]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blob - drivers/vhost/vhost.c
projects / mirror_ubuntu-bionic-kernel.git / blob
? search:


7cbc18bdaf672f4b5b9c48c9f8f04fe0220eaf8a
[mirror_ubuntu-bionic-kernel.git] / drivers / vhost / vhost.c
1 /* Copyright (C) 2009 Red Hat, Inc.
2 * Copyright (C) 2006 Rusty Russell IBM Corporation
3 *
4 * Author: Michael S. Tsirkin <mst@redhat.com>
5 *
6 * Inspiration, some code, and most witty comments come from
7 * Documentation/virtual/lguest/lguest.c, by Rusty Russell
8 *
9 * This work is licensed under the terms of the GNU GPL, version 2.
10 *
11 * Generic code for virtio server in host kernel.
12 */
13
14 #include <linux/eventfd.h>
15 #include <linux/vhost.h>
16 #include <linux/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 #include <linux/nospec.h>
34
35 #include "vhost.h"
36
37 static ushort max_mem_regions = 64;
38 module_param(max_mem_regions, ushort, 0444);
39 MODULE_PARM_DESC(max_mem_regions,
40 "Maximum number of memory regions in memory map. (default: 64)");
41 static int max_iotlb_entries = 2048;
42 module_param(max_iotlb_entries, int, 0444);
43 MODULE_PARM_DESC(max_iotlb_entries,
44 "Maximum number of iotlb entries. (default: 2048)");
45
46 enum {
47 VHOST_MEMORY_F_LOG = 0x1,
48 };
49
50 #define vhost_used_event(vq) ((__virtio16 __user *)&vq->avail->ring[vq->num])
51 #define vhost_avail_event(vq) ((__virtio16 __user *)&vq->used->ring[vq->num])
52
53 INTERVAL_TREE_DEFINE(struct vhost_umem_node,
54 rb, __u64, __subtree_last,
55 START, LAST, static inline, vhost_umem_interval_tree);
56
57 #ifdef CONFIG_VHOST_CROSS_ENDIAN_LEGACY
58 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
59 {
60 vq->user_be = !virtio_legacy_is_little_endian();
61 }
62
63 static void vhost_enable_cross_endian_big(struct vhost_virtqueue *vq)
64 {
65 vq->user_be = true;
66 }
67
68 static void vhost_enable_cross_endian_little(struct vhost_virtqueue *vq)
69 {
70 vq->user_be = false;
71 }
72
73 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
74 {
75 struct vhost_vring_state s;
76
77 if (vq->private_data)
78 return -EBUSY;
79
80 if (copy_from_user(&s, argp, sizeof(s)))
81 return -EFAULT;
82
83 if (s.num != VHOST_VRING_LITTLE_ENDIAN &&
84 s.num != VHOST_VRING_BIG_ENDIAN)
85 return -EINVAL;
86
87 if (s.num == VHOST_VRING_BIG_ENDIAN)
88 vhost_enable_cross_endian_big(vq);
89 else
90 vhost_enable_cross_endian_little(vq);
91
92 return 0;
93 }
94
95 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
96 int __user *argp)
97 {
98 struct vhost_vring_state s = {
99 .index = idx,
100 .num = vq->user_be
101 };
102
103 if (copy_to_user(argp, &s, sizeof(s)))
104 return -EFAULT;
105
106 return 0;
107 }
108
109 static void vhost_init_is_le(struct vhost_virtqueue *vq)
110 {
111 /* Note for legacy virtio: user_be is initialized at reset time
112 * according to the host endianness. If userspace does not set an
113 * explicit endianness, the default behavior is native endian, as
114 * expected by legacy virtio.
115 */
116 vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1) || !vq->user_be;
117 }
118 #else
119 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
120 {
121 }
122
123 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
124 {
125 return -ENOIOCTLCMD;
126 }
127
128 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
129 int __user *argp)
130 {
131 return -ENOIOCTLCMD;
132 }
133
134 static void vhost_init_is_le(struct vhost_virtqueue *vq)
135 {
136 vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1)
137 || virtio_legacy_is_little_endian();
138 }
139 #endif /* CONFIG_VHOST_CROSS_ENDIAN_LEGACY */
140
141 static void vhost_reset_is_le(struct vhost_virtqueue *vq)
142 {
143 vhost_init_is_le(vq);
144 }
145
146 struct vhost_flush_struct {
147 struct vhost_work work;
148 struct completion wait_event;
149 };
150
151 static void vhost_flush_work(struct vhost_work *work)
152 {
153 struct vhost_flush_struct *s;
154
155 s = container_of(work, struct vhost_flush_struct, work);
156 complete(&s->wait_event);
157 }
158
159 static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
160 poll_table *pt)
161 {
162 struct vhost_poll *poll;
163
164 poll = container_of(pt, struct vhost_poll, table);
165 poll->wqh = wqh;
166 add_wait_queue(wqh, &poll->wait);
167 }
168
169 static int vhost_poll_wakeup(wait_queue_entry_t *wait, unsigned mode, int sync,
170 void *key)
171 {
172 struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
173
174 if (!((unsigned long)key & poll->mask))
175 return 0;
176
177 vhost_poll_queue(poll);
178 return 0;
179 }
180
181 void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
182 {
183 clear_bit(VHOST_WORK_QUEUED, &work->flags);
184 work->fn = fn;
185 init_waitqueue_head(&work->done);
186 }
187 EXPORT_SYMBOL_GPL(vhost_work_init);
188
189 /* Init poll structure */
190 void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
191 unsigned long mask, struct vhost_dev *dev)
192 {
193 init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
194 init_poll_funcptr(&poll->table, vhost_poll_func);
195 poll->mask = mask;
196 poll->dev = dev;
197 poll->wqh = NULL;
198
199 vhost_work_init(&poll->work, fn);
200 }
201 EXPORT_SYMBOL_GPL(vhost_poll_init);
202
203 /* Start polling a file. We add ourselves to file's wait queue. The caller must
204 * keep a reference to a file until after vhost_poll_stop is called. */
