drivers/vhost/vhost.c: delete double assignment
[linux-2.6.git] / 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/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/miscdevice.h>
19 #include <linux/mutex.h>
20 #include <linux/rcupdate.h>
21 #include <linux/poll.h>
22 #include <linux/file.h>
23 #include <linux/highmem.h>
24 #include <linux/slab.h>
25 #include <linux/kthread.h>
26 #include <linux/cgroup.h>
27
28 #include <linux/net.h>
29 #include <linux/if_packet.h>
30 #include <linux/if_arp.h>
31
32 #include <net/sock.h>
33
34 #include "vhost.h"
35
36 enum {
37         VHOST_MEMORY_MAX_NREGIONS = 64,
38         VHOST_MEMORY_F_LOG = 0x1,
39 };
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         poll = container_of(pt, struct vhost_poll, table);
46
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 static 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
81         vhost_work_init(&poll->work, fn);
82 }
83
84 /* Start polling a file. We add ourselves to file's wait queue. The caller must
85  * keep a reference to a file until after vhost_poll_stop is called. */
86 void vhost_poll_start(struct vhost_poll *poll, struct file *file)
87 {
88         unsigned long mask;
89         mask = file->f_op->poll(file, &poll->table);
90         if (mask)
91                 vhost_poll_wakeup(&poll->wait, 0, 0, (void *)mask);
92 }
93
94 /* Stop polling a file. After this function returns, it becomes safe to drop the
95  * file reference. You must also flush afterwards. */
96 void vhost_poll_stop(struct vhost_poll *poll)
97 {
98         remove_wait_queue(poll->wqh, &poll->wait);
99 }
100
101 static void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work)
102 {
103         unsigned seq;
104         int left;
105         int flushing;
106
107         spin_lock_irq(&dev->work_lock);
108         seq = work->queue_seq;
109         work->flushing++;
110         spin_unlock_irq(&dev->work_lock);
111         wait_event(work->done, ({
112                    spin_lock_irq(&dev->work_lock);
113                    left = seq - work->done_seq <= 0;
114                    spin_unlock_irq(&dev->work_lock);
115                    left;
116         }));
117         spin_lock_irq(&dev->work_lock);
118         flushing = --work->flushing;
119         spin_unlock_irq(&dev->work_lock);
120         BUG_ON(flushing < 0);
121 }
122
123 /* Flush any work that has been scheduled. When calling this, don't hold any
124  * locks that are also used by the callback. */
125 void vhost_poll_flush(struct vhost_poll *poll)
126 {
127         vhost_work_flush(poll->dev, &poll->work);
128 }
129
130 static inline void vhost_work_queue(struct vhost_dev *dev,
131                                     struct vhost_work *work)
132 {
133         unsigned long flags;
134
135         spin_lock_irqsave(&dev->work_lock, flags);
136         if (list_empty(&work->node)) {
137                 list_add_tail(&work->node, &dev->work_list);
138                 work->queue_seq++;
139                 wake_up_process(dev->worker);
140         }
141         spin_unlock_irqrestore(&dev->work_lock, flags);
142 }
143
144 void vhost_poll_queue(struct vhost_poll *poll)
145 {
146         vhost_work_queue(poll->dev, &poll->work);
147 }
148
149 static void vhost_vq_reset(struct vhost_dev *dev,
150                            struct vhost_virtqueue *vq)
151 {
152         vq->num = 1;
153         vq->desc = NULL;
154         vq->avail = NULL;
155         vq->used = NULL;
156         vq->last_avail_idx = 0;
157         vq->avail_idx = 0;
158         vq->last_used_idx = 0;
159         vq->used_flags = 0;
160         vq->log_used = false;
161         vq->log_addr = -1ull;
162         vq->vhost_hlen = 0;
163         vq->sock_hlen = 0;
164         vq->private_data = NULL;
165         vq->log_base = NULL;
166         vq->error_ctx = NULL;
167         vq->error = NULL;
168         vq->kick = NULL;
169         vq->call_ctx = NULL;
170         vq->call = NULL;
171         vq->log_ctx = NULL;
172 }
173
174 static int vhost_worker(void *data)
175 {
176         struct vhost_dev *dev = data;
177         struct vhost_work *work = NULL;
178         unsigned uninitialized_var(seq);
179
180         for (;;) {
181                 /* mb paired w/ kthread_stop */
182                 set_current_state(TASK_INTERRUPTIBLE);
183
184                 spin_lock_irq(&dev->work_lock);
185                 if (work) {
186                         work->done_seq = seq;
187                         if (work->flushing)
188                                 wake_up_all(&work->done);
189                 }
190
191                 if (kthread_should_stop()) {
192                         spin_unlock_irq(&dev->work_lock);
193                         __set_current_state(TASK_RUNNING);
194                         return 0;
195                 }
196                 if (!list_empty(&dev->work_list)) {
197                         work = list_first_entry(&dev->work_list,
198                                                 struct vhost_work, node);
199                         list_del_init(&work->node);
200                         seq = work->queue_seq;
201                 } else
202                         work = NULL;
203                 spin_unlock_irq(&dev->work_lock);
204
205                 if (work) {
206                         __set_current_state(TASK_RUNNING);
207                         work->fn(work);
208                 } else
209                         schedule();
210
211         }
212 }
213
214 /* Helper to allocate iovec buffers for all vqs. */
215 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
216 {
217         int i;
218         for (i = 0; i < dev->nvqs; ++i) {
219                 dev->vqs[i].indirect = kmalloc(sizeof *dev->vqs[i].indirect *
220                                                UIO_MAXIOV, GFP_KERNEL);
221                 dev->vqs[i].log = kmalloc(sizeof *dev->vqs[i].log * UIO_MAXIOV,
222                                           GFP_KERNEL);
223                 dev->vqs[i].heads = kmalloc(sizeof *dev->vqs[i].heads *
224                                             UIO_MAXIOV, GFP_KERNEL);
225
226                 if (!dev->vqs[i].indirect || !dev->vqs[i].log ||
227                         !dev->vqs[i].heads)
228                         goto err_nomem;
229         }
230         return 0;
231 err_nomem:
232         for (; i >= 0; --i) {
233                 kfree(dev->vqs[i].indirect);
