mm: thp: set the accessed flag for old pages on access fault
[pandora-kernel.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/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 <linux/net.h>
30 #include <linux/if_packet.h>
31 #include <linux/if_arp.h>
32
33 #include "vhost.h"
34
35 enum {
36         VHOST_MEMORY_MAX_NREGIONS = 64,
37         VHOST_MEMORY_F_LOG = 0x1,
38 };
39
40 static unsigned vhost_zcopy_mask __read_mostly;
41
42 #define vhost_used_event(vq) ((u16 __user *)&vq->avail->ring[vq->num])
43 #define vhost_avail_event(vq) ((u16 __user *)&vq->used->ring[vq->num])
44
45 static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
46                             poll_table *pt)
47 {
48         struct vhost_poll *poll;
49
50         poll = container_of(pt, struct vhost_poll, table);
51         poll->wqh = wqh;
52         add_wait_queue(wqh, &poll->wait);
53 }
54
55 static int vhost_poll_wakeup(wait_queue_t *wait, unsigned mode, int sync,
56                              void *key)
57 {
58         struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
59
60         if (!((unsigned long)key & poll->mask))
61                 return 0;
62
63         vhost_poll_queue(poll);
64         return 0;
65 }
66
67 static void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
68 {
69         INIT_LIST_HEAD(&work->node);
70         work->fn = fn;
71         init_waitqueue_head(&work->done);
72         work->flushing = 0;
73         work->queue_seq = work->done_seq = 0;
74 }
75
76 /* Init poll structure */
77 void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
78                      unsigned long mask, struct vhost_dev *dev)
79 {
80         init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
81         init_poll_funcptr(&poll->table, vhost_poll_func);
82         poll->mask = mask;
83         poll->dev = dev;
84
85         vhost_work_init(&poll->work, fn);
86 }
87
88 /* Start polling a file. We add ourselves to file's wait queue. The caller must
89  * keep a reference to a file until after vhost_poll_stop is called. */
90 void vhost_poll_start(struct vhost_poll *poll, struct file *file)
91 {
92         unsigned long mask;
93
94         mask = file->f_op->poll(file, &poll->table);
95         if (mask)
96                 vhost_poll_wakeup(&poll->wait, 0, 0, (void *)mask);
97 }
98
99 /* Stop polling a file. After this function returns, it becomes safe to drop the
100  * file reference. You must also flush afterwards. */
101 void vhost_poll_stop(struct vhost_poll *poll)
102 {
103         remove_wait_queue(poll->wqh, &poll->wait);
104 }
105
106 static bool vhost_work_seq_done(struct vhost_dev *dev, struct vhost_work *work,
107                                 unsigned seq)
108 {
109         int left;
110
111         spin_lock_irq(&dev->work_lock);
112         left = seq - work->done_seq;
113         spin_unlock_irq(&dev->work_lock);
114         return left <= 0;
115 }
116
117 static void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work)
118 {
119         unsigned seq;
120         int flushing;
121
122         spin_lock_irq(&dev->work_lock);
123         seq = work->queue_seq;
124         work->flushing++;
125         spin_unlock_irq(&dev->work_lock);
126         wait_event(work->done, vhost_work_seq_done(dev, work, seq));
127         spin_lock_irq(&dev->work_lock);
128         flushing = --work->flushing;
129         spin_unlock_irq(&dev->work_lock);
130         BUG_ON(flushing < 0);
131 }
132
133 /* Flush any work that has been scheduled. When calling this, don't hold any
134  * locks that are also used by the callback. */
135 void vhost_poll_flush(struct vhost_poll *poll)
136 {
137         vhost_work_flush(poll->dev, &poll->work);
138 }
139
140 static inline void vhost_work_queue(struct vhost_dev *dev,
141                                     struct vhost_work *work)
142 {
143         unsigned long flags;
144
145         spin_lock_irqsave(&dev->work_lock, flags);
146         if (list_empty(&work->node)) {
147                 list_add_tail(&work->node, &dev->work_list);
148                 work->queue_seq++;
149                 wake_up_process(dev->worker);
150         }
151         spin_unlock_irqrestore(&dev->work_lock, flags);
152 }
153
154 void vhost_poll_queue(struct vhost_poll *poll)
155 {
156         vhost_work_queue(poll->dev, &poll->work);
157 }
158
159 static void vhost_vq_reset(struct vhost_dev *dev,
160                            struct vhost_virtqueue *vq)
161 {
162         vq->num = 1;
163         vq->desc = NULL;
164         vq->avail = NULL;
165         vq->used = NULL;
166         vq->last_avail_idx = 0;
167         vq->avail_idx = 0;
168         vq->last_used_idx = 0;
169         vq->signalled_used = 0;
170         vq->signalled_used_valid = false;
171         vq->used_flags = 0;
172         vq->log_used = false;
173         vq->log_addr = -1ull;
174         vq->vhost_hlen = 0;
175         vq->sock_hlen = 0;
176         vq->private_data = NULL;
177         vq->log_base = NULL;
178         vq->error_ctx = NULL;
179         vq->error = NULL;
180         vq->kick = NULL;
181         vq->call_ctx = NULL;
182         vq->call = NULL;
183         vq->log_ctx = NULL;
184         vq->upend_idx = 0;
185         vq->done_idx = 0;
186         vq->ubufs = NULL;
187 }
188
189 static int vhost_worker(void *data)
190 {
191         struct vhost_dev *dev = data;
192         struct vhost_work *work = NULL;
193         unsigned uninitialized_var(seq);
194
195         use_mm(dev->mm);
196
197         for (;;) {
198                 /* mb paired w/ kthread_stop */
199                 set_current_state(TASK_INTERRUPTIBLE);
200
201                 spin_lock_irq(&dev->work_lock);
202                 if (work) {
203                         work->done_seq = seq;
204                         if (work->flushing)
205                                 wake_up_all(&work->done);
206                 }
207
208                 if (kthread_should_stop()) {
209                         spin_unlock_irq(&dev->work_lock);
210                         __set_current_state(TASK_RUNNING);
211                         break;
212                 }
213                 if (!list_empty(&dev->work_list)) {
214                         work = list_first_entry(&dev->work_list,
215                                                 struct vhost_work, node);
216                         list_del_init(&work->node);
217                         seq = work->queue_seq;
218                 } else
219                         work = NULL;