205 int vhost_poll_start(struct vhost_poll *poll, struct file *file)
206 {
207 unsigned long mask;
208 int ret = 0;
209
210 if (poll->wqh)
211 return 0;
212
213 mask = file->f_op->poll(file, &poll->table);
214 if (mask)
215 vhost_poll_wakeup(&poll->wait, 0, 0, (void *)mask);
216 if (mask & POLLERR) {
217 vhost_poll_stop(poll);
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_USED);
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 mutex_lock(&dev->mutex);
998 vhost_dev_lock_vqs(dev);
999 switch (msg->type) {
1000 case VHOST_IOTLB_UPDATE:
1001 if (!dev->iotlb) {
1002 ret = -EFAULT;
1003 break;
1004 }
1005 if (umem_access_ok(msg->uaddr, msg->size, msg->perm)) {
1006 ret = -EFAULT;
1007 break;
1008 }
1009 vhost_vq_meta_reset(dev);
1010 if (vhost_new_umem_range(dev->iotlb, msg->iova, msg->size,
1011 msg->iova + msg->size - 1,
1012 msg->uaddr, msg->perm)) {
1013 ret = -ENOMEM;
1014 break;
1015 }
1016 vhost_iotlb_notify_vq(dev, msg);
1017 break;
1018 case VHOST_IOTLB_INVALIDATE:
1019 if (!dev->iotlb) {
1020 ret = -EFAULT;
1021 break;
1022 }
1023 vhost_vq_meta_reset(dev);
1024 vhost_del_umem_range(dev->iotlb, msg->iova,
1025 msg->iova + msg->size - 1);
1026 break;
1027 default:
1028 ret = -EINVAL;
1029 break;
1030 }
1031
1032 vhost_dev_unlock_vqs(dev);
1033 mutex_unlock(&dev->mutex);
1034
1035 return ret;
1036 }
1037 ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
1038 struct iov_iter *from)
1039 {
1040 struct vhost_msg_node node;
1041 unsigned size = sizeof(struct vhost_msg);
1042 size_t ret;
1043 int err;
1044
1045 if (iov_iter_count(from) < size)
1046 return 0;
1047 ret = copy_from_iter(&node.msg, size, from);
1048 if (ret != size)
1049 goto done;
1050
1051 switch (node.msg.type) {
1052 case VHOST_IOTLB_MSG:
1053 err = vhost_process_iotlb_msg(dev, &node.msg.iotlb);
1054 if (err)
1055 ret = err;
1056 break;
1057 default:
1058 ret = -EINVAL;
1059 break;
1060 }
1061
1062 done:
1063 return ret;
1064 }
1065 EXPORT_SYMBOL(vhost_chr_write_iter);
1066
1067 unsigned int vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1068 poll_table *wait)
1069 {
1070 unsigned int mask = 0;
1071
1072 poll_wait(file, &dev->wait, wait);
1073
1074 if (!list_empty(&dev->read_list))
1075 mask |= POLLIN | POLLRDNORM;
1076
1077 return mask;
1078 }
1079 EXPORT_SYMBOL(vhost_chr_poll);
1080
1081 ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1082 int noblock)
1083 {
1084 DEFINE_WAIT(wait);
1085 struct vhost_msg_node *node;
1086 ssize_t ret = 0;
1087 unsigned size = sizeof(struct vhost_msg);
1088
1089 if (iov_iter_count(to) < size)
1090 return 0;
1091
1092 while (1) {
1093 if (!noblock)
1094 prepare_to_wait(&dev->wait, &wait,
1095 TASK_INTERRUPTIBLE);
1096
1097 node = vhost_dequeue_msg(dev, &dev->read_list);
1098 if (node)
1099 break;
1100 if (noblock) {
1101 ret = -EAGAIN;
1102 break;
1103 }
1104 if (signal_pending(current)) {
1105 ret = -ERESTARTSYS;
1106 break;
1107 }
1108 if (!dev->iotlb) {
1109 ret = -EBADFD;
1110 break;
1111 }
1112
1113 schedule();
1114 }
1115
1116 if (!noblock)
1117 finish_wait(&dev->wait, &wait);
1118
1119 if (node) {
1120 ret = copy_to_iter(&node->msg, size, to);
1121
1122 if (ret != size || node->msg.type != VHOST_IOTLB_MISS) {
1123 kfree(node);
1124 return ret;
1125 }
1126
1127 vhost_enqueue_msg(dev, &dev->pending_list, node);
1128 }
1129
1130 return ret;
1131 }
1132 EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1133
1134 static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1135 {
1136 struct vhost_dev *dev = vq->dev;
1137 struct vhost_msg_node *node;
1138 struct vhost_iotlb_msg *msg;
1139
1140 node = vhost_new_msg(vq, VHOST_IOTLB_MISS);
1141 if (!node)
1142 return -ENOMEM;
1143
1144 msg = &node->msg.iotlb;
1145 msg->type = VHOST_IOTLB_MISS;
1146 msg->iova = iova;
1147 msg->perm = access;
1148
1149 vhost_enqueue_msg(dev, &dev->read_list, node);
1150
1151 return 0;
1152 }
1153
1154 static int vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1155 struct vring_desc __user *desc,
1156 struct vring_avail __user *avail,
1157 struct vring_used __user *used)
1158
1159 {
1160 size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1161
1162 return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
1163 access_ok(VERIFY_READ, avail,
1164 sizeof *avail + num * sizeof *avail->ring + s) &&
1165 access_ok(VERIFY_WRITE, used,
1166 sizeof *used + num * sizeof *used->ring + s);
1167 }
1168
1169 static void vhost_vq_meta_update(struct vhost_virtqueue *vq,
1170 const struct vhost_umem_node *node,
1171 int type)
1172 {
1173 int access = (type == VHOST_ADDR_USED) ?
1174 VHOST_ACCESS_WO : VHOST_ACCESS_RO;
1175
1176 if (likely(node->perm & access))
1177 vq->meta_iotlb[type] = node;
1178 }
1179
1180 static int iotlb_access_ok(struct vhost_virtqueue *vq,
1181 int access, u64 addr, u64 len, int type)
1182 {
1183 const struct vhost_umem_node *node;
1184 struct vhost_umem *umem = vq->iotlb;
1185 u64 s = 0, size, orig_addr = addr, last = addr + len - 1;
1186
1187 if (vhost_vq_meta_fetch(vq, addr, len, type))
1188 return true;
1189
1190 while (len > s) {
1191 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1192 addr,
1193 last);
1194 if (node == NULL || node->start > addr) {
1195 vhost_iotlb_miss(vq, addr, access);
1196 return false;
1197 } else if (!(node->perm & access)) {
1198 /* Report the possible access violation by
1199 * request another translation from userspace.