234                 kfree(dev->vqs[i].log);
235                 kfree(dev->vqs[i].heads);
236         }
237         return -ENOMEM;
238 }
239
240 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
241 {
242         int i;
243         for (i = 0; i < dev->nvqs; ++i) {
244                 kfree(dev->vqs[i].indirect);
245                 dev->vqs[i].indirect = NULL;
246                 kfree(dev->vqs[i].log);
247                 dev->vqs[i].log = NULL;
248                 kfree(dev->vqs[i].heads);
249                 dev->vqs[i].heads = NULL;
250         }
251 }
252
253 long vhost_dev_init(struct vhost_dev *dev,
254                     struct vhost_virtqueue *vqs, int nvqs)
255 {
256         int i;
257
258         dev->vqs = vqs;
259         dev->nvqs = nvqs;
260         mutex_init(&dev->mutex);
261         dev->log_ctx = NULL;
262         dev->log_file = NULL;
263         dev->memory = NULL;
264         dev->mm = NULL;
265         spin_lock_init(&dev->work_lock);
266         INIT_LIST_HEAD(&dev->work_list);
267         dev->worker = NULL;
268
269         for (i = 0; i < dev->nvqs; ++i) {
270                 dev->vqs[i].log = NULL;
271                 dev->vqs[i].indirect = NULL;
272                 dev->vqs[i].heads = NULL;
273                 dev->vqs[i].dev = dev;
274                 mutex_init(&dev->vqs[i].mutex);
275                 vhost_vq_reset(dev, dev->vqs + i);
276                 if (dev->vqs[i].handle_kick)
277                         vhost_poll_init(&dev->vqs[i].poll,
278                                         dev->vqs[i].handle_kick, POLLIN, dev);
279         }
280
281         return 0;
282 }
283
284 /* Caller should have device mutex */
285 long vhost_dev_check_owner(struct vhost_dev *dev)
286 {
287         /* Are you the owner? If not, I don't think you mean to do that */
288         return dev->mm == current->mm ? 0 : -EPERM;
289 }
290
291 struct vhost_attach_cgroups_struct {
292         struct vhost_work work;
293         struct task_struct *owner;
294         int ret;
295 };
296
297 static void vhost_attach_cgroups_work(struct vhost_work *work)
298 {
299         struct vhost_attach_cgroups_struct *s;
300         s = container_of(work, struct vhost_attach_cgroups_struct, work);
301         s->ret = cgroup_attach_task_all(s->owner, current);
302 }
303
304 static int vhost_attach_cgroups(struct vhost_dev *dev)
305 {
306         struct vhost_attach_cgroups_struct attach;
307         attach.owner = current;
308         vhost_work_init(&attach.work, vhost_attach_cgroups_work);
309         vhost_work_queue(dev, &attach.work);
310         vhost_work_flush(dev, &attach.work);
311         return attach.ret;
312 }
313
314 /* Caller should have device mutex */
315 static long vhost_dev_set_owner(struct vhost_dev *dev)
316 {
317         struct task_struct *worker;
318         int err;
319         /* Is there an owner already? */
320         if (dev->mm) {
321                 err = -EBUSY;
322                 goto err_mm;
323         }
324         /* No owner, become one */
325         dev->mm = get_task_mm(current);
326         worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
327         if (IS_ERR(worker)) {
328                 err = PTR_ERR(worker);
329                 goto err_worker;
330         }
331
332         dev->worker = worker;
333         wake_up_process(worker);        /* avoid contributing to loadavg */
334
335         err = vhost_attach_cgroups(dev);
336         if (err)
337                 goto err_cgroup;
338
339         err = vhost_dev_alloc_iovecs(dev);
340         if (err)
341                 goto err_cgroup;
342
343         return 0;
344 err_cgroup:
345         kthread_stop(worker);
346         dev->worker = NULL;
347 err_worker:
348         if (dev->mm)
349                 mmput(dev->mm);
350         dev->mm = NULL;
351 err_mm:
352         return err;
353 }
354
355 /* Caller should have device mutex */
356 long vhost_dev_reset_owner(struct vhost_dev *dev)
357 {
358         struct vhost_memory *memory;
359
360         /* Restore memory to default empty mapping. */
361         memory = kmalloc(offsetof(struct vhost_memory, regions), GFP_KERNEL);
362         if (!memory)
363                 return -ENOMEM;
364
365         vhost_dev_cleanup(dev);
366
367         memory->nregions = 0;
368         RCU_INIT_POINTER(dev->memory, memory);
369         return 0;
370 }
371
372 /* Caller should have device mutex */
373 void vhost_dev_cleanup(struct vhost_dev *dev)
374 {
375         int i;
376         for (i = 0; i < dev->nvqs; ++i) {
377                 if (dev->vqs[i].kick && dev->vqs[i].handle_kick) {
378                         vhost_poll_stop(&dev->vqs[i].poll);
379                         vhost_poll_flush(&dev->vqs[i].poll);
380                 }
381                 if (dev->vqs[i].error_ctx)
382                         eventfd_ctx_put(dev->vqs[i].error_ctx);
383                 if (dev->vqs[i].error)
384                         fput(dev->vqs[i].error);
385                 if (dev->vqs[i].kick)
386                         fput(dev->vqs[i].kick);
387                 if (dev->vqs[i].call_ctx)
388                         eventfd_ctx_put(dev->vqs[i].call_ctx);
389                 if (dev->vqs[i].call)
390                         fput(dev->vqs[i].call);
391                 vhost_vq_reset(dev, dev->vqs + i);
392         }
393         vhost_dev_free_iovecs(dev);
394         if (dev->log_ctx)
395                 eventfd_ctx_put(dev->log_ctx);
396         dev->log_ctx = NULL;
397         if (dev->log_file)
398                 fput(dev->log_file);
399         dev->log_file = NULL;
400         /* No one will access memory at this point */
401         kfree(rcu_dereference_protected(dev->memory,
402                                         lockdep_is_held(&dev->mutex)));
403         RCU_INIT_POINTER(dev->memory, NULL);
404         if (dev->mm)
405                 mmput(dev->mm);
406         dev->mm = NULL;
407
408         WARN_ON(!list_empty(&dev->work_list));
409         if (dev->worker) {
410                 kthread_stop(dev->worker);
411                 dev->worker = NULL;
412         }
413 }
414
415 static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
416 {
417         u64 a = addr / VHOST_PAGE_SIZE / 8;
418         /* Make sure 64 bit math will not overflow. */
419         if (a > ULONG_MAX - (unsigned long)log_base ||
420             a + (unsigned long)log_base > ULONG_MAX)
