xref: /openbmc/linux/drivers/vhost/vhost.c (revision 174cd4b1)
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_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 (!((unsigned long)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 		     unsigned long 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 	unsigned long 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, (void *)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_reset(struct vhost_dev *dev,
288 			   struct vhost_virtqueue *vq)
289 {
290 	vq->num = 1;
291 	vq->desc = NULL;
292 	vq->avail = NULL;
293 	vq->used = NULL;
294 	vq->last_avail_idx = 0;
295 	vq->last_used_event = 0;
296 	vq->avail_idx = 0;
297 	vq->last_used_idx = 0;
298 	vq->signalled_used = 0;
299 	vq->signalled_used_valid = false;
300 	vq->used_flags = 0;
301 	vq->log_used = false;
302 	vq->log_addr = -1ull;
303 	vq->private_data = NULL;
304 	vq->acked_features = 0;
305 	vq->log_base = NULL;
306 	vq->error_ctx = NULL;
307 	vq->error = NULL;
308 	vq->kick = NULL;
309 	vq->call_ctx = NULL;
310 	vq->call = NULL;
311 	vq->log_ctx = NULL;
312 	vhost_reset_is_le(vq);
313 	vhost_disable_cross_endian(vq);
314 	vq->busyloop_timeout = 0;
315 	vq->umem = NULL;
316 	vq->iotlb = NULL;
317 }
318 
319 static int vhost_worker(void *data)
320 {
321 	struct vhost_dev *dev = data;
322 	struct vhost_work *work, *work_next;
323 	struct llist_node *node;
324 	mm_segment_t oldfs = get_fs();
325 
326 	set_fs(USER_DS);
327 	use_mm(dev->mm);
328 
329 	for (;;) {
330 		/* mb paired w/ kthread_stop */
331 		set_current_state(TASK_INTERRUPTIBLE);
332 
333 		if (kthread_should_stop()) {
334 			__set_current_state(TASK_RUNNING);
335 			break;
336 		}
337 
338 		node = llist_del_all(&dev->work_list);
339 		if (!node)
340 			schedule();
341 
342 		node = llist_reverse_order(node);
343 		/* make sure flag is seen after deletion */
344 		smp_wmb();
345 		llist_for_each_entry_safe(work, work_next, node, node) {
346 			clear_bit(VHOST_WORK_QUEUED, &work->flags);
347 			__set_current_state(TASK_RUNNING);
348 			work->fn(work);
349 			if (need_resched())
350 				schedule();
351 		}
352 	}
353 	unuse_mm(dev->mm);
354 	set_fs(oldfs);
355 	return 0;
356 }
357 
358 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
359 {
360 	kfree(vq->indirect);
361 	vq->indirect = NULL;
362 	kfree(vq->log);
363 	vq->log = NULL;
364 	kfree(vq->heads);
365 	vq->heads = NULL;
366 }
367 
368 /* Helper to allocate iovec buffers for all vqs. */
369 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
370 {
371 	struct vhost_virtqueue *vq;
372 	int i;
373 
374 	for (i = 0; i < dev->nvqs; ++i) {
375 		vq = dev->vqs[i];
376 		vq->indirect = kmalloc(sizeof *vq->indirect * UIO_MAXIOV,
377 				       GFP_KERNEL);
378 		vq->log = kmalloc(sizeof *vq->log * UIO_MAXIOV, GFP_KERNEL);
379 		vq->heads = kmalloc(sizeof *vq->heads * UIO_MAXIOV, GFP_KERNEL);
380 		if (!vq->indirect || !vq->log || !vq->heads)
381 			goto err_nomem;
382 	}
383 	return 0;
384 
385 err_nomem:
386 	for (; i >= 0; --i)
387 		vhost_vq_free_iovecs(dev->vqs[i]);
388 	return -ENOMEM;
389 }
390 
391 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
392 {
393 	int i;
394 
395 	for (i = 0; i < dev->nvqs; ++i)
396 		vhost_vq_free_iovecs(dev->vqs[i]);
397 }
398 
399 void vhost_dev_init(struct vhost_dev *dev,
400 		    struct vhost_virtqueue **vqs, int nvqs)
401 {
402 	struct vhost_virtqueue *vq;
403 	int i;
404 
405 	dev->vqs = vqs;
406 	dev->nvqs = nvqs;
407 	mutex_init(&dev->mutex);
408 	dev->log_ctx = NULL;
409 	dev->log_file = NULL;
410 	dev->umem = NULL;
411 	dev->iotlb = NULL;
412 	dev->mm = NULL;
413 	dev->worker = NULL;
414 	init_llist_head(&dev->work_list);
415 	init_waitqueue_head(&dev->wait);
416 	INIT_LIST_HEAD(&dev->read_list);
417 	INIT_LIST_HEAD(&dev->pending_list);
418 	spin_lock_init(&dev->iotlb_lock);
419 
420 
421 	for (i = 0; i < dev->nvqs; ++i) {
422 		vq = dev->vqs[i];
423 		vq->log = NULL;
424 		vq->indirect = NULL;
425 		vq->heads = NULL;
426 		vq->dev = dev;
427 		mutex_init(&vq->mutex);
428 		vhost_vq_reset(dev, vq);
429 		if (vq->handle_kick)
430 			vhost_poll_init(&vq->poll, vq->handle_kick,
431 					POLLIN, dev);
432 	}
433 }
434 EXPORT_SYMBOL_GPL(vhost_dev_init);
435 
436 /* Caller should have device mutex */
437 long vhost_dev_check_owner(struct vhost_dev *dev)
438 {
439 	/* Are you the owner? If not, I don't think you mean to do that */
440 	return dev->mm == current->mm ? 0 : -EPERM;
441 }
442 EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
443 
444 struct vhost_attach_cgroups_struct {
445 	struct vhost_work work;
446 	struct task_struct *owner;
447 	int ret;
448 };
449 
450 static void vhost_attach_cgroups_work(struct vhost_work *work)
451 {
452 	struct vhost_attach_cgroups_struct *s;
453 
454 	s = container_of(work, struct vhost_attach_cgroups_struct, work);
455 	s->ret = cgroup_attach_task_all(s->owner, current);
456 }
457 
458 static int vhost_attach_cgroups(struct vhost_dev *dev)
459 {
460 	struct vhost_attach_cgroups_struct attach;
461 
462 	attach.owner = current;
463 	vhost_work_init(&attach.work, vhost_attach_cgroups_work);
464 	vhost_work_queue(dev, &attach.work);
465 	vhost_work_flush(dev, &attach.work);
466 	return attach.ret;
467 }
468 
469 /* Caller should have device mutex */
470 bool vhost_dev_has_owner(struct vhost_dev *dev)
471 {
472 	return dev->mm;
473 }
474 EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
475 
476 /* Caller should have device mutex */
477 long vhost_dev_set_owner(struct vhost_dev *dev)
478 {
479 	struct task_struct *worker;
480 	int err;
481 
482 	/* Is there an owner already? */
483 	if (vhost_dev_has_owner(dev)) {
484 		err = -EBUSY;
485 		goto err_mm;
486 	}
487 
488 	/* No owner, become one */
489 	dev->mm = get_task_mm(current);
490 	worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
491 	if (IS_ERR(worker)) {
492 		err = PTR_ERR(worker);
493 		goto err_worker;
494 	}
495 
496 	dev->worker = worker;
497 	wake_up_process(worker);	/* avoid contributing to loadavg */
498 
499 	err = vhost_attach_cgroups(dev);
500 	if (err)
501 		goto err_cgroup;
502 
503 	err = vhost_dev_alloc_iovecs(dev);
504 	if (err)
505 		goto err_cgroup;
506 
507 	return 0;
508 err_cgroup:
509 	kthread_stop(worker);
510 	dev->worker = NULL;
511 err_worker:
512 	if (dev->mm)
513 		mmput(dev->mm);
514 	dev->mm = NULL;
515 err_mm:
516 	return err;
517 }
518 EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
519 
520 static void *vhost_kvzalloc(unsigned long size)
521 {
522 	void *n = kzalloc(size, GFP_KERNEL | __GFP_NOWARN | __GFP_REPEAT);
523 
524 	if (!n)
525 		n = vzalloc(size);
526 	return n;
527 }
528 
529 struct vhost_umem *vhost_dev_reset_owner_prepare(void)
530 {
531 	return vhost_kvzalloc(sizeof(struct vhost_umem));
532 }
533 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
534 
535 /* Caller should have device mutex */
536 void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_umem *umem)
537 {
538 	int i;
539 
540 	vhost_dev_cleanup(dev, true);
541 
542 	/* Restore memory to default empty mapping. */
543 	INIT_LIST_HEAD(&umem->umem_list);
544 	dev->umem = umem;
545 	/* We don't need VQ locks below since vhost_dev_cleanup makes sure
546 	 * VQs aren't running.