1200 */
1201 return false;
1202 }
1203
1204 size = node->size - addr + node->start;
1205
1206 if (orig_addr == addr && size >= len)
1207 vhost_vq_meta_update(vq, node, type);
1208
1209 s += size;
1210 addr += size;
1211 }
1212
1213 return true;
1214 }
1215
1216 int vq_iotlb_prefetch(struct vhost_virtqueue *vq)
1217 {
1218 size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1219 unsigned int num = vq->num;
1220
1221 if (!vq->iotlb)
1222 return 1;
1223
1224 return iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->desc,
1225 num * sizeof(*vq->desc), VHOST_ADDR_DESC) &&
1226 iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->avail,
1227 sizeof *vq->avail +
1228 num * sizeof(*vq->avail->ring) + s,
1229 VHOST_ADDR_AVAIL) &&
1230 iotlb_access_ok(vq, VHOST_ACCESS_WO, (u64)(uintptr_t)vq->used,
1231 sizeof *vq->used +
1232 num * sizeof(*vq->used->ring) + s,
1233 VHOST_ADDR_USED);
1234 }
1235 EXPORT_SYMBOL_GPL(vq_iotlb_prefetch);
1236
1237 /* Can we log writes? */
1238 /* Caller should have device mutex but not vq mutex */
1239 int vhost_log_access_ok(struct vhost_dev *dev)
1240 {
1241 return memory_access_ok(dev, dev->umem, 1);
1242 }
1243 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1244
1245 /* Verify access for write logging. */
1246 /* Caller should have vq mutex and device mutex */
1247 static int vq_log_access_ok(struct vhost_virtqueue *vq,
1248 void __user *log_base)
1249 {
1250 size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1251
1252 return vq_memory_access_ok(log_base, vq->umem,
1253 vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1254 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
1255 sizeof *vq->used +
1256 vq->num * sizeof *vq->used->ring + s));
1257 }
1258
1259 /* Can we start vq? */
1260 /* Caller should have vq mutex and device mutex */
1261 int vhost_vq_access_ok(struct vhost_virtqueue *vq)
1262 {
1263 if (!vq_log_access_ok(vq, vq->log_base))
1264 return 0;
1265
1266 /* Access validation occurs at prefetch time with IOTLB */
1267 if (vq->iotlb)
1268 return 1;
1269
1270 return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used);
1271 }
1272 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1273
1274 static struct vhost_umem *vhost_umem_alloc(void)
1275 {
1276 struct vhost_umem *umem = kvzalloc(sizeof(*umem), GFP_KERNEL);
1277
1278 if (!umem)
1279 return NULL;
1280
1281 umem->umem_tree = RB_ROOT_CACHED;
1282 umem->numem = 0;
1283 INIT_LIST_HEAD(&umem->umem_list);
1284
1285 return umem;
1286 }
1287
1288 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1289 {
1290 struct vhost_memory mem, *newmem;
1291 struct vhost_memory_region *region;
1292 struct vhost_umem *newumem, *oldumem;
1293 unsigned long size = offsetof(struct vhost_memory, regions);
1294 int i;
1295
1296 if (copy_from_user(&mem, m, size))
1297 return -EFAULT;
1298 if (mem.padding)
1299 return -EOPNOTSUPP;
1300 if (mem.nregions > max_mem_regions)
1301 return -E2BIG;
1302 newmem = kvzalloc(size + mem.nregions * sizeof(*m->regions), GFP_KERNEL);
1303 if (!newmem)
1304 return -ENOMEM;
1305
1306 memcpy(newmem, &mem, size);
1307 if (copy_from_user(newmem->regions, m->regions,
1308 mem.nregions * sizeof *m->regions)) {
1309 kvfree(newmem);
1310 return -EFAULT;
1311 }
1312
1313 newumem = vhost_umem_alloc();
1314 if (!newumem) {
1315 kvfree(newmem);
1316 return -ENOMEM;
1317 }
1318
1319 for (region = newmem->regions;
1320 region < newmem->regions + mem.nregions;
1321 region++) {
1322 if (vhost_new_umem_range(newumem,
1323 region->guest_phys_addr,
1324 region->memory_size,
1325 region->guest_phys_addr +
1326 region->memory_size - 1,
1327 region->userspace_addr,
1328 VHOST_ACCESS_RW))
1329 goto err;
1330 }
1331
1332 if (!memory_access_ok(d, newumem, 0))
1333 goto err;
1334
1335 oldumem = d->umem;
1336 d->umem = newumem;
1337
1338 /* All memory accesses are done under some VQ mutex. */
1339 for (i = 0; i < d->nvqs; ++i) {
1340 mutex_lock(&d->vqs[i]->mutex);
1341 d->vqs[i]->umem = newumem;
1342 mutex_unlock(&d->vqs[i]->mutex);
1343 }
1344
1345 kvfree(newmem);
1346 vhost_umem_clean(oldumem);
1347 return 0;
1348
1349 err:
1350 vhost_umem_clean(newumem);
1351 kvfree(newmem);
1352 return -EFAULT;
1353 }
1354
1355 long vhost_vring_ioctl(struct vhost_dev *d, int ioctl, void __user *argp)
1356 {
1357 struct file *eventfp, *filep = NULL;
1358 bool pollstart = false, pollstop = false;
1359 struct eventfd_ctx *ctx = NULL;
1360 u32 __user *idxp = argp;
1361 struct vhost_virtqueue *vq;
1362 struct vhost_vring_state s;
1363 struct vhost_vring_file f;
1364 struct vhost_vring_addr a;
1365 u32 idx;
1366 long r;
1367
1368 r = get_user(idx, idxp);
1369 if (r < 0)
1370 return r;
1371 if (idx >= d->nvqs)
1372 return -ENOBUFS;
1373
1374 idx = array_index_nospec(idx, d->nvqs);
1375 vq = d->vqs[idx];
1376
1377 mutex_lock(&vq->mutex);
1378
1379 switch (ioctl) {
1380 case VHOST_SET_VRING_NUM:
1381 /* Resizing ring with an active backend?
1382 * You don't want to do that. */
1383 if (vq->private_data) {
1384 r = -EBUSY;
1385 break;
1386 }
1387 if (copy_from_user(&s, argp, sizeof s)) {
1388 r = -EFAULT;
1389 break;
1390 }
1391 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
1392 r = -EINVAL;
1393 break;
1394 }
1395 vq->num = s.num;
1396 break;
1397 case VHOST_SET_VRING_BASE:
1398 /* Moving base with an active backend?
1399 * You don't want to do that. */
1400 if (vq->private_data) {
1401 r = -EBUSY;
1402 break;
1403 }
1404 if (copy_from_user(&s, argp, sizeof s)) {
1405 r = -EFAULT;
1406 break;
1407 }
1408 if (s.num > 0xffff) {
1409 r = -EINVAL;
1410 break;
1411 }
1412 vq->last_avail_idx = s.num;
1413 /* Forget the cached index value. */
1414 vq->avail_idx = vq->last_avail_idx;
1415 break;
1416 case VHOST_GET_VRING_BASE:
1417 s.index = idx;
1418 s.num = vq->last_avail_idx;
1419 if (copy_to_user(argp, &s, sizeof s))
1420 r = -EFAULT;
1421 break;
1422 case VHOST_SET_VRING_ADDR:
1423 if (copy_from_user(&a, argp, sizeof a)) {
1424 r = -EFAULT;
1425 break;
1426 }
1427 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
1428 r = -EOPNOTSUPP;
1429 break;
1430 }
1431 /* For 32bit, verify that the top 32bits of the user
1432 data are set to zero. */
1433 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1434 (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1435 (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
1436 r = -EFAULT;
1437 break;
1438 }
1439
1440 /* Make sure it's safe to cast pointers to vring types. */
1441 BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1442 BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1443 if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1444 (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1445 (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1))) {
1446 r = -EINVAL;
1447 break;
1448 }
1449
1450 /* We only verify access here if backend is configured.
1451 * If it is not, we don't as size might not have been setup.