421                 return 0;
422
423         return access_ok(VERIFY_WRITE, log_base + a,
424                          (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
425 }
426
427 /* Caller should have vq mutex and device mutex. */
428 static int vq_memory_access_ok(void __user *log_base, struct vhost_memory *mem,
429                                int log_all)
430 {
431         int i;
432
433         if (!mem)
434                 return 0;
435
436         for (i = 0; i < mem->nregions; ++i) {
437                 struct vhost_memory_region *m = mem->regions + i;
438                 unsigned long a = m->userspace_addr;
439                 if (m->memory_size > ULONG_MAX)
440                         return 0;
441                 else if (!access_ok(VERIFY_WRITE, (void __user *)a,
442                                     m->memory_size))
443                         return 0;
444                 else if (log_all && !log_access_ok(log_base,
445                                                    m->guest_phys_addr,
446                                                    m->memory_size))
447                         return 0;
448         }
449         return 1;
450 }
451
452 /* Can we switch to this memory table? */
453 /* Caller should have device mutex but not vq mutex */
454 static int memory_access_ok(struct vhost_dev *d, struct vhost_memory *mem,
455                             int log_all)
456 {
457         int i;
458         for (i = 0; i < d->nvqs; ++i) {
459                 int ok;
460                 mutex_lock(&d->vqs[i].mutex);
461                 /* If ring is inactive, will check when it's enabled. */
462                 if (d->vqs[i].private_data)
463                         ok = vq_memory_access_ok(d->vqs[i].log_base, mem,
464                                                  log_all);
465                 else
466                         ok = 1;
467                 mutex_unlock(&d->vqs[i].mutex);
468                 if (!ok)
469                         return 0;
470         }
471         return 1;
472 }
473
474 static int vq_access_ok(unsigned int num,
475                         struct vring_desc __user *desc,
476                         struct vring_avail __user *avail,
477                         struct vring_used __user *used)
478 {
479         return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
480                access_ok(VERIFY_READ, avail,
481                          sizeof *avail + num * sizeof *avail->ring) &&
482                access_ok(VERIFY_WRITE, used,
483                         sizeof *used + num * sizeof *used->ring);
484 }
485
486 /* Can we log writes? */
487 /* Caller should have device mutex but not vq mutex */
488 int vhost_log_access_ok(struct vhost_dev *dev)
489 {
490         struct vhost_memory *mp;
491
492         mp = rcu_dereference_protected(dev->memory,
493                                        lockdep_is_held(&dev->mutex));
494         return memory_access_ok(dev, mp, 1);
495 }
496
497 /* Verify access for write logging. */
498 /* Caller should have vq mutex and device mutex */
499 static int vq_log_access_ok(struct vhost_virtqueue *vq, void __user *log_base)
500 {
501         struct vhost_memory *mp;
502
503         mp = rcu_dereference_protected(vq->dev->memory,
504                                        lockdep_is_held(&vq->mutex));
505         return vq_memory_access_ok(log_base, mp,
506                             vhost_has_feature(vq->dev, VHOST_F_LOG_ALL)) &&
507                 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
508                                         sizeof *vq->used +
509                                         vq->num * sizeof *vq->used->ring));
510 }
511
512 /* Can we start vq? */
513 /* Caller should have vq mutex and device mutex */
514 int vhost_vq_access_ok(struct vhost_virtqueue *vq)
515 {
516         return vq_access_ok(vq->num, vq->desc, vq->avail, vq->used) &&
517                 vq_log_access_ok(vq, vq->log_base);
518 }
519
520 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
521 {
522         struct vhost_memory mem, *newmem, *oldmem;
523         unsigned long size = offsetof(struct vhost_memory, regions);
524         if (copy_from_user(&mem, m, size))
525                 return -EFAULT;
526         if (mem.padding)
527                 return -EOPNOTSUPP;
528         if (mem.nregions > VHOST_MEMORY_MAX_NREGIONS)
529                 return -E2BIG;
530         newmem = kmalloc(size + mem.nregions * sizeof *m->regions, GFP_KERNEL);
531         if (!newmem)
532                 return -ENOMEM;
533
534         memcpy(newmem, &mem, size);
535         if (copy_from_user(newmem->regions, m->regions,
536                            mem.nregions * sizeof *m->regions)) {
537                 kfree(newmem);
538                 return -EFAULT;
539         }
540
541         if (!memory_access_ok(d, newmem, vhost_has_feature(d, VHOST_F_LOG_ALL))) {
542                 kfree(newmem);
543                 return -EFAULT;
544         }
545         oldmem = rcu_dereference_protected(d->memory,
546                                            lockdep_is_held(&d->mutex));
547         rcu_assign_pointer(d->memory, newmem);
548         synchronize_rcu();
549         kfree(oldmem);
550         return 0;
551 }
552
553 static int init_used(struct vhost_virtqueue *vq,
554                      struct vring_used __user *used)
555 {
556         int r = put_user(vq->used_flags, &used->flags);
557         if (r)
558                 return r;
559         return get_user(vq->last_used_idx, &used->idx);
560 }
561
562 static long vhost_set_vring(struct vhost_dev *d, int ioctl, void __user *argp)
563 {
564         struct file *eventfp, *filep = NULL,
565                     *pollstart = NULL, *pollstop = NULL;
566         struct eventfd_ctx *ctx = NULL;
567         u32 __user *idxp = argp;
568         struct vhost_virtqueue *vq;
569         struct vhost_vring_state s;
570         struct vhost_vring_file f;
571         struct vhost_vring_addr a;
572         u32 idx;
573         long r;
574
575         r = get_user(idx, idxp);
576         if (r < 0)
577                 return r;
578         if (idx >= d->nvqs)
579                 return -ENOBUFS;
580
581         vq = d->vqs + idx;
582
583         mutex_lock(&vq->mutex);
584
585         switch (ioctl) {
586         case VHOST_SET_VRING_NUM:
587                 /* Resizing ring with an active backend?