220                 spin_unlock_irq(&dev->work_lock);
221
222                 if (work) {
223                         __set_current_state(TASK_RUNNING);
224                         work->fn(work);
225                         if (need_resched())
226                                 schedule();
227                 } else
228                         schedule();
229
230         }
231         unuse_mm(dev->mm);
232         return 0;
233 }
234
235 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
236 {
237         kfree(vq->indirect);
238         vq->indirect = NULL;
239         kfree(vq->log);
240         vq->log = NULL;
241         kfree(vq->heads);
242         vq->heads = NULL;
243         kfree(vq->ubuf_info);
244         vq->ubuf_info = NULL;
245 }
246
247 void vhost_enable_zcopy(int vq)
248 {
249         vhost_zcopy_mask |= 0x1 << vq;
250 }
251
252 /* Helper to allocate iovec buffers for all vqs. */
253 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
254 {
255         int i;
256         bool zcopy;
257
258         for (i = 0; i < dev->nvqs; ++i) {
259                 dev->vqs[i].indirect = kmalloc(sizeof *dev->vqs[i].indirect *
260                                                UIO_MAXIOV, GFP_KERNEL);
261                 dev->vqs[i].log = kmalloc(sizeof *dev->vqs[i].log * UIO_MAXIOV,
262                                           GFP_KERNEL);
263                 dev->vqs[i].heads = kmalloc(sizeof *dev->vqs[i].heads *
264                                             UIO_MAXIOV, GFP_KERNEL);
265                 zcopy = vhost_zcopy_mask & (0x1 << i);
266                 if (zcopy)
267                         dev->vqs[i].ubuf_info =
268                                 kmalloc(sizeof *dev->vqs[i].ubuf_info *
269                                         UIO_MAXIOV, GFP_KERNEL);
270                 if (!dev->vqs[i].indirect || !dev->vqs[i].log ||
271                         !dev->vqs[i].heads ||
272                         (zcopy && !dev->vqs[i].ubuf_info))
273                         goto err_nomem;
274         }
275         return 0;
276
277 err_nomem:
278         for (; i >= 0; --i)
279                 vhost_vq_free_iovecs(&dev->vqs[i]);
280         return -ENOMEM;
281 }
282
283 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
284 {
285         int i;
286
287         for (i = 0; i < dev->nvqs; ++i)
288                 vhost_vq_free_iovecs(&dev->vqs[i]);
289 }
290
291 long vhost_dev_init(struct vhost_dev *dev,
292                     struct vhost_virtqueue *vqs, int nvqs)
293 {
294         int i;
295
296         dev->vqs = vqs;
297         dev->nvqs = nvqs;
298         mutex_init(&dev->mutex);
299         dev->log_ctx = NULL;
300         dev->log_file = NULL;
301         dev->memory = NULL;
302         dev->mm = NULL;
303         spin_lock_init(&dev->work_lock);
304         INIT_LIST_HEAD(&dev->work_list);
305         dev->worker = NULL;
306
307         for (i = 0; i < dev->nvqs; ++i) {
308                 dev->vqs[i].log = NULL;
309                 dev->vqs[i].indirect = NULL;
310                 dev->vqs[i].heads = NULL;
311                 dev->vqs[i].ubuf_info = NULL;
312                 dev->vqs[i].dev = dev;
313                 mutex_init(&dev->vqs[i].mutex);
314                 vhost_vq_reset(dev, dev->vqs + i);
315                 if (dev->vqs[i].handle_kick)
316                         vhost_poll_init(&dev->vqs[i].poll,
317                                         dev->vqs[i].handle_kick, POLLIN, dev);
318         }
319
320         return 0;
321 }
322
323 /* Caller should have device mutex */
324 long vhost_dev_check_owner(struct vhost_dev *dev)
325 {
326         /* Are you the owner? If not, I don't think you mean to do that */
327         return dev->mm == current->mm ? 0 : -EPERM;
328 }
329
330 struct vhost_attach_cgroups_struct {
331         struct vhost_work work;
332         struct task_struct *owner;
333         int ret;
334 };
335
336 static void vhost_attach_cgroups_work(struct vhost_work *work)
337 {
338         struct vhost_attach_cgroups_struct *s;
339
340         s = container_of(work, struct vhost_attach_cgroups_struct, work);
341         s->ret = cgroup_attach_task_all(s->owner, current);
342 }
343
344 static int vhost_attach_cgroups(struct vhost_dev *dev)
345 {
346         struct vhost_attach_cgroups_struct attach;
347
348         attach.owner = current;
349         vhost_work_init(&attach.work, vhost_attach_cgroups_work);
350         vhost_work_queue(dev, &attach.work);
351         vhost_work_flush(dev, &attach.work);
352         return attach.ret;
353 }
354
355 /* Caller should have device mutex */
356 static long vhost_dev_set_owner(struct vhost_dev *dev)
357 {
358         struct task_struct *worker;
359         int err;
360
361         /* Is there an owner already? */
362         if (dev->mm) {
363                 err = -EBUSY;
364                 goto err_mm;
365         }
366
367         /* No owner, become one */
368         dev->mm = get_task_mm(current);
369         worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
370         if (IS_ERR(worker)) {
371                 err = PTR_ERR(worker);
372                 goto err_worker;
373         }
374
375         dev->worker = worker;
376         wake_up_process(worker);        /* avoid contributing to loadavg */
377
378         err = vhost_attach_cgroups(dev);
379         if (err)
380                 goto err_cgroup;
381
382         err = vhost_dev_alloc_iovecs(dev);
383         if (err)
384                 goto err_cgroup;
385
386         return 0;
387 err_cgroup:
388         kthread_stop(worker);
389         dev->worker = NULL;
390 err_worker:
391         if (dev->mm)
392                 mmput(dev->mm);
393         dev->mm = NULL;
394 err_mm:
395         return err;
396 }
397
398 /* Caller should have device mutex */
399 long vhost_dev_reset_owner(struct vhost_dev *dev)
400 {
401         struct vhost_memory *memory;
402
403         /* Restore memory to default empty mapping. */
404         memory = kmalloc(offsetof(struct vhost_memory, regions), GFP_KERNEL);
405         if (!memory)
406                 return -ENOMEM;
407
408         vhost_dev_cleanup(dev);
409
410         memory->nregions = 0;
411         RCU_INIT_POINTER(dev->memory, memory);
412         return 0;
413 }
414
415 /* In case of DMA done not in order in lower device driver for some reason.
416  * upend_idx is used to track end of used idx, done_idx is used to track head
417  * of used idx. Once lower device DMA done contiguously, we will signal KVM
418  * guest used idx.
