        "len 0x%llx not multiple of 0x%zx\n",
1113                        (unsigned long long)indirect->len,
1114                        sizeof desc);
1115                 return -EINVAL;
1116         }
1117
1118         ret = translate_desc(dev, indirect->addr, indirect->len, vq->indirect,
1119                              UIO_MAXIOV);
1120         if (unlikely(ret < 0)) {
1121                 vq_err(vq, "Translation failure %d in indirect.\n", ret);
1122                 return ret;
1123         }
1124
1125         /* We will use the result as an address to read from, so most
1126          * architectures only need a compiler barrier here. */
1127         read_barrier_depends();
1128
1129         count = indirect->len / sizeof desc;
1130         /* Buffers are chained via a 16 bit next field, so
1131          * we can have at most 2^16 of these. */
1132         if (unlikely(count > USHRT_MAX + 1)) {
1133                 vq_err(vq, "Indirect buffer length too big: %d\n",
1134                        indirect->len);
1135                 return -E2BIG;
1136         }
1137
1138         do {
1139                 unsigned iov_count = *in_num + *out_num;
1140                 if (unlikely(++found > count)) {
1141                         vq_err(vq, "Loop detected: last one at %u "
1142                                "indirect size %u\n",
1143                                i, count);
1144                         return -EINVAL;
1145                 }
1146                 if (unlikely(memcpy_fromiovec((unsigned char *)&desc,
1147                                               vq->indirect, sizeof desc))) {
1148                         vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1149                                i, (size_t)indirect->addr + i * sizeof desc);
1150                         return -EINVAL;
1151                 }
1152                 if (unlikely(desc.flags & VRING_DESC_F_INDIRECT)) {
1153                         vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1154                                i, (size_t)indirect->addr + i * sizeof desc);
1155                         return -EINVAL;
1156                 }
1157
1158                 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
1159                                      iov_size - iov_count);
1160                 if (unlikely(ret < 0)) {
1161                         vq_err(vq, "Translation failure %d indirect idx %d\n",
1162                                ret, i);
1163                         return ret;
1164                 }
1165                 /* If this is an input descriptor, increment that count. */
1166                 if (desc.flags & VRING_DESC_F_WRITE) {
1167                         *in_num += ret;
1168                         if (unlikely(log)) {
1169                                 log[*log_num].addr = desc.addr;
1170                                 log[*log_num].len = desc.len;
1171                                 ++*log_num;
1172                         }
1173                 } else {
1174                         /* If it's an output descriptor, they're all supposed
1175                          * to come before any input descriptors. */
1176                         if (unlikely(*in_num)) {
1177                                 vq_err(vq, "Indirect descriptor "
1178                                        "has out after in: idx %d\n", i);
1179                                 return -EINVAL;
1180                         }
1181                         *out_num += ret;
1182                 }
1183         } while ((i = next_desc(&desc)) != -1);
1184         return 0;
1185 }
1186
1187 /* This looks in the virtqueue and for the first available buffer, and converts
1188  * it to an iovec for convenient access.  Since descriptors consist of some
1189  * number of output then some number of input descriptors, it's actually two
1190  * iovecs, but we pack them into one and note how many of each there were.