588                  * You don't want to do that. */
589                 if (vq->private_data) {
590                         r = -EBUSY;
591                         break;
592                 }
593                 if (copy_from_user(&s, argp, sizeof s)) {
594                         r = -EFAULT;
595                         break;
596                 }
597                 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
598                         r = -EINVAL;
599                         break;
600                 }
601                 vq->num = s.num;
602                 break;
603         case VHOST_SET_VRING_BASE:
604                 /* Moving base with an active backend?
605                  * You don't want to do that. */
606                 if (vq->private_data) {
607                         r = -EBUSY;
608                         break;
609                 }
610                 if (copy_from_user(&s, argp, sizeof s)) {
611                         r = -EFAULT;
612                         break;
613                 }
614                 if (s.num > 0xffff) {
615                         r = -EINVAL;
616                         break;
617                 }
618                 vq->last_avail_idx = s.num;
619                 /* Forget the cached index value. */
620                 vq->avail_idx = vq->last_avail_idx;
621                 break;
622         case VHOST_GET_VRING_BASE:
623                 s.index = idx;
624                 s.num = vq->last_avail_idx;
625                 if (copy_to_user(argp, &s, sizeof s))
626                         r = -EFAULT;
627                 break;
628         case VHOST_SET_VRING_ADDR:
629                 if (copy_from_user(&a, argp, sizeof a)) {
630                         r = -EFAULT;
631                         break;
632                 }
633                 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
634                         r = -EOPNOTSUPP;
635                         break;
636                 }
637                 /* For 32bit, verify that the top 32bits of the user
638                    data are set to zero. */
639                 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
640                     (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
641                     (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
642                         r = -EFAULT;
643                         break;
644                 }
645                 if ((a.avail_user_addr & (sizeof *vq->avail->ring - 1)) ||
646                     (a.used_user_addr & (sizeof *vq->used->ring - 1)) ||
647                     (a.log_guest_addr & (sizeof *vq->used->ring - 1))) {
648                         r = -EINVAL;
649                         break;
650                 }
651
652                 /* We only verify access here if backend is configured.
653                  * If it is not, we don't as size might not have been setup.
654                  * We will verify when backend is configured. */
655                 if (vq->private_data) {
656                         if (!vq_access_ok(vq->num,
657                                 (void __user *)(unsigned long)a.desc_user_addr,
658                                 (void __user *)(unsigned long)a.avail_user_addr,
659                                 (void __user *)(unsigned long)a.used_user_addr)) {
660                                 r = -EINVAL;
661                                 break;
662                         }
663
664                         /* Also validate log access for used ring if enabled. */
665                         if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
666                             !log_access_ok(vq->log_base, a.log_guest_addr,
667                                            sizeof *vq->used +
668                                            vq->num * sizeof *vq->used->ring)) {
669                                 r = -EINVAL;
670                                 break;
671                         }
672                 }
673
674                 r = init_used(vq, (struct vring_used __user *)(unsigned long)
675                               a.used_user_addr);
676                 if (r)
677                         break;
678                 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
679                 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
680                 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
681                 vq->log_addr = a.log_guest_addr;
682                 vq->used = (void __user *)(unsigned long)a.used_user_addr;
683                 break;
684         case VHOST_SET_VRING_KICK:
685                 if (copy_from_user(&f, argp, sizeof f)) {
686                         r = -EFAULT;
687                         break;
688                 }
689                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
690                 if (IS_ERR(eventfp)) {
691                         r = PTR_ERR(eventfp);
692                         break;
693                 }
694                 if (eventfp != vq->kick) {
695                         pollstop = filep = vq->kick;
696                         pollstart = vq->kick = eventfp;
697                 } else
698                         filep = eventfp;
699                 break;
700         case VHOST_SET_VRING_CALL:
701                 if (copy_from_user(&f, argp, sizeof f)) {
702                         r = -EFAULT;
703                         break;
704                 }
705                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
706                 if (IS_ERR(eventfp)) {
707                         r = PTR_ERR(eventfp);
708                         break;
709                 }
710                 if (eventfp != vq->call) {
711                         filep = vq->call;
712                         ctx = vq->call_ctx;
713                         vq->call = eventfp;
714                         vq->call_ctx = eventfp ?
715                                 eventfd_ctx_fileget(eventfp) : NULL;
716                 } else
717                         filep = eventfp;
718                 break;
719         case VHOST_SET_VRING_ERR:
720                 if (copy_from_user(&f, argp, sizeof f)) {
721                         r = -EFAULT;
722                         break;
723                 }
724                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
725                 if (IS_ERR(eventfp)) {
726                         r = PTR_ERR(eventfp);
727                         break;
728                 }
729                 if (eventfp != vq->error) {
730                         filep = vq->error;
731                         vq->error = eventfp;
732                         ctx = vq->error_ctx;
733                         vq->error_ctx = eventfp ?
