Merge tags 'cris-for-4.16' and 'cris-for-4.16-urgent' of git://git.kernel.org/pub...
[muen/linux.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/uio.h>
17 #include <linux/mm.h>
18 #include <linux/mmu_context.h>
19 #include <linux/miscdevice.h>
20 #include <linux/mutex.h>
21 #include <linux/poll.h>
22 #include <linux/file.h>
23 #include <linux/highmem.h>
24 #include <linux/slab.h>
25 #include <linux/vmalloc.h>
26 #include <linux/kthread.h>
27 #include <linux/cgroup.h>
28 #include <linux/module.h>
29 #include <linux/sort.h>
30 #include <linux/sched/mm.h>
31 #include <linux/sched/signal.h>
32 #include <linux/interval_tree_generic.h>
33
34 #include "vhost.h"
35
36 static ushort max_mem_regions = 64;
37 module_param(max_mem_regions, ushort, 0444);
38 MODULE_PARM_DESC(max_mem_regions,
39         "Maximum number of memory regions in memory map. (default: 64)");
40 static int max_iotlb_entries = 2048;
41 module_param(max_iotlb_entries, int, 0444);
42 MODULE_PARM_DESC(max_iotlb_entries,
43         "Maximum number of iotlb entries. (default: 2048)");
44
45 enum {
46         VHOST_MEMORY_F_LOG = 0x1,
47 };
48
49 #define vhost_used_event(vq) ((__virtio16 __user *)&vq->avail->ring[vq->num])
50 #define vhost_avail_event(vq) ((__virtio16 __user *)&vq->used->ring[vq->num])
51
52 INTERVAL_TREE_DEFINE(struct vhost_umem_node,
53                      rb, __u64, __subtree_last,
54                      START, LAST, static inline, vhost_umem_interval_tree);
55
56 #ifdef CONFIG_VHOST_CROSS_ENDIAN_LEGACY
57 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
58 {
59         vq->user_be = !virtio_legacy_is_little_endian();
60 }
61
62 static void vhost_enable_cross_endian_big(struct vhost_virtqueue *vq)
63 {
64         vq->user_be = true;
65 }
66
67 static void vhost_enable_cross_endian_little(struct vhost_virtqueue *vq)
68 {
69         vq->user_be = false;
70 }
71
72 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
73 {
74         struct vhost_vring_state s;
75
76         if (vq->private_data)
77                 return -EBUSY;
78
79         if (copy_from_user(&s, argp, sizeof(s)))
80                 return -EFAULT;
81
82         if (s.num != VHOST_VRING_LITTLE_ENDIAN &&
83             s.num != VHOST_VRING_BIG_ENDIAN)
84                 return -EINVAL;
85
86         if (s.num == VHOST_VRING_BIG_ENDIAN)
87                 vhost_enable_cross_endian_big(vq);
88         else
89                 vhost_enable_cross_endian_little(vq);
90
91         return 0;
92 }
93
94 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
95                                    int __user *argp)
96 {
97         struct vhost_vring_state s = {
98                 .index = idx,
99                 .num = vq->user_be
100         };
101
102         if (copy_to_user(argp, &s, sizeof(s)))
103                 return -EFAULT;
104
105         return 0;
106 }
107
108 static void vhost_init_is_le(struct vhost_virtqueue *vq)
109 {
110         /* Note for legacy virtio: user_be is initialized at reset time
111          * according to the host endianness. If userspace does not set an
112          * explicit endianness, the default behavior is native endian, as
113          * expected by legacy virtio.
114          */
115         vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1) || !vq->user_be;
116 }
117 #else
118 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
119 {
120 }
121
122 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
123 {
124         return -ENOIOCTLCMD;
125 }
126
127 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
128                                    int __user *argp)
129 {
130         return -ENOIOCTLCMD;
131 }
132
133 static void vhost_init_is_le(struct vhost_virtqueue *vq)
134 {
135         vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1)
136                 || virtio_legacy_is_little_endian();
137 }
138 #endif /* CONFIG_VHOST_CROSS_ENDIAN_LEGACY */
139
140 static void vhost_reset_is_le(struct vhost_virtqueue *vq)
141 {
142         vhost_init_is_le(vq);
143 }
144
145 struct vhost_flush_struct {
146         struct vhost_work work;
147         struct completion wait_event;
148 };
149
150 static void vhost_flush_work(struct vhost_work *work)
151 {
152         struct vhost_flush_struct *s;
153
154         s = container_of(work, struct vhost_flush_struct, work);
155         complete(&s->wait_event);
156 }
157
158 static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
159                             poll_table *pt)
160 {
161         struct vhost_poll *poll;
162
163         poll = container_of(pt, struct vhost_poll, table);
164         poll->wqh = wqh;
165         add_wait_queue(wqh, &poll->wait);
166 }
167
168 static int vhost_poll_wakeup(wait_queue_entry_t *wait, unsigned mode, int sync,
169                              void *key)
170 {
171         struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
172
173         if (!(key_to_poll(key) & poll->mask))
174                 return 0;
175
176         vhost_poll_queue(poll);
177         return 0;
178 }
179
180 void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
181 {
182         clear_bit(VHOST_WORK_QUEUED, &work->flags);
183         work->fn = fn;
184         init_waitqueue_head(&work->done);
185 }
186 EXPORT_SYMBOL_GPL(vhost_work_init);
187
188 /* Init poll structure */
189 void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
190                      __poll_t mask, struct vhost_dev *dev)
191 {
192         init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
193         init_poll_funcptr(&poll->table, vhost_poll_func);
194         poll->mask = mask;
195         poll->dev = dev;
196         poll->wqh = NULL;
197
198         vhost_work_init(&poll->work, fn);
199 }
200 EXPORT_SYMBOL_GPL(vhost_poll_init);
201
202 /* Start polling a file. We add ourselves to file's wait queue. The caller must
203  * keep a reference to a file until after vhost_poll_stop is called. */
204 int vhost_poll_start(struct vhost_poll *poll, struct file *file)
205 {
206         __poll_t mask;
207         int ret = 0;
208
209         if (poll->wqh)
210                 return 0;
211
212         mask = file->f_op->poll(file, &poll->table);
213         if (mask)
214                 vhost_poll_wakeup(&poll->wait, 0, 0, poll_to_key(mask));
215         if (mask & POLLERR) {
216                 if (poll->wqh)
217                         remove_wait_queue(poll->wqh, &poll->wait);
218                 ret = -EINVAL;
219         }
220
221         return ret;
222 }
223 EXPORT_SYMBOL_GPL(vhost_poll_start);
224
225 /* Stop polling a file. After this function returns, it becomes safe to drop the
226  * file reference. You must also flush afterwards. */
227 void vhost_poll_stop(struct vhost_poll *poll)
228 {
229         if (poll->wqh) {
230                 remove_wait_queue(poll->wqh, &poll->wait);
231                 poll->wqh = NULL;
232         }
233 }
234 EXPORT_SYMBOL_GPL(vhost_poll_stop);
235
236 void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work)
237 {
238         struct vhost_flush_struct flush;
239
240         if (dev->worker) {
241                 init_completion(&flush.wait_event);
242                 vhost_work_init(&flush.work, vhost_flush_work);
243
244                 vhost_work_queue(dev, &flush.work);
245                 wait_for_completion(&flush.wait_event);
246         }
247 }
248 EXPORT_SYMBOL_GPL(vhost_work_flush);
249
250 /* Flush any work that has been scheduled. When calling this, don't hold any
251  * locks that are also used by the callback. */
252 void vhost_poll_flush(struct vhost_poll *poll)
253 {
254         vhost_work_flush(poll->dev, &poll->work);
255 }
256 EXPORT_SYMBOL_GPL(vhost_poll_flush);
257
258 void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work)
259 {
260         if (!dev->worker)
261                 return;
262
263         if (!test_and_set_bit(VHOST_WORK_QUEUED, &work->flags)) {
264                 /* We can only add the work to the list after we're
265                  * sure it was not in the list.
266                  * test_and_set_bit() implies a memory barrier.