1452 * We will verify when backend is configured. */
1453 if (vq->private_data) {
1454 if (!vq_access_ok(vq, vq->num,
1455 (void __user *)(unsigned long)a.desc_user_addr,
1456 (void __user *)(unsigned long)a.avail_user_addr,
1457 (void __user *)(unsigned long)a.used_user_addr)) {
1458 r = -EINVAL;
1459 break;
1460 }
1461
1462 /* Also validate log access for used ring if enabled. */
1463 if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
1464 !log_access_ok(vq->log_base, a.log_guest_addr,
1465 sizeof *vq->used +
1466 vq->num * sizeof *vq->used->ring)) {
1467 r = -EINVAL;
1468 break;
1469 }
1470 }
1471
1472 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1473 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1474 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1475 vq->log_addr = a.log_guest_addr;
1476 vq->used = (void __user *)(unsigned long)a.used_user_addr;
1477 break;
1478 case VHOST_SET_VRING_KICK:
1479 if (copy_from_user(&f, argp, sizeof f)) {
1480 r = -EFAULT;
1481 break;
1482 }
1483 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1484 if (IS_ERR(eventfp)) {
1485 r = PTR_ERR(eventfp);
1486 break;
1487 }
1488 if (eventfp != vq->kick) {
1489 pollstop = (filep = vq->kick) != NULL;
1490 pollstart = (vq->kick = eventfp) != NULL;
1491 } else
1492 filep = eventfp;
1493 break;
1494 case VHOST_SET_VRING_CALL:
1495 if (copy_from_user(&f, argp, sizeof f)) {
1496 r = -EFAULT;
1497 break;
1498 }
1499 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1500 if (IS_ERR(eventfp)) {
1501 r = PTR_ERR(eventfp);
1502 break;
1503 }
1504 if (eventfp != vq->call) {
1505 filep = vq->call;
1506 ctx = vq->call_ctx;
1507 vq->call = eventfp;
1508 vq->call_ctx = eventfp ?
1509 eventfd_ctx_fileget(eventfp) : NULL;
1510 } else
1511 filep = eventfp;
1512 break;
1513 case VHOST_SET_VRING_ERR:
1514 if (copy_from_user(&f, argp, sizeof f)) {
1515 r = -EFAULT;
1516 break;
1517 }
1518 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1519 if (IS_ERR(eventfp)) {
1520 r = PTR_ERR(eventfp);
1521 break;
1522 }
1523 if (eventfp != vq->error) {
1524 filep = vq->error;
1525 vq->error = eventfp;
1526 ctx = vq->error_ctx;
1527 vq->error_ctx = eventfp ?
1528 eventfd_ctx_fileget(eventfp) : NULL;
1529 } else
1530 filep = eventfp;
1531 break;
1532 case VHOST_SET_VRING_ENDIAN:
1533 r = vhost_set_vring_endian(vq, argp);
1534 break;
1535 case VHOST_GET_VRING_ENDIAN:
1536 r = vhost_get_vring_endian(vq, idx, argp);
1537 break;
1538 case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
1539 if (copy_from_user(&s, argp, sizeof(s))) {
1540 r = -EFAULT;
1541 break;
1542 }
1543 vq->busyloop_timeout = s.num;
1544 break;
1545 case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
1546 s.index = idx;
1547 s.num = vq->busyloop_timeout;
1548 if (copy_to_user(argp, &s, sizeof(s)))
1549 r = -EFAULT;
1550 break;
1551 default:
1552 r = -ENOIOCTLCMD;
1553 }
1554
1555 if (pollstop && vq->handle_kick)
1556 vhost_poll_stop(&vq->poll);
1557
1558 if (ctx)
1559 eventfd_ctx_put(ctx);
1560 if (filep)
1561 fput(filep);
1562
1563 if (pollstart && vq->handle_kick)
1564 r = vhost_poll_start(&vq->poll, vq->kick);
1565
1566 mutex_unlock(&vq->mutex);
1567
1568 if (pollstop && vq->handle_kick)
1569 vhost_poll_flush(&vq->poll);
1570 return r;
1571 }
1572 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
1573
1574 int vhost_init_device_iotlb(struct vhost_dev *d, bool enabled)
1575 {
1576 struct vhost_umem *niotlb, *oiotlb;
1577 int i;
1578
1579 niotlb = vhost_umem_alloc();
1580 if (!niotlb)
1581 return -ENOMEM;
1582
1583 oiotlb = d->iotlb;
1584 d->iotlb = niotlb;
1585
1586 for (i = 0; i < d->nvqs; ++i) {
1587 struct vhost_virtqueue *vq = d->vqs[i];
1588
1589 mutex_lock(&vq->mutex);
1590 vq->iotlb = niotlb;
1591 __vhost_vq_meta_reset(vq);
1592 mutex_unlock(&vq->mutex);
1593 }
1594
1595 vhost_umem_clean(oiotlb);
1596
1597 return 0;
1598 }
1599 EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
1600
1601 /* Caller must have device mutex */
1602 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1603 {
1604 struct file *eventfp, *filep = NULL;
1605 struct eventfd_ctx *ctx = NULL;
1606 u64 p;
1607 long r;
1608 int i, fd;
1609
1610 /* If you are not the owner, you can become one */
1611 if (ioctl == VHOST_SET_OWNER) {
1612 r = vhost_dev_set_owner(d);
1613 goto done;
1614 }
1615
1616 /* You must be the owner to do anything else */
1617 r = vhost_dev_check_owner(d);
1618 if (r)
1619 goto done;
1620
1621 switch (ioctl) {
1622 case VHOST_SET_MEM_TABLE:
1623 r = vhost_set_memory(d, argp);
1624 break;
1625 case VHOST_SET_LOG_BASE:
1626 if (copy_from_user(&p, argp, sizeof p)) {
1627 r = -EFAULT;
1628 break;
1629 }
1630 if ((u64)(unsigned long)p != p) {
1631 r = -EFAULT;
1632 break;
1633 }
1634 for (i = 0; i < d->nvqs; ++i) {
1635 struct vhost_virtqueue *vq;
1636 void __user *base = (void __user *)(unsigned long)p;
1637 vq = d->vqs[i];
1638 mutex_lock(&vq->mutex);
1639 /* If ring is inactive, will check when it's enabled. */
1640 if (vq->private_data && !vq_log_access_ok(vq, base))
1641 r = -EFAULT;
1642 else
1643 vq->log_base = base;
1644 mutex_unlock(&vq->mutex);
1645 }
1646 break;
1647 case VHOST_SET_LOG_FD:
1648 r = get_user(fd, (int __user *)argp);
1649 if (r < 0)
1650 break;
1651 eventfp = fd == -1 ? NULL : eventfd_fget(fd);
1652 if (IS_ERR(eventfp)) {
1653 r = PTR_ERR(eventfp);
1654 break;
1655 }
1656 if (eventfp != d->log_file) {
1657 filep = d->log_file;
1658 d->log_file = eventfp;
1659 ctx = d->log_ctx;
1660 d->log_ctx = eventfp ?