547 	 */
548 	for (i = 0; i < dev->nvqs; ++i)
549 		dev->vqs[i]->umem = umem;
550 }
551 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
552 
553 void vhost_dev_stop(struct vhost_dev *dev)
554 {
555 	int i;
556 
557 	for (i = 0; i < dev->nvqs; ++i) {
558 		if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick) {
559 			vhost_poll_stop(&dev->vqs[i]->poll);
560 			vhost_poll_flush(&dev->vqs[i]->poll);
561 		}
562 	}
563 }
564 EXPORT_SYMBOL_GPL(vhost_dev_stop);
565 
566 static void vhost_umem_free(struct vhost_umem *umem,
567 			    struct vhost_umem_node *node)
568 {
569 	vhost_umem_interval_tree_remove(node, &umem->umem_tree);
570 	list_del(&node->link);
571 	kfree(node);
572 	umem->numem--;
573 }
574 
575 static void vhost_umem_clean(struct vhost_umem *umem)
576 {
577 	struct vhost_umem_node *node, *tmp;
578 
579 	if (!umem)
580 		return;
581 
582 	list_for_each_entry_safe(node, tmp, &umem->umem_list, link)
583 		vhost_umem_free(umem, node);
584 
585 	kvfree(umem);
586 }
587 
588 static void vhost_clear_msg(struct vhost_dev *dev)
589 {
590 	struct vhost_msg_node *node, *n;
591 
592 	spin_lock(&dev->iotlb_lock);
593 
594 	list_for_each_entry_safe(node, n, &dev->read_list, node) {
595 		list_del(&node->node);
596 		kfree(node);
597 	}
598 
599 	list_for_each_entry_safe(node, n, &dev->pending_list, node) {
600 		list_del(&node->node);
601 		kfree(node);
602 	}
603 
604 	spin_unlock(&dev->iotlb_lock);
605 }
606 
607 /* Caller should have device mutex if and only if locked is set */
608 void vhost_dev_cleanup(struct vhost_dev *dev, bool locked)
609 {
610 	int i;
611 
612 	for (i = 0; i < dev->nvqs; ++i) {
613 		if (dev->vqs[i]->error_ctx)
614 			eventfd_ctx_put(dev->vqs[i]->error_ctx);
615 		if (dev->vqs[i]->error)
616 			fput(dev->vqs[i]->error);
617 		if (dev->vqs[i]->kick)
618 			fput(dev->vqs[i]->kick);
619 		if (dev->vqs[i]->call_ctx)
620 			eventfd_ctx_put(dev->vqs[i]->call_ctx);
621 		if (dev->vqs[i]->call)
622 			fput(dev->vqs[i]->call);
623 		vhost_vq_reset(dev, dev->vqs[i]);
624 	}
625 	vhost_dev_free_iovecs(dev);
626 	if (dev->log_ctx)
627 		eventfd_ctx_put(dev->log_ctx);
628 	dev->log_ctx = NULL;
629 	if (dev->log_file)
630 		fput(dev->log_file);
631 	dev->log_file = NULL;
632 	/* No one will access memory at this point */
633 	vhost_umem_clean(dev->umem);
634 	dev->umem = NULL;
635 	vhost_umem_clean(dev->iotlb);
636 	dev->iotlb = NULL;
637 	vhost_clear_msg(dev);
638 	wake_up_interruptible_poll(&dev->wait, POLLIN | POLLRDNORM);
639 	WARN_ON(!llist_empty(&dev->work_list));
640 	if (dev->worker) {
641 		kthread_stop(dev->worker);
642 		dev->worker = NULL;
643 	}
644 	if (dev->mm)
645 		mmput(dev->mm);
646 	dev->mm = NULL;
647 }
648 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
649 
650 static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
651 {
652 	u64 a = addr / VHOST_PAGE_SIZE / 8;
653 
654 	/* Make sure 64 bit math will not overflow. */
655 	if (a > ULONG_MAX - (unsigned long)log_base ||
656 	    a + (unsigned long)log_base > ULONG_MAX)
657 		return 0;
658 
659 	return access_ok(VERIFY_WRITE, log_base + a,
660 			 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
661 }
662 
663 static bool vhost_overflow(u64 uaddr, u64 size)
664 {
665 	/* Make sure 64 bit math will not overflow. */
666 	return uaddr > ULONG_MAX || size > ULONG_MAX || uaddr > ULONG_MAX - size;
667 }
668 
669 /* Caller should have vq mutex and device mutex. */
670 static int vq_memory_access_ok(void __user *log_base, struct vhost_umem *umem,
671 			       int log_all)
672 {
673 	struct vhost_umem_node *node;
674 
675 	if (!umem)
676 		return 0;
677 
678 	list_for_each_entry(node, &umem->umem_list, link) {
679 		unsigned long a = node->userspace_addr;
680 
681 		if (vhost_overflow(node->userspace_addr, node->size))
682 			return 0;
683 
684 
685 		if (!access_ok(VERIFY_WRITE, (void __user *)a,
686 				    node->size))
687 			return 0;
688 		else if (log_all && !log_access_ok(log_base,
689 						   node->start,
690 						   node->size))
691 			return 0;
692 	}
693 	return 1;
694 }
695 
696 /* Can we switch to this memory table? */
697 /* Caller should have device mutex but not vq mutex */
698 static int memory_access_ok(struct vhost_dev *d, struct vhost_umem *umem,
699 			    int log_all)
700 {
701 	int i;
702 
703 	for (i = 0; i < d->nvqs; ++i) {
704 		int ok;
705 		bool log;
706 
707 		mutex_lock(&d->vqs[i]->mutex);
708 		log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
709 		/* If ring is inactive, will check when it's enabled. */
710 		if (d->vqs[i]->private_data)
711 			ok = vq_memory_access_ok(d->vqs[i]->log_base,
712 						 umem, log);
713 		else
714 			ok = 1;
715 		mutex_unlock(&d->vqs[i]->mutex);
716 		if (!ok)
717 			return 0;
718 	}
719 	return 1;
720 }
721 
722 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
723 			  struct iovec iov[], int iov_size, int access);
724 
725 static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
726 			      const void *from, unsigned size)
727 {
728 	int ret;
729 
730 	if (!vq->iotlb)
731 		return __copy_to_user(to, from, size);
732 	else {
733 		/* This function should be called after iotlb
734 		 * prefetch, which means we're sure that all vq
735 		 * could be access through iotlb. So -EAGAIN should
736 		 * not happen in this case.
737 		 */
738 		/* TODO: more fast path */
739 		struct iov_iter t;
740 		ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
741 				     ARRAY_SIZE(vq->iotlb_iov),
742 				     VHOST_ACCESS_WO);
743 		if (ret < 0)
744 			goto out;
745 		iov_iter_init(&t, WRITE, vq->iotlb_iov, ret, size);
746 		ret = copy_to_iter(from, size, &t);
747 		if (ret == size)
748 			ret = 0;
749 	}
750 out:
751 	return ret;
752 }
753 
754 static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
755 				void __user *from, unsigned size)
756 {
757 	int ret;
758 
759 	if (!vq->iotlb)
760 		return __copy_from_user(to, from, size);
761 	else {
762 		/* This function should be called after iotlb
763 		 * prefetch, which means we're sure that vq
764 		 * could be access through iotlb. So -EAGAIN should
765 		 * not happen in this case.
766 		 */
767 		/* TODO: more fast path */
768 		struct iov_iter f;
769 		ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
770 				     ARRAY_SIZE(vq->iotlb_iov),
771 				     VHOST_ACCESS_RO);
772 		if (ret < 0) {
773 			vq_err(vq, "IOTLB translation failure: uaddr "
774 			       "%p size 0x%llx\n", from,
775 			       (unsigned long long) size);
776 			goto out;
777 		}
778 		iov_iter_init(&f, READ, vq->iotlb_iov, ret, size);
779 		ret = copy_from_iter(to, size, &f);
780 		if (ret == size)
781 			ret = 0;
782 	}
783 
784 out:
785 	return ret;
786 }
787 
788 static void __user *__vhost_get_user(struct vhost_virtqueue *vq,
789 				     void __user *addr, unsigned size)
790 {
791 	int ret;
792 
793 	/* This function should be called after iotlb
794 	 * prefetch, which means we're sure that vq
795 	 * could be access through iotlb. So -EAGAIN should
796 	 * not happen in this case.