267                  */
268                 llist_add(&work->node, &dev->work_list);
269                 wake_up_process(dev->worker);
270         }
271 }
272 EXPORT_SYMBOL_GPL(vhost_work_queue);
273
274 /* A lockless hint for busy polling code to exit the loop */
275 bool vhost_has_work(struct vhost_dev *dev)
276 {
277         return !llist_empty(&dev->work_list);
278 }
279 EXPORT_SYMBOL_GPL(vhost_has_work);
280
281 void vhost_poll_queue(struct vhost_poll *poll)
282 {
283         vhost_work_queue(poll->dev, &poll->work);
284 }
285 EXPORT_SYMBOL_GPL(vhost_poll_queue);
286
287 static void __vhost_vq_meta_reset(struct vhost_virtqueue *vq)
288 {
289         int j;
290
291         for (j = 0; j < VHOST_NUM_ADDRS; j++)
292                 vq->meta_iotlb[j] = NULL;
293 }
294
295 static void vhost_vq_meta_reset(struct vhost_dev *d)
296 {
297         int i;
298
299         for (i = 0; i < d->nvqs; ++i)
300                 __vhost_vq_meta_reset(d->vqs[i]);
301 }
302
303 static void vhost_vq_reset(struct vhost_dev *dev,
304                            struct vhost_virtqueue *vq)
305 {
306         vq->num = 1;
307         vq->desc = NULL;
308         vq->avail = NULL;
309         vq->used = NULL;
310         vq->last_avail_idx = 0;
311         vq->avail_idx = 0;
312         vq->last_used_idx = 0;
313         vq->signalled_used = 0;
314         vq->signalled_used_valid = false;
315         vq->used_flags = 0;
316         vq->log_used = false;
317         vq->log_addr = -1ull;
318         vq->private_data = NULL;
319         vq->acked_features = 0;
320         vq->log_base = NULL;
321         vq->error_ctx = NULL;
322         vq->error = NULL;
323         vq->kick = NULL;
324         vq->call_ctx = NULL;
325         vq->call = NULL;
326         vq->log_ctx = NULL;
327         vhost_reset_is_le(vq);
328         vhost_disable_cross_endian(vq);
329         vq->busyloop_timeout = 0;
330         vq->umem = NULL;
331         vq->iotlb = NULL;
332         __vhost_vq_meta_reset(vq);
333 }
334
335 static int vhost_worker(void *data)
336 {
337         struct vhost_dev *dev = data;
338         struct vhost_work *work, *work_next;
339         struct llist_node *node;
340         mm_segment_t oldfs = get_fs();
341
342         set_fs(USER_DS);
343         use_mm(dev->mm);
344
345         for (;;) {
346                 /* mb paired w/ kthread_stop */
347                 set_current_state(TASK_INTERRUPTIBLE);
348
349                 if (kthread_should_stop()) {
350                         __set_current_state(TASK_RUNNING);
351                         break;
352                 }
353
354                 node = llist_del_all(&dev->work_list);
355                 if (!node)
356                         schedule();
357
358                 node = llist_reverse_order(node);
359                 /* make sure flag is seen after deletion */
360                 smp_wmb();
361                 llist_for_each_entry_safe(work, work_next, node, node) {
362                         clear_bit(VHOST_WORK_QUEUED, &work->flags);
363                         __set_current_state(TASK_RUNNING);
364                         work->fn(work);
365                         if (need_resched())
366                                 schedule();
367                 }
368         }
369         unuse_mm(dev->mm);
370         set_fs(oldfs);
371         return 0;
372 }
373
374 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
375 {
376         kfree(vq->indirect);
377         vq->indirect = NULL;
378         kfree(vq->log);
379         vq->log = NULL;
380         kfree(vq->heads);
381         vq->heads = NULL;
382 }
383
384 /* Helper to allocate iovec buffers for all vqs. */
385 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
386 {
387         struct vhost_virtqueue *vq;
388         int i;
389
390         for (i = 0; i < dev->nvqs; ++i) {
391                 vq = dev->vqs[i];
392                 vq->indirect = kmalloc(sizeof *vq->indirect * UIO_MAXIOV,
393                                        GFP_KERNEL);
394                 vq->log = kmalloc(sizeof *vq->log * UIO_MAXIOV, GFP_KERNEL);
395                 vq->heads = kmalloc(sizeof *vq->heads * UIO_MAXIOV, GFP_KERNEL);
396                 if (!vq->indirect || !vq->log || !vq->heads)
397                         goto err_nomem;
398         }
399         return 0;
400
401 err_nomem:
402         for (; i >= 0; --i)
403                 vhost_vq_free_iovecs(dev->vqs[i]);
404         return -ENOMEM;
405 }
406
407 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
408 {
409         int i;
410
411         for (i = 0; i < dev->nvqs; ++i)
412                 vhost_vq_free_iovecs(dev->vqs[i]);
413 }
414
415 void vhost_dev_init(struct vhost_dev *dev,
416                     struct vhost_virtqueue **vqs, int nvqs)
417 {
418         struct vhost_virtqueue *vq;
419         int i;
420
421         dev->vqs = vqs;
422         dev->nvqs = nvqs;
423         mutex_init(&dev->mutex);
424         dev->log_ctx = NULL;
425         dev->log_file = NULL;
426         dev->umem = NULL;
427         dev->iotlb = NULL;
428         dev->mm = NULL;
429         dev->worker = NULL;
430         init_llist_head(&dev->work_list);
431         init_waitqueue_head(&dev->wait);
432         INIT_LIST_HEAD(&dev->read_list);
433         INIT_LIST_HEAD(&dev->pending_list);
434         spin_lock_init(&dev->iotlb_lock);
435
436
437         for (i = 0; i < dev->nvqs; ++i) {
438                 vq = dev->vqs[i];
439                 vq->log = NULL;
440                 vq->indirect = NULL;
441                 vq->heads = NULL;
442                 vq->dev = dev;
443                 mutex_init(&vq->mutex);
444                 vhost_vq_reset(dev, vq);
445                 if (vq->handle_kick)
446                         vhost_poll_init(&vq->poll, vq->handle_kick,
447                                         POLLIN, dev);
448         }
449 }
450 EXPORT_SYMBOL_GPL(vhost_dev_init);
451
452 /* Caller should have device mutex */
453 long vhost_dev_check_owner(struct vhost_dev *dev)
454 {
455         /* Are you the owner? If not, I don't think you mean to do that */
456         return dev->mm == current->mm ? 0 : -EPERM;
457 }
458 EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
459
460 struct vhost_attach_cgroups_struct {
461         struct vhost_work work;
462         struct task_struct *owner;
463         int ret;
464 };
465
466 static void vhost_attach_cgroups_work(struct vhost_work *work)
467 {
468         struct vhost_attach_cgroups_struct *s;
469
470         s = container_of(work, struct vhost_attach_cgroups_struct, work);
471         s->ret = cgroup_attach_task_all(s->owner, current);
472 }
473
474 static int vhost_attach_cgroups(struct vhost_dev *dev)
475 {
476         struct vhost_attach_cgroups_struct attach;
477
478         attach.owner = current;
479         vhost_work_init(&attach.work, vhost_attach_cgroups_work);
480         vhost_work_queue(dev, &attach.work);
481         vhost_work_flush(dev, &attach.work);
482         return attach.ret;
483 }
484
485 /* Caller should have device mutex */
486 bool vhost_dev_has_owner(struct vhost_dev *dev)
487 {
488         return dev->mm;
489 }
490 EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
491
492 /* Caller should have device mutex */
493 long vhost_dev_set_owner(struct vhost_dev *dev)
494 {
495         struct task_struct *worker;
496         int err;
497
498         /* Is there an owner already? */
499         if (vhost_dev_has_owner(dev)) {
500                 err = -EBUSY;
501                 goto err_mm;
502         }
503
504         /* No owner, become one */
505         dev->mm = get_task_mm(current);
506         worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
507         if (IS_ERR(worker)) {
508                 err = PTR_ERR(worker);
509                 goto err_worker;
510         }
511
512         dev->worker = worker;
513         wake_up_process(worker);        /* avoid contributing to loadavg */
514
515         err = vhost_attach_cgroups(dev);
516         if (err)
517                 goto err_cgroup;
518
519         err = vhost_dev_alloc_iovecs(dev);
520         if (err)
521                 goto err_cgroup;
522
523         return 0;
524 err_cgroup:
525         kthread_stop(worker);
526         dev->worker = NULL;
527 err_worker:
528         if (dev->mm)
529                 mmput(dev->mm);
530         dev->mm = NULL;
531 err_mm:
532         return err;
533 }
534 EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
535
536 struct vhost_umem *vhost_dev_reset_owner_prepare(void)
537 {
538         return kvzalloc(sizeof(struct vhost_umem), GFP_KERNEL);
539 }
540 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
541
542 /* Caller should have device mutex */
543 void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_umem *umem)
544 {
545         int i;
546
547         vhost_dev_cleanup(dev, true);
548
549         /* Restore memory to default empty mapping. */
550         INIT_LIST_HEAD(&umem->umem_list);
551         dev->umem = umem;
552         /* We don't need VQ locks below since vhost_dev_cleanup makes sure
553          * VQs aren't running.