419  */
420 int vhost_zerocopy_signal_used(struct vhost_virtqueue *vq)
421 {
422         int i;
423         int j = 0;
424
425         for (i = vq->done_idx; i != vq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
426                 if ((vq->heads[i].len == VHOST_DMA_DONE_LEN)) {
427                         vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
428                         vhost_add_used_and_signal(vq->dev, vq,
429                                                   vq->heads[i].id, 0);
430                         ++j;
431                 } else
432                         break;
433         }
434         if (j)
435                 vq->done_idx = i;
436         return j;
437 }
438
439 /* Caller should have device mutex */
440 void vhost_dev_cleanup(struct vhost_dev *dev)
441 {
442         int i;
443
444         for (i = 0; i < dev->nvqs; ++i) {
445                 if (dev->vqs[i].kick && dev->vqs[i].handle_kick) {
446                         vhost_poll_stop(&dev->vqs[i].poll);
447                         vhost_poll_flush(&dev->vqs[i].poll);
448                 }
449                 /* Wait for all lower device DMAs done. */
450                 if (dev->vqs[i].ubufs)
451                         vhost_ubuf_put_and_wait(dev->vqs[i].ubufs);
452
453                 /* Signal guest as appropriate. */
454                 vhost_zerocopy_signal_used(&dev->vqs[i]);
455
456                 if (dev->vqs[i].error_ctx)
457                         eventfd_ctx_put(dev->vqs[i].error_ctx);
458                 if (dev->vqs[i].error)
459                         fput(dev->vqs[i].error);
460                 if (dev->vqs[i].kick)
461                         fput(dev->vqs[i].kick);
462                 if (dev->vqs[i].call_ctx)
463                         eventfd_ctx_put(dev->vqs[i].call_ctx);
464                 if (dev->vqs[i].call)
465                         fput(dev->vqs[i].call);
466                 vhost_vq_reset(dev, dev->vqs + i);
467         }
468         vhost_dev_free_iovecs(dev);
469         if (dev->log_ctx)
470                 eventfd_ctx_put(dev->log_ctx);
471         dev->log_ctx = NULL;
472         if (dev->log_file)
473                 fput(dev->log_file);
474         dev->log_file = NULL;
475         /* No one will access memory at this point */
476         kfree(rcu_dereference_protected(dev->memory,
477                                         lockdep_is_held(&dev->mutex)));
478         RCU_INIT_POINTER(dev->memory, NULL);
479         WARN_ON(!list_empty(&dev->work_list));
480         if (dev->worker) {
481                 kthread_stop(dev->worker);
482                 dev->worker = NULL;
483         }
484         if (dev->mm)
485                 mmput(dev->mm);
486         dev->mm = NULL;
487 }
488
489 static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
490 {
491         u64 a = addr / VHOST_PAGE_SIZE / 8;
492
493         /* Make sure 64 bit math will not overflow. */
494         if (a > ULONG_MAX - (unsigned long)log_base ||
495             a + (unsigned long)log_base > ULONG_MAX)
496                 return 0;
497
498         return access_ok(VERIFY_WRITE, log_base + a,
499                          (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
500 }
501
502 /* Caller should have vq mutex and device mutex. */
503 static int vq_memory_access_ok(void __user *log_base, struct vhost_memory *mem,
504                                int log_all)
505 {
506         int i;
507
508         if (!mem)
509                 return 0;
510
511         for (i = 0; i < mem->nregions; ++i) {
512                 struct vhost_memory_region *m = mem->regions + i;
513                 unsigned long a = m->userspace_addr;
514                 if (m->memory_size > ULONG_MAX)
515                         return 0;
516                 else if (!access_ok(VERIFY_WRITE, (void __user *)a,
517                                     m->memory_size))
518                         return 0;
519                 else if (log_all && !log_access_ok(log_base,
520                                                    m->guest_phys_addr,
521                                                    m->memory_size))
522                         return 0;
523         }
524         return 1;
525 }
526
527 /* Can we switch to this memory table? */
528 /* Caller should have device mutex but not vq mutex */
529 static int memory_access_ok(struct vhost_dev *d, struct vhost_memory *mem,
530                             int log_all)
531 {
532         int i;
533
534         for (i = 0; i < d->nvqs; ++i) {
535                 int ok;
536                 mutex_lock(&d->vqs[i].mutex);
537                 /* If ring is inactive, will check when it's enabled. */
538                 if (d->vqs[i].private_data)
539                         ok = vq_memory_access_ok(d->vqs[i].log_base, mem,
540                                                  log_all);
541                 else
542                         ok = 1;
543                 mutex_unlock(&d->vqs[i].mutex);
544                 if (!ok)
545                         return 0;
546         }
547         return 1;
548 }
549
550 static int vq_access_ok(struct vhost_dev *d, unsigned int num,
551                         struct vring_desc __user *desc,
552                         struct vring_avail __user *avail,
553                         struct vring_used __user *used)
554 {
555         size_t s = vhost_has_feature(d, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
556         return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
557                access_ok(VERIFY_READ, avail,
558                          sizeof *avail + num * sizeof *avail->ring + s) &&
559                access_ok(VERIFY_WRITE, used,
560                         sizeof *used + num * sizeof *used->ring + s);
561 }
562
563 /* Can we log writes? */
564 /* Caller should have device mutex but not vq mutex */
565 int vhost_log_access_ok(struct vhost_dev *dev)
566 {
567         struct vhost_memory *mp;
568
569         mp = rcu_dereference_protected(dev->memory,
570                                        lockdep_is_held(&dev->mutex));
571         return memory_access_ok(dev, mp, 1);
572 }
573
574 /* Verify access for write logging. */
575 /* Caller should have vq mutex and device mutex */
576 static int vq_log_access_ok(struct vhost_dev *d, struct vhost_virtqueue *vq,
577                             void __user *log_base)
578 {
579         struct vhost_memory *mp;
580         size_t s = vhost_has_feature(d, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
581
582         mp = rcu_dereference_protected(vq->dev->memory,
583                                        lockdep_is_held(&vq->mutex));
584         return vq_memory_access_ok(log_base, mp,
585                             vhost_has_feature(vq->dev, VHOST_F_LOG_ALL)) &&
586                 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
587                                         sizeof *vq->used +
588                                         vq->num * sizeof *vq->used->ring + s));
589 }
590
591 /* Can we start vq? */
592 /* Caller should have vq mutex and device mutex */
593 int vhost_vq_access_ok(struct vhost_virtqueue *vq)
594 {
595         return vq_access_ok(vq->dev, vq->num, vq->desc, vq->avail, vq->used) &&
596                 vq_log_access_ok(vq->dev, vq, vq->log_base);
597 }
598
599 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
600 {
601         struct vhost_memory mem, *newmem, *oldmem;
602         unsigned long size = offsetof(struct vhost_memory, regions);
603
604         if (copy_from_user(&mem, m, size))
605                 return -EFAULT;
606         if (mem.padding)
607                 return -EOPNOTSUPP;
608         if (mem.nregions > VHOST_MEMORY_MAX_NREGIONS)
609                 return -E2BIG;
610         newmem = kmalloc(size + mem.nregions * sizeof *m->regions, GFP_KERNEL);
611         if (!newmem)
612                 return -ENOMEM;
613
614         memcpy(newmem, &mem, size);
615         if (copy_from_user(newmem->regions, m->regions,
616                            mem.nregions * sizeof *m->regions)) {
617                 kfree(newmem);
618                 return -EFAULT;
619         }
620
621         if (!memory_access_ok(d, newmem,
622                               vhost_has_feature(d, VHOST_F_LOG_ALL))) {
623                 kfree(newmem);
624                 return -EFAULT;
625         }
626         oldmem = rcu_dereference_protected(d->memory,
627                                            lockdep_is_held(&d->mutex));
628         rcu_assign_pointer(d->memory, newmem);
629         synchronize_rcu();
630         kfree(oldmem);
631         return 0;
632 }
633
634 static long vhost_set_vring(struct vhost_dev *d, int ioctl, void __user *argp)
635 {
636         struct file *eventfp, *filep = NULL,
637                     *pollstart = NULL, *pollstop = NULL;
638         struct eventfd_ctx *ctx = NULL;
639         u32 __user *idxp = argp;
640         struct vhost_virtqueue *vq;
641         struct vhost_vring_state s;
642         struct vhost_vring_file f;
643         struct vhost_vring_addr a;
644         u32 idx;
645         long r;
646
647         r = get_user(idx, idxp);
648         if (r < 0)
649                 return r;
650         if (idx >= d->nvqs)
651                 return -ENOBUFS;
652
653         vq = d->vqs + idx;
654
655         mutex_lock(&vq->mutex);
656
657         switch (ioctl) {
658         case VHOST_SET_VRING_NUM:
659                 /* Resizing ring with an active backend?