734                                 eventfd_ctx_fileget(eventfp) : NULL;
735                 } else
736                         filep = eventfp;
737                 break;
738         default:
739                 r = -ENOIOCTLCMD;
740         }
741
742         if (pollstop && vq->handle_kick)
743                 vhost_poll_stop(&vq->poll);
744
745         if (ctx)
746                 eventfd_ctx_put(ctx);
747         if (filep)
748                 fput(filep);
749
750         if (pollstart && vq->handle_kick)
751                 vhost_poll_start(&vq->poll, vq->kick);
752
753         mutex_unlock(&vq->mutex);
754
755         if (pollstop && vq->handle_kick)
756                 vhost_poll_flush(&vq->poll);
757         return r;
758 }
759
760 /* Caller must have device mutex */
761 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, unsigned long arg)
762 {
763         void __user *argp = (void __user *)arg;
764         struct file *eventfp, *filep = NULL;
765         struct eventfd_ctx *ctx = NULL;
766         u64 p;
767         long r;
768         int i, fd;
769
770         /* If you are not the owner, you can become one */
771         if (ioctl == VHOST_SET_OWNER) {
772                 r = vhost_dev_set_owner(d);
773                 goto done;
774         }
775
776         /* You must be the owner to do anything else */
777         r = vhost_dev_check_owner(d);
778         if (r)
779                 goto done;
780
781         switch (ioctl) {
782         case VHOST_SET_MEM_TABLE:
783                 r = vhost_set_memory(d, argp);
784                 break;
785         case VHOST_SET_LOG_BASE:
786                 if (copy_from_user(&p, argp, sizeof p)) {
787                         r = -EFAULT;
788                         break;
789                 }
790                 if ((u64)(unsigned long)p != p) {
791                         r = -EFAULT;
792                         break;
793                 }
794                 for (i = 0; i < d->nvqs; ++i) {
795                         struct vhost_virtqueue *vq;
796                         void __user *base = (void __user *)(unsigned long)p;
797                         vq = d->vqs + i;
798                         mutex_lock(&vq->mutex);
799                         /* If ring is inactive, will check when it's enabled. */
800                         if (vq->private_data && !vq_log_access_ok(vq, base))
801                                 r = -EFAULT;
802                         else
803                                 vq->log_base = base;
804                         mutex_unlock(&vq->mutex);
805                 }
806                 break;
807         case VHOST_SET_LOG_FD:
808                 r = get_user(fd, (int __user *)argp);
809                 if (r < 0)
810                         break;
811                 eventfp = fd == -1 ? NULL : eventfd_fget(fd);
812                 if (IS_ERR(eventfp)) {
813                         r = PTR_ERR(eventfp);
814                         break;
815                 }
816                 if (eventfp != d->log_file) {
817                         filep = d->log_file;
818                         ctx = d->log_ctx;
819                         d->log_ctx = eventfp ?
820                                 eventfd_ctx_fileget(eventfp) : NULL;
821                 } else
822                         filep = eventfp;
823                 for (i = 0; i < d->nvqs; ++i) {
824                         mutex_lock(&d->vqs[i].mutex);
825                         d->vqs[i].log_ctx = d->log_ctx;
826                         mutex_unlock(&d->vqs[i].mutex);
827                 }
828                 if (ctx)
829                         eventfd_ctx_put(ctx);
830                 if (filep)
831                         fput(filep);
832                 break;
833         default:
834                 r = vhost_set_vring(d, ioctl, argp);
835                 break;
836         }
837 done:
838         return r;
839 }
840
841 static const struct vhost_memory_region *find_region(struct vhost_memory *mem,
842                                                      __u64 addr, __u32 len)
843 {
844         struct vhost_memory_region *reg;
845         int i;
846         /* linear search is not brilliant, but we really have on the order of 6
847          * regions in practice */
848         for (i = 0; i < mem->nregions; ++i) {
849                 reg = mem->regions + i;
850                 if (reg->guest_phys_addr <= addr &&
851                     reg->guest_phys_addr + reg->memory_size - 1 >= addr)
852                         return reg;
853         }
854         return NULL;
855 }
856
857 /* TODO: This is really inefficient.  We need something like get_user()
858  * (instruction directly accesses the data, with an exception table entry
859  * returning -EFAULT). See Documentation/x86/exception-tables.txt.