797 	 */
798 	/* TODO: more fast path */
799 	ret = translate_desc(vq, (u64)(uintptr_t)addr, size, vq->iotlb_iov,
800 			     ARRAY_SIZE(vq->iotlb_iov),
801 			     VHOST_ACCESS_RO);
802 	if (ret < 0) {
803 		vq_err(vq, "IOTLB translation failure: uaddr "
804 			"%p size 0x%llx\n", addr,
805 			(unsigned long long) size);
806 		return NULL;
807 	}
808 
809 	if (ret != 1 || vq->iotlb_iov[0].iov_len != size) {
810 		vq_err(vq, "Non atomic userspace memory access: uaddr "
811 			"%p size 0x%llx\n", addr,
812 			(unsigned long long) size);
813 		return NULL;
814 	}
815 
816 	return vq->iotlb_iov[0].iov_base;
817 }
818 
819 #define vhost_put_user(vq, x, ptr) \
820 ({ \
821 	int ret = -EFAULT; \
822 	if (!vq->iotlb) { \
823 		ret = __put_user(x, ptr); \
824 	} else { \
825 		__typeof__(ptr) to = \
826 			(__typeof__(ptr)) __vhost_get_user(vq, ptr, sizeof(*ptr)); \
827 		if (to != NULL) \
828 			ret = __put_user(x, to); \
829 		else \
830 			ret = -EFAULT;	\
831 	} \
832 	ret; \
833 })
834 
835 #define vhost_get_user(vq, x, ptr) \
836 ({ \
837 	int ret; \
838 	if (!vq->iotlb) { \
839 		ret = __get_user(x, ptr); \
840 	} else { \
841 		__typeof__(ptr) from = \
842 			(__typeof__(ptr)) __vhost_get_user(vq, ptr, sizeof(*ptr)); \
843 		if (from != NULL) \
844 			ret = __get_user(x, from); \
845 		else \
846 			ret = -EFAULT; \
847 	} \
848 	ret; \
849 })
850 
851 static void vhost_dev_lock_vqs(struct vhost_dev *d)
852 {
853 	int i = 0;
854 	for (i = 0; i < d->nvqs; ++i)
855 		mutex_lock(&d->vqs[i]->mutex);
856 }
857 
858 static void vhost_dev_unlock_vqs(struct vhost_dev *d)
859 {
860 	int i = 0;
861 	for (i = 0; i < d->nvqs; ++i)
862 		mutex_unlock(&d->vqs[i]->mutex);
863 }
864 
865 static int vhost_new_umem_range(struct vhost_umem *umem,
866 				u64 start, u64 size, u64 end,
867 				u64 userspace_addr, int perm)
868 {
869 	struct vhost_umem_node *tmp, *node = kmalloc(sizeof(*node), GFP_ATOMIC);
870 
871 	if (!node)
872 		return -ENOMEM;
873 
874 	if (umem->numem == max_iotlb_entries) {
875 		tmp = list_first_entry(&umem->umem_list, typeof(*tmp), link);
876 		vhost_umem_free(umem, tmp);
877 	}
878 
879 	node->start = start;
880 	node->size = size;
881 	node->last = end;
882 	node->userspace_addr = userspace_addr;
883 	node->perm = perm;
884 	INIT_LIST_HEAD(&node->link);
885 	list_add_tail(&node->link, &umem->umem_list);
886 	vhost_umem_interval_tree_insert(node, &umem->umem_tree);
887 	umem->numem++;
888 
889 	return 0;
890 }
891 
892 static void vhost_del_umem_range(struct vhost_umem *umem,
893 				 u64 start, u64 end)
894 {
895 	struct vhost_umem_node *node;
896 
897 	while ((node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
898 							   start, end)))
899 		vhost_umem_free(umem, node);
900 }
901 
902 static void vhost_iotlb_notify_vq(struct vhost_dev *d,
903 				  struct vhost_iotlb_msg *msg)
904 {
905 	struct vhost_msg_node *node, *n;
906 
907 	spin_lock(&d->iotlb_lock);
908 
909 	list_for_each_entry_safe(node, n, &d->pending_list, node) {
910 		struct vhost_iotlb_msg *vq_msg = &node->msg.iotlb;
911 		if (msg->iova <= vq_msg->iova &&
912 		    msg->iova + msg->size - 1 > vq_msg->iova &&
913 		    vq_msg->type == VHOST_IOTLB_MISS) {
914 			vhost_poll_queue(&node->vq->poll);
915 			list_del(&node->node);
916 			kfree(node);
917 		}
918 	}
919 
920 	spin_unlock(&d->iotlb_lock);
921 }
922 
923 static int umem_access_ok(u64 uaddr, u64 size, int access)
924 {
925 	unsigned long a = uaddr;
926 
927 	/* Make sure 64 bit math will not overflow. */
928 	if (vhost_overflow(uaddr, size))
929 		return -EFAULT;
930 
931 	if ((access & VHOST_ACCESS_RO) &&
932 	    !access_ok(VERIFY_READ, (void __user *)a, size))
933 		return -EFAULT;
934 	if ((access & VHOST_ACCESS_WO) &&
935 	    !access_ok(VERIFY_WRITE, (void __user *)a, size))
936 		return -EFAULT;
937 	return 0;
938 }
939 
940 static int vhost_process_iotlb_msg(struct vhost_dev *dev,
941 				   struct vhost_iotlb_msg *msg)
942 {
943 	int ret = 0;
944 
945 	vhost_dev_lock_vqs(dev);
946 	switch (msg->type) {
947 	case VHOST_IOTLB_UPDATE:
948 		if (!dev->iotlb) {
949 			ret = -EFAULT;
950 			break;
951 		}
952 		if (umem_access_ok(msg->uaddr, msg->size, msg->perm)) {
953 			ret = -EFAULT;
954 			break;
955 		}
956 		if (vhost_new_umem_range(dev->iotlb, msg->iova, msg->size,
957 					 msg->iova + msg->size - 1,
958 					 msg->uaddr, msg->perm)) {
959 			ret = -ENOMEM;
960 			break;
961 		}
962 		vhost_iotlb_notify_vq(dev, msg);
963 		break;
964 	case VHOST_IOTLB_INVALIDATE:
965 		vhost_del_umem_range(dev->iotlb, msg->iova,
966 				     msg->iova + msg->size - 1);
967 		break;
968 	default:
969 		ret = -EINVAL;
970 		break;
971 	}
972 
973 	vhost_dev_unlock_vqs(dev);
974 	return ret;
975 }
976 ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
977 			     struct iov_iter *from)
978 {
979 	struct vhost_msg_node node;
980 	unsigned size = sizeof(struct vhost_msg);
981 	size_t ret;
982 	int err;
983 
984 	if (iov_iter_count(from) < size)
985 		return 0;
986 	ret = copy_from_iter(&node.msg, size, from);
987 	if (ret != size)
988 		goto done;
989 
990 	switch (node.msg.type) {
991 	case VHOST_IOTLB_MSG:
992 		err = vhost_process_iotlb_msg(dev, &node.msg.iotlb);
993 		if (err)
994 			ret = err;
995 		break;
996 	default:
997 		ret = -EINVAL;
998 		break;
999 	}
1000 
1001 done:
1002 	return ret;
1003 }
1004 EXPORT_SYMBOL(vhost_chr_write_iter);
1005 
1006 unsigned int vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1007 			    poll_table *wait)
1008 {
1009 	unsigned int mask = 0;
1010 
1011 	poll_wait(file, &dev->wait, wait);
1012 
1013 	if (!list_empty(&dev->read_list))
1014 		mask |= POLLIN | POLLRDNORM;
1015 
1016 	return mask;
1017 }
1018 EXPORT_SYMBOL(vhost_chr_poll);
1019 
1020 ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1021 			    int noblock)
1022 {
1023 	DEFINE_WAIT(wait);
1024 	struct vhost_msg_node *node;
1025 	ssize_t ret = 0;
1026 	unsigned size = sizeof(struct vhost_msg);
1027 
1028 	if (iov_iter_count(to) < size)
1029 		return 0;
1030 
1031 	while (1) {
1032 		if (!noblock)
1033 			prepare_to_wait(&dev->wait, &wait,
1034 					TASK_INTERRUPTIBLE);
1035 
1036 		node = vhost_dequeue_msg(dev, &dev->read_list);
1037 		if (node)
1038 			break;
1039 		if (noblock) {
1040 			ret = -EAGAIN;
1041 			break;
1042 		}
1043 		if (signal_pending(current)) {
1044 			ret = -ERESTARTSYS;
1045 			break;
1046 		}
1047 		if (!dev->iotlb) {
1048 			ret = -EBADFD;
1049 			break;
1050 		}
1051 
1052 		schedule();
1053 	}
1054 
1055 	if (!noblock)
1056 		finish_wait(&dev->wait, &wait);
1057 
1058 	if (node) {
1059 		ret = copy_to_iter(&node->msg, size, to);
1060 
1061 		if (ret != size || node->msg.type != VHOST_IOTLB_MISS) {
1062 			kfree(node);
1063 			return ret;
1064 		}
1065 
1066 		vhost_enqueue_msg(dev, &dev->pending_list, node);
1067 	}
1068 
1069 	return ret;
1070 }
1071 EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1072 
1073 static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1074 {
1075 	struct vhost_dev *dev = vq->dev;
1076 	struct vhost_msg_node *node;
1077 	struct vhost_iotlb_msg *msg;
1078 
1079 	node = vhost_new_msg(vq, VHOST_IOTLB_MISS);
1080 	if (!node)
1081 		return -ENOMEM;
1082 
1083 	msg = &node->msg.iotlb;
1084 	msg->type = VHOST_IOTLB_MISS;
1085 	msg->iova = iova;
1086 	msg->perm = access;
1087 
1088 	vhost_enqueue_msg(dev, &dev->read_list, node);
1089 
1090 	return 0;
1091 }
1092 
1093 static int vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1094 			struct vring_desc __user *desc,
1095 			struct vring_avail __user *avail,
1096 			struct vring_used __user *used)
1097 
1098 {
1099 	size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1100 
1101 	return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
1102 	       access_ok(VERIFY_READ, avail,
1103 			 sizeof *avail + num * sizeof *avail->ring + s) &&
1104 	       access_ok(VERIFY_WRITE, used,
1105 			sizeof *used + num * sizeof *used->ring + s);
1106 }
1107 
1108 static int iotlb_access_ok(struct vhost_virtqueue *vq,
1109 			   int access, u64 addr, u64 len)
1110 {
1111 	const struct vhost_umem_node *node;
1112 	struct vhost_umem *umem = vq->iotlb;
1113 	u64 s = 0, size;
1114 
1115 	while (len > s) {
1116 		node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1117 							   addr,
1118 							   addr + len - 1);
1119 		if (node == NULL || node->start > addr) {
1120 			vhost_iotlb_miss(vq, addr, access);
1121 			return false;
1122 		} else if (!(node->perm & access)) {
1123 			/* Report the possible access violation by
1124 			 * request another translation from userspace.