1661 eventfd_ctx_fileget(eventfp) : NULL;
1662 } else
1663 filep = eventfp;
1664 for (i = 0; i < d->nvqs; ++i) {
1665 mutex_lock(&d->vqs[i]->mutex);
1666 d->vqs[i]->log_ctx = d->log_ctx;
1667 mutex_unlock(&d->vqs[i]->mutex);
1668 }
1669 if (ctx)
1670 eventfd_ctx_put(ctx);
1671 if (filep)
1672 fput(filep);
1673 break;
1674 default:
1675 r = -ENOIOCTLCMD;
1676 break;
1677 }
1678 done:
1679 return r;
1680 }
1681 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1682
1683 /* TODO: This is really inefficient. We need something like get_user()
1684 * (instruction directly accesses the data, with an exception table entry
1685 * returning -EFAULT). See Documentation/x86/exception-tables.txt.
1686 */
1687 static int set_bit_to_user(int nr, void __user *addr)
1688 {
1689 unsigned long log = (unsigned long)addr;
1690 struct page *page;
1691 void *base;
1692 int bit = nr + (log % PAGE_SIZE) * 8;
1693 int r;
1694
1695 r = get_user_pages_fast(log, 1, 1, &page);
1696 if (r < 0)
1697 return r;
1698 BUG_ON(r != 1);
1699 base = kmap_atomic(page);
1700 set_bit(bit, base);
1701 kunmap_atomic(base);
1702 set_page_dirty_lock(page);
1703 put_page(page);
1704 return 0;
1705 }
1706
1707 static int log_write(void __user *log_base,
1708 u64 write_address, u64 write_length)
1709 {
1710 u64 write_page = write_address / VHOST_PAGE_SIZE;
1711 int r;
1712
1713 if (!write_length)
1714 return 0;
1715 write_length += write_address % VHOST_PAGE_SIZE;
1716 for (;;) {
1717 u64 base = (u64)(unsigned long)log_base;
1718 u64 log = base + write_page / 8;
1719 int bit = write_page % 8;
1720 if ((u64)(unsigned long)log != log)
1721 return -EFAULT;
1722 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1723 if (r < 0)
1724 return r;
1725 if (write_length <= VHOST_PAGE_SIZE)
1726 break;
1727 write_length -= VHOST_PAGE_SIZE;
1728 write_page += 1;
1729 }
1730 return r;
1731 }
1732
1733 static int log_write_hva(struct vhost_virtqueue *vq, u64 hva, u64 len)
1734 {
1735 struct vhost_umem *umem = vq->umem;
1736 struct vhost_umem_node *u;
1737 u64 start, end, l, min;
1738 int r;
1739 bool hit = false;
1740
1741 while (len) {
1742 min = len;
1743 /* More than one GPAs can be mapped into a single HVA. So
1744 * iterate all possible umems here to be safe.
1745 */
1746 list_for_each_entry(u, &umem->umem_list, link) {
1747 if (u->userspace_addr > hva - 1 + len ||
1748 u->userspace_addr - 1 + u->size < hva)
1749 continue;
1750 start = max(u->userspace_addr, hva);
1751 end = min(u->userspace_addr - 1 + u->size,
1752 hva - 1 + len);
1753 l = end - start + 1;
1754 r = log_write(vq->log_base,
1755 u->start + start - u->userspace_addr,
1756 l);
1757 if (r < 0)
1758 return r;
1759 hit = true;
1760 min = min(l, min);
1761 }
1762
1763 if (!hit)
1764 return -EFAULT;
1765
1766 len -= min;
1767 hva += min;
1768 }
1769
1770 return 0;
1771 }
1772
1773 static int log_used(struct vhost_virtqueue *vq, u64 used_offset, u64 len)
1774 {
1775 struct iovec iov[64];
1776 int i, ret;
1777
1778 if (!vq->iotlb)
1779 return log_write(vq->log_base, vq->log_addr + used_offset, len);
1780
1781 ret = translate_desc(vq, (uintptr_t)vq->used + used_offset,
1782 len, iov, 64, VHOST_ACCESS_WO);
1783 if (ret < 0)
1784 return ret;
1785
1786 for (i = 0; i < ret; i++) {
1787 ret = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1788 iov[i].iov_len);
1789 if (ret)
1790 return ret;
1791 }
1792
1793 return 0;
1794 }
1795
1796 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1797 unsigned int log_num, u64 len, struct iovec *iov, int count)
1798 {
1799 int i, r;
1800
1801 /* Make sure data written is seen before log. */
1802 smp_wmb();
1803
1804 if (vq->iotlb) {
1805 for (i = 0; i < count; i++) {
1806 r = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1807 iov[i].iov_len);
1808 if (r < 0)
1809 return r;
1810 }
1811 return 0;
1812 }
1813
1814 for (i = 0; i < log_num; ++i) {
1815 u64 l = min(log[i].len, len);
1816 r = log_write(vq->log_base, log[i].addr, l);
1817 if (r < 0)
1818 return r;
1819 len -= l;
1820 if (!len) {
1821 if (vq->log_ctx)
1822 eventfd_signal(vq->log_ctx, 1);
1823 return 0;
1824 }
1825 }
1826 /* Length written exceeds what we have stored. This is a bug. */
1827 BUG();
1828 return 0;
1829 }
1830 EXPORT_SYMBOL_GPL(vhost_log_write);
1831
1832 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1833 {
1834 void __user *used;
1835 if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
1836 &vq->used->flags) < 0)
1837 return -EFAULT;
1838 if (unlikely(vq->log_used)) {
1839 /* Make sure the flag is seen before log. */
1840 smp_wmb();
1841 /* Log used flag write. */
1842 used = &vq->used->flags;
1843 log_used(vq, (used - (void __user *)vq->used),
1844 sizeof vq->used->flags);
1845 if (vq->log_ctx)
1846 eventfd_signal(vq->log_ctx, 1);
1847 }
1848 return 0;
1849 }
1850
1851 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1852 {
1853 if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
1854 vhost_avail_event(vq)))
1855 return -EFAULT;
1856 if (unlikely(vq->log_used)) {
1857 void __user *used;
1858 /* Make sure the event is seen before log. */
1859 smp_wmb();
1860 /* Log avail event write */
1861 used = vhost_avail_event(vq);
1862 log_used(vq, (used - (void __user *)vq->used),
1863 sizeof *vhost_avail_event(vq));
1864 if (vq->log_ctx)
1865 eventfd_signal(vq->log_ctx, 1);
1866 }
1867 return 0;
1868 }
1869
1870 int vhost_vq_init_access(struct vhost_virtqueue *vq)
1871 {
1872 __virtio16 last_used_idx;
1873 int r;
1874 bool is_le = vq->is_le;
1875
1876 if (!vq->private_data)
1877 return 0;
1878
1879 vhost_init_is_le(vq);
1880
1881 r = vhost_update_used_flags(vq);
1882 if (r)
1883 goto err;
1884 vq->signalled_used_valid = false;
1885 if (!vq->iotlb &&
1886 !access_ok(VERIFY_READ, &vq->used->idx, sizeof vq->used->idx)) {
1887 r = -EFAULT;
1888 goto err;
1889 }
1890 r = vhost_get_used(vq, last_used_idx, &vq->used->idx);
1891 if (r) {
1892 vq_err(vq, "Can't access used idx at %p\n",
1893 &vq->used->idx);
1894 goto err;
1895 }
1896 vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
1897 return 0;
1898
1899 err:
1900 vq->is_le = is_le;
1901 return r;
1902 }
1903 EXPORT_SYMBOL_GPL(vhost_vq_init_access);
1904
1905 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
1906 struct iovec iov[], int iov_size, int access)
1907 {
1908 const struct vhost_umem_node *node;
1909 struct vhost_dev *dev = vq->dev;
1910 struct vhost_umem *umem = dev->iotlb ? dev->iotlb : dev->umem;