554          */
555         for (i = 0; i < dev->nvqs; ++i)
556                 dev->vqs[i]->umem = umem;
557 }
558 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
559
560 void vhost_dev_stop(struct vhost_dev *dev)
561 {
562         int i;
563
564         for (i = 0; i < dev->nvqs; ++i) {
565                 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick) {
566                         vhost_poll_stop(&dev->vqs[i]->poll);
567                         vhost_poll_flush(&dev->vqs[i]->poll);
568                 }
569         }
570 }
571 EXPORT_SYMBOL_GPL(vhost_dev_stop);
572
573 static void vhost_umem_free(struct vhost_umem *umem,
574                             struct vhost_umem_node *node)
575 {
576         vhost_umem_interval_tree_remove(node, &umem->umem_tree);
577         list_del(&node->link);
578         kfree(node);
579         umem->numem--;
580 }
581
582 static void vhost_umem_clean(struct vhost_umem *umem)
583 {
584         struct vhost_umem_node *node, *tmp;
585
586         if (!umem)
587                 return;
588
589         list_for_each_entry_safe(node, tmp, &umem->umem_list, link)
590                 vhost_umem_free(umem, node);
591
592         kvfree(umem);
593 }
594
595 static void vhost_clear_msg(struct vhost_dev *dev)
596 {
597         struct vhost_msg_node *node, *n;
598
599         spin_lock(&dev->iotlb_lock);
600
601         list_for_each_entry_safe(node, n, &dev->read_list, node) {
602                 list_del(&node->node);
603                 kfree(node);
604         }
605
606         list_for_each_entry_safe(node, n, &dev->pending_list, node) {
607                 list_del(&node->node);
608                 kfree(node);
609         }
610
611         spin_unlock(&dev->iotlb_lock);
612 }
613
614 /* Caller should have device mutex if and only if locked is set */
615 void vhost_dev_cleanup(struct vhost_dev *dev, bool locked)
616 {
617         int i;
618
619         for (i = 0; i < dev->nvqs; ++i) {
620                 if (dev->vqs[i]->error_ctx)
621                         eventfd_ctx_put(dev->vqs[i]->error_ctx);
622                 if (dev->vqs[i]->error)
623                         fput(dev->vqs[i]->error);
624                 if (dev->vqs[i]->kick)
625                         fput(dev->vqs[i]->kick);
626                 if (dev->vqs[i]->call_ctx)
627                         eventfd_ctx_put(dev->vqs[i]->call_ctx);
628                 if (dev->vqs[i]->call)
629                         fput(dev->vqs[i]->call);
630                 vhost_vq_reset(dev, dev->vqs[i]);
631         }
632         vhost_dev_free_iovecs(dev);
633         if (dev->log_ctx)
634                 eventfd_ctx_put(dev->log_ctx);
635         dev->log_ctx = NULL;
636         if (dev->log_file)
637                 fput(dev->log_file);
638         dev->log_file = NULL;
639         /* No one will access memory at this point */
640         vhost_umem_clean(dev->umem);
641         dev->umem = NULL;
642         vhost_umem_clean(dev->iotlb);
643         dev->iotlb = NULL;
644         vhost_clear_msg(dev);
645         wake_up_interruptible_poll(&dev->wait, POLLIN | POLLRDNORM);
646         WARN_ON(!llist_empty(&dev->work_list));
647         if (dev->worker) {
648                 kthread_stop(dev->worker);
649                 dev->worker = NULL;
650         }
651         if (dev->mm)
652                 mmput(dev->mm);
653         dev->mm = NULL;
654 }
655 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
656
657 static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
658 {
659         u64 a = addr / VHOST_PAGE_SIZE / 8;
660
661         /* Make sure 64 bit math will not overflow. */
662         if (a > ULONG_MAX - (unsigned long)log_base ||
663             a + (unsigned long)log_base > ULONG_MAX)
664                 return 0;
665
666         return access_ok(VERIFY_WRITE, log_base + a,
667                          (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
668 }
669
670 static bool vhost_overflow(u64 uaddr, u64 size)
671 {
672         /* Make sure 64 bit math will not overflow. */
673         return uaddr > ULONG_MAX || size > ULONG_MAX || uaddr > ULONG_MAX - size;
674 }
675
676 /* Caller should have vq mutex and device mutex. */
677 static int vq_memory_access_ok(void __user *log_base, struct vhost_umem *umem,
678                                int log_all)
679 {
680         struct vhost_umem_node *node;
681
682         if (!umem)
683                 return 0;
684
685         list_for_each_entry(node, &umem->umem_list, link) {
686                 unsigned long a = node->userspace_addr;
687
688                 if (vhost_overflow(node->userspace_addr, node->size))
689                         return 0;
690
691
692                 if (!access_ok(VERIFY_WRITE, (void __user *)a,
693                                     node->size))
694                         return 0;
695                 else if (log_all && !log_access_ok(log_base,
696                                                    node->start,
697                                                    node->size))
698                         return 0;
699         }
700         return 1;
701 }
702
703 static inline void __user *vhost_vq_meta_fetch(struct vhost_virtqueue *vq,
704                                                u64 addr, unsigned int size,
705                                                int type)
706 {
707         const struct vhost_umem_node *node = vq->meta_iotlb[type];
708
709         if (!node)
710                 return NULL;
711
712         return (void *)(uintptr_t)(node->userspace_addr + addr - node->start);
713 }
714
715 /* Can we switch to this memory table? */
716 /* Caller should have device mutex but not vq mutex */
717 static int memory_access_ok(struct vhost_dev *d, struct vhost_umem *umem,
718                             int log_all)
719 {
720         int i;
721
722         for (i = 0; i < d->nvqs; ++i) {
723                 int ok;
724                 bool log;
725
726                 mutex_lock(&d->vqs[i]->mutex);
727                 log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
728                 /* If ring is inactive, will check when it's enabled. */
729                 if (d->vqs[i]->private_data)
730                         ok = vq_memory_access_ok(d->vqs[i]->log_base,
731                                                  umem, log);
732                 else
733                         ok = 1;
734                 mutex_unlock(&d->vqs[i]->mutex);
735                 if (!ok)
736                         return 0;
737         }
738         return 1;
739 }
740
741 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
742                           struct iovec iov[], int iov_size, int access);
743
744 static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
745                               const void *from, unsigned size)
746 {
747         int ret;
748
749         if (!vq->iotlb)
750                 return __copy_to_user(to, from, size);
751         else {
752                 /* This function should be called after iotlb
753                  * prefetch, which means we're sure that all vq
754                  * could be access through iotlb. So -EAGAIN should
755                  * not happen in this case.
756                  */
757                 struct iov_iter t;
758                 void __user *uaddr = vhost_vq_meta_fetch(vq,
759                                      (u64)(uintptr_t)to, size,
760                                      VHOST_ADDR_DESC);
761
762                 if (uaddr)
763                         return __copy_to_user(uaddr, from, size);
764
765                 ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
766                                      ARRAY_SIZE(vq->iotlb_iov),
767                                      VHOST_ACCESS_WO);
768                 if (ret < 0)
769                         goto out;
770                 iov_iter_init(&t, WRITE, vq->iotlb_iov, ret, size);
771                 ret = copy_to_iter(from, size, &t);
772                 if (ret == size)
773                         ret = 0;
774         }
775 out:
776         return ret;
777 }
778
779 static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
780                                 void __user *from, unsigned size)
781 {
782         int ret;
783
784         if (!vq->iotlb)
785                 return __copy_from_user(to, from, size);
786         else {
787                 /* This function should be called after iotlb
788                  * prefetch, which means we're sure that vq
789                  * could be access through iotlb. So -EAGAIN should
790                  * not happen in this case.
791                  */
792                 void __user *uaddr = vhost_vq_meta_fetch(vq,
793                                      (u64)(uintptr_t)from, size,
794                                      VHOST_ADDR_DESC);
795                 struct iov_iter f;
796
797                 if (uaddr)
798                         return __copy_from_user(to, uaddr, size);
799
800                 ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
801                                      ARRAY_SIZE(vq->iotlb_iov),
802                                      VHOST_ACCESS_RO);
803                 if (ret < 0) {
804                         vq_err(vq, "IOTLB translation failure: uaddr "
805                                "%p size 0x%llx\n", from,
806                                (unsigned long long) size);
807                         goto out;
808                 }
809                 iov_iter_init(&f, READ, vq->iotlb_iov, ret, size);
810                 ret = copy_from_iter(to, size, &f);
811                 if (ret == size)
812                         ret = 0;
813         }
814
815 out:
816         return ret;
817 }
818
819 static void __user *__vhost_get_user_slow(struct vhost_virtqueue *vq,
820                                           void __user *addr, unsigned int size,
821                                           int type)
822 {
823         int ret;
824
825         ret = translate_desc(vq, (u64)(uintptr_t)addr, size, vq->iotlb_iov,
826                              ARRAY_SIZE(vq->iotlb_iov),
827                              VHOST_ACCESS_RO);
828         if (ret < 0) {
829                 vq_err(vq, "IOTLB translation failure: uaddr "
830                         "%p size 0x%llx\n", addr,
831                         (unsigned long long) size);
832                 return NULL;
833         }
834
835         if (ret != 1 || vq->iotlb_iov[0].iov_len != size) {
836                 vq_err(vq, "Non atomic userspace memory access: uaddr "
837                         "%p size 0x%llx\n", addr,
838                         (unsigned long long) size);
839                 return NULL;
840         }
841
842         return vq->iotlb_iov[0].iov_base;
843 }
844
845 /* This function should be called after iotlb
846  * prefetch, which means we're sure that vq
847  * could be access through iotlb. So -EAGAIN should
848  * not happen in this case.
849  */
850 static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
851                                             void *addr, unsigned int size,
852                                             int type)
853 {
854         void __user *uaddr = vhost_vq_meta_fetch(vq,
855                              (u64)(uintptr_t)addr, size, type);
856         if (uaddr)
857                 return uaddr;
858
859         return __vhost_get_user_slow(vq, addr, size, type);
860 }
861
862 #define vhost_put_user(vq, x, ptr)              \
863 ({ \
864         int ret = -EFAULT; \
865         if (!vq->iotlb) { \
866                 ret = __put_user(x, ptr); \
867         } else { \
868                 __typeof__(ptr) to = \
869                         (__typeof__(ptr)) __vhost_get_user(vq, ptr,     \
870                                           sizeof(*ptr), VHOST_ADDR_USED); \
871                 if (to != NULL) \
872                         ret = __put_user(x, to); \
873                 else \
874                         ret = -EFAULT;  \
875         } \
876         ret; \
877 })
878
879 #define vhost_get_user(vq, x, ptr, type)                \
880 ({ \
881         int ret; \
882         if (!vq->iotlb) { \
883                 ret = __get_user(x, ptr); \
884         } else { \
885                 __typeof__(ptr) from = \
886                         (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
887                                                            sizeof(*ptr), \
888                                                            type); \
889                 if (from != NULL) \
890                         ret = __get_user(x, from); \
891                 else \
892                         ret = -EFAULT; \
893         } \
894         ret; \
895 })
896
897 #define vhost_get_avail(vq, x, ptr) \
898         vhost_get_user(vq, x, ptr, VHOST_ADDR_AVAIL)
899
900 #define vhost_get_used(vq, x, ptr) \
901         vhost_get_user(vq, x, ptr, VHOST_ADDR_USED)
902
903 static void vhost_dev_lock_vqs(struct vhost_dev *d)
904 {
905         int i = 0;