1191  *
1192  * This function returns the descriptor number found, or vq->num (which is
1193  * never a valid descriptor number) if none was found.  A negative code is
1194  * returned on error. */
1195 int vhost_get_vq_desc(struct vhost_dev *dev, struct vhost_virtqueue *vq,
1196                       struct iovec iov[], unsigned int iov_size,
1197                       unsigned int *out_num, unsigned int *in_num,
1198                       struct vhost_log *log, unsigned int *log_num)
1199 {
1200         struct vring_desc desc;
1201         unsigned int i, head, found = 0;
1202         u16 last_avail_idx;
1203         int ret;
1204
1205         /* Check it isn't doing very strange things with descriptor numbers. */
1206         last_avail_idx = vq->last_avail_idx;
1207         if (unlikely(__get_user(vq->avail_idx, &vq->avail->idx))) {
1208                 vq_err(vq, "Failed to access avail idx at %p\n",
1209                        &vq->avail->idx);
1210                 return -EFAULT;
1211         }
1212
1213         if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
1214                 vq_err(vq, "Guest moved used index from %u to %u",
1215                        last_avail_idx, vq->avail_idx);
1216                 return -EFAULT;
1217         }
1218
1219         /* If there's nothing new since last we looked, return invalid. */
1220         if (vq->avail_idx == last_avail_idx)
1221                 return vq->num;
1222
1223         /* Only get avail ring entries after they have been exposed by guest. */
1224         smp_rmb();
1225
1226         /* Grab the next descriptor number they're advertising, and increment
1227          * the index we've seen. */
1228         if (unlikely(__get_user(head,
1229                                 &vq->avail->ring[last_avail_idx % vq->num]))) {
1230                 vq_err(vq, "Failed to read head: idx %d address %p\n",
1231                        last_avail_idx,
1232                        &vq->avail->ring[last_avail_idx % vq->num]);
1233                 return -EFAULT;
1234         }
1235
1236         /* If their number is silly, that's an error. */
1237         if (unlikely(head >= vq->num)) {
1238                 vq_err(vq, "Guest says index %u > %u is available",
1239                        head, vq->num);
1240                 return -EINVAL;
1241         }
1242
1243         /* When we start there are none of either input nor output. */
1244         *out_num = *in_num = 0;
1245         if (unlikely(log))
1246                 *log_num = 0;
1247
1248         i = head;
1249         do {
1250                 unsigned iov_count = *in_num + *out_num;
1251                 if (unlikely(i >= vq->num)) {
1252                         vq_err(vq, "Desc index is %u > %u, head = %u",
1253                                i, vq->num, head);
1254                         return -EINVAL;
1255                 }
1256                 if (unlikely(++found > vq->num)) {
1257                         vq_err(vq, "Loop detected: last one at %u "
1258                                "vq size %u head %u\n",
1259                                i, vq->num, head);
1260                         return -EINVAL;
1261                 }
1262                 ret = __copy_from_user(&desc, vq->desc + i, sizeof desc);
1263                 if (unlikely(ret)) {
1264                         vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
1265                                i, vq->desc + i);
1266                         return -EFAULT;
1267                 }
1268                 if (desc.flags & VRING_DESC_F_INDIRECT) {
1269                         ret = get_indirect(dev, vq, iov, iov_size,
1270                                            out_num, in_num,
1271                                            log, log_num, &desc);
1272                         if (unlikely(ret < 0)) {
1273                                 vq_err(vq, "Failure detected "
1274                                        "in indirect descriptor at idx %d\n", i);
1275                                 return ret;
1276                         }
1277                         continue;
1278                 }
1279
1280                 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
1281                                      iov_size - iov_count);
1282                 if (unlikely(ret < 0)) {
1283                         vq_err(vq, "Translation failure %d descriptor idx %d\n",
1284                                ret, i);
1285                         return ret;
1286                 }
1287                 if (desc.flags & VRING_DESC_F_WRITE) {
1288                         /* If this is an input descriptor,
1289                          * increment that count. */
1290                         *in_num += ret;
1291                         if (unlikely(log)) {
1292                                 log[*log_num].addr = desc.addr;
1293                                 log[*log_num].len = desc.len;
1294                                 ++*log_num;
1295                         }
1296                 } else {
1297                         /* If it's an output descriptor, they're all supposed
1298                          * to come before any input descriptors. */
1299                         if (unlikely(*in_num)) {
1300                                 vq_err(vq, "Descriptor has out after in: "
1301                                        "idx %d\n", i);
1302                                 return -EINVAL;
1303                         }
1304                         *out_num += ret;
1305                 }
1306         } while ((i = next_desc(&desc)) != -1);
1307
1308         /* On success, increment avail index. */
1309         vq->last_avail_idx++;