660                  * You don't want to do that. */
661                 if (vq->private_data) {
662                         r = -EBUSY;
663                         break;
664                 }
665                 if (copy_from_user(&s, argp, sizeof s)) {
666                         r = -EFAULT;
667                         break;
668                 }
669                 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
670                         r = -EINVAL;
671                         break;
672                 }
673                 vq->num = s.num;
674                 break;
675         case VHOST_SET_VRING_BASE:
676                 /* Moving base with an active backend?
677                  * You don't want to do that. */
678                 if (vq->private_data) {
679                         r = -EBUSY;
680                         break;
681                 }
682                 if (copy_from_user(&s, argp, sizeof s)) {
683                         r = -EFAULT;
684                         break;
685                 }
686                 if (s.num > 0xffff) {
687                         r = -EINVAL;
688                         break;
689                 }
690                 vq->last_avail_idx = s.num;
691                 /* Forget the cached index value. */
692                 vq->avail_idx = vq->last_avail_idx;
693                 break;
694         case VHOST_GET_VRING_BASE:
695                 s.index = idx;
696                 s.num = vq->last_avail_idx;
697                 if (copy_to_user(argp, &s, sizeof s))
698                         r = -EFAULT;
699                 break;
700         case VHOST_SET_VRING_ADDR:
701                 if (copy_from_user(&a, argp, sizeof a)) {
702                         r = -EFAULT;
703                         break;
704                 }
705                 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
706                         r = -EOPNOTSUPP;
707                         break;
708                 }
709                 /* For 32bit, verify that the top 32bits of the user
710                    data are set to zero. */
711                 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
712                     (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
713                     (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
714                         r = -EFAULT;
715                         break;
716                 }
717                 if ((a.avail_user_addr & (sizeof *vq->avail->ring - 1)) ||
718                     (a.used_user_addr & (sizeof *vq->used->ring - 1)) ||
719                     (a.log_guest_addr & (sizeof *vq->used->ring - 1))) {
720                         r = -EINVAL;
721                         break;
722                 }
723
724                 /* We only verify access here if backend is configured.
725                  * If it is not, we don't as size might not have been setup.
726                  * We will verify when backend is configured. */
727                 if (vq->private_data) {
728                         if (!vq_access_ok(d, vq->num,
729                                 (void __user *)(unsigned long)a.desc_user_addr,
730                                 (void __user *)(unsigned long)a.avail_user_addr,
731                                 (void __user *)(unsigned long)a.used_user_addr)) {
732                                 r = -EINVAL;
733                                 break;
734                         }
735
736                         /* Also validate log access for used ring if enabled. */
737                         if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
738                             !log_access_ok(vq->log_base, a.log_guest_addr,
739                                            sizeof *vq->used +
740                                            vq->num * sizeof *vq->used->ring)) {
741                                 r = -EINVAL;
742                                 break;
743                         }
744                 }
745
746                 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
747                 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
748                 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
749                 vq->log_addr = a.log_guest_addr;
750                 vq->used = (void __user *)(unsigned long)a.used_user_addr;
751                 break;
752         case VHOST_SET_VRING_KICK:
753                 if (copy_from_user(&f, argp, sizeof f)) {
754                         r = -EFAULT;
755                         break;
756                 }
757                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
758                 if (IS_ERR(eventfp)) {
759                         r = PTR_ERR(eventfp);
760                         break;
761                 }
762                 if (eventfp != vq->kick) {
763                         pollstop = filep = vq->kick;
764                         pollstart = vq->kick = eventfp;
765                 } else
766                         filep = eventfp;
767                 break;
768         case VHOST_SET_VRING_CALL:
769                 if (copy_from_user(&f, argp, sizeof f)) {
770                         r = -EFAULT;
771                         break;
772                 }
773                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
774                 if (IS_ERR(eventfp)) {
775                         r = PTR_ERR(eventfp);
776                         break;
777                 }
778                 if (eventfp != vq->call) {
779                         filep = vq->call;
780                         ctx = vq->call_ctx;
781                         vq->call = eventfp;
782                         vq->call_ctx = eventfp ?
783                                 eventfd_ctx_fileget(eventfp) : NULL;
784                 } else
785                         filep = eventfp;
786                 break;
787         case VHOST_SET_VRING_ERR:
788                 if (copy_from_user(&f, argp, sizeof f)) {
789                         r = -EFAULT;
790                         break;
791                 }
792                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
793                 if (IS_ERR(eventfp)) {
794                         r = PTR_ERR(eventfp);
795                         break;
796                 }
797                 if (eventfp != vq->error) {
798                         filep = vq->error;
799                         vq->error = eventfp;
800                         ctx = vq->error_ctx;
801                         vq->error_ctx = eventfp ?