860  */
861 static int set_bit_to_user(int nr, void __user *addr)
862 {
863         unsigned long log = (unsigned long)addr;
864         struct page *page;
865         void *base;
866         int bit = nr + (log % PAGE_SIZE) * 8;
867         int r;
868         r = get_user_pages_fast(log, 1, 1, &page);
869         if (r < 0)
870                 return r;
871         BUG_ON(r != 1);
872         base = kmap_atomic(page, KM_USER0);
873         set_bit(bit, base);
874         kunmap_atomic(base, KM_USER0);
875         set_page_dirty_lock(page);
876         put_page(page);
877         return 0;
878 }
879
880 static int log_write(void __user *log_base,
881                      u64 write_address, u64 write_length)
882 {
883         int r;
884         if (!write_length)
885                 return 0;
886         write_address /= VHOST_PAGE_SIZE;
887         for (;;) {
888                 u64 base = (u64)(unsigned long)log_base;
889                 u64 log = base + write_address / 8;
890                 int bit = write_address % 8;
891                 if ((u64)(unsigned long)log != log)
892                         return -EFAULT;
893                 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
894                 if (r < 0)
895                         return r;
896                 if (write_length <= VHOST_PAGE_SIZE)
897                         break;
898                 write_length -= VHOST_PAGE_SIZE;
899                 write_address += VHOST_PAGE_SIZE;
900         }
901         return r;
902 }
903
904 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
905                     unsigned int log_num, u64 len)
906 {
907         int i, r;
908
909         /* Make sure data written is seen before log. */
910         smp_wmb();
911         for (i = 0; i < log_num; ++i) {
912                 u64 l = min(log[i].len, len);
913                 r = log_write(vq->log_base, log[i].addr, l);
914                 if (r < 0)
915                         return r;
916                 len -= l;
917                 if (!len) {
918                         if (vq->log_ctx)
919                                 eventfd_signal(vq->log_ctx, 1);
920                         return 0;
921                 }
922         }
923         /* Length written exceeds what we have stored. This is a bug. */
924         BUG();
925         return 0;
926 }
927
928 static int translate_desc(struct vhost_dev *dev, u64 addr, u32 len,
929                           struct iovec iov[], int iov_size)
930 {
931         const struct vhost_memory_region *reg;
932         struct vhost_memory *mem;
933         struct iovec *_iov;
934         u64 s = 0;
935         int ret = 0;
936
937         rcu_read_lock();
938
939         mem = rcu_dereference(dev->memory);
940         while ((u64)len > s) {
941                 u64 size;
942                 if (unlikely(ret >= iov_size)) {
943                         ret = -ENOBUFS;
944                         break;
945                 }
946                 reg = find_region(mem, addr, len);
947                 if (unlikely(!reg)) {
948                         ret = -EFAULT;
949                         break;
950                 }
951                 _iov = iov + ret;
952                 size = reg->memory_size - addr + reg->guest_phys_addr;
953                 _iov->iov_len = min((u64)len, size);
954                 _iov->iov_base = (void __user *)(unsigned long)
955                         (reg->userspace_addr + addr - reg->guest_phys_addr);
956                 s += size;
957                 addr += size;
958                 ++ret;
959         }
960
961         rcu_read_unlock();
962         return ret;
963 }
964
965 /* Each buffer in the virtqueues is actually a chain of descriptors.  This
966  * function returns the next descriptor in the chain,
967  * or -1U if we're at the end. */
968 static unsigned next_desc(struct vring_desc *desc)
969 {
970         unsigned int next;
971
972         /* If this descriptor says it doesn't chain, we're done. */
973         if (!(desc->flags & VRING_DESC_F_NEXT))
974                 return -1U;
975
976         /* Check they're not leading us off end of descriptors. */
977         next = desc->next;
978         /* Make sure compiler knows to grab that: we don't want it changing! */
979         /* We will use the result as an index in an array, so most
980          * architectures only need a compiler barrier here. */
981         read_barrier_depends();
982
983         return next;
984 }
985
986 static int get_indirect(struct vhost_dev *dev, struct vhost_virtqueue *vq,
987                         struct iovec iov[], unsigned int iov_size,
988                         unsigned int *out_num, unsigned int *in_num,
989                         struct vhost_log *log, unsigned int *log_num,
990                         struct vring_desc *indirect)
991 {
992         struct vring_desc desc;
993         unsigned int i = 0, count, found = 0;
994         int ret;
995
996         /* Sanity check */
997         if (unlikely(indirect->len % sizeof desc)) {
998                 vq_err(vq, "Invalid length in indirect descriptor: "
999                        "len 0x%llx not multiple of 0x%zx\n",
1000                        (unsigned long long)indirect->len,
1001                        sizeof desc);
1002                 return -EINVAL;
1003         }
1004
1005         ret = translate_desc(dev, indirect->addr, indirect->len, vq->indirect,
1006                              UIO_MAXIOV);
1007         if (unlikely(ret < 0)) {
1008                 vq_err(vq, "Translation failure %d in indirect.\n", ret);
1009                 return ret;
1010         }
1011
1012         /* We will use the result as an address to read from, so most
1013          * architectures only need a compiler barrier here. */
1014         read_barrier_depends();
1015
1016         count = indirect->len / sizeof desc;
1017         /* Buffers are chained via a 16 bit next field, so
1018          * we can have at most 2^16 of these. */
1019         if (unlikely(count > USHRT_MAX + 1)) {
1020                 vq_err(vq, "Indirect buffer length too big: %d\n",
1021                        indirect->len);
1022                 return -E2BIG;
1023         }
1024
1025         do {
1026                 unsigned iov_count = *in_num + *out_num;
1027                 if (unlikely(++found > count)) {
1028                         vq_err(vq, "Loop detected: last one at %u "
1029                                "indirect size %u\n",
1030                                i, count);
1031                         return -EINVAL;
1032                 }
1033                 if (unlikely(memcpy_fromiovec((unsigned char *)&desc, vq->indirect,
1034                                               sizeof desc))) {
1035                         vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1036                                i, (size_t)indirect->addr + i * sizeof desc);
1037                         return -EINVAL;
1038                 }
1039                 if (unlikely(desc.flags & VRING_DESC_F_INDIRECT)) {
1040                         vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1041                                i, (size_t)indirect->addr + i * sizeof desc);
1042                         return -EINVAL;
1043                 }
1044
1045                 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
1046                                      iov_size - iov_count);
1047                 if (unlikely(ret < 0)) {
1048                         vq_err(vq, "Translation failure %d indirect idx %d\n",
1049                                ret, i);
1050                         return ret;
1051                 }
1052                 /* If this is an input descriptor, increment that count. */
1053                 if (desc.flags & VRING_DESC_F_WRITE) {
1054                         *in_num += ret;
1055                         if (unlikely(log)) {
1056                                 log[*log_num].addr = desc.addr;
1057                                 log[*log_num].len = desc.len;
1058                                 ++*log_num;
1059                         }
1060                 } else {
1061                         /* If it's an output descriptor, they're all supposed
1062                          * to come before any input descriptors. */
1063                         if (unlikely(*in_num)) {
1064                                 vq_err(vq, "Indirect descriptor "
1065                                        "has out after in: idx %d\n", i);
1066                                 return -EINVAL;
1067                         }
1068                         *out_num += ret;
1069                 }
1070         } while ((i = next_desc(&desc)) != -1);
1071         return 0;
1072 }
1073
1074 /* This looks in the virtqueue and for the first available buffer, and converts
1075  * it to an iovec for convenient access.  Since descriptors consist of some
1076  * number of output then some number of input descriptors, it's actually two
1077  * iovecs, but we pack them into one and note how many of each there were.