1125 			 */
1126 			return false;
1127 		}
1128 
1129 		size = node->size - addr + node->start;
1130 		s += size;
1131 		addr += size;
1132 	}
1133 
1134 	return true;
1135 }
1136 
1137 int vq_iotlb_prefetch(struct vhost_virtqueue *vq)
1138 {
1139 	size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1140 	unsigned int num = vq->num;
1141 
1142 	if (!vq->iotlb)
1143 		return 1;
1144 
1145 	return iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->desc,
1146 			       num * sizeof *vq->desc) &&
1147 	       iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->avail,
1148 			       sizeof *vq->avail +
1149 			       num * sizeof *vq->avail->ring + s) &&
1150 	       iotlb_access_ok(vq, VHOST_ACCESS_WO, (u64)(uintptr_t)vq->used,
1151 			       sizeof *vq->used +
1152 			       num * sizeof *vq->used->ring + s);
1153 }
1154 EXPORT_SYMBOL_GPL(vq_iotlb_prefetch);
1155 
1156 /* Can we log writes? */
1157 /* Caller should have device mutex but not vq mutex */
1158 int vhost_log_access_ok(struct vhost_dev *dev)
1159 {
1160 	return memory_access_ok(dev, dev->umem, 1);
1161 }
1162 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1163 
1164 /* Verify access for write logging. */
1165 /* Caller should have vq mutex and device mutex */
1166 static int vq_log_access_ok(struct vhost_virtqueue *vq,
1167 			    void __user *log_base)
1168 {
1169 	size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1170 
1171 	return vq_memory_access_ok(log_base, vq->umem,
1172 				   vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1173 		(!vq->log_used || log_access_ok(log_base, vq->log_addr,
1174 					sizeof *vq->used +
1175 					vq->num * sizeof *vq->used->ring + s));
1176 }
1177 
1178 /* Can we start vq? */
1179 /* Caller should have vq mutex and device mutex */
1180 int vhost_vq_access_ok(struct vhost_virtqueue *vq)
1181 {
1182 	if (vq->iotlb) {
1183 		/* When device IOTLB was used, the access validation
1184 		 * will be validated during prefetching.
1185 		 */
1186 		return 1;
1187 	}
1188 	return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used) &&
1189 		vq_log_access_ok(vq, vq->log_base);
1190 }
1191 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1192 
1193 static struct vhost_umem *vhost_umem_alloc(void)
1194 {
1195 	struct vhost_umem *umem = vhost_kvzalloc(sizeof(*umem));
1196 
1197 	if (!umem)
1198 		return NULL;
1199 
1200 	umem->umem_tree = RB_ROOT;
1201 	umem->numem = 0;
1202 	INIT_LIST_HEAD(&umem->umem_list);
1203 
1204 	return umem;
1205 }
1206 
1207 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1208 {
1209 	struct vhost_memory mem, *newmem;
1210 	struct vhost_memory_region *region;
1211 	struct vhost_umem *newumem, *oldumem;
1212 	unsigned long size = offsetof(struct vhost_memory, regions);
1213 	int i;
1214 
1215 	if (copy_from_user(&mem, m, size))
1216 		return -EFAULT;
1217 	if (mem.padding)
1218 		return -EOPNOTSUPP;
1219 	if (mem.nregions > max_mem_regions)
1220 		return -E2BIG;
1221 	newmem = vhost_kvzalloc(size + mem.nregions * sizeof(*m->regions));
1222 	if (!newmem)
1223 		return -ENOMEM;
1224 
1225 	memcpy(newmem, &mem, size);
1226 	if (copy_from_user(newmem->regions, m->regions,
1227 			   mem.nregions * sizeof *m->regions)) {
1228 		kvfree(newmem);
1229 		return -EFAULT;
1230 	}
1231 
1232 	newumem = vhost_umem_alloc();
1233 	if (!newumem) {
1234 		kvfree(newmem);
1235 		return -ENOMEM;
1236 	}
1237 
1238 	for (region = newmem->regions;
1239 	     region < newmem->regions + mem.nregions;
1240 	     region++) {
1241 		if (vhost_new_umem_range(newumem,
1242 					 region->guest_phys_addr,
1243 					 region->memory_size,
1244 					 region->guest_phys_addr +
1245 					 region->memory_size - 1,
1246 					 region->userspace_addr,
1247 					 VHOST_ACCESS_RW))
1248 			goto err;
1249 	}
1250 
1251 	if (!memory_access_ok(d, newumem, 0))
1252 		goto err;
1253 
1254 	oldumem = d->umem;
1255 	d->umem = newumem;
1256 
1257 	/* All memory accesses are done under some VQ mutex. */
1258 	for (i = 0; i < d->nvqs; ++i) {
1259 		mutex_lock(&d->vqs[i]->mutex);
1260 		d->vqs[i]->umem = newumem;
1261 		mutex_unlock(&d->vqs[i]->mutex);
1262 	}
1263 
1264 	kvfree(newmem);
1265 	vhost_umem_clean(oldumem);
1266 	return 0;
1267 
1268 err:
1269 	vhost_umem_clean(newumem);
1270 	kvfree(newmem);
1271 	return -EFAULT;
1272 }
1273 
1274 long vhost_vring_ioctl(struct vhost_dev *d, int ioctl, void __user *argp)
1275 {
1276 	struct file *eventfp, *filep = NULL;
1277 	bool pollstart = false, pollstop = false;
1278 	struct eventfd_ctx *ctx = NULL;
1279 	u32 __user *idxp = argp;
1280 	struct vhost_virtqueue *vq;
1281 	struct vhost_vring_state s;
1282 	struct vhost_vring_file f;
1283 	struct vhost_vring_addr a;
1284 	u32 idx;
1285 	long r;
1286 
1287 	r = get_user(idx, idxp);
1288 	if (r < 0)
1289 		return r;
1290 	if (idx >= d->nvqs)
1291 		return -ENOBUFS;
1292 
1293 	vq = d->vqs[idx];
1294 
1295 	mutex_lock(&vq->mutex);
1296 
1297 	switch (ioctl) {
1298 	case VHOST_SET_VRING_NUM:
1299 		/* Resizing ring with an active backend?
1300 		 * You don't want to do that. */
1301 		if (vq->private_data) {
1302 			r = -EBUSY;
1303 			break;
1304 		}
1305 		if (copy_from_user(&s, argp, sizeof s)) {
1306 			r = -EFAULT;
1307 			break;
1308 		}
1309 		if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
1310 			r = -EINVAL;
1311 			break;
1312 		}
1313 		vq->num = s.num;
1314 		break;
1315 	case VHOST_SET_VRING_BASE:
1316 		/* Moving base with an active backend?