1911 struct iovec *_iov;
1912 u64 s = 0;
1913 int ret = 0;
1914
1915 while ((u64)len > s) {
1916 u64 size;
1917 if (unlikely(ret >= iov_size)) {
1918 ret = -ENOBUFS;
1919 break;
1920 }
1921
1922 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1923 addr, addr + len - 1);
1924 if (node == NULL || node->start > addr) {
1925 if (umem != dev->iotlb) {
1926 ret = -EFAULT;
1927 break;
1928 }
1929 ret = -EAGAIN;
1930 break;
1931 } else if (!(node->perm & access)) {
1932 ret = -EPERM;
1933 break;
1934 }
1935
1936 _iov = iov + ret;
1937 size = node->size - addr + node->start;
1938 _iov->iov_len = min((u64)len - s, size);
1939 _iov->iov_base = (void __user *)(unsigned long)
1940 (node->userspace_addr + addr - node->start);
1941 s += size;
1942 addr += size;
1943 ++ret;
1944 }
1945
1946 if (ret == -EAGAIN)
1947 vhost_iotlb_miss(vq, addr, access);
1948 return ret;
1949 }
1950
1951 /* Each buffer in the virtqueues is actually a chain of descriptors. This
1952 * function returns the next descriptor in the chain,
1953 * or -1U if we're at the end. */
1954 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
1955 {
1956 unsigned int next;
1957
1958 /* If this descriptor says it doesn't chain, we're done. */
1959 if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
1960 return -1U;
1961
1962 /* Check they're not leading us off end of descriptors. */
1963 next = vhost16_to_cpu(vq, desc->next);
1964 /* Make sure compiler knows to grab that: we don't want it changing! */
1965 /* We will use the result as an index in an array, so most
1966 * architectures only need a compiler barrier here. */
1967 read_barrier_depends();
1968
1969 return next;
1970 }
1971
1972 static int get_indirect(struct vhost_virtqueue *vq,
1973 struct iovec iov[], unsigned int iov_size,
1974 unsigned int *out_num, unsigned int *in_num,
1975 struct vhost_log *log, unsigned int *log_num,
1976 struct vring_desc *indirect)
1977 {
1978 struct vring_desc desc;
1979 unsigned int i = 0, count, found = 0;
1980 u32 len = vhost32_to_cpu(vq, indirect->len);
1981 struct iov_iter from;
1982 int ret, access;
1983
1984 /* Sanity check */
1985 if (unlikely(len % sizeof desc)) {
1986 vq_err(vq, "Invalid length in indirect descriptor: "
1987 "len 0x%llx not multiple of 0x%zx\n",
1988 (unsigned long long)len,
1989 sizeof desc);
1990 return -EINVAL;
1991 }
1992
1993 ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
1994 UIO_MAXIOV, VHOST_ACCESS_RO);
1995 if (unlikely(ret < 0)) {
1996 if (ret != -EAGAIN)
1997 vq_err(vq, "Translation failure %d in indirect.\n", ret);
1998 return ret;
1999 }
2000 iov_iter_init(&from, READ, vq->indirect, ret, len);
2001
2002 /* We will use the result as an address to read from, so most
2003 * architectures only need a compiler barrier here. */
2004 read_barrier_depends();
2005
2006 count = len / sizeof desc;
2007 /* Buffers are chained via a 16 bit next field, so
2008 * we can have at most 2^16 of these. */
2009 if (unlikely(count > USHRT_MAX + 1)) {
2010 vq_err(vq, "Indirect buffer length too big: %d\n",
2011 indirect->len);
2012 return -E2BIG;
2013 }
2014
2015 do {
2016 unsigned iov_count = *in_num + *out_num;
2017 if (unlikely(++found > count)) {
2018 vq_err(vq, "Loop detected: last one at %u "
2019 "indirect size %u\n",
2020 i, count);
2021 return -EINVAL;
2022 }
2023 if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
2024 vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
2025 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2026 return -EINVAL;
2027 }
2028 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
2029 vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
2030 i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2031 return -EINVAL;
2032 }
2033
2034 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2035 access = VHOST_ACCESS_WO;
2036 else
2037 access = VHOST_ACCESS_RO;
2038
2039 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2040 vhost32_to_cpu(vq, desc.len), iov + iov_count,
2041 iov_size - iov_count, access);
2042 if (unlikely(ret < 0)) {
2043 if (ret != -EAGAIN)
2044 vq_err(vq, "Translation failure %d indirect idx %d\n",
2045 ret, i);
2046 return ret;
2047 }
2048 /* If this is an input descriptor, increment that count. */
2049 if (access == VHOST_ACCESS_WO) {
2050 *in_num += ret;
2051 if (unlikely(log)) {
2052 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2053 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2054 ++*log_num;
2055 }
2056 } else {
2057 /* If it's an output descriptor, they're all supposed
2058 * to come before any input descriptors. */
2059 if (unlikely(*in_num)) {
2060 vq_err(vq, "Indirect descriptor "
2061 "has out after in: idx %d\n", i);
2062 return -EINVAL;
2063 }
2064 *out_num += ret;
2065 }
2066 } while ((i = next_desc(vq, &desc)) != -1);
2067 return 0;
2068 }
2069
2070 /* This looks in the virtqueue and for the first available buffer, and converts
2071 * it to an iovec for convenient access. Since descriptors consist of some
2072 * number of output then some number of input descriptors, it's actually two
2073 * iovecs, but we pack them into one and note how many of each there were.
2074 *
2075 * This function returns the descriptor number found, or vq->num (which is
2076 * never a valid descriptor number) if none was found. A negative code is
2077 * returned on error. */
2078 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
2079 struct iovec iov[], unsigned int iov_size,
2080 unsigned int *out_num, unsigned int *in_num,
2081 struct vhost_log *log, unsigned int *log_num)
2082 {
2083 struct vring_desc desc;
2084 unsigned int i, head, found = 0;
2085 u16 last_avail_idx;
2086 __virtio16 avail_idx;
2087 __virtio16 ring_head;
2088 int ret, access;
2089
2090 /* Check it isn't doing very strange things with descriptor numbers. */
2091 last_avail_idx = vq->last_avail_idx;
2092
2093 if (vq->avail_idx == vq->last_avail_idx) {
2094 if (unlikely(vhost_get_avail(vq, avail_idx, &vq->avail->idx))) {
2095 vq_err(vq, "Failed to access avail idx at %p\n",
2096 &vq->avail->idx);
2097 return -EFAULT;
2098 }
2099 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2100
2101 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
2102 vq_err(vq, "Guest moved used index from %u to %u",
2103 last_avail_idx, vq->avail_idx);
2104 return -EFAULT;
2105 }
2106
2107 /* If there's nothing new since last we looked, return
2108 * invalid.