906         for (i = 0; i < d->nvqs; ++i)
907                 mutex_lock_nested(&d->vqs[i]->mutex, i);
908 }
909
910 static void vhost_dev_unlock_vqs(struct vhost_dev *d)
911 {
912         int i = 0;
913         for (i = 0; i < d->nvqs; ++i)
914                 mutex_unlock(&d->vqs[i]->mutex);
915 }
916
917 static int vhost_new_umem_range(struct vhost_umem *umem,
918                                 u64 start, u64 size, u64 end,
919                                 u64 userspace_addr, int perm)
920 {
921         struct vhost_umem_node *tmp, *node = kmalloc(sizeof(*node), GFP_ATOMIC);
922
923         if (!node)
924                 return -ENOMEM;
925
926         if (umem->numem == max_iotlb_entries) {
927                 tmp = list_first_entry(&umem->umem_list, typeof(*tmp), link);
928                 vhost_umem_free(umem, tmp);
929         }
930
931         node->start = start;
932         node->size = size;
933         node->last = end;
934         node->userspace_addr = userspace_addr;
935         node->perm = perm;
936         INIT_LIST_HEAD(&node->link);
937         list_add_tail(&node->link, &umem->umem_list);
938         vhost_umem_interval_tree_insert(node, &umem->umem_tree);
939         umem->numem++;
940
941         return 0;
942 }
943
944 static void vhost_del_umem_range(struct vhost_umem *umem,
945                                  u64 start, u64 end)
946 {
947         struct vhost_umem_node *node;
948
949         while ((node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
950                                                            start, end)))
951                 vhost_umem_free(umem, node);
952 }
953
954 static void vhost_iotlb_notify_vq(struct vhost_dev *d,
955                                   struct vhost_iotlb_msg *msg)
956 {
957         struct vhost_msg_node *node, *n;
958
959         spin_lock(&d->iotlb_lock);
960
961         list_for_each_entry_safe(node, n, &d->pending_list, node) {
962                 struct vhost_iotlb_msg *vq_msg = &node->msg.iotlb;
963                 if (msg->iova <= vq_msg->iova &&
964                     msg->iova + msg->size - 1 > vq_msg->iova &&
965                     vq_msg->type == VHOST_IOTLB_MISS) {
966                         vhost_poll_queue(&node->vq->poll);
967                         list_del(&node->node);
968                         kfree(node);
969                 }
970         }
971
972         spin_unlock(&d->iotlb_lock);
973 }
974
975 static int umem_access_ok(u64 uaddr, u64 size, int access)
976 {
977         unsigned long a = uaddr;
978
979         /* Make sure 64 bit math will not overflow. */
980         if (vhost_overflow(uaddr, size))
981                 return -EFAULT;
982
983         if ((access & VHOST_ACCESS_RO) &&
984             !access_ok(VERIFY_READ, (void __user *)a, size))
985                 return -EFAULT;
986         if ((access & VHOST_ACCESS_WO) &&
987             !access_ok(VERIFY_WRITE, (void __user *)a, size))
988                 return -EFAULT;
989         return 0;
990 }
991
992 static int vhost_process_iotlb_msg(struct vhost_dev *dev,
993                                    struct vhost_iotlb_msg *msg)
994 {
995         int ret = 0;
996
997         vhost_dev_lock_vqs(dev);
998         switch (msg->type) {
999         case VHOST_IOTLB_UPDATE:
1000                 if (!dev->iotlb) {
1001                         ret = -EFAULT;
1002                         break;
1003                 }
1004                 if (umem_access_ok(msg->uaddr, msg->size, msg->perm)) {
1005                         ret = -EFAULT;
1006                         break;
1007                 }
1008                 vhost_vq_meta_reset(dev);
1009                 if (vhost_new_umem_range(dev->iotlb, msg->iova, msg->size,
1010                                          msg->iova + msg->size - 1,
1011                                          msg->uaddr, msg->perm)) {
1012                         ret = -ENOMEM;
1013                         break;
1014                 }
1015                 vhost_iotlb_notify_vq(dev, msg);
1016                 break;
1017         case VHOST_IOTLB_INVALIDATE:
1018                 if (!dev->iotlb) {
1019                         ret = -EFAULT;
1020                         break;
1021                 }
1022                 vhost_vq_meta_reset(dev);
1023                 vhost_del_umem_range(dev->iotlb, msg->iova,
1024                                      msg->iova + msg->size - 1);
1025                 break;
1026         default:
1027                 ret = -EINVAL;
1028                 break;
1029         }
1030
1031         vhost_dev_unlock_vqs(dev);
1032         return ret;
1033 }
1034 ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
1035                              struct iov_iter *from)
1036 {
1037         struct vhost_msg_node node;
1038         unsigned size = sizeof(struct vhost_msg);
1039         size_t ret;
1040         int err;
1041
1042         if (iov_iter_count(from) < size)
1043                 return 0;
1044         ret = copy_from_iter(&node.msg, size, from);
1045         if (ret != size)
1046                 goto done;
1047
1048         switch (node.msg.type) {
1049         case VHOST_IOTLB_MSG:
1050                 err = vhost_process_iotlb_msg(dev, &node.msg.iotlb);
1051                 if (err)
1052                         ret = err;
1053                 break;
1054         default:
1055                 ret = -EINVAL;
1056                 break;
1057         }
1058
1059 done:
1060         return ret;
1061 }
1062 EXPORT_SYMBOL(vhost_chr_write_iter);
1063
1064 __poll_t vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1065                             poll_table *wait)
1066 {
1067         __poll_t mask = 0;
1068
1069         poll_wait(file, &dev->wait, wait);
1070
1071         if (!list_empty(&dev->read_list))
1072                 mask |= POLLIN | POLLRDNORM;
1073
1074         return mask;
1075 }
1076 EXPORT_SYMBOL(vhost_chr_poll);
1077
1078 ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1079                             int noblock)
1080 {
1081         DEFINE_WAIT(wait);
1082         struct vhost_msg_node *node;
1083         ssize_t ret = 0;
1084         unsigned size = sizeof(struct vhost_msg);
1085
1086         if (iov_iter_count(to) < size)
1087                 return 0;
1088
1089         while (1) {
1090                 if (!noblock)
1091                         prepare_to_wait(&dev->wait, &wait,
1092                                         TASK_INTERRUPTIBLE);
1093
1094                 node = vhost_dequeue_msg(dev, &dev->read_list);
1095                 if (node)
1096                         break;
1097                 if (noblock) {
1098                         ret = -EAGAIN;
1099                         break;
1100                 }
1101                 if (signal_pending(current)) {
1102                         ret = -ERESTARTSYS;
1103                         break;
1104                 }
1105                 if (!dev->iotlb) {
1106                         ret = -EBADFD;
1107                         break;
1108                 }
1109
1110                 schedule();
1111         }
1112
1113         if (!noblock)
1114                 finish_wait(&dev->wait, &wait);
1115
1116         if (node) {
1117                 ret = copy_to_iter(&node->msg, size, to);
1118
1119                 if (ret != size || node->msg.type != VHOST_IOTLB_MISS) {
1120                         kfree(node);
1121                         return ret;
1122                 }
1123
1124                 vhost_enqueue_msg(dev, &dev->pending_list, node);
1125         }
1126
1127         return ret;
1128 }
1129 EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1130
1131 static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1132 {
1133         struct vhost_dev *dev = vq->dev;
1134         struct vhost_msg_node *node;
1135         struct vhost_iotlb_msg *msg;
1136
1137         node = vhost_new_msg(vq, VHOST_IOTLB_MISS);
1138         if (!node)
1139                 return -ENOMEM;
1140
1141         msg = &node->msg.iotlb;
1142         msg->type = VHOST_IOTLB_MISS;
1143         msg->iova = iova;
1144         msg->perm = access;
1145
1146         vhost_enqueue_msg(dev, &dev->read_list, node);
1147
1148         return 0;
1149 }
1150
1151 static int vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1152                         struct vring_desc __user *desc,
1153                         struct vring_avail __user *avail,
1154                         struct vring_used __user *used)
1155
1156 {
1157         size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1158
1159         return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
1160                access_ok(VERIFY_READ, avail,
1161                          sizeof *avail + num * sizeof *avail->ring + s) &&
1162                access_ok(VERIFY_WRITE, used,
1163                         sizeof *used + num * sizeof *used->ring + s);
1164 }
1165
1166 static void vhost_vq_meta_update(struct vhost_virtqueue *vq,
1167                                  const struct vhost_umem_node *node,
1168                                  int type)
1169 {
1170         int access = (type == VHOST_ADDR_USED) ?
1171                      VHOST_ACCESS_WO : VHOST_ACCESS_RO;
1172
1173         if (likely(node->perm & access))
1174                 vq->meta_iotlb[type] = node;
1175 }
1176
1177 static int iotlb_access_ok(struct vhost_virtqueue *vq,
1178                            int access, u64 addr, u64 len, int type)
1179 {
1180         const struct vhost_umem_node *node;
1181         struct vhost_umem *umem = vq->iotlb;
1182         u64 s = 0, size, orig_addr = addr, last = addr + len - 1;
1183
1184         if (vhost_vq_meta_fetch(vq, addr, len, type))
1185                 return true;
1186
1187         while (len > s) {
1188                 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1189                                                            addr,
1190                                                            last);
1191                 if (node == NULL || node->start > addr) {
1192                         vhost_iotlb_miss(vq, addr, access);
1193                         return false;
1194                 } else if (!(node->perm & access)) {
1195                         /* Report the possible access violation by
1196                          * request another translation from userspace.
1197                          */
1198                         return false;
1199                 }
1200
1201                 size = node->size - addr + node->start;
1202
1203                 if (orig_addr == addr && size >= len)
1204                         vhost_vq_meta_update(vq, node, type);
1205
1206                 s += size;
1207                 addr += size;
1208         }
1209
1210         return true;
1211 }
1212
1213 int vq_iotlb_prefetch(struct vhost_virtqueue *vq)
1214 {
1215         size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1216         unsigned int num = vq->num;
1217
1218         if (!vq->iotlb)
1219                 return 1;
1220
1221         return iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->desc,
1222                                num * sizeof(*vq->desc), VHOST_ADDR_DESC) &&
1223                iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->avail,
1224                                sizeof *vq->avail +
1225                                num * sizeof(*vq->avail->ring) + s,
1226                                VHOST_ADDR_AVAIL) &&
1227                iotlb_access_ok(vq, VHOST_ACCESS_WO, (u64)(uintptr_t)vq->used,
1228                                sizeof *vq->used +
1229                                num * sizeof(*vq->used->ring) + s,
1230                                VHOST_ADDR_USED);
1231 }
1232 EXPORT_SYMBOL_GPL(vq_iotlb_prefetch);
1233
1234 /* Can we log writes? */
1235 /* Caller should have device mutex but not vq mutex */
1236 int vhost_log_access_ok(struct vhost_dev *dev)
1237 {
1238         return memory_access_ok(dev, dev->umem, 1);
1239 }
1240 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1241
1242 /* Verify access for write logging. */
1243 /* Caller should have vq mutex and device mutex */
1244 static int vq_log_access_ok(struct vhost_virtqueue *vq,
1245                             void __user *log_base)
1246 {
1247         size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1248
1249         return vq_memory_access_ok(log_base, vq->umem,
1250                                    vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1251                 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
1252                                         sizeof *vq->used +
1253                                         vq->num * sizeof *vq->used->ring + s));
1254 }
1255
1256 /* Can we start vq? */
1257 /* Caller should have vq mutex and device mutex */
1258 int vhost_vq_access_ok(struct vhost_virtqueue *vq)
1259 {
1260         if (vq->iotlb) {
1261                 /* When device IOTLB was used, the access validation
1262                  * will be validated during prefetching.