1317 		 * You don't want to do that. */
1318 		if (vq->private_data) {
1319 			r = -EBUSY;
1320 			break;
1321 		}
1322 		if (copy_from_user(&s, argp, sizeof s)) {
1323 			r = -EFAULT;
1324 			break;
1325 		}
1326 		if (s.num > 0xffff) {
1327 			r = -EINVAL;
1328 			break;
1329 		}
1330 		vq->last_avail_idx = vq->last_used_event = s.num;
1331 		/* Forget the cached index value. */
1332 		vq->avail_idx = vq->last_avail_idx;
1333 		break;
1334 	case VHOST_GET_VRING_BASE:
1335 		s.index = idx;
1336 		s.num = vq->last_avail_idx;
1337 		if (copy_to_user(argp, &s, sizeof s))
1338 			r = -EFAULT;
1339 		break;
1340 	case VHOST_SET_VRING_ADDR:
1341 		if (copy_from_user(&a, argp, sizeof a)) {
1342 			r = -EFAULT;
1343 			break;
1344 		}
1345 		if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
1346 			r = -EOPNOTSUPP;
1347 			break;
1348 		}
1349 		/* For 32bit, verify that the top 32bits of the user
1350 		   data are set to zero. */
1351 		if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1352 		    (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1353 		    (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
1354 			r = -EFAULT;
1355 			break;
1356 		}
1357 
1358 		/* Make sure it's safe to cast pointers to vring types. */
1359 		BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1360 		BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1361 		if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1362 		    (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1363 		    (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1))) {
1364 			r = -EINVAL;
1365 			break;
1366 		}
1367 
1368 		/* We only verify access here if backend is configured.
1369 		 * If it is not, we don't as size might not have been setup.
1370 		 * We will verify when backend is configured. */
1371 		if (vq->private_data) {
1372 			if (!vq_access_ok(vq, vq->num,
1373 				(void __user *)(unsigned long)a.desc_user_addr,
1374 				(void __user *)(unsigned long)a.avail_user_addr,
1375 				(void __user *)(unsigned long)a.used_user_addr)) {
1376 				r = -EINVAL;
1377 				break;
1378 			}
1379 
1380 			/* Also validate log access for used ring if enabled. */
1381 			if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
1382 			    !log_access_ok(vq->log_base, a.log_guest_addr,
1383 					   sizeof *vq->used +
1384 					   vq->num * sizeof *vq->used->ring)) {
1385 				r = -EINVAL;
1386 				break;
1387 			}
1388 		}
1389 
1390 		vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1391 		vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1392 		vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1393 		vq->log_addr = a.log_guest_addr;
1394 		vq->used = (void __user *)(unsigned long)a.used_user_addr;
1395 		break;
1396 	case VHOST_SET_VRING_KICK:
1397 		if (copy_from_user(&f, argp, sizeof f)) {
1398 			r = -EFAULT;
1399 			break;
1400 		}
1401 		eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1402 		if (IS_ERR(eventfp)) {
1403 			r = PTR_ERR(eventfp);
1404 			break;
1405 		}
1406 		if (eventfp != vq->kick) {
1407 			pollstop = (filep = vq->kick) != NULL;
1408 			pollstart = (vq->kick = eventfp) != NULL;
1409 		} else
1410 			filep = eventfp;
1411 		break;
1412 	case VHOST_SET_VRING_CALL:
1413 		if (copy_from_user(&f, argp, sizeof f)) {
1414 			r = -EFAULT;
1415 			break;
1416 		}
1417 		eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1418 		if (IS_ERR(eventfp)) {
1419 			r = PTR_ERR(eventfp);
1420 			break;
1421 		}
1422 		if (eventfp != vq->call) {
1423 			filep = vq->call;
1424 			ctx = vq->call_ctx;
1425 			vq->call = eventfp;
1426 			vq->call_ctx = eventfp ?
1427 				eventfd_ctx_fileget(eventfp) : NULL;
1428 		} else
1429 			filep = eventfp;
1430 		break;
1431 	case VHOST_SET_VRING_ERR:
1432 		if (copy_from_user(&f, argp, sizeof f)) {
1433 			r = -EFAULT;
1434 			break;
1435 		}
1436 		eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1437 		if (IS_ERR(eventfp)) {
1438 			r = PTR_ERR(eventfp);
1439 			break;
1440 		}
1441 		if (eventfp != vq->error) {
1442 			filep = vq->error;
1443 			vq->error = eventfp;
1444 			ctx = vq->error_ctx;
1445 			vq->error_ctx = eventfp ?
1446 				eventfd_ctx_fileget(eventfp) : NULL;
1447 		} else
1448 			filep = eventfp;
1449 		break;
1450 	case VHOST_SET_VRING_ENDIAN:
1451 		r = vhost_set_vring_endian(vq, argp);
1452 		break;
1453 	case VHOST_GET_VRING_ENDIAN:
1454 		r = vhost_get_vring_endian(vq, idx, argp);
1455 		break;
1456 	case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
1457 		if (copy_from_user(&s, argp, sizeof(s))) {
1458 			r = -EFAULT;
1459 			break;
1460 		}
1461 		vq->busyloop_timeout = s.num;
1462 		break;
1463 	case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
1464 		s.index = idx;
1465 		s.num = vq->busyloop_timeout;
1466 		if (copy_to_user(argp, &s, sizeof(s)))
1467 			r = -EFAULT;
1468 		break;
1469 	default:
1470 		r = -ENOIOCTLCMD;
1471 	}
1472 
1473 	if (pollstop && vq->handle_kick)
1474 		vhost_poll_stop(&vq->poll);
1475 
1476 	if (ctx)
1477 		eventfd_ctx_put(ctx);
1478 	if (filep)
1479 		fput(filep);
1480 
1481 	if (pollstart && vq->handle_kick)
1482 		r = vhost_poll_start(&vq->poll, vq->kick);
1483 
1484 	mutex_unlock(&vq->mutex);
1485 
1486 	if (pollstop && vq->handle_kick)
1487 		vhost_poll_flush(&vq->poll);
1488 	return r;
1489 }
1490 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
1491 
1492 int vhost_init_device_iotlb(struct vhost_dev *d, bool enabled)
1493 {
1494 	struct vhost_umem *niotlb, *oiotlb;
1495 	int i;
1496 
1497 	niotlb = vhost_umem_alloc();
1498 	if (!niotlb)
1499 		return -ENOMEM;
1500 
1501 	oiotlb = d->iotlb;
1502 	d->iotlb = niotlb;
1503 
1504 	for (i = 0; i < d->nvqs; ++i) {
1505 		mutex_lock(&d->vqs[i]->mutex);
1506 		d->vqs[i]->iotlb = niotlb;
1507 		mutex_unlock(&d->vqs[i]->mutex);
1508 	}
1509 
1510 	vhost_umem_clean(oiotlb);
1511 
1512 	return 0;
1513 }
1514 EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
1515 
1516 /* Caller must have device mutex */
1517 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1518 {
1519 	struct file *eventfp, *filep = NULL;
1520 	struct eventfd_ctx *ctx = NULL;
1521 	u64 p;
1522 	long r;
1523 	int i, fd;
1524 
1525 	/* If you are not the owner, you can become one */
1526 	if (ioctl == VHOST_SET_OWNER) {
1527 		r = vhost_dev_set_owner(d);
1528 		goto done;
1529 	}
1530 
1531 	/* You must be the owner to do anything else */
1532 	r = vhost_dev_check_owner(d);
1533 	if (r)
1534 		goto done;
1535 
1536 	switch (ioctl) {
1537 	case VHOST_SET_MEM_TABLE:
1538 		r = vhost_set_memory(d, argp);
1539 		break;
1540 	case VHOST_SET_LOG_BASE:
1541 		if (copy_from_user(&p, argp, sizeof p)) {
1542 			r = -EFAULT;
1543 			break;
1544 		}
1545 		if ((u64)(unsigned long)p != p) {
1546 			r = -EFAULT;
1547 			break;
1548 		}
1549 		for (i = 0; i < d->nvqs; ++i) {
1550 			struct vhost_virtqueue *vq;
1551 			void __user *base = (void __user *)(unsigned long)p;
1552 			vq = d->vqs[i];
1553 			mutex_lock(&vq->mutex);
1554 			/* If ring is inactive, will check when it's enabled. */
1555 			if (vq->private_data && !vq_log_access_ok(vq, base))
1556 				r = -EFAULT;
1557 			else
1558 				vq->log_base = base;
1559 			mutex_unlock(&vq->mutex);
1560 		}
1561 		break;
1562 	case VHOST_SET_LOG_FD:
1563 		r = get_user(fd, (int __user *)argp);
1564 		if (r < 0)
1565 			break;
1566 		eventfp = fd == -1 ? NULL : eventfd_fget(fd);
1567 		if (IS_ERR(eventfp)) {
1568 			r = PTR_ERR(eventfp);
1569 			break;
1570 		}
1571 		if (eventfp != d->log_file) {
1572 			filep = d->log_file;
1573 			d->log_file = eventfp;
1574 			ctx = d->log_ctx;
1575 			d->log_ctx = eventfp ?