802                                 eventfd_ctx_fileget(eventfp) : NULL;
803                 } else
804                         filep = eventfp;
805                 break;
806         default:
807                 r = -ENOIOCTLCMD;
808         }
809
810         if (pollstop && vq->handle_kick)
811                 vhost_poll_stop(&vq->poll);
812
813         if (ctx)
814                 eventfd_ctx_put(ctx);
815         if (filep)
816                 fput(filep);
817
818         if (pollstart && vq->handle_kick)
819                 vhost_poll_start(&vq->poll, vq->kick);
820
821         mutex_unlock(&vq->mutex);
822
823         if (pollstop && vq->handle_kick)
824                 vhost_poll_flush(&vq->poll);
825         return r;
826 }
827
828 /* Caller must have device mutex */
829 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, unsigned long arg)
830 {
831         void __user *argp = (void __user *)arg;
832         struct file *eventfp, *filep = NULL;
833         struct eventfd_ctx *ctx = NULL;
834         u64 p;
835         long r;
836         int i, fd;
837
838         /* If you are not the owner, you can become one */
839         if (ioctl == VHOST_SET_OWNER) {
840                 r = vhost_dev_set_owner(d);
841                 goto done;
842         }
843
844         /* You must be the owner to do anything else */
845         r = vhost_dev_check_owner(d);
846         if (r)
847                 goto done;
848
849         switch (ioctl) {
850         case VHOST_SET_MEM_TABLE:
851                 r = vhost_set_memory(d, argp);
852                 break;
853         case VHOST_SET_LOG_BASE:
854                 if (copy_from_user(&p, argp, sizeof p)) {
855                         r = -EFAULT;
856                         break;
857                 }
858                 if ((u64)(unsigned long)p != p) {
859                         r = -EFAULT;
860                         break;
861                 }
862                 for (i = 0; i < d->nvqs; ++i) {
863                         struct vhost_virtqueue *vq;
864                         void __user *base = (void __user *)(unsigned long)p;
865                         vq = d->vqs + i;
866                         mutex_lock(&vq->mutex);
867                         /* If ring is inactive, will check when it's enabled. */
868                         if (vq->private_data && !vq_log_access_ok(d, vq, base))
869                                 r = -EFAULT;
870                         else
871                                 vq->log_base = base;
872                         mutex_unlock(&vq->mutex);
873                 }
874                 break;
875         case VHOST_SET_LOG_FD:
876                 r = get_user(fd, (int __user *)argp);
877                 if (r < 0)
878                         break;
879                 eventfp = fd == -1 ? NULL : eventfd_fget(fd);
880                 if (IS_ERR(eventfp)) {
881                         r = PTR_ERR(eventfp);
882                         break;
883                 }
884                 if (eventfp != d->log_file) {
885                         filep = d->log_file;
886                         ctx = d->log_ctx;
887                         d->log_ctx = eventfp ?
888                                 eventfd_ctx_fileget(eventfp) : NULL;
889                 } else
890                         filep = eventfp;
891                 for (i = 0; i < d->nvqs; ++i) {
892                         mutex_lock(&d->vqs[i].mutex);
893                         d->vqs[i].log_ctx = d->log_ctx;
894                         mutex_unlock(&d->vqs[i].mutex);
895                 }
896                 if (ctx)
897                         eventfd_ctx_put(ctx);
898                 if (filep)
899                         fput(filep);
900                 break;
901         default:
902                 r = vhost_set_vring(d, ioctl, argp);
903                 break;
904         }
905 done:
906         return r;
907 }
908
909 static const struct vhost_memory_region *find_region(struct vhost_memory *mem,
910                                                      __u64 addr, __u32 len)
911 {
912         struct vhost_memory_region *reg;
913         int i;
914
915         /* linear search is not brilliant, but we really have on the order of 6
916          * regions in practice */
917         for (i = 0; i < mem->nregions; ++i) {
918                 reg = mem->regions + i;
919                 if (reg->guest_phys_addr <= addr &&
920                     reg->guest_phys_addr + reg->memory_size - 1 >= addr)
921                         return reg;
922         }
923         return NULL;
924 }
925
926 /* TODO: This is really inefficient.  We need something like get_user()
927  * (instruction directly accesses the data, with an exception table entry
928  * returning -EFAULT). See Documentation/x86/exception-tables.txt.
929  */
930 static int set_bit_to_user(int nr, void __user *addr)
931 {
932         unsigned long log = (unsigned long)addr;
933         struct page *page;
934         void *base;
935         int bit = nr + (log % PAGE_SIZE) * 8;
936         int r;
937
938         r = get_user_pages_fast(log, 1, 1, &page);
939         if (r < 0)
940                 return r;
941         BUG_ON(r != 1);
942         base = kmap_atomic(page, KM_USER0);
943         set_bit(bit, base);
944         kunmap_atomic(base, KM_USER0);
945         set_page_dirty_lock(page);
946         put_page(page);
947         return 0;
948 }
949
950 static int log_write(void __user *log_base,
951                      u64 write_address, u64 write_length)
952 {
953         u64 write_page = write_address / VHOST_PAGE_SIZE;
954         int r;
955
956         if (!write_length)
957                 return 0;
958         write_length += write_address % VHOST_PAGE_SIZE;
959         for (;;) {
960                 u64 base = (u64)(unsigned long)log_base;
961                 u64 log = base + write_page / 8;
962                 int bit = write_page % 8;
963                 if ((u64)(unsigned long)log != log)
964                         return -EFAULT;
965                 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
966                 if (r < 0)
967                         return r;
968                 if (write_length <= VHOST_PAGE_SIZE)
969                         break;
970                 write_length -= VHOST_PAGE_SIZE;
971                 write_page += 1;
972         }
973         return r;
974 }
975
976 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
977                     unsigned int log_num, u64 len)
978 {
979         int i, r;
980
981         /* Make sure data written is seen before log. */
982         smp_wmb();
983         for (i = 0; i < log_num; ++i) {
984                 u64 l = min(log[i].len, len);
985                 r = log_write(vq->log_base, log[i].addr, l);
986                 if (r < 0)
987                         return r;
988                 len -= l;
989                 if (!len) {
990                         if (vq->log_ctx)
991                                 eventfd_signal(vq->log_ctx, 1);
992                         return 0;
993                 }
994         }
995         /* Length written exceeds what we have stored. This is a bug. */
996         BUG();
997         return 0;
998 }
999
1000 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1001 {
1002         void __user *used;
1003         if (__put_user(vq->used_flags, &vq->used->flags) < 0)
1004                 return -EFAULT;
1005         if (unlikely(vq->log_used)) {
1006                 /* Make sure the flag is seen before log. */
1007                 smp_wmb();
1008                 /* Log used flag write. */
1009                 used = &vq->used->flags;
1010                 log_write(vq->log_base, vq->log_addr +