1263                  */
1264                 return 1;
1265         }
1266         return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used) &&
1267                 vq_log_access_ok(vq, vq->log_base);
1268 }
1269 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1270
1271 static struct vhost_umem *vhost_umem_alloc(void)
1272 {
1273         struct vhost_umem *umem = kvzalloc(sizeof(*umem), GFP_KERNEL);
1274
1275         if (!umem)
1276                 return NULL;
1277
1278         umem->umem_tree = RB_ROOT_CACHED;
1279         umem->numem = 0;
1280         INIT_LIST_HEAD(&umem->umem_list);
1281
1282         return umem;
1283 }
1284
1285 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1286 {
1287         struct vhost_memory mem, *newmem;
1288         struct vhost_memory_region *region;
1289         struct vhost_umem *newumem, *oldumem;
1290         unsigned long size = offsetof(struct vhost_memory, regions);
1291         int i;
1292
1293         if (copy_from_user(&mem, m, size))
1294                 return -EFAULT;
1295         if (mem.padding)
1296                 return -EOPNOTSUPP;
1297         if (mem.nregions > max_mem_regions)
1298                 return -E2BIG;
1299         newmem = kvzalloc(size + mem.nregions * sizeof(*m->regions), GFP_KERNEL);
1300         if (!newmem)
1301                 return -ENOMEM;
1302
1303         memcpy(newmem, &mem, size);
1304         if (copy_from_user(newmem->regions, m->regions,
1305                            mem.nregions * sizeof *m->regions)) {
1306                 kvfree(newmem);
1307                 return -EFAULT;
1308         }
1309
1310         newumem = vhost_umem_alloc();
1311         if (!newumem) {
1312                 kvfree(newmem);
1313                 return -ENOMEM;
1314         }
1315
1316         for (region = newmem->regions;
1317              region < newmem->regions + mem.nregions;
1318              region++) {
1319                 if (vhost_new_umem_range(newumem,
1320                                          region->guest_phys_addr,
1321                                          region->memory_size,
1322                                          region->guest_phys_addr +
1323                                          region->memory_size - 1,
1324                                          region->userspace_addr,
1325                                          VHOST_ACCESS_RW))
1326                         goto err;
1327         }
1328
1329         if (!memory_access_ok(d, newumem, 0))
1330                 goto err;
1331
1332         oldumem = d->umem;
1333         d->umem = newumem;
1334
1335         /* All memory accesses are done under some VQ mutex. */
1336         for (i = 0; i < d->nvqs; ++i) {
1337                 mutex_lock(&d->vqs[i]->mutex);
1338                 d->vqs[i]->umem = newumem;
1339                 mutex_unlock(&d->vqs[i]->mutex);
1340         }
1341
1342         kvfree(newmem);
1343         vhost_umem_clean(oldumem);
1344         return 0;
1345
1346 err:
1347         vhost_umem_clean(newumem);
1348         kvfree(newmem);
1349         return -EFAULT;
1350 }
1351
1352 long vhost_vring_ioctl(struct vhost_dev *d, int ioctl, void __user *argp)
1353 {
1354         struct file *eventfp, *filep = NULL;
1355         bool pollstart = false, pollstop = false;
1356         struct eventfd_ctx *ctx = NULL;
1357         u32 __user *idxp = argp;
1358         struct vhost_virtqueue *vq;
1359         struct vhost_vring_state s;
1360         struct vhost_vring_file f;
1361         struct vhost_vring_addr a;
1362         u32 idx;
1363         long r;
1364
1365         r = get_user(idx, idxp);
1366         if (r < 0)
1367                 return r;
1368         if (idx >= d->nvqs)
1369                 return -ENOBUFS;
1370
1371         vq = d->vqs[idx];
1372
1373         mutex_lock(&vq->mutex);
1374
1375         switch (ioctl) {
1376         case VHOST_SET_VRING_NUM:
1377                 /* Resizing ring with an active backend?
1378                  * You don't want to do that. */
1379                 if (vq->private_data) {
1380                         r = -EBUSY;
1381                         break;
1382                 }
1383                 if (copy_from_user(&s, argp, sizeof s)) {
1384                         r = -EFAULT;
1385                         break;
1386                 }
1387                 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
1388                         r = -EINVAL;
1389                         break;
1390                 }
1391                 vq->num = s.num;
1392                 break;
1393         case VHOST_SET_VRING_BASE:
1394                 /* Moving base with an active backend?
1395                  * You don't want to do that. */
1396                 if (vq->private_data) {
1397                         r = -EBUSY;
1398                         break;
1399                 }
1400                 if (copy_from_user(&s, argp, sizeof s)) {
1401                         r = -EFAULT;
1402                         break;
1403                 }
1404                 if (s.num > 0xffff) {
1405                         r = -EINVAL;
1406                         break;
1407                 }
1408                 vq->last_avail_idx = s.num;
1409                 /* Forget the cached index value. */
1410                 vq->avail_idx = vq->last_avail_idx;
1411                 break;
1412         case VHOST_GET_VRING_BASE:
1413                 s.index = idx;
1414                 s.num = vq->last_avail_idx;
1415                 if (copy_to_user(argp, &s, sizeof s))
1416                         r = -EFAULT;
1417                 break;
1418         case VHOST_SET_VRING_ADDR:
1419                 if (copy_from_user(&a, argp, sizeof a)) {
1420                         r = -EFAULT;
1421                         break;
1422                 }
1423                 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
1424                         r = -EOPNOTSUPP;
1425                         break;
1426                 }
1427                 /* For 32bit, verify that the top 32bits of the user
1428                    data are set to zero. */
1429                 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1430                     (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1431                     (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
1432                         r = -EFAULT;
1433                         break;
1434                 }
1435
1436                 /* Make sure it's safe to cast pointers to vring types. */
1437                 BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1438                 BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1439                 if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1440                     (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1441                     (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1))) {
1442                         r = -EINVAL;
1443                         break;
1444                 }
1445
1446                 /* We only verify access here if backend is configured.
1447                  * If it is not, we don't as size might not have been setup.
1448                  * We will verify when backend is configured. */
1449                 if (vq->private_data) {
1450                         if (!vq_access_ok(vq, vq->num,
1451                                 (void __user *)(unsigned long)a.desc_user_addr,
1452                                 (void __user *)(unsigned long)a.avail_user_addr,
1453                                 (void __user *)(unsigned long)a.used_user_addr)) {
1454                                 r = -EINVAL;
1455                                 break;
1456                         }
1457
1458                         /* Also validate log access for used ring if enabled. */
1459                         if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
1460                             !log_access_ok(vq->log_base, a.log_guest_addr,
1461                                            sizeof *vq->used +
1462                                            vq->num * sizeof *vq->used->ring)) {
1463                                 r = -EINVAL;
1464                                 break;
1465                         }
1466                 }
1467
1468                 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1469                 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1470                 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1471                 vq->log_addr = a.log_guest_addr;
1472                 vq->used = (void __user *)(unsigned long)a.used_user_addr;
1473                 break;
1474         case VHOST_SET_VRING_KICK:
1475                 if (copy_from_user(&f, argp, sizeof f)) {
1476                         r = -EFAULT;
1477                         break;
1478                 }
1479                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1480                 if (IS_ERR(eventfp)) {
1481                         r = PTR_ERR(eventfp);
1482                         break;
1483                 }
1484                 if (eventfp != vq->kick) {
1485                         pollstop = (filep = vq->kick) != NULL;
1486                         pollstart = (vq->kick = eventfp) != NULL;
1487                 } else
1488                         filep = eventfp;
1489                 break;
1490         case VHOST_SET_VRING_CALL:
1491                 if (copy_from_user(&f, argp, sizeof f)) {
1492                         r = -EFAULT;
1493                         break;
1494                 }
1495                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1496                 if (IS_ERR(eventfp)) {
1497                         r = PTR_ERR(eventfp);
1498                         break;
1499                 }
1500                 if (eventfp != vq->call) {
1501                         filep = vq->call;
1502                         ctx = vq->call_ctx;
1503                         vq->call = eventfp;
1504                         vq->call_ctx = eventfp ?
1505                                 eventfd_ctx_fileget(eventfp) : NULL;
1506                 } else
1507                         filep = eventfp;
1508                 break;
1509         case VHOST_SET_VRING_ERR:
1510                 if (copy_from_user(&f, argp, sizeof f)) {
1511                         r = -EFAULT;
1512                         break;
1513                 }
1514                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1515                 if (IS_ERR(eventfp)) {
1516                         r = PTR_ERR(eventfp);
1517                         break;
1518                 }
1519                 if (eventfp != vq->error) {
1520                         filep = vq->error;
1521                         vq->error = eventfp;
1522                         ctx = vq->error_ctx;
1523                         vq->error_ctx = eventfp ?