1310
1311         /* Assume notifications from guest are disabled at this point,
1312          * if they aren't we would need to update avail_event index. */
1313         BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
1314         return head;
1315 }
1316
1317 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
1318 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
1319 {
1320         vq->last_avail_idx -= n;
1321 }
1322
1323 /* After we've used one of their buffers, we tell them about it.  We'll then
1324  * want to notify the guest, using eventfd. */
1325 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
1326 {
1327         struct vring_used_elem __user *used;
1328
1329         /* The virtqueue contains a ring of used buffers.  Get a pointer to the
1330          * next entry in that used ring. */
1331         used = &vq->used->ring[vq->last_used_idx % vq->num];
1332         if (__put_user(head, &used->id)) {
1333                 vq_err(vq, "Failed to write used id");
1334                 return -EFAULT;
1335         }
1336         if (__put_user(len, &used->len)) {
1337                 vq_err(vq, "Failed to write used len");
1338                 return -EFAULT;
1339         }
1340         /* Make sure buffer is written before we update index. */
1341         smp_wmb();
1342         if (__put_user(vq->last_used_idx + 1, &vq->used->idx)) {
1343                 vq_err(vq, "Failed to increment used idx");
1344                 return -EFAULT;
1345         }
1346         if (unlikely(vq->log_used)) {
1347                 /* Make sure data is seen before log. */
1348                 smp_wmb();
1349                 /* Log used ring entry write. */
1350                 log_write(vq->log_base,
1351                           vq->log_addr +
1352                            ((void __user *)used - (void __user *)vq->used),
1353                           sizeof *used);
1354                 /* Log used index update. */
1355                 log_write(vq->log_base,
1356                           vq->log_addr + offsetof(struct vring_used, idx),
1357                           sizeof vq->used->idx);
1358                 if (vq->log_ctx)
1359                         eventfd_signal(vq->log_ctx, 1);
1360         }
1361         vq->last_used_idx++;
1362         /* If the driver never bothers to signal in a very long while,
1363          * used index might wrap around. If that happens, invalidate
1364          * signalled_used index we stored. TODO: make sure driver
1365          * signals at least once in 2^16 and remove this. */
1366         if (unlikely(vq->last_used_idx == vq->signalled_used))
1367                 vq->signalled_used_valid = false;
1368         return 0;
1369 }
1370
1371 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
1372                             struct vring_used_elem *heads,
1373                             unsigned count)
1374 {
1375         struct vring_used_elem __user *used;
1376         u16 old, new;
1377         int start;
1378
1379         start = vq->last_used_idx % vq->num;
1380         used = vq->used->ring + start;
1381         if (__copy_to_user(used, heads, count * sizeof *used)) {
1382                 vq_err(vq, "Failed to write used");
1383                 return -EFAULT;
1384         }
1385         if (unlikely(vq->log_used)) {
1386                 /* Make sure data is seen before log. */
1387                 smp_wmb();
1388                 /* Log used ring entry write. */
1389                 log_write(vq->log_base,
1390                           vq->log_addr +
1391                            ((void __user *)used - (void __user *)vq->used),
1392                           count * sizeof *used);
1393         }
1394         old = vq->last_used_idx;
1395         new = (vq->last_used_idx += count);
1396         /* If the driver never bothers to signal in a very long while,
1397          * used index might wrap around. If that happens, invalidate
1398          * signalled_used index we stored. TODO: make sure driver
1399          * signals at least once in 2^16 and remove this. */
1400         if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
1401                 vq->signalled_used_valid = false;
1402         return 0;
1403 }
1404
1405 /* After we've used one of their buffers, we tell them about it.  We'll then
1406  * want to notify the guest, using eventfd. */
1407 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
1408                      unsigned count)
1409 {
1410         int start, n, r;
1411
1412         start = vq->last_used_idx % vq->num;
1413         n = vq->num - start;
1414         if (n < count) {
1415                 r = __vhost_add_used_n(vq, heads, n);
1416                 if (r < 0)
1417                         return r;
1418                 heads += n;
1419                 count -= n;
1420         }
1421         r = __vhost_add_used_n(vq, heads, count);
1422
1423         /* Make sure buffer is written before we update index. */
1424         smp_wmb();
1425         if (put_user(vq->last_used_idx, &vq->used->idx)) {
1426                 vq_err(vq, "Failed to increment used idx");
1427                 return -EFAULT;
1428         }
1429         if (unlikely(vq->log_used)) {
1430                 /* Log used index update. */
1431                 log_write(vq->log_base,
1432                           vq->log_addr + offsetof(struct vring_used, idx),
1433                           sizeof vq->used->idx);
1434                 if (vq->log_ctx)
1435                         eventfd_signal(vq->log_ctx, 1);
1436         }
1437         return r;
1438 }
1439
1440 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1441 {