1078  *
1079  * This function returns the descriptor number found, or vq->num (which is
1080  * never a valid descriptor number) if none was found.  A negative code is
1081  * returned on error. */
1082 int vhost_get_vq_desc(struct vhost_dev *dev, struct vhost_virtqueue *vq,
1083                       struct iovec iov[], unsigned int iov_size,
1084                       unsigned int *out_num, unsigned int *in_num,
1085                       struct vhost_log *log, unsigned int *log_num)
1086 {
1087         struct vring_desc desc;
1088         unsigned int i, head, found = 0;
1089         u16 last_avail_idx;
1090         int ret;
1091
1092         /* Check it isn't doing very strange things with descriptor numbers. */
1093         last_avail_idx = vq->last_avail_idx;
1094         if (unlikely(get_user(vq->avail_idx, &vq->avail->idx))) {
1095                 vq_err(vq, "Failed to access avail idx at %p\n",
1096                        &vq->avail->idx);
1097                 return -EFAULT;
1098         }
1099
1100         if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
1101                 vq_err(vq, "Guest moved used index from %u to %u",
1102                        last_avail_idx, vq->avail_idx);
1103                 return -EFAULT;
1104         }
1105
1106         /* If there's nothing new since last we looked, return invalid. */
1107         if (vq->avail_idx == last_avail_idx)
1108                 return vq->num;
1109
1110         /* Only get avail ring entries after they have been exposed by guest. */
1111         smp_rmb();
1112
1113         /* Grab the next descriptor number they're advertising, and increment
1114          * the index we've seen. */
1115         if (unlikely(get_user(head,
1116                               &vq->avail->ring[last_avail_idx % vq->num]))) {
1117                 vq_err(vq, "Failed to read head: idx %d address %p\n",
1118                        last_avail_idx,
1119                        &vq->avail->ring[last_avail_idx % vq->num]);
1120                 return -EFAULT;
1121         }
1122
1123         /* If their number is silly, that's an error. */
1124         if (unlikely(head >= vq->num)) {
1125                 vq_err(vq, "Guest says index %u > %u is available",
1126                        head, vq->num);
1127                 return -EINVAL;
1128         }
1129
1130         /* When we start there are none of either input nor output. */
1131         *out_num = *in_num = 0;
1132         if (unlikely(log))
1133                 *log_num = 0;
1134
1135         i = head;
1136         do {
1137                 unsigned iov_count = *in_num + *out_num;
1138                 if (unlikely(i >= vq->num)) {
1139                         vq_err(vq, "Desc index is %u > %u, head = %u",
1140                                i, vq->num, head);
1141                         return -EINVAL;
1142                 }
1143                 if (unlikely(++found > vq->num)) {
1144                         vq_err(vq, "Loop detected: last one at %u "
1145                                "vq size %u head %u\n",
1146                                i, vq->num, head);
1147                         return -EINVAL;
1148                 }
1149                 ret = copy_from_user(&desc, vq->desc + i, sizeof desc);
1150                 if (unlikely(ret)) {
1151                         vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
1152                                i, vq->desc + i);
1153                         return -EFAULT;
1154                 }
1155                 if (desc.flags & VRING_DESC_F_INDIRECT) {
1156                         ret = get_indirect(dev, vq, iov, iov_size,
1157                                            out_num, in_num,
1158                                            log, log_num, &desc);
1159                         if (unlikely(ret < 0)) {
1160                                 vq_err(vq, "Failure detected "
1161                                        "in indirect descriptor at idx %d\n", i);
1162                                 return ret;
1163                         }
1164                         continue;
1165                 }
1166
1167                 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
1168                                      iov_size - iov_count);
1169                 if (unlikely(ret < 0)) {
1170                         vq_err(vq, "Translation failure %d descriptor idx %d\n",
1171                                ret, i);
1172                         return ret;
1173                 }
1174                 if (desc.flags & VRING_DESC_F_WRITE) {
1175                         /* If this is an input descriptor,
1176                          * increment that count. */
1177                         *in_num += ret;
1178                         if (unlikely(log)) {
1179                                 log[*log_num].addr = desc.addr;
1180                                 log[*log_num].len = desc.len;
1181                                 ++*log_num;
1182                         }
1183                 } else {
1184                         /* If it's an output descriptor, they're all supposed
1185                          * to come before any input descriptors. */
1186                         if (unlikely(*in_num)) {
1187                                 vq_err(vq, "Descriptor has out after in: "
1188                                        "idx %d\n", i);
1189                                 return -EINVAL;
1190                         }
1191                         *out_num += ret;
1192                 }
1193         } while ((i = next_desc(&desc)) != -1);
1194
1195         /* On success, increment avail index. */
1196         vq->last_avail_idx++;
1197         return head;
1198 }
1199
1200 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
1201 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
1202 {
1203         vq->last_avail_idx -= n;
1204 }
1205
1206 /* After we've used one of their buffers, we tell them about it.  We'll then
1207  * want to notify the guest, using eventfd. */
1208 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
1209 {
1210         struct vring_used_elem __user *used;
1211
1212         /* The virtqueue contains a ring of used buffers.  Get a pointer to the
1213          * next entry in that used ring. */
1214         used = &vq->used->ring[vq->last_used_idx % vq->num];
1215         if (put_user(head, &used->id)) {
1216                 vq_err(vq, "Failed to write used id");
1217                 return -EFAULT;
1218         }
1219         if (put_user(len, &used->len)) {
1220                 vq_err(vq, "Failed to write used len");
1221                 return -EFAULT;
1222         }