1524                                 eventfd_ctx_fileget(eventfp) : NULL;
1525                 } else
1526                         filep = eventfp;
1527                 break;
1528         case VHOST_SET_VRING_ENDIAN:
1529                 r = vhost_set_vring_endian(vq, argp);
1530                 break;
1531         case VHOST_GET_VRING_ENDIAN:
1532                 r = vhost_get_vring_endian(vq, idx, argp);
1533                 break;
1534         case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
1535                 if (copy_from_user(&s, argp, sizeof(s))) {
1536                         r = -EFAULT;
1537                         break;
1538                 }
1539                 vq->busyloop_timeout = s.num;
1540                 break;
1541         case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
1542                 s.index = idx;
1543                 s.num = vq->busyloop_timeout;
1544                 if (copy_to_user(argp, &s, sizeof(s)))
1545                         r = -EFAULT;
1546                 break;
1547         default:
1548                 r = -ENOIOCTLCMD;
1549         }
1550
1551         if (pollstop && vq->handle_kick)
1552                 vhost_poll_stop(&vq->poll);
1553
1554         if (ctx)
1555                 eventfd_ctx_put(ctx);
1556         if (filep)
1557                 fput(filep);
1558
1559         if (pollstart && vq->handle_kick)
1560                 r = vhost_poll_start(&vq->poll, vq->kick);
1561
1562         mutex_unlock(&vq->mutex);
1563
1564         if (pollstop && vq->handle_kick)
1565                 vhost_poll_flush(&vq->poll);
1566         return r;
1567 }
1568 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
1569
1570 int vhost_init_device_iotlb(struct vhost_dev *d, bool enabled)
1571 {
1572         struct vhost_umem *niotlb, *oiotlb;
1573         int i;
1574
1575         niotlb = vhost_umem_alloc();
1576         if (!niotlb)
1577                 return -ENOMEM;
1578
1579         oiotlb = d->iotlb;
1580         d->iotlb = niotlb;
1581
1582         for (i = 0; i < d->nvqs; ++i) {
1583                 mutex_lock(&d->vqs[i]->mutex);
1584                 d->vqs[i]->iotlb = niotlb;
1585                 mutex_unlock(&d->vqs[i]->mutex);
1586         }
1587
1588         vhost_umem_clean(oiotlb);
1589
1590         return 0;
1591 }
1592 EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
1593
1594 /* Caller must have device mutex */
1595 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1596 {
1597         struct file *eventfp, *filep = NULL;
1598         struct eventfd_ctx *ctx = NULL;
1599         u64 p;
1600         long r;
1601         int i, fd;
1602
1603         /* If you are not the owner, you can become one */
1604         if (ioctl == VHOST_SET_OWNER) {
1605                 r = vhost_dev_set_owner(d);
1606                 goto done;
1607         }
1608
1609         /* You must be the owner to do anything else */
1610         r = vhost_dev_check_owner(d);
1611         if (r)
1612                 goto done;
1613
1614         switch (ioctl) {
1615         case VHOST_SET_MEM_TABLE:
1616                 r = vhost_set_memory(d, argp);
1617                 break;
1618         case VHOST_SET_LOG_BASE:
1619                 if (copy_from_user(&p, argp, sizeof p)) {
1620                         r = -EFAULT;
1621                         break;
1622                 }
1623                 if ((u64)(unsigned long)p != p) {
1624                         r = -EFAULT;
1625                         break;
1626                 }
1627                 for (i = 0; i < d->nvqs; ++i) {
1628                         struct vhost_virtqueue *vq;
1629                         void __user *base = (void __user *)(unsigned long)p;
1630                         vq = d->vqs[i];
1631                         mutex_lock(&vq->mutex);
1632                         /* If ring is inactive, will check when it's enabled. */
1633                         if (vq->private_data && !vq_log_access_ok(vq, base))
1634                                 r = -EFAULT;
1635                         else
1636                                 vq->log_base = base;
1637                         mutex_unlock(&vq->mutex);
1638                 }
1639                 break;
1640         case VHOST_SET_LOG_FD:
1641                 r = get_user(fd, (int __user *)argp);
1642                 if (r < 0)
1643                         break;
1644                 eventfp = fd == -1 ? NULL : eventfd_fget(fd);
1645                 if (IS_ERR(eventfp)) {
1646                         r = PTR_ERR(eventfp);
1647                         break;
1648                 }
1649                 if (eventfp != d->log_file) {
1650                         filep = d->log_file;
1651                         d->log_file = eventfp;
1652                         ctx = d->log_ctx;
1653                         d->log_ctx = eventfp ?
1654                                 eventfd_ctx_fileget(eventfp) : NULL;
1655                 } else
1656                         filep = eventfp;
1657                 for (i = 0; i < d->nvqs; ++i) {
1658                         mutex_lock(&d->vqs[i]->mutex);
1659                         d->vqs[i]->log_ctx = d->log_ctx;
1660                         mutex_unlock(&d->vqs[i]->mutex);
1661                 }
1662                 if (ctx)
1663                         eventfd_ctx_put(ctx);
1664                 if (filep)
1665                         fput(filep);
1666                 break;
1667         default:
1668                 r = -ENOIOCTLCMD;
1669                 break;
1670         }
1671 done:
1672         return r;
1673 }
1674 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1675
1676 /* TODO: This is really inefficient.  We need something like get_user()
1677  * (instruction directly accesses the data, with an exception table entry
1678  * returning -EFAULT). See Documentation/x86/exception-tables.txt.
1679  */
1680 static int set_bit_to_user(int nr, void __user *addr)
1681 {
1682         unsigned long log = (unsigned long)addr;
1683         struct page *page;
1684         void *base;
1685         int bit = nr + (log % PAGE_SIZE) * 8;
1686         int r;
1687
1688         r = get_user_pages_fast(log, 1, 1, &page);
1689         if (r < 0)
1690                 return r;
1691         BUG_ON(r != 1);
1692         base = kmap_atomic(page);
1693         set_bit(bit, base);
1694         kunmap_atomic(base);
1695         set_page_dirty_lock(page);
1696         put_page(page);
1697         return 0;
1698 }
1699
1700 static int log_write(void __user *log_base,
1701                      u64 write_address, u64 write_length)
1702 {
1703         u64 write_page = write_address / VHOST_PAGE_SIZE;
1704         int r;
1705
1706         if (!write_length)
1707                 return 0;
1708         write_length += write_address % VHOST_PAGE_SIZE;
1709         for (;;) {
1710                 u64 base = (u64)(unsigned long)log_base;
1711                 u64 log = base + write_page / 8;
1712                 int bit = write_page % 8;
1713                 if ((u64)(unsigned long)log != log)
1714                         return -EFAULT;
1715                 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1716                 if (r < 0)
1717                         return r;
1718                 if (write_length <= VHOST_PAGE_SIZE)
1719                         break;
1720                 write_length -= VHOST_PAGE_SIZE;
1721                 write_page += 1;
1722         }
1723         return r;
1724 }
1725
1726 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1727                     unsigned int log_num, u64 len)
1728 {
1729         int i, r;
1730
1731         /* Make sure data written is seen before log. */
1732         smp_wmb();
1733         for (i = 0; i < log_num; ++i) {
1734                 u64 l = min(log[i].len, len);
1735                 r = log_write(vq->log_base, log[i].addr, l);
1736                 if (r < 0)
1737                         return r;
1738                 len -= l;
1739                 if (!len) {
1740                         if (vq->log_ctx)
1741                                 eventfd_signal(vq->log_ctx, 1);
1742                         return 0;
1743                 }
1744         }
1745         /* Length written exceeds what we have stored. This is a bug. */
1746         BUG();
1747         return 0;
1748 }
1749 EXPORT_SYMBOL_GPL(vhost_log_write);
1750
1751 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1752 {
1753         void __user *used;
1754         if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
1755                            &vq->used->flags) < 0)
1756                 return -EFAULT;
1757         if (unlikely(vq->log_used)) {
1758                 /* Make sure the flag is seen before log. */
1759                 smp_wmb();
1760                 /* Log used flag write. */
1761                 used = &vq->used->flags;
1762                 log_write(vq->log_base, vq->log_addr +
1763                           (used - (void __user *)vq->used),
1764                           sizeof vq->used->flags);
1765                 if (vq->log_ctx)
1766                         eventfd_signal(vq->log_ctx, 1);
1767         }
1768         return 0;
1769 }
1770
1771 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1772 {
1773         if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
1774                            vhost_avail_event(vq)))
1775                 return -EFAULT;
1776         if (unlikely(vq->log_used)) {
1777                 void __user *used;
1778                 /* Make sure the event is seen before log. */
1779                 smp_wmb();
1780                 /* Log avail event write */
1781                 used = vhost_avail_event(vq);
1782                 log_write(vq->log_base, vq->log_addr +
1783                           (used - (void __user *)vq->used),
1784                           sizeof *vhost_avail_event(vq));
1785                 if (vq->log_ctx)
1786                         eventfd_signal(vq->log_ctx, 1);
1787         }
1788         return 0;
1789 }
1790
1791 int vhost_vq_init_access(struct vhost_virtqueue *vq)
1792 {
1793         __virtio16 last_used_idx;
1794         int r;
1795         bool is_le = vq->is_le;
1796
1797         if (!vq->private_data)
1798                 return 0;
1799
1800         vhost_init_is_le(vq);
1801
1802         r = vhost_update_used_flags(vq);
1803         if (r)
1804                 goto err;
1805         vq->signalled_used_valid = false;
1806         if (!vq->iotlb &&
1807             !access_ok(VERIFY_READ, &vq->used->idx, sizeof vq->used->idx)) {
1808                 r = -EFAULT;
1809                 goto err;
1810         }
1811         r = vhost_get_used(vq, last_used_idx, &vq->used->idx);
1812         if (r) {
1813                 vq_err(vq, "Can't access used idx at %p\n",
1814                        &vq->used->idx);
1815                 goto err;
1816         }
1817         vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
1818         return 0;
1819
1820 err:
1821         vq->is_le = is_le;
1822         return r;
1823 }
1824 EXPORT_SYMBOL_GPL(vhost_vq_init_access);
1825
1826 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
1827                           struct iovec iov[], int iov_size, int access)
1828 {
1829         const struct vhost_umem_node *node;
1830         struct vhost_dev *dev = vq->dev;
1831         struct vhost_umem *umem = dev->iotlb ? dev->iotlb : dev->umem;
1832         struct iovec *_iov;
1833         u64 s = 0;
1834         int ret = 0;
1835
1836         while ((u64)len > s) {
1837                 u64 size;
1838                 if (unlikely(ret >= iov_size)) {
1839                         ret = -ENOBUFS;
1840                         break;
1841                 }
1842
1843                 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1844                                                         addr, addr + len - 1);
1845                 if (node == NULL || node->start > addr) {
1846                         if (umem != dev->iotlb) {
1847                                 ret = -EFAULT;
1848                                 break;
1849                         }
1850                         ret = -EAGAIN;
1851                         break;
1852                 } else if (!(node->perm & access)) {
1853                         ret = -EPERM;