1442         __u16 old, new, event;
1443         bool v;
1444         /* Flush out used index updates. This is paired
1445          * with the barrier that the Guest executes when enabling
1446          * interrupts. */
1447         smp_mb();
1448
1449         if (vhost_has_feature(dev, VIRTIO_F_NOTIFY_ON_EMPTY) &&
1450             unlikely(vq->avail_idx == vq->last_avail_idx))
1451                 return true;
1452
1453         if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1454                 __u16 flags;
1455                 if (__get_user(flags, &vq->avail->flags)) {
1456                         vq_err(vq, "Failed to get flags");
1457                         return true;
1458                 }
1459                 return !(flags & VRING_AVAIL_F_NO_INTERRUPT);
1460         }
1461         old = vq->signalled_used;
1462         v = vq->signalled_used_valid;
1463         new = vq->signalled_used = vq->last_used_idx;
1464         vq->signalled_used_valid = true;
1465
1466         if (unlikely(!v))
1467                 return true;
1468
1469         if (get_user(event, vhost_used_event(vq))) {
1470                 vq_err(vq, "Failed to get used event idx");
1471                 return true;
1472         }
1473         return vring_need_event(event, new, old);
1474 }
1475
1476 /* This actually signals the guest, using eventfd. */
1477 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1478 {
1479         /* Signal the Guest tell them we used something up. */
1480         if (vq->call_ctx && vhost_notify(dev, vq))
1481                 eventfd_signal(vq->call_ctx, 1);
1482 }
1483
1484 /* And here's the combo meal deal.  Supersize me! */
1485 void vhost_add_used_and_signal(struct vhost_dev *dev,
1486                                struct vhost_virtqueue *vq,
1487                                unsigned int head, int len)
1488 {
1489         vhost_add_used(vq, head, len);
1490         vhost_signal(dev, vq);
1491 }
1492
1493 /* multi-buffer version of vhost_add_used_and_signal */
1494 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
1495                                  struct vhost_virtqueue *vq,
1496                                  struct vring_used_elem *heads, unsigned count)
1497 {
1498         vhost_add_used_n(vq, heads, count);
1499         vhost_signal(dev, vq);
1500 }
1501
1502 /* OK, now we need to know about added descriptors. */
1503 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1504 {
1505         u16 avail_idx;
1506         int r;
1507
1508         if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
1509                 return false;
1510         vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
1511         if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1512                 r = vhost_update_used_flags(vq);
1513                 if (r) {
1514                         vq_err(vq, "Failed to enable notification at %p: %d\n",
1515                                &vq->used->flags, r);
1516                         return false;
1517                 }
1518         } else {
1519                 r = vhost_update_avail_event(vq, vq->avail_idx);
1520                 if (r) {
1521                         vq_err(vq, "Failed to update avail event index at %p: %d\n",
1522                                vhost_avail_event(vq), r);
1523                         return false;
1524                 }
1525         }
1526         /* They could have slipped one in as we were doing that: make
1527          * sure it's written, then check again. */
1528         smp_mb();
1529         r = __get_user(avail_idx, &vq->avail->idx);
1530         if (r) {
1531                 vq_err(vq, "Failed to check avail idx at %p: %d\n",
1532                        &vq->avail->idx, r);
1533                 return false;
1534         }
1535
1536         return avail_idx != vq->avail_idx;
1537 }
1538
1539 /* We don't need to be notified again. */
1540 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1541 {
1542         int r;
1543
1544         if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
1545                 return;
1546         vq->used_flags |= VRING_USED_F_NO_NOTIFY;
1547         if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1548                 r = vhost_update_used_flags(vq);
1549                 if (r)
1550                         vq_err(vq, "Failed to enable notification at %p: %d\n",
1551                                &vq->used->flags, r);
1552         }
1553 }
1554
1555 static void vhost_zerocopy_done_signal(struct kref *kref)
1556 {
1557         struct vhost_ubuf_ref *ubufs = container_of(kref, struct vhost_ubuf_ref,
1558                                                     kref);
1559         wake_up(&ubufs->wait);
1560 }
1561
1562 struct vhost_ubuf_ref *vhost_ubuf_alloc(struct vhost_virtqueue *vq,
1563                                         bool zcopy)
1564 {
1565         struct vhost_ubuf_ref *ubufs;
1566         /* No zero copy backend? Nothing to count. */
1567         if (!zcopy)
1568                 return NULL;
1569         ubufs = kmalloc(sizeof *ubufs, GFP_KERNEL);
1570         if (!ubufs)
1571                 return ERR_PTR(-ENOMEM);
1572         kref_init(&ubufs->kref);
1573         init_waitqueue_head(&ubufs->wait);
1574         ubufs->vq = vq;
1575         return ubufs;
1576 }
1577
1578 void vhost_ubuf_put(struct vhost_ubuf_ref *ubufs)
1579 {
1580         kref_put(&ubufs->kref, vhost_zerocopy_done_signal);
1581 }
1582
1583 void vhost_ubuf_put_and_wait(struct vhost_ubuf_ref *ubufs)
1584 {
1585         kref_put(&ubufs->kref, vhost_zerocopy_done_signal);
1586         wait_event(ubufs->wait, !atomic_read(&ubufs->kref.refcount));
1587         kfree(ubufs);
1588 }