1223         /* Make sure buffer is written before we update index. */
1224         smp_wmb();
1225         if (put_user(vq->last_used_idx + 1, &vq->used->idx)) {
1226                 vq_err(vq, "Failed to increment used idx");
1227                 return -EFAULT;
1228         }
1229         if (unlikely(vq->log_used)) {
1230                 /* Make sure data is seen before log. */
1231                 smp_wmb();
1232                 /* Log used ring entry write. */
1233                 log_write(vq->log_base,
1234                           vq->log_addr +
1235                            ((void __user *)used - (void __user *)vq->used),
1236                           sizeof *used);
1237                 /* Log used index update. */
1238                 log_write(vq->log_base,
1239                           vq->log_addr + offsetof(struct vring_used, idx),
1240                           sizeof vq->used->idx);
1241                 if (vq->log_ctx)
1242                         eventfd_signal(vq->log_ctx, 1);
1243         }
1244         vq->last_used_idx++;
1245         return 0;
1246 }
1247
1248 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
1249                             struct vring_used_elem *heads,
1250                             unsigned count)
1251 {
1252         struct vring_used_elem __user *used;
1253         int start;
1254
1255         start = vq->last_used_idx % vq->num;
1256         used = vq->used->ring + start;
1257         if (copy_to_user(used, heads, count * sizeof *used)) {
1258                 vq_err(vq, "Failed to write used");
1259                 return -EFAULT;
1260         }
1261         if (unlikely(vq->log_used)) {
1262                 /* Make sure data is seen before log. */
1263                 smp_wmb();
1264                 /* Log used ring entry write. */
1265                 log_write(vq->log_base,
1266                           vq->log_addr +
1267                            ((void __user *)used - (void __user *)vq->used),
1268                           count * sizeof *used);
1269         }
1270         vq->last_used_idx += count;
1271         return 0;
1272 }
1273
1274 /* After we've used one of their buffers, we tell them about it.  We'll then
1275  * want to notify the guest, using eventfd. */
1276 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
1277                      unsigned count)
1278 {
1279         int start, n, r;
1280
1281         start = vq->last_used_idx % vq->num;
1282         n = vq->num - start;
1283         if (n < count) {
1284                 r = __vhost_add_used_n(vq, heads, n);
1285                 if (r < 0)
1286                         return r;
1287                 heads += n;
1288                 count -= n;
1289         }
1290         r = __vhost_add_used_n(vq, heads, count);
1291
1292         /* Make sure buffer is written before we update index. */
1293         smp_wmb();
1294         if (put_user(vq->last_used_idx, &vq->used->idx)) {
1295                 vq_err(vq, "Failed to increment used idx");
1296                 return -EFAULT;
1297         }
1298         if (unlikely(vq->log_used)) {
1299                 /* Log used index update. */
1300                 log_write(vq->log_base,
1301                           vq->log_addr + offsetof(struct vring_used, idx),
1302                           sizeof vq->used->idx);
1303                 if (vq->log_ctx)
1304                         eventfd_signal(vq->log_ctx, 1);
1305         }
1306         return r;
1307 }
1308
1309 /* This actually signals the guest, using eventfd. */
1310 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1311 {
1312         __u16 flags;
1313         /* Flush out used index updates. This is paired
1314          * with the barrier that the Guest executes when enabling
1315          * interrupts. */
1316         smp_mb();
1317
1318         if (get_user(flags, &vq->avail->flags)) {
1319                 vq_err(vq, "Failed to get flags");
1320                 return;
1321         }
1322
1323         /* If they don't want an interrupt, don't signal, unless empty. */
1324         if ((flags & VRING_AVAIL_F_NO_INTERRUPT) &&
1325             (vq->avail_idx != vq->last_avail_idx ||
1326              !vhost_has_feature(dev, VIRTIO_F_NOTIFY_ON_EMPTY)))
1327                 return;
1328
1329         /* Signal the Guest tell them we used something up. */
1330         if (vq->call_ctx)
1331                 eventfd_signal(vq->call_ctx, 1);
1332 }
1333
1334 /* And here's the combo meal deal.  Supersize me! */
1335 void vhost_add_used_and_signal(struct vhost_dev *dev,
1336                                struct vhost_virtqueue *vq,
1337                                unsigned int head, int len)
1338 {
1339         vhost_add_used(vq, head, len);
1340         vhost_signal(dev, vq);
1341 }
1342
1343 /* multi-buffer version of vhost_add_used_and_signal */
1344 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
1345                                  struct vhost_virtqueue *vq,
1346                                  struct vring_used_elem *heads, unsigned count)
1347 {
1348         vhost_add_used_n(vq, heads, count);
1349         vhost_signal(dev, vq);
1350 }
1351
1352 /* OK, now we need to know about added descriptors. */
1353 bool vhost_enable_notify(struct vhost_virtqueue *vq)
1354 {
1355         u16 avail_idx;
1356         int r;
1357         if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
1358                 return false;
1359         vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
1360         r = put_user(vq->used_flags, &vq->used->flags);
1361         if (r) {
1362                 vq_err(vq, "Failed to enable notification at %p: %d\n",
1363                        &vq->used->flags, r);
1364                 return false;
1365         }
1366         /* They could have slipped one in as we were doing that: make
1367          * sure it's written, then check again. */
1368         smp_mb();
1369         r = get_user(avail_idx, &vq->avail->idx);
1370         if (r) {
1371                 vq_err(vq, "Failed to check avail idx at %p: %d\n",
1372                        &vq->avail->idx, r);
1373                 return false;
1374         }
1375
1376         return avail_idx != vq->avail_idx;
1377 }
1378
1379 /* We don't need to be notified again. */
1380 void vhost_disable_notify(struct vhost_virtqueue *vq)
1381 {
1382         int r;
1383         if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
1384                 return;
1385         vq->used_flags |= VRING_USED_F_NO_NOTIFY;
1386         r = put_user(vq->used_flags, &vq->used->flags);
1387         if (r)
1388                 vq_err(vq, "Failed to enable notification at %p: %d\n",
1389                        &vq->used->flags, r);
1390 }