1011                           (used - (void __user *)vq->used),
1012                           sizeof vq->used->flags);
1013                 if (vq->log_ctx)
1014                         eventfd_signal(vq->log_ctx, 1);
1015         }
1016         return 0;
1017 }
1018
1019 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1020 {
1021         if (__put_user(vq->avail_idx, vhost_avail_event(vq)))
1022                 return -EFAULT;
1023         if (unlikely(vq->log_used)) {
1024                 void __user *used;
1025                 /* Make sure the event is seen before log. */
1026                 smp_wmb();
1027                 /* Log avail event write */
1028                 used = vhost_avail_event(vq);
1029                 log_write(vq->log_base, vq->log_addr +
1030                           (used - (void __user *)vq->used),
1031                           sizeof *vhost_avail_event(vq));
1032                 if (vq->log_ctx)
1033                         eventfd_signal(vq->log_ctx, 1);
1034         }
1035         return 0;
1036 }
1037
1038 int vhost_init_used(struct vhost_virtqueue *vq)
1039 {
1040         int r;
1041         if (!vq->private_data)
1042                 return 0;
1043
1044         r = vhost_update_used_flags(vq);
1045         if (r)
1046                 return r;
1047         vq->signalled_used_valid = false;
1048         return get_user(vq->last_used_idx, &vq->used->idx);
1049 }
1050
1051 static int translate_desc(struct vhost_dev *dev, u64 addr, u32 len,
1052                           struct iovec iov[], int iov_size)
1053 {
1054         const struct vhost_memory_region *reg;
1055         struct vhost_memory *mem;
1056         struct iovec *_iov;
1057         u64 s = 0;
1058         int ret = 0;
1059
1060         rcu_read_lock();
1061
1062         mem = rcu_dereference(dev->memory);
1063         while ((u64)len > s) {
1064                 u64 size;
1065                 if (unlikely(ret >= iov_size)) {
1066                         ret = -ENOBUFS;
1067                         break;
1068                 }
1069                 reg = find_region(mem, addr, len);
1070                 if (unlikely(!reg)) {
1071                         ret = -EFAULT;
1072                         break;
1073                 }
1074                 _iov = iov + ret;
1075                 size = reg->memory_size - addr + reg->guest_phys_addr;
1076                 _iov->iov_len = min((u64)len, size);
1077                 _iov->iov_base = (void __user *)(unsigned long)
1078                         (reg->userspace_addr + addr - reg->guest_phys_addr);
1079                 s += size;
1080                 addr += size;
1081                 ++ret;
1082         }
1083
1084         rcu_read_unlock();
1085         return ret;
1086 }
1087
1088 /* Each buffer in the virtqueues is actually a chain of descriptors.  This
1089  * function returns the next descriptor in the chain,
1090  * or -1U if we're at the end. */
1091 static unsigned next_desc(struct vring_desc *desc)
1092 {
1093         unsigned int next;
1094
1095         /* If this descriptor says it doesn't chain, we're done. */
1096         if (!(desc->flags & VRING_DESC_F_NEXT))
1097                 return -1U;
1098
1099         /* Check they're not leading us off end of descriptors. */
1100         next = desc->next;
1101         /* Make sure compiler knows to grab that: we don't want it changing! */
1102         /* We will use the result as an index in an array, so most
1103          * architectures only need a compiler barrier here. */
1104         read_barrier_depends();
1105
1106         return next;
1107 }
1108
1109 static int get_indirect(struct vhost_dev *dev, struct vhost_virtqueue *vq,
1110                         struct iovec iov[], unsigned int iov_size,
1111                         unsigned int *out_num, unsigned int *in_num,
1112                         struct vhost_log *log, unsigned int *log_num,
1113                         struct vring_desc *indirect)
1114 {
1115         struct vring_desc desc;
1116         unsigned int i = 0, count, found = 0;
1117         int ret;
1118
1119         /* Sanity check */
1120         if (unlikely(indirect->len % sizeof desc)) {
1121                 vq_err(vq, "Invalid length in indirect descriptor: "
1122                        "len 0x%llx not multiple of 0x%zx\n",
1123                        (unsigned long long)indirect->len,
1124                        sizeof desc);
1125                 return -EINVAL;
1126         }
1127
1128         ret = translate_desc(dev, indirect->addr, indirect->len, vq->indirect,
1129                              UIO_MAXIOV);
1130         if (unlikely(ret < 0)) {
1131                 vq_err(vq, "Translation failure %d in indirect.\n", ret);
1132                 return ret;
1133         }
1134
1135         /* We will use the result as an address to read from, so most
1136          * architectures only need a compiler barrier here. */
1137         read_barrier_depends();
1138
1139         count = indirect->len / sizeof desc;
1140         /* Buffers are chained via a 16 bit next field, so
1141          * we can have at most 2^16 of these. */
1142         if (unlikely(count > USHRT_MAX + 1)) {
1143                 vq_err(vq, "Indirect buffer length too big: %d\n",
1144                        indirect->len);
1145                 return -E2BIG;
1146         }
1147
1148         do {
1149                 unsigned iov_count = *in_num + *out_num;
1150                 if (unlikely(++found > count)) {
1151                         vq_err(vq, "Loop detected: last one at %u "
1152                                "indirect size %u\n",
1153                                i, count);
1154                         return -EINVAL;
1155                 }
1156                 if (unlikely(memcpy_fromiovec((unsigned char *)&desc,
1157                                               vq->indirect, sizeof desc))) {
1158                         vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1159                                i, (size_t)indirect->addr + i * sizeof desc);
1160                         return -EINVAL;
1161                 }
1162                 if (unlikely(desc.flags & VRING_DESC_F_INDIRECT)) {
1163                         vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1164                                i, (size_t)indirect->addr + i * sizeof desc);
1165                         return -EINVAL;
1166                 }
1167
1168                 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
1169                                      iov_size - iov_count);
1170                 if (unlikely(ret < 0)) {
1171                         vq_err(vq, "Translation failure %d indirect idx %d\n",
1172                                ret, i);
1173                         return ret;
1174                 }
1175                 /* If this is an input descriptor, increment that count. */
1176                 if (desc.flags & VRING_DESC_F_WRITE) {
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, "Indirect descriptor "
1188                                        "has out after in: idx %d\n", i);
1189                                 return -EINVAL;
1190                         }
1191                         *out_num += ret;
1192                 }
1193         } while ((i = next_desc(&desc)) != -1);
1194         return 0;
1195 }
1196
1197 /* This looks in the virtqueue and for the first available buffer, and converts
1198  * it to an iovec for convenient access.  Since descriptors consist of some
1199  * number of output then some number of input descriptors, it's actually two
1200  * iovecs, but we pack them into one and note how many of each there were.