1854                         break;
1855                 }
1856
1857                 _iov = iov + ret;
1858                 size = node->size - addr + node->start;
1859                 _iov->iov_len = min((u64)len - s, size);
1860                 _iov->iov_base = (void __user *)(unsigned long)
1861                         (node->userspace_addr + addr - node->start);
1862                 s += size;
1863                 addr += size;
1864                 ++ret;
1865         }
1866
1867         if (ret == -EAGAIN)
1868                 vhost_iotlb_miss(vq, addr, access);
1869         return ret;
1870 }
1871
1872 /* Each buffer in the virtqueues is actually a chain of descriptors.  This
1873  * function returns the next descriptor in the chain,
1874  * or -1U if we're at the end. */
1875 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
1876 {
1877         unsigned int next;
1878
1879         /* If this descriptor says it doesn't chain, we're done. */
1880         if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
1881                 return -1U;
1882
1883         /* Check they're not leading us off end of descriptors. */
1884         next = vhost16_to_cpu(vq, READ_ONCE(desc->next));
1885         return next;
1886 }
1887
1888 static int get_indirect(struct vhost_virtqueue *vq,
1889                         struct iovec iov[], unsigned int iov_size,
1890                         unsigned int *out_num, unsigned int *in_num,
1891                         struct vhost_log *log, unsigned int *log_num,
1892                         struct vring_desc *indirect)
1893 {
1894         struct vring_desc desc;
1895         unsigned int i = 0, count, found = 0;
1896         u32 len = vhost32_to_cpu(vq, indirect->len);
1897         struct iov_iter from;
1898         int ret, access;
1899
1900         /* Sanity check */
1901         if (unlikely(len % sizeof desc)) {
1902                 vq_err(vq, "Invalid length in indirect descriptor: "
1903                        "len 0x%llx not multiple of 0x%zx\n",
1904                        (unsigned long long)len,
1905                        sizeof desc);
1906                 return -EINVAL;
1907         }
1908
1909         ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
1910                              UIO_MAXIOV, VHOST_ACCESS_RO);
1911         if (unlikely(ret < 0)) {
1912                 if (ret != -EAGAIN)
1913                         vq_err(vq, "Translation failure %d in indirect.\n", ret);
1914                 return ret;
1915         }
1916         iov_iter_init(&from, READ, vq->indirect, ret, len);
1917
1918         /* We will use the result as an address to read from, so most
1919          * architectures only need a compiler barrier here. */
1920         read_barrier_depends();
1921
1922         count = len / sizeof desc;
1923         /* Buffers are chained via a 16 bit next field, so
1924          * we can have at most 2^16 of these. */
1925         if (unlikely(count > USHRT_MAX + 1)) {
1926                 vq_err(vq, "Indirect buffer length too big: %d\n",
1927                        indirect->len);
1928                 return -E2BIG;
1929         }
1930
1931         do {
1932                 unsigned iov_count = *in_num + *out_num;
1933                 if (unlikely(++found > count)) {
1934                         vq_err(vq, "Loop detected: last one at %u "
1935                                "indirect size %u\n",
1936                                i, count);
1937                         return -EINVAL;
1938                 }
1939                 if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
1940                         vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1941                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1942                         return -EINVAL;
1943                 }
1944                 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
1945                         vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1946                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1947                         return -EINVAL;
1948                 }
1949
1950                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
1951                         access = VHOST_ACCESS_WO;
1952                 else
1953                         access = VHOST_ACCESS_RO;
1954
1955                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
1956                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
1957                                      iov_size - iov_count, access);
1958                 if (unlikely(ret < 0)) {
1959                         if (ret != -EAGAIN)
1960                                 vq_err(vq, "Translation failure %d indirect idx %d\n",
1961                                         ret, i);
1962                         return ret;
1963                 }
1964                 /* If this is an input descriptor, increment that count. */
1965                 if (access == VHOST_ACCESS_WO) {
1966                         *in_num += ret;
1967                         if (unlikely(log)) {
1968                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
1969                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
1970                                 ++*log_num;
1971                         }
1972                 } else {
1973                         /* If it's an output descriptor, they're all supposed
1974                          * to come before any input descriptors. */
1975                         if (unlikely(*in_num)) {
1976                                 vq_err(vq, "Indirect descriptor "
1977                                        "has out after in: idx %d\n", i);
1978                                 return -EINVAL;
1979                         }
1980                         *out_num += ret;
1981                 }
1982         } while ((i = next_desc(vq, &desc)) != -1);
1983         return 0;
1984 }
1985
1986 /* This looks in the virtqueue and for the first available buffer, and converts
1987  * it to an iovec for convenient access.  Since descriptors consist of some
1988  * number of output then some number of input descriptors, it's actually two
1989  * iovecs, but we pack them into one and note how many of each there were.
1990  *
1991  * This function returns the descriptor number found, or vq->num (which is
1992  * never a valid descriptor number) if none was found.  A negative code is
1993  * returned on error. */
1994 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
1995                       struct iovec iov[], unsigned int iov_size,
1996                       unsigned int *out_num, unsigned int *in_num,
1997                       struct vhost_log *log, unsigned int *log_num)
1998 {
1999         struct vring_desc desc;
2000         unsigned int i, head, found = 0;
2001         u16 last_avail_idx;
2002         __virtio16 avail_idx;
2003         __virtio16 ring_head;
2004         int ret, access;
2005
2006         /* Check it isn't doing very strange things with descriptor numbers. */
2007         last_avail_idx = vq->last_avail_idx;
2008
2009         if (vq->avail_idx == vq->last_avail_idx) {
2010                 if (unlikely(vhost_get_avail(vq, avail_idx, &vq->avail->idx))) {
2011                         vq_err(vq, "Failed to access avail idx at %p\n",
2012                                 &vq->avail->idx);
2013                         return -EFAULT;
2014                 }
2015                 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2016
2017                 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
2018                         vq_err(vq, "Guest moved used index from %u to %u",
2019                                 last_avail_idx, vq->avail_idx);
2020                         return -EFAULT;
2021                 }
2022
2023                 /* If there's nothing new since last we looked, return
2024                  * invalid.
2025                  */
2026                 if (vq->avail_idx == last_avail_idx)
2027                         return vq->num;
2028
2029                 /* Only get avail ring entries after they have been
2030                  * exposed by guest.
2031                  */
2032                 smp_rmb();
2033         }
2034
2035         /* Grab the next descriptor number they're advertising, and increment
2036          * the index we've seen. */
2037         if (unlikely(vhost_get_avail(vq, ring_head,
2038                      &vq->avail->ring[last_avail_idx & (vq->num - 1)]))) {
2039                 vq_err(vq, "Failed to read head: idx %d address %p\n",
2040                        last_avail_idx,
2041                        &vq->avail->ring[last_avail_idx % vq->num]);
2042                 return -EFAULT;
2043         }
2044
2045         head = vhost16_to_cpu(vq, ring_head);
2046
2047         /* If their number is silly, that's an error. */
2048         if (unlikely(head >= vq->num)) {
2049                 vq_err(vq, "Guest says index %u > %u is available",
2050                        head, vq->num);
2051                 return -EINVAL;
2052         }
2053
2054         /* When we start there are none of either input nor output. */
2055         *out_num = *in_num = 0;
2056         if (unlikely(log))
2057                 *log_num = 0;
2058
2059         i = head;
2060         do {
2061                 unsigned iov_count = *in_num + *out_num;
2062                 if (unlikely(i >= vq->num)) {
2063                         vq_err(vq, "Desc index is %u > %u, head = %u",
2064                                i, vq->num, head);
2065                         return -EINVAL;
2066                 }
2067                 if (unlikely(++found > vq->num)) {
2068                         vq_err(vq, "Loop detected: last one at %u "
2069                                "vq size %u head %u\n",
2070                                i, vq->num, head);
2071                         return -EINVAL;
2072                 }
2073                 ret = vhost_copy_from_user(vq, &desc, vq->desc + i,
2074                                            sizeof desc);
2075                 if (unlikely(ret)) {
2076                         vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
2077                                i, vq->desc + i);
2078                         return -EFAULT;
2079                 }
2080                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2081                         ret = get_indirect(vq, iov, iov_size,
2082                                            out_num, in_num,
2083                                            log, log_num, &desc);
2084                         if (unlikely(ret < 0)) {
2085                                 if (ret != -EAGAIN)
2086                                         vq_err(vq, "Failure detected "
2087                                                 "in indirect descriptor at idx %d\n", i);
2088                                 return ret;
2089                         }
2090                         continue;
2091                 }
2092
2093                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2094                         access = VHOST_ACCESS_WO;
2095                 else
2096                         access = VHOST_ACCESS_RO;
2097                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2098                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
2099                                      iov_size - iov_count, access);
2100                 if (unlikely(ret < 0)) {
2101                         if (ret != -EAGAIN)
2102                                 vq_err(vq, "Translation failure %d descriptor idx %d\n",
2103                                         ret, i);
2104                         return ret;
2105                 }
2106                 if (access == VHOST_ACCESS_WO) {
2107                         /* If this is an input descriptor,
2108                          * increment that count. */
2109                         *in_num += ret;
2110                         if (unlikely(log)) {
2111                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2112                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2113                                 ++*log_num;
2114                         }
2115                 } else {