2109 */
2110 if (vq->avail_idx == last_avail_idx)
2111 return vq->num;
2112
2113 /* Only get avail ring entries after they have been
2114 * exposed by guest.
2115 */
2116 smp_rmb();
2117 }
2118
2119 /* Grab the next descriptor number they're advertising, and increment
2120 * the index we've seen. */
2121 if (unlikely(vhost_get_avail(vq, ring_head,
2122 &vq->avail->ring[last_avail_idx & (vq->num - 1)]))) {
2123 vq_err(vq, "Failed to read head: idx %d address %p\n",
2124 last_avail_idx,
2125 &vq->avail->ring[last_avail_idx % vq->num]);
2126 return -EFAULT;
2127 }
2128
2129 head = vhost16_to_cpu(vq, ring_head);
2130
2131 /* If their number is silly, that's an error. */
2132 if (unlikely(head >= vq->num)) {
2133 vq_err(vq, "Guest says index %u > %u is available",
2134 head, vq->num);
2135 return -EINVAL;
2136 }
2137
2138 /* When we start there are none of either input nor output. */
2139 *out_num = *in_num = 0;
2140 if (unlikely(log))
2141 *log_num = 0;
2142
2143 i = head;
2144 do {
2145 unsigned iov_count = *in_num + *out_num;
2146 if (unlikely(i >= vq->num)) {
2147 vq_err(vq, "Desc index is %u > %u, head = %u",
2148 i, vq->num, head);
2149 return -EINVAL;
2150 }
2151 if (unlikely(++found > vq->num)) {
2152 vq_err(vq, "Loop detected: last one at %u "
2153 "vq size %u head %u\n",
2154 i, vq->num, head);
2155 return -EINVAL;
2156 }
2157 ret = vhost_copy_from_user(vq, &desc, vq->desc + i,
2158 sizeof desc);
2159 if (unlikely(ret)) {
2160 vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
2161 i, vq->desc + i);
2162 return -EFAULT;
2163 }
2164 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2165 ret = get_indirect(vq, iov, iov_size,
2166 out_num, in_num,
2167 log, log_num, &desc);
2168 if (unlikely(ret < 0)) {
2169 if (ret != -EAGAIN)
2170 vq_err(vq, "Failure detected "
2171 "in indirect descriptor at idx %d\n", i);
2172 return ret;
2173 }
2174 continue;
2175 }
2176
2177 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2178 access = VHOST_ACCESS_WO;
2179 else
2180 access = VHOST_ACCESS_RO;
2181 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2182 vhost32_to_cpu(vq, desc.len), iov + iov_count,
2183 iov_size - iov_count, access);
2184 if (unlikely(ret < 0)) {
2185 if (ret != -EAGAIN)
2186 vq_err(vq, "Translation failure %d descriptor idx %d\n",
2187 ret, i);
2188 return ret;
2189 }
2190 if (access == VHOST_ACCESS_WO) {
2191 /* If this is an input descriptor,
2192 * increment that count. */
2193 *in_num += ret;
2194 if (unlikely(log)) {
2195 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2196 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2197 ++*log_num;
2198 }
2199 } else {
2200 /* If it's an output descriptor, they're all supposed
2201 * to come before any input descriptors. */
2202 if (unlikely(*in_num)) {
2203 vq_err(vq, "Descriptor has out after in: "
2204 "idx %d\n", i);
2205 return -EINVAL;
2206 }
2207 *out_num += ret;
2208 }
2209 } while ((i = next_desc(vq, &desc)) != -1);
2210
2211 /* On success, increment avail index. */
2212 vq->last_avail_idx++;
2213
2214 /* Assume notifications from guest are disabled at this point,
2215 * if they aren't we would need to update avail_event index. */
2216 BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2217 return head;
2218 }
2219 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2220
2221 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2222 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2223 {
2224 vq->last_avail_idx -= n;
2225 }
2226 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2227
2228 /* After we've used one of their buffers, we tell them about it. We'll then
2229 * want to notify the guest, using eventfd. */
2230 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2231 {
2232 struct vring_used_elem heads = {
2233 cpu_to_vhost32(vq, head),
2234 cpu_to_vhost32(vq, len)
2235 };
2236
2237 return vhost_add_used_n(vq, &heads, 1);
2238 }
2239 EXPORT_SYMBOL_GPL(vhost_add_used);
2240
2241 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2242 struct vring_used_elem *heads,
2243 unsigned count)
2244 {
2245 struct vring_used_elem __user *used;
2246 u16 old, new;
2247 int start;
2248
2249 start = vq->last_used_idx & (vq->num - 1);
2250 used = vq->used->ring + start;
2251 if (count == 1) {
2252 if (vhost_put_user(vq, heads[0].id, &used->id)) {
2253 vq_err(vq, "Failed to write used id");
2254 return -EFAULT;
2255 }
2256 if (vhost_put_user(vq, heads[0].len, &used->len)) {
2257 vq_err(vq, "Failed to write used len");
2258 return -EFAULT;
2259 }
2260 } else if (vhost_copy_to_user(vq, used, heads, count * sizeof *used)) {
2261 vq_err(vq, "Failed to write used");
2262 return -EFAULT;
2263 }
2264 if (unlikely(vq->log_used)) {
2265 /* Make sure data is seen before log. */
2266 smp_wmb();
2267 /* Log used ring entry write. */
2268 log_used(vq, ((void __user *)used - (void __user *)vq->used),
2269 count * sizeof *used);
2270 }
2271 old = vq->last_used_idx;
2272 new = (vq->last_used_idx += count);
2273 /* If the driver never bothers to signal in a very long while,
2274 * used index might wrap around. If that happens, invalidate
2275 * signalled_used index we stored. TODO: make sure driver
2276 * signals at least once in 2^16 and remove this. */
2277 if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2278 vq->signalled_used_valid = false;
2279 return 0;
2280 }
2281
2282 /* After we've used one of their buffers, we tell them about it. We'll then
2283 * want to notify the guest, using eventfd. */
2284 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2285 unsigned count)
2286 {
2287 int start, n, r;
2288
2289 start = vq->last_used_idx & (vq->num - 1);
2290 n = vq->num - start;
2291 if (n < count) {
2292 r = __vhost_add_used_n(vq, heads, n);
2293 if (r < 0)
2294 return r;
2295 heads += n;
2296 count -= n;
2297 }
2298 r = __vhost_add_used_n(vq, heads, count);
2299
2300 /* Make sure buffer is written before we update index. */
2301 smp_wmb();
2302 if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
2303 &vq->used->idx)) {
2304 vq_err(vq, "Failed to increment used idx");
2305 return -EFAULT;
2306 }
2307 if (unlikely(vq->log_used)) {
2308 /* Make sure used idx is seen before log. */
2309 smp_wmb();