1201  *
1202  * This function returns the descriptor number found, or vq->num (which is
1203  * never a valid descriptor number) if none was found.  A negative code is
1204  * returned on error. */
1205 int vhost_get_vq_desc(struct vhost_dev *dev, struct vhost_virtqueue *vq,
1206                       struct iovec iov[], unsigned int iov_size,
1207                       unsigned int *out_num, unsigned int *in_num,
1208                       struct vhost_log *log, unsigned int *log_num)
1209 {
1210         struct vring_desc desc;
1211         unsigned int i, head, found = 0;
1212         u16 last_avail_idx;
1213         int ret;
1214
1215         /* Check it isn't doing very strange things with descriptor numbers. */
1216         last_avail_idx = vq->last_avail_idx;
1217         if (unlikely(__get_user(vq->avail_idx, &vq->avail->idx))) {
1218                 vq_err(vq, "Failed to access avail idx at %p\n",
1219                        &vq->avail->idx);
1220                 return -EFAULT;
1221         }
1222
1223         if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
1224                 vq_err(vq, "Guest moved used index from %u to %u",
1225                        last_avail_idx, vq->avail_idx);
1226                 return -EFAULT;
1227         }
1228
1229         /* If there's nothing new since last we looked, return invalid. */
1230         if (vq->avail_idx == last_avail_idx)
1231                 return vq->num;
1232
1233         /* Only get avail ring entries after they have been exposed by guest. */
1234         smp_rmb();
1235
1236         /* Grab the next descriptor number they're advertising, and increment
1237          * the index we've seen. */
1238         if (unlikely(__get_user(head,
1239                                 &vq->avail->ring[last_avail_idx % vq->num]))) {
1240                 vq_err(vq, "Failed to read head: idx %d address %p\n",
1241                        last_avail_idx,
1242                        &vq->avail->ring[last_avail_idx % vq->num]);
1243                 return -EFAULT;
1244         }
1245
1246         /* If their number is silly, that's an error. */
1247         if (unlikely(head >= vq->num)) {
1248                 vq_err(vq, "Guest says index %u > %u is available",
1249                        head, vq->num);
1250                 return -EINVAL;
1251         }
1252
1253         /* When we start there are none of either input nor output. */
1254         *out_num = *in_num = 0;
1255         if (unlikely(log))
1256                 *log_num = 0;
1257
1258         i = head;
1259         do {
1260                 unsigned iov_count = *in_num + *out_num;
1261                 if (unlikely(i >= vq->num)) {
1262                         vq_err(vq, "Desc index is %u > %u, head = %u",
1263                                i, vq->num, head);
1264                         return -EINVAL;
1265                 }
1266                 if (unlikely(++found > vq->num)) {
1267                         vq_err(vq, "Loop detected: last one at %u "
1268                                "vq size %u head %u\n",
1269                                i, vq->num, head);
1270                         return -EINVAL;
1271                 }
1272                 ret = __copy_from_user(&desc, vq->desc + i, sizeof desc);
1273                 if (unlikely(ret)) {
1274                         vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
1275                                i, vq->desc + i);
1276                         return -EFAULT;
1277                 }
1278                 if (desc.flags & VRING_DESC_F_INDIRECT) {
1279                         ret = get_indirect(dev, vq, iov, iov_size,
1280                                            out_num, in_num,
1281                                            log, log_num, &desc);
1282                         if (unlikely(ret < 0)) {
1283                                 vq_err(vq, "Failure detected "
1284                                        "in indirect descriptor at idx %d\n", i);
1285                                 return ret;
1286                         }
1287                         continue;
1288                 }
1289
1290                 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
1291                                      iov_size - iov_count);
1292                 if (unlikely(ret < 0)) {
1293                         vq_err(vq, "Translation failure %d descriptor idx %d\n",
1294                                ret, i);
1295                         return ret;
1296                 }
1297                 if (desc.flags & VRING_DESC_F_WRITE) {
1298                         /* If this is an input descriptor,
1299                          * increment that count. */
1300                         *in_num += ret;
1301                         if (unlikely(log)) {
1302                                 log[*log_num].addr = desc.addr;
1303                                 log[*log_num].len = desc.len;
1304                                 ++*log_num;
1305                         }
1306                 } else {
1307                         /* If it's an output descriptor, they're all supposed
1308                          * to come before any input descriptors. */
1309                         if (unlikely(*in_num)) {
1310                                 vq_err(vq, "Descriptor has out after in: "
1311                                        "idx %d\n", i);
1312                                 return -EINVAL;
1313                         }
1314                         *out_num += ret;
1315                 }
1316         } while ((i = next_desc(&desc)) != -1);
1317
1318         /* On success, increment avail index. */
1319         vq->last_avail_idx++;
1320
1321         /* Assume notifications from guest are disabled at this point,
1322          * if they aren't we would need to update avail_event index. */
1323         BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
1324         return head;
1325 }
1326
1327 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
1328 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
1329 {
1330         vq->last_avail_idx -= n;
1331 }
1332
1333 /* After we've used one of their buffers, we tell them about it.  We'll then
1334  * want to notify the guest, using eventfd. */
1335 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
1336 {
1337         struct vring_used_elem __user *used;
1338
1339         /* The virtqueue contains a ring of used buffers.  Get a pointer to the
1340          * next entry in that used ring. */
1341         used = &vq->used->ring[vq->last_used_idx % vq->num];
1342         if (__put_user(head, &used->id)) {
1343                 vq_err(vq, "Failed to write used id");
1344                 return -EFAULT;
1345         }
1346         if (__put_user(len, &used->len)) {
1347                 vq_err(vq, "Failed to write used len");
1348                 return -EFAULT;
1349         }
1350         /* Make sure buffer is written before we update index. */
1351         smp_wmb();
1352         if (__put_user(vq->last_used_idx + 1, &vq->used->idx)) {
1353                 vq_err(vq, "Failed to increment used idx");
1354                 return -EFAULT;
1355         }
1356         if (unlikely(vq->log_used)) {
1357                 /* Make sure data is seen before log. */
1358                 smp_wmb();
1359                 /* Log used ring entry write. */
1360                 log_write(vq->log_base,
1361                           vq->log_addr +
1362                            ((void __user *)used - (void __user *)vq->used),
1363                           sizeof *used);
1364                 /* Log used index update. */
1365                 log_write(vq->log_base,
1366                           vq->log_addr + offsetof(struct vring_used, idx),
1367                           sizeof vq->used->idx);
1368                 if (vq->log_ctx)
1369                         eventfd_signal(vq->log_ctx, 1);
1370         }
1371         vq->last_used_idx++;
1372         /* If the driver never bothers to signal in a very long while,
1373          * used index might wrap around. If that happens, invalidate
1374          * signalled_used index we stored. TODO: make sure driver
1375          * signals at least once in 2^16 and remove this. */
1376         if (unlikely(vq->last_used_idx == vq->signalled_used))
1377                 vq->signalled_used_valid = false;
1378         return 0;
1379 }
1380
1381 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
1382                             struct vring_used_elem *heads,
1383                             unsigned count)
1384 {
1385         struct vring_used_elem __user *used;
1386         u16 old, new;
1387         int start;
1388
1389         start = vq->last_used_idx % vq->num;