1576 				eventfd_ctx_fileget(eventfp) : NULL;
1577 		} else
1578 			filep = eventfp;
1579 		for (i = 0; i < d->nvqs; ++i) {
1580 			mutex_lock(&d->vqs[i]->mutex);
1581 			d->vqs[i]->log_ctx = d->log_ctx;
1582 			mutex_unlock(&d->vqs[i]->mutex);
1583 		}
1584 		if (ctx)
1585 			eventfd_ctx_put(ctx);
1586 		if (filep)
1587 			fput(filep);
1588 		break;
1589 	default:
1590 		r = -ENOIOCTLCMD;
1591 		break;
1592 	}
1593 done:
1594 	return r;
1595 }
1596 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1597 
1598 /* TODO: This is really inefficient.  We need something like get_user()
1599  * (instruction directly accesses the data, with an exception table entry
1600  * returning -EFAULT). See Documentation/x86/exception-tables.txt.
1601  */
1602 static int set_bit_to_user(int nr, void __user *addr)
1603 {
1604 	unsigned long log = (unsigned long)addr;
1605 	struct page *page;
1606 	void *base;
1607 	int bit = nr + (log % PAGE_SIZE) * 8;
1608 	int r;
1609 
1610 	r = get_user_pages_fast(log, 1, 1, &page);
1611 	if (r < 0)
1612 		return r;
1613 	BUG_ON(r != 1);
1614 	base = kmap_atomic(page);
1615 	set_bit(bit, base);
1616 	kunmap_atomic(base);
1617 	set_page_dirty_lock(page);
1618 	put_page(page);
1619 	return 0;
1620 }
1621 
1622 static int log_write(void __user *log_base,
1623 		     u64 write_address, u64 write_length)
1624 {
1625 	u64 write_page = write_address / VHOST_PAGE_SIZE;
1626 	int r;
1627 
1628 	if (!write_length)
1629 		return 0;
1630 	write_length += write_address % VHOST_PAGE_SIZE;
1631 	for (;;) {
1632 		u64 base = (u64)(unsigned long)log_base;
1633 		u64 log = base + write_page / 8;
1634 		int bit = write_page % 8;
1635 		if ((u64)(unsigned long)log != log)
1636 			return -EFAULT;
1637 		r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1638 		if (r < 0)
1639 			return r;
1640 		if (write_length <= VHOST_PAGE_SIZE)
1641 			break;
1642 		write_length -= VHOST_PAGE_SIZE;
1643 		write_page += 1;
1644 	}
1645 	return r;
1646 }
1647 
1648 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1649 		    unsigned int log_num, u64 len)
1650 {
1651 	int i, r;
1652 
1653 	/* Make sure data written is seen before log. */
1654 	smp_wmb();
1655 	for (i = 0; i < log_num; ++i) {
1656 		u64 l = min(log[i].len, len);
1657 		r = log_write(vq->log_base, log[i].addr, l);
1658 		if (r < 0)
1659 			return r;
1660 		len -= l;
1661 		if (!len) {
1662 			if (vq->log_ctx)
1663 				eventfd_signal(vq->log_ctx, 1);
1664 			return 0;
1665 		}
1666 	}
1667 	/* Length written exceeds what we have stored. This is a bug. */
1668 	BUG();
1669 	return 0;
1670 }
1671 EXPORT_SYMBOL_GPL(vhost_log_write);
1672 
1673 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1674 {
1675 	void __user *used;
1676 	if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
1677 			   &vq->used->flags) < 0)
1678 		return -EFAULT;
1679 	if (unlikely(vq->log_used)) {
1680 		/* Make sure the flag is seen before log. */
1681 		smp_wmb();
1682 		/* Log used flag write. */
1683 		used = &vq->used->flags;
1684 		log_write(vq->log_base, vq->log_addr +
1685 			  (used - (void __user *)vq->used),
1686 			  sizeof vq->used->flags);
1687 		if (vq->log_ctx)
1688 			eventfd_signal(vq->log_ctx, 1);
1689 	}
1690 	return 0;
1691 }
1692 
1693 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1694 {
1695 	if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
1696 			   vhost_avail_event(vq)))
1697 		return -EFAULT;
1698 	if (unlikely(vq->log_used)) {
1699 		void __user *used;
1700 		/* Make sure the event is seen before log. */
1701 		smp_wmb();
1702 		/* Log avail event write */
1703 		used = vhost_avail_event(vq);
1704 		log_write(vq->log_base, vq->log_addr +
1705 			  (used - (void __user *)vq->used),
1706 			  sizeof *vhost_avail_event(vq));
1707 		if (vq->log_ctx)
1708 			eventfd_signal(vq->log_ctx, 1);
1709 	}
1710 	return 0;
1711 }
1712 
1713 int vhost_vq_init_access(struct vhost_virtqueue *vq)
1714 {
1715 	__virtio16 last_used_idx;
1716 	int r;
1717 	bool is_le = vq->is_le;
1718 
1719 	if (!vq->private_data)
1720 		return 0;
1721 
1722 	vhost_init_is_le(vq);
1723 
1724 	r = vhost_update_used_flags(vq);
1725 	if (r)
1726 		goto err;
1727 	vq->signalled_used_valid = false;
1728 	if (!vq->iotlb &&
1729 	    !access_ok(VERIFY_READ, &vq->used->idx, sizeof vq->used->idx)) {
1730 		r = -EFAULT;
1731 		goto err;
1732 	}
1733 	r = vhost_get_user(vq, last_used_idx, &vq->used->idx);
1734 	if (r) {
1735 		vq_err(vq, "Can't access used idx at %p\n",
1736 		       &vq->used->idx);
1737 		goto err;
1738 	}
1739 	vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
1740 	return 0;
1741 
1742 err:
1743 	vq->is_le = is_le;
1744 	return r;
1745 }
1746 EXPORT_SYMBOL_GPL(vhost_vq_init_access);
1747 
1748 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
1749 			  struct iovec iov[], int iov_size, int access)
1750 {
1751 	const struct vhost_umem_node *node;
1752 	struct vhost_dev *dev = vq->dev;
1753 	struct vhost_umem *umem = dev->iotlb ? dev->iotlb : dev->umem;
1754 	struct iovec *_iov;
1755 	u64 s = 0;
1756 	int ret = 0;
1757 
1758 	while ((u64)len > s) {
1759 		u64 size;
1760 		if (unlikely(ret >= iov_size)) {
1761 			ret = -ENOBUFS;
1762 			break;
1763 		}
1764 
1765 		node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1766 							addr, addr + len - 1);
1767 		if (node == NULL || node->start > addr) {
1768 			if (umem != dev->iotlb) {
1769 				ret = -EFAULT;
1770 				break;
1771 			}
1772 			ret = -EAGAIN;
1773 			break;
1774 		} else if (!(node->perm & access)) {
1775 			ret = -EPERM;
1776 			break;
1777 		}
1778 
1779 		_iov = iov + ret;
1780 		size = node->size - addr + node->start;
1781 		_iov->iov_len = min((u64)len - s, size);
1782 		_iov->iov_base = (void __user *)(unsigned long)
1783 			(node->userspace_addr + addr - node->start);
1784 		s += size;
1785 		addr += size;
1786 		++ret;
1787 	}
1788 
1789 	if (ret == -EAGAIN)
1790 		vhost_iotlb_miss(vq, addr, access);
1791 	return ret;
1792 }
1793 
1794 /* Each buffer in the virtqueues is actually a chain of descriptors.  This
1795  * function returns the next descriptor in the chain,
1796  * or -1U if we're at the end. */
1797 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
1798 {
1799 	unsigned int next;
1800 
1801 	/* If this descriptor says it doesn't chain, we're done. */
1802 	if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
1803 		return -1U;
1804 
1805 	/* Check they're not leading us off end of descriptors. */
1806 	next = vhost16_to_cpu(vq, desc->next);
1807 	/* Make sure compiler knows to grab that: we don't want it changing! */
1808 	/* We will use the result as an index in an array, so most
1809 	 * architectures only need a compiler barrier here. */
1810 	read_barrier_depends();
1811 
1812 	return next;
1813 }
1814 
1815 static int get_indirect(struct vhost_virtqueue *vq,
1816 			struct iovec iov[], unsigned int iov_size,
1817 			unsigned int *out_num, unsigned int *in_num,
1818 			struct vhost_log *log, unsigned int *log_num,
1819 			struct vring_desc *indirect)
1820 {
1821 	struct vring_desc desc;
1822 	unsigned int i = 0, count, found = 0;
1823 	u32 len = vhost32_to_cpu(vq, indirect->len);
1824 	struct iov_iter from;
1825 	int ret, access;
1826 
1827 	/* Sanity check */
1828 	if (unlikely(len % sizeof desc)) {
1829 		vq_err(vq, "Invalid length in indirect descriptor: "
1830 		       "len 0x%llx not multiple of 0x%zx\n",
1831 		       (unsigned long long)len,
1832 		       sizeof desc);
1833 		return -EINVAL;
1834 	}
1835 
1836 	ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
1837 			     UIO_MAXIOV, VHOST_ACCESS_RO);
1838 	if (unlikely(ret < 0)) {
1839 		if (ret != -EAGAIN)
1840 			vq_err(vq, "Translation failure %d in indirect.\n", ret);
1841 		return ret;
1842 	}