2116                         /* If it's an output descriptor, they're all supposed
2117                          * to come before any input descriptors. */
2118                         if (unlikely(*in_num)) {
2119                                 vq_err(vq, "Descriptor has out after in: "
2120                                        "idx %d\n", i);
2121                                 return -EINVAL;
2122                         }
2123                         *out_num += ret;
2124                 }
2125         } while ((i = next_desc(vq, &desc)) != -1);
2126
2127         /* On success, increment avail index. */
2128         vq->last_avail_idx++;
2129
2130         /* Assume notifications from guest are disabled at this point,
2131          * if they aren't we would need to update avail_event index. */
2132         BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2133         return head;
2134 }
2135 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2136
2137 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2138 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2139 {
2140         vq->last_avail_idx -= n;
2141 }
2142 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2143
2144 /* After we've used one of their buffers, we tell them about it.  We'll then
2145  * want to notify the guest, using eventfd. */
2146 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2147 {
2148         struct vring_used_elem heads = {
2149                 cpu_to_vhost32(vq, head),
2150                 cpu_to_vhost32(vq, len)
2151         };
2152
2153         return vhost_add_used_n(vq, &heads, 1);
2154 }
2155 EXPORT_SYMBOL_GPL(vhost_add_used);
2156
2157 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2158                             struct vring_used_elem *heads,
2159                             unsigned count)
2160 {
2161         struct vring_used_elem __user *used;
2162         u16 old, new;
2163         int start;
2164
2165         start = vq->last_used_idx & (vq->num - 1);
2166         used = vq->used->ring + start;
2167         if (count == 1) {
2168                 if (vhost_put_user(vq, heads[0].id, &used->id)) {
2169                         vq_err(vq, "Failed to write used id");
2170                         return -EFAULT;
2171                 }
2172                 if (vhost_put_user(vq, heads[0].len, &used->len)) {
2173                         vq_err(vq, "Failed to write used len");
2174                         return -EFAULT;
2175                 }
2176         } else if (vhost_copy_to_user(vq, used, heads, count * sizeof *used)) {
2177                 vq_err(vq, "Failed to write used");
2178                 return -EFAULT;
2179         }
2180         if (unlikely(vq->log_used)) {
2181                 /* Make sure data is seen before log. */
2182                 smp_wmb();
2183                 /* Log used ring entry write. */
2184                 log_write(vq->log_base,
2185                           vq->log_addr +
2186                            ((void __user *)used - (void __user *)vq->used),
2187                           count * sizeof *used);
2188         }
2189         old = vq->last_used_idx;
2190         new = (vq->last_used_idx += count);
2191         /* If the driver never bothers to signal in a very long while,
2192          * used index might wrap around. If that happens, invalidate
2193          * signalled_used index we stored. TODO: make sure driver
2194          * signals at least once in 2^16 and remove this. */
2195         if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2196                 vq->signalled_used_valid = false;
2197         return 0;
2198 }
2199
2200 /* After we've used one of their buffers, we tell them about it.  We'll then
2201  * want to notify the guest, using eventfd. */
2202 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2203                      unsigned count)
2204 {
2205         int start, n, r;
2206
2207         start = vq->last_used_idx & (vq->num - 1);
2208         n = vq->num - start;
2209         if (n < count) {
2210                 r = __vhost_add_used_n(vq, heads, n);
2211                 if (r < 0)
2212                         return r;
2213                 heads += n;
2214                 count -= n;
2215         }
2216         r = __vhost_add_used_n(vq, heads, count);
2217
2218         /* Make sure buffer is written before we update index. */
2219         smp_wmb();
2220         if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
2221                            &vq->used->idx)) {
2222                 vq_err(vq, "Failed to increment used idx");
2223                 return -EFAULT;
2224         }
2225         if (unlikely(vq->log_used)) {
2226                 /* Log used index update. */
2227                 log_write(vq->log_base,
2228                           vq->log_addr + offsetof(struct vring_used, idx),
2229                           sizeof vq->used->idx);
2230                 if (vq->log_ctx)
2231                         eventfd_signal(vq->log_ctx, 1);
2232         }
2233         return r;
2234 }
2235 EXPORT_SYMBOL_GPL(vhost_add_used_n);
2236
2237 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2238 {
2239         __u16 old, new;
2240         __virtio16 event;
2241         bool v;
2242         /* Flush out used index updates. This is paired
2243          * with the barrier that the Guest executes when enabling
2244          * interrupts. */
2245         smp_mb();
2246
2247         if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2248             unlikely(vq->avail_idx == vq->last_avail_idx))
2249                 return true;
2250
2251         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2252                 __virtio16 flags;
2253                 if (vhost_get_avail(vq, flags, &vq->avail->flags)) {
2254                         vq_err(vq, "Failed to get flags");
2255                         return true;
2256                 }
2257                 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2258         }
2259         old = vq->signalled_used;
2260         v = vq->signalled_used_valid;
2261         new = vq->signalled_used = vq->last_used_idx;
2262         vq->signalled_used_valid = true;
2263
2264         if (unlikely(!v))
2265                 return true;
2266
2267         if (vhost_get_avail(vq, event, vhost_used_event(vq))) {
2268                 vq_err(vq, "Failed to get used event idx");
2269                 return true;
2270         }
2271         return vring_need_event(vhost16_to_cpu(vq, event), new, old);
2272 }
2273
2274 /* This actually signals the guest, using eventfd. */
2275 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2276 {
2277         /* Signal the Guest tell them we used something up. */
2278         if (vq->call_ctx && vhost_notify(dev, vq))
2279                 eventfd_signal(vq->call_ctx, 1);
2280 }
2281 EXPORT_SYMBOL_GPL(vhost_signal);
2282
2283 /* And here's the combo meal deal.  Supersize me! */
2284 void vhost_add_used_and_signal(struct vhost_dev *dev,
2285                                struct vhost_virtqueue *vq,
2286                                unsigned int head, int len)
2287 {
2288         vhost_add_used(vq, head, len);
2289         vhost_signal(dev, vq);
2290 }
2291 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2292
2293 /* multi-buffer version of vhost_add_used_and_signal */
2294 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2295                                  struct vhost_virtqueue *vq,
2296                                  struct vring_used_elem *heads, unsigned count)
2297 {
2298         vhost_add_used_n(vq, heads, count);
2299         vhost_signal(dev, vq);
2300 }
2301 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2302
2303 /* return true if we're sure that avaiable ring is empty */
2304 bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2305 {
2306         __virtio16 avail_idx;
2307         int r;
2308
2309         if (vq->avail_idx != vq->last_avail_idx)
2310                 return false;
2311
2312         r = vhost_get_avail(vq, avail_idx, &vq->avail->idx);
2313         if (unlikely(r))
2314                 return false;
2315         vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2316
2317         return vq->avail_idx == vq->last_avail_idx;
2318 }
2319 EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2320
2321 /* OK, now we need to know about added descriptors. */
2322 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2323 {
2324         __virtio16 avail_idx;
2325         int r;
2326
2327         if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2328                 return false;
2329         vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2330         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2331                 r = vhost_update_used_flags(vq);
2332                 if (r) {
2333                         vq_err(vq, "Failed to enable notification at %p: %d\n",
2334                                &vq->used->flags, r);
2335                         return false;
2336                 }
2337         } else {
2338                 r = vhost_update_avail_event(vq, vq->avail_idx);
2339                 if (r) {
2340                         vq_err(vq, "Failed to update avail event index at %p: %d\n",
2341                                vhost_avail_event(vq), r);
2342                         return false;
2343                 }
2344         }
2345         /* They could have slipped one in as we were doing that: make
2346          * sure it's written, then check again. */
2347         smp_mb();
2348         r = vhost_get_avail(vq, avail_idx, &vq->avail->idx);
2349         if (r) {
2350                 vq_err(vq, "Failed to check avail idx at %p: %d\n",
2351                        &vq->avail->idx, r);
2352                 return false;
2353         }
2354
2355         return vhost16_to_cpu(vq, avail_idx) != vq->avail_idx;
2356 }
2357 EXPORT_SYMBOL_GPL(vhost_enable_notify);
2358
2359 /* We don't need to be notified again. */
2360 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2361 {
2362         int r;
2363
2364         if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2365                 return;
2366         vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2367         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2368                 r = vhost_update_used_flags(vq);
2369                 if (r)
2370                         vq_err(vq, "Failed to enable notification at %p: %d\n",
2371                                &vq->used->flags, r);
2372         }
2373 }
2374 EXPORT_SYMBOL_GPL(vhost_disable_notify);
2375
2376 /* Create a new message. */
2377 struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2378 {
2379         struct vhost_msg_node *node = kmalloc(sizeof *node, GFP_KERNEL);
2380         if (!node)
2381                 return NULL;
2382         node->vq = vq;
2383         node->msg.type = type;
2384         return node;
2385 }
2386 EXPORT_SYMBOL_GPL(vhost_new_msg);
2387
2388 void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2389                        struct vhost_msg_node *node)
2390 {
2391         spin_lock(&dev->iotlb_lock);
2392         list_add_tail(&node->node, head);
2393         spin_unlock(&dev->iotlb_lock);
2394
2395         wake_up_interruptible_poll(&dev->wait, POLLIN | POLLRDNORM);
2396 }
2397 EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2398
2399 struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2400                                          struct list_head *head)
2401 {
2402         struct vhost_msg_node *node = NULL;
2403
2404         spin_lock(&dev->iotlb_lock);
2405         if (!list_empty(head)) {
2406                 node = list_first_entry(head, struct vhost_msg_node,
2407                                         node);
2408                 list_del(&node->node);
2409         }
2410         spin_unlock(&dev->iotlb_lock);
2411
2412         return node;
2413 }
2414 EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2415
2416
2417 static int __init vhost_init(void)
2418 {
2419         return 0;
2420 }
2421
2422 static void __exit vhost_exit(void)
2423 {
2424 }
2425
2426 module_init(vhost_init);
2427 module_exit(vhost_exit);
2428
2429 MODULE_VERSION("0.0.1");
2430 MODULE_LICENSE("GPL v2");
2431 MODULE_AUTHOR("Michael S. Tsirkin");
2432 MODULE_DESCRIPTION("Host kernel accelerator for virtio");