1390         used = vq->used->ring + start;
1391         if (__copy_to_user(used, heads, count * sizeof *used)) {
1392                 vq_err(vq, "Failed to write used");
1393                 return -EFAULT;
1394         }
1395         if (unlikely(vq->log_used)) {
1396                 /* Make sure data is seen before log. */
1397                 smp_wmb();
1398                 /* Log used ring entry write. */
1399                 log_write(vq->log_base,
1400                           vq->log_addr +
1401                            ((void __user *)used - (void __user *)vq->used),
1402                           count * sizeof *used);
1403         }
1404         old = vq->last_used_idx;
1405         new = (vq->last_used_idx += count);
1406         /* If the driver never bothers to signal in a very long while,
1407          * used index might wrap around. If that happens, invalidate
1408          * signalled_used index we stored. TODO: make sure driver
1409          * signals at least once in 2^16 and remove this. */
1410         if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
1411                 vq->signalled_used_valid = false;
1412         return 0;
1413 }
1414
1415 /* After we've used one of their buffers, we tell them about it.  We'll then
1416  * want to notify the guest, using eventfd. */
1417 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
1418                      unsigned count)
1419 {
1420         int start, n, r;
1421
1422         start = vq->last_used_idx % vq->num;
1423         n = vq->num - start;
1424         if (n < count) {
1425                 r = __vhost_add_used_n(vq, heads, n);
1426                 if (r < 0)
1427                         return r;
1428                 heads += n;
1429                 count -= n;
1430         }
1431         r = __vhost_add_used_n(vq, heads, count);
1432
1433         /* Make sure buffer is written before we update index. */
1434         smp_wmb();
1435         if (put_user(vq->last_used_idx, &vq->used->idx)) {
1436                 vq_err(vq, "Failed to increment used idx");
1437                 return -EFAULT;
1438         }
1439         if (unlikely(vq->log_used)) {
1440                 /* Log used index update. */
1441                 log_write(vq->log_base,
1442                           vq->log_addr + offsetof(struct vring_used, idx),
1443                           sizeof vq->used->idx);
1444                 if (vq->log_ctx)
1445                         eventfd_signal(vq->log_ctx, 1);
1446         }
1447         return r;
1448 }
1449
1450 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1451 {
1452         __u16 old, new, event;
1453         bool v;
1454         /* Flush out used index updates. This is paired
1455          * with the barrier that the Guest executes when enabling
1456          * interrupts. */
1457         smp_mb();
1458
1459         if (vhost_has_feature(dev, VIRTIO_F_NOTIFY_ON_EMPTY) &&
1460             unlikely(vq->avail_idx == vq->last_avail_idx))
1461                 return true;
1462
1463         if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1464                 __u16 flags;
1465                 if (__get_user(flags, &vq->avail->flags)) {
1466                         vq_err(vq, "Failed to get flags");
1467                         return true;
1468                 }
1469                 return !(flags & VRING_AVAIL_F_NO_INTERRUPT);
1470         }
1471         old = vq->signalled_used;
1472         v = vq->signalled_used_valid;
1473         new = vq->signalled_used = vq->last_used_idx;
1474         vq->signalled_used_valid = true;
1475
1476         if (unlikely(!v))
1477                 return true;
1478
1479         if (get_user(event, vhost_used_event(vq))) {
1480                 vq_err(vq, "Failed to get used event idx");
1481                 return true;
1482         }
1483         return vring_need_event(event, new, old);
1484 }
1485
1486 /* This actually signals the guest, using eventfd. */
1487 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1488 {
1489         /* Signal the Guest tell them we used something up. */
1490         if (vq->call_ctx && vhost_notify(dev, vq))
1491                 eventfd_signal(vq->call_ctx, 1);
1492 }
1493
1494 /* And here's the combo meal deal.  Supersize me! */
1495 void vhost_add_used_and_signal(struct vhost_dev *dev,
1496                                struct vhost_virtqueue *vq,
1497                                unsigned int head, int len)
1498 {
1499         vhost_add_used(vq, head, len);
1500         vhost_signal(dev, vq);
1501 }
1502
1503 /* multi-buffer version of vhost_add_used_and_signal */
1504 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
1505                                  struct vhost_virtqueue *vq,
1506                                  struct vring_used_elem *heads, unsigned count)
1507 {
1508         vhost_add_used_n(vq, heads, count);
1509         vhost_signal(dev, vq);
1510 }
1511
1512 /* OK, now we need to know about added descriptors. */
1513 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1514 {
1515         u16 avail_idx;
1516         int r;
1517
1518         if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
1519                 return false;
1520         vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
1521         if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1522                 r = vhost_update_used_flags(vq);
1523                 if (r) {
1524                         vq_err(vq, "Failed to enable notification at %p: %d\n",
1525                                &vq->used->flags, r);
1526                         return false;
1527                 }
1528         } else {
1529                 r = vhost_update_avail_event(vq, vq->avail_idx);
1530                 if (r) {
1531                         vq_err(vq, "Failed to update avail event index at %p: %d\n",
1532                                vhost_avail_event(vq), r);
1533                         return false;
1534                 }
1535         }
1536         /* They could have slipped one in as we were doing that: make
1537          * sure it's written, then check again. */
1538         smp_mb();
1539         r = __get_user(avail_idx, &vq->avail->idx);
1540         if (r) {
1541                 vq_err(vq, "Failed to check avail idx at %p: %d\n",
1542                        &vq->avail->idx, r);
1543                 return false;
1544         }
1545
1546         return avail_idx != vq->avail_idx;
1547 }
1548
1549 /* We don't need to be notified again. */
1550 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1551 {
1552         int r;
1553
1554         if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
1555                 return;
1556         vq->used_flags |= VRING_USED_F_NO_NOTIFY;
1557         if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1558                 r = vhost_update_used_flags(vq);
1559                 if (r)
1560                         vq_err(vq, "Failed to enable notification at %p: %d\n",
1561                                &vq->used->flags, r);
1562         }
1563 }
1564
1565 static void vhost_zerocopy_done_signal(struct kref *kref)
1566 {
1567         struct vhost_ubuf_ref *ubufs = container_of(kref, struct vhost_ubuf_ref,
1568                                                     kref);
1569         wake_up(&ubufs->wait);
1570 }
1571
1572 struct vhost_ubuf_ref *vhost_ubuf_alloc(struct vhost_virtqueue *vq,
1573                                         bool zcopy)
1574 {
1575         struct vhost_ubuf_ref *ubufs;
1576         /* No zero copy backend? Nothing to count. */
1577         if (!zcopy)
1578                 return NULL;
1579         ubufs = kmalloc(sizeof *ubufs, GFP_KERNEL);
1580         if (!ubufs)
1581                 return ERR_PTR(-ENOMEM);
1582         kref_init(&ubufs->kref);
1583         init_waitqueue_head(&ubufs->wait);
1584         ubufs->vq = vq;
1585         return ubufs;
1586 }
1587
1588 void vhost_ubuf_put(struct vhost_ubuf_ref *ubufs)
1589 {
1590         kref_put(&ubufs->kref, vhost_zerocopy_done_signal);
1591 }
1592
1593 void vhost_ubuf_put_and_wait(struct vhost_ubuf_ref *ubufs)
1594 {
1595         kref_put(&ubufs->kref, vhost_zerocopy_done_signal);
1596         wait_event(ubufs->wait, !atomic_read(&ubufs->kref.refcount));
1597         kfree(ubufs);
1598 }
1599
1600 void vhost_zerocopy_callback(void *arg)
1601 {
1602         struct ubuf_info *ubuf = arg;
1603         struct vhost_ubuf_ref *ubufs = ubuf->arg;
1604         struct vhost_virtqueue *vq = ubufs->vq;
1605
1606         /* set len = 1 to mark this desc buffers done DMA */
1607         vq->heads[ubuf->desc].len = VHOST_DMA_DONE_LEN;
1608         kref_put(&ubufs->kref, vhost_zerocopy_done_signal);
1609 }