1843 	iov_iter_init(&from, READ, vq->indirect, ret, len);
1844 
1845 	/* We will use the result as an address to read from, so most
1846 	 * architectures only need a compiler barrier here. */
1847 	read_barrier_depends();
1848 
1849 	count = len / sizeof desc;
1850 	/* Buffers are chained via a 16 bit next field, so
1851 	 * we can have at most 2^16 of these. */
1852 	if (unlikely(count > USHRT_MAX + 1)) {
1853 		vq_err(vq, "Indirect buffer length too big: %d\n",
1854 		       indirect->len);
1855 		return -E2BIG;
1856 	}
1857 
1858 	do {
1859 		unsigned iov_count = *in_num + *out_num;
1860 		if (unlikely(++found > count)) {
1861 			vq_err(vq, "Loop detected: last one at %u "
1862 			       "indirect size %u\n",
1863 			       i, count);
1864 			return -EINVAL;
1865 		}
1866 		if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
1867 			vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1868 			       i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1869 			return -EINVAL;
1870 		}
1871 		if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
1872 			vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1873 			       i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1874 			return -EINVAL;
1875 		}
1876 
1877 		if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
1878 			access = VHOST_ACCESS_WO;
1879 		else
1880 			access = VHOST_ACCESS_RO;
1881 
1882 		ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
1883 				     vhost32_to_cpu(vq, desc.len), iov + iov_count,
1884 				     iov_size - iov_count, access);
1885 		if (unlikely(ret < 0)) {
1886 			if (ret != -EAGAIN)
1887 				vq_err(vq, "Translation failure %d indirect idx %d\n",
1888 					ret, i);
1889 			return ret;
1890 		}
1891 		/* If this is an input descriptor, increment that count. */
1892 		if (access == VHOST_ACCESS_WO) {
1893 			*in_num += ret;
1894 			if (unlikely(log)) {
1895 				log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
1896 				log[*log_num].len = vhost32_to_cpu(vq, desc.len);
1897 				++*log_num;
1898 			}
1899 		} else {
1900 			/* If it's an output descriptor, they're all supposed
1901 			 * to come before any input descriptors. */
1902 			if (unlikely(*in_num)) {
1903 				vq_err(vq, "Indirect descriptor "
1904 				       "has out after in: idx %d\n", i);
1905 				return -EINVAL;
1906 			}
1907 			*out_num += ret;
1908 		}
1909 	} while ((i = next_desc(vq, &desc)) != -1);
1910 	return 0;
1911 }
1912 
1913 /* This looks in the virtqueue and for the first available buffer, and converts
1914  * it to an iovec for convenient access.  Since descriptors consist of some
1915  * number of output then some number of input descriptors, it's actually two
1916  * iovecs, but we pack them into one and note how many of each there were.
1917  *
1918  * This function returns the descriptor number found, or vq->num (which is
1919  * never a valid descriptor number) if none was found.  A negative code is
1920  * returned on error. */
1921 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
1922 		      struct iovec iov[], unsigned int iov_size,
1923 		      unsigned int *out_num, unsigned int *in_num,
1924 		      struct vhost_log *log, unsigned int *log_num)
1925 {
1926 	struct vring_desc desc;
1927 	unsigned int i, head, found = 0;
1928 	u16 last_avail_idx;
1929 	__virtio16 avail_idx;
1930 	__virtio16 ring_head;
1931 	int ret, access;
1932 
1933 	/* Check it isn't doing very strange things with descriptor numbers. */
1934 	last_avail_idx = vq->last_avail_idx;
1935 	if (unlikely(vhost_get_user(vq, avail_idx, &vq->avail->idx))) {
1936 		vq_err(vq, "Failed to access avail idx at %p\n",
1937 		       &vq->avail->idx);
1938 		return -EFAULT;
1939 	}
1940 	vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
1941 
1942 	if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
1943 		vq_err(vq, "Guest moved used index from %u to %u",
1944 		       last_avail_idx, vq->avail_idx);
1945 		return -EFAULT;
1946 	}
1947 
1948 	/* If there's nothing new since last we looked, return invalid. */
1949 	if (vq->avail_idx == last_avail_idx)
1950 		return vq->num;
1951 
1952 	/* Only get avail ring entries after they have been exposed by guest. */
1953 	smp_rmb();
1954 
1955 	/* Grab the next descriptor number they're advertising, and increment
1956 	 * the index we've seen. */
1957 	if (unlikely(vhost_get_user(vq, ring_head,
1958 		     &vq->avail->ring[last_avail_idx & (vq->num - 1)]))) {
1959 		vq_err(vq, "Failed to read head: idx %d address %p\n",
1960 		       last_avail_idx,
1961 		       &vq->avail->ring[last_avail_idx % vq->num]);
1962 		return -EFAULT;
1963 	}
1964 
1965 	head = vhost16_to_cpu(vq, ring_head);
1966 
1967 	/* If their number is silly, that's an error. */
1968 	if (unlikely(head >= vq->num)) {
1969 		vq_err(vq, "Guest says index %u > %u is available",
1970 		       head, vq->num);
1971 		return -EINVAL;
1972 	}
1973 
1974 	/* When we start there are none of either input nor output. */
1975 	*out_num = *in_num = 0;
1976 	if (unlikely(log))
1977 		*log_num = 0;
1978 
1979 	i = head;
1980 	do {
1981 		unsigned iov_count = *in_num + *out_num;
1982 		if (unlikely(i >= vq->num)) {
1983 			vq_err(vq, "Desc index is %u > %u, head = %u",
1984 			       i, vq->num, head);
1985 			return -EINVAL;
1986 		}
1987 		if (unlikely(++found > vq->num)) {
1988 			vq_err(vq, "Loop detected: last one at %u "
1989 			       "vq size %u head %u\n",
1990 			       i, vq->num, head);
1991 			return -EINVAL;
1992 		}
1993 		ret = vhost_copy_from_user(vq, &desc, vq->desc + i,
1994 					   sizeof desc);
1995 		if (unlikely(ret)) {
1996 			vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
1997 			       i, vq->desc + i);
1998 			return -EFAULT;
1999 		}
2000 		if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2001 			ret = get_indirect(vq, iov, iov_size,
2002 					   out_num, in_num,
2003 					   log, log_num, &desc);
2004 			if (unlikely(ret < 0)) {
2005 				if (ret != -EAGAIN)
2006 					vq_err(vq, "Failure detected "
2007 						"in indirect descriptor at idx %d\n", i);
2008 				return ret;
2009 			}
2010 			continue;
2011 		}
2012 
2013 		if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2014 			access = VHOST_ACCESS_WO;
2015 		else
2016 			access = VHOST_ACCESS_RO;
2017 		ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2018 				     vhost32_to_cpu(vq, desc.len), iov + iov_count,
2019 				     iov_size - iov_count, access);
2020 		if (unlikely(ret < 0)) {
2021 			if (ret != -EAGAIN)
2022 				vq_err(vq, "Translation failure %d descriptor idx %d\n",
2023 					ret, i);
2024 			return ret;
2025 		}
2026 		if (access == VHOST_ACCESS_WO) {
2027 			/* If this is an input descriptor,
2028 			 * increment that count. */
2029 			*in_num += ret;
2030 			if (unlikely(log)) {
2031 				log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2032 				log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2033 				++*log_num;
2034 			}
2035 		} else {
2036 			/* If it's an output descriptor, they're all supposed
2037 			 * to come before any input descriptors. */
2038 			if (unlikely(*in_num)) {
2039 				vq_err(vq, "Descriptor has out after in: "
2040 				       "idx %d\n", i);
2041 				return -EINVAL;
2042 			}
2043 			*out_num += ret;
2044 		}
2045 	} while ((i = next_desc(vq, &desc)) != -1);
2046 
2047 	/* On success, increment avail index. */
2048 	vq->last_avail_idx++;
2049 
2050 	/* Assume notifications from guest are disabled at this point,
2051 	 * if they aren't we would need to update avail_event index. */
2052 	BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2053 	return head;
2054 }
2055 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2056 
2057 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2058 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2059 {