2310 /* Log used index update. */
2311 log_used(vq, offsetof(struct vring_used, idx),
2312 sizeof vq->used->idx);
2313 if (vq->log_ctx)
2314 eventfd_signal(vq->log_ctx, 1);
2315 }
2316 return r;
2317 }
2318 EXPORT_SYMBOL_GPL(vhost_add_used_n);
2319
2320 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2321 {
2322 __u16 old, new;
2323 __virtio16 event;
2324 bool v;
2325 /* Flush out used index updates. This is paired
2326 * with the barrier that the Guest executes when enabling
2327 * interrupts. */
2328 smp_mb();
2329
2330 if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2331 unlikely(vq->avail_idx == vq->last_avail_idx))
2332 return true;
2333
2334 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2335 __virtio16 flags;
2336 if (vhost_get_avail(vq, flags, &vq->avail->flags)) {
2337 vq_err(vq, "Failed to get flags");
2338 return true;
2339 }
2340 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2341 }
2342 old = vq->signalled_used;
2343 v = vq->signalled_used_valid;
2344 new = vq->signalled_used = vq->last_used_idx;
2345 vq->signalled_used_valid = true;
2346
2347 if (unlikely(!v))
2348 return true;
2349
2350 if (vhost_get_avail(vq, event, vhost_used_event(vq))) {
2351 vq_err(vq, "Failed to get used event idx");
2352 return true;
2353 }
2354 return vring_need_event(vhost16_to_cpu(vq, event), new, old);
2355 }
2356
2357 /* This actually signals the guest, using eventfd. */
2358 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2359 {
2360 /* Signal the Guest tell them we used something up. */
2361 if (vq->call_ctx && vhost_notify(dev, vq))
2362 eventfd_signal(vq->call_ctx, 1);
2363 }
2364 EXPORT_SYMBOL_GPL(vhost_signal);
2365
2366 /* And here's the combo meal deal. Supersize me! */
2367 void vhost_add_used_and_signal(struct vhost_dev *dev,
2368 struct vhost_virtqueue *vq,
2369 unsigned int head, int len)
2370 {
2371 vhost_add_used(vq, head, len);
2372 vhost_signal(dev, vq);
2373 }
2374 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2375
2376 /* multi-buffer version of vhost_add_used_and_signal */
2377 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2378 struct vhost_virtqueue *vq,
2379 struct vring_used_elem *heads, unsigned count)
2380 {
2381 vhost_add_used_n(vq, heads, count);
2382 vhost_signal(dev, vq);
2383 }
2384 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2385
2386 /* return true if we're sure that avaiable ring is empty */
2387 bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2388 {
2389 __virtio16 avail_idx;
2390 int r;
2391
2392 if (vq->avail_idx != vq->last_avail_idx)
2393 return false;
2394
2395 r = vhost_get_avail(vq, avail_idx, &vq->avail->idx);
2396 if (unlikely(r))
2397 return false;
2398 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2399
2400 return vq->avail_idx == vq->last_avail_idx;
2401 }
2402 EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2403
2404 /* OK, now we need to know about added descriptors. */
2405 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2406 {
2407 __virtio16 avail_idx;
2408 int r;
2409
2410 if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2411 return false;
2412 vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2413 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2414 r = vhost_update_used_flags(vq);
2415 if (r) {
2416 vq_err(vq, "Failed to enable notification at %p: %d\n",
2417 &vq->used->flags, r);
2418 return false;
2419 }
2420 } else {
2421 r = vhost_update_avail_event(vq, vq->avail_idx);
2422 if (r) {
2423 vq_err(vq, "Failed to update avail event index at %p: %d\n",
2424 vhost_avail_event(vq), r);
2425 return false;
2426 }
2427 }
2428 /* They could have slipped one in as we were doing that: make
2429 * sure it's written, then check again. */
2430 smp_mb();
2431 r = vhost_get_avail(vq, avail_idx, &vq->avail->idx);
2432 if (r) {
2433 vq_err(vq, "Failed to check avail idx at %p: %d\n",
2434 &vq->avail->idx, r);
2435 return false;
2436 }
2437
2438 return vhost16_to_cpu(vq, avail_idx) != vq->avail_idx;
2439 }
2440 EXPORT_SYMBOL_GPL(vhost_enable_notify);
2441
2442 /* We don't need to be notified again. */
2443 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2444 {
2445 int r;
2446
2447 if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2448 return;
2449 vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2450 if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2451 r = vhost_update_used_flags(vq);
2452 if (r)
2453 vq_err(vq, "Failed to enable notification at %p: %d\n",
2454 &vq->used->flags, r);
2455 }
2456 }
2457 EXPORT_SYMBOL_GPL(vhost_disable_notify);
2458
2459 /* Create a new message. */
2460 struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2461 {
2462 struct vhost_msg_node *node = kmalloc(sizeof *node, GFP_KERNEL);
2463 if (!node)
2464 return NULL;
2465
2466 /* Make sure all padding within the structure is initialized. */
2467 memset(&node->msg, 0, sizeof node->msg);
2468 node->vq = vq;
2469 node->msg.type = type;
2470 return node;
2471 }
2472 EXPORT_SYMBOL_GPL(vhost_new_msg);
2473
2474 void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2475 struct vhost_msg_node *node)
2476 {
2477 spin_lock(&dev->iotlb_lock);
2478 list_add_tail(&node->node, head);
2479 spin_unlock(&dev->iotlb_lock);
2480
2481 wake_up_interruptible_poll(&dev->wait, POLLIN | POLLRDNORM);
2482 }
2483 EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2484
2485 struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2486 struct list_head *head)
2487 {
2488 struct vhost_msg_node *node = NULL;
2489
2490 spin_lock(&dev->iotlb_lock);
2491 if (!list_empty(head)) {
2492 node = list_first_entry(head, struct vhost_msg_node,
2493 node);
2494 list_del(&node->node);
2495 }
2496 spin_unlock(&dev->iotlb_lock);
2497
2498 return node;
2499 }
2500 EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2501
2502
2503 static int __init vhost_init(void)
2504 {
2505 return 0;
2506 }
2507
2508 static void __exit vhost_exit(void)
2509 {
2510 }
2511
2512 module_init(vhost_init);
2513 module_exit(vhost_exit);
2514
2515 MODULE_VERSION("0.0.1");
2516 MODULE_LICENSE("GPL v2");
2517 MODULE_AUTHOR("Michael S. Tsirkin");
2518 MODULE_DESCRIPTION("Host kernel accelerator for virtio");
ubuntu-bionic-kernel mirror