2060 	vq->last_avail_idx -= n;
2061 }
2062 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2063 
2064 /* After we've used one of their buffers, we tell them about it.  We'll then
2065  * want to notify the guest, using eventfd. */
2066 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2067 {
2068 	struct vring_used_elem heads = {
2069 		cpu_to_vhost32(vq, head),
2070 		cpu_to_vhost32(vq, len)
2071 	};
2072 
2073 	return vhost_add_used_n(vq, &heads, 1);
2074 }
2075 EXPORT_SYMBOL_GPL(vhost_add_used);
2076 
2077 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2078 			    struct vring_used_elem *heads,
2079 			    unsigned count)
2080 {
2081 	struct vring_used_elem __user *used;
2082 	u16 old, new;
2083 	int start;
2084 
2085 	start = vq->last_used_idx & (vq->num - 1);
2086 	used = vq->used->ring + start;
2087 	if (count == 1) {
2088 		if (vhost_put_user(vq, heads[0].id, &used->id)) {
2089 			vq_err(vq, "Failed to write used id");
2090 			return -EFAULT;
2091 		}
2092 		if (vhost_put_user(vq, heads[0].len, &used->len)) {
2093 			vq_err(vq, "Failed to write used len");
2094 			return -EFAULT;
2095 		}
2096 	} else if (vhost_copy_to_user(vq, used, heads, count * sizeof *used)) {
2097 		vq_err(vq, "Failed to write used");
2098 		return -EFAULT;
2099 	}
2100 	if (unlikely(vq->log_used)) {
2101 		/* Make sure data is seen before log. */
2102 		smp_wmb();
2103 		/* Log used ring entry write. */
2104 		log_write(vq->log_base,
2105 			  vq->log_addr +
2106 			   ((void __user *)used - (void __user *)vq->used),
2107 			  count * sizeof *used);
2108 	}
2109 	old = vq->last_used_idx;
2110 	new = (vq->last_used_idx += count);
2111 	/* If the driver never bothers to signal in a very long while,
2112 	 * used index might wrap around. If that happens, invalidate
2113 	 * signalled_used index we stored. TODO: make sure driver
2114 	 * signals at least once in 2^16 and remove this. */
2115 	if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2116 		vq->signalled_used_valid = false;
2117 	return 0;
2118 }
2119 
2120 /* After we've used one of their buffers, we tell them about it.  We'll then
2121  * want to notify the guest, using eventfd. */
2122 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2123 		     unsigned count)
2124 {
2125 	int start, n, r;
2126 
2127 	start = vq->last_used_idx & (vq->num - 1);
2128 	n = vq->num - start;
2129 	if (n < count) {
2130 		r = __vhost_add_used_n(vq, heads, n);
2131 		if (r < 0)
2132 			return r;
2133 		heads += n;
2134 		count -= n;
2135 	}
2136 	r = __vhost_add_used_n(vq, heads, count);
2137 
2138 	/* Make sure buffer is written before we update index. */
2139 	smp_wmb();
2140 	if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
2141 			   &vq->used->idx)) {
2142 		vq_err(vq, "Failed to increment used idx");
2143 		return -EFAULT;
2144 	}
2145 	if (unlikely(vq->log_used)) {
2146 		/* Log used index update. */
2147 		log_write(vq->log_base,
2148 			  vq->log_addr + offsetof(struct vring_used, idx),
2149 			  sizeof vq->used->idx);
2150 		if (vq->log_ctx)
2151 			eventfd_signal(vq->log_ctx, 1);
2152 	}
2153 	return r;
2154 }
2155 EXPORT_SYMBOL_GPL(vhost_add_used_n);
2156 
2157 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2158 {
2159 	__u16 old, new;
2160 	__virtio16 event;
2161 	bool v;
2162 
2163 	if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2164 	    unlikely(vq->avail_idx == vq->last_avail_idx))
2165 		return true;
2166 
2167 	if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2168 		__virtio16 flags;
2169 		/* Flush out used index updates. This is paired
2170 		 * with the barrier that the Guest executes when enabling
2171 		 * interrupts. */
2172 		smp_mb();
2173 		if (vhost_get_user(vq, flags, &vq->avail->flags)) {
2174 			vq_err(vq, "Failed to get flags");
2175 			return true;
2176 		}
2177 		return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2178 	}
2179 	old = vq->signalled_used;
2180 	v = vq->signalled_used_valid;
2181 	new = vq->signalled_used = vq->last_used_idx;
2182 	vq->signalled_used_valid = true;
2183 
2184 	if (unlikely(!v))
2185 		return true;
2186 
2187 	/* We're sure if the following conditions are met, there's no
2188 	 * need to notify guest:
2189 	 * 1) cached used event is ahead of new
2190 	 * 2) old to new updating does not cross cached used event. */
2191 	if (vring_need_event(vq->last_used_event, new + vq->num, new) &&
2192 	    !vring_need_event(vq->last_used_event, new, old))
2193 		return false;
2194 
2195 	/* Flush out used index updates. This is paired
2196 	 * with the barrier that the Guest executes when enabling
2197 	 * interrupts. */
2198 	smp_mb();
2199 
2200 	if (vhost_get_user(vq, event, vhost_used_event(vq))) {
2201 		vq_err(vq, "Failed to get used event idx");
2202 		return true;
2203 	}
2204 	vq->last_used_event = vhost16_to_cpu(vq, event);
2205 
2206 	return vring_need_event(vq->last_used_event, new, old);
2207 }
2208 
2209 /* This actually signals the guest, using eventfd. */
2210 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2211 {
2212 	/* Signal the Guest tell them we used something up. */
2213 	if (vq->call_ctx && vhost_notify(dev, vq))
2214 		eventfd_signal(vq->call_ctx, 1);
2215 }
2216 EXPORT_SYMBOL_GPL(vhost_signal);
2217 
2218 /* And here's the combo meal deal.  Supersize me! */
2219 void vhost_add_used_and_signal(struct vhost_dev *dev,
2220 			       struct vhost_virtqueue *vq,
2221 			       unsigned int head, int len)
2222 {
2223 	vhost_add_used(vq, head, len);
2224 	vhost_signal(dev, vq);
2225 }
2226 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2227 
2228 /* multi-buffer version of vhost_add_used_and_signal */
2229 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2230 				 struct vhost_virtqueue *vq,
2231 				 struct vring_used_elem *heads, unsigned count)
2232 {
2233 	vhost_add_used_n(vq, heads, count);
2234 	vhost_signal(dev, vq);
2235 }
2236 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2237 
2238 /* return true if we're sure that avaiable ring is empty */
2239 bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2240 {
2241 	__virtio16 avail_idx;
2242 	int r;
2243 
2244 	if (vq->avail_idx != vq->last_avail_idx)
2245 		return false;
2246 
2247 	r = vhost_get_user(vq, avail_idx, &vq->avail->idx);
2248 	if (unlikely(r))
2249 		return false;
2250 	vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2251 
2252 	return vq->avail_idx == vq->last_avail_idx;
2253 }
2254 EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2255 
2256 /* OK, now we need to know about added descriptors. */
2257 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2258 {
2259 	__virtio16 avail_idx;
2260 	int r;
2261 
2262 	if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2263 		return false;
2264 	vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2265 	if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2266 		r = vhost_update_used_flags(vq);
2267 		if (r) {
2268 			vq_err(vq, "Failed to enable notification at %p: %d\n",
2269 			       &vq->used->flags, r);
2270 			return false;
2271 		}
2272 	} else {
2273 		r = vhost_update_avail_event(vq, vq->avail_idx);
2274 		if (r) {
2275 			vq_err(vq, "Failed to update avail event index at %p: %d\n",
2276 			       vhost_avail_event(vq), r);
2277 			return false;
2278 		}
2279 	}
2280 	/* They could have slipped one in as we were doing that: make
2281 	 * sure it's written, then check again. */
2282 	smp_mb();
2283 	r = vhost_get_user(vq, avail_idx, &vq->avail->idx);
2284 	if (r) {
2285 		vq_err(vq, "Failed to check avail idx at %p: %d\n",
2286 		       &vq->avail->idx, r);
2287 		return false;
2288 	}
2289 
2290 	return vhost16_to_cpu(vq, avail_idx) != vq->avail_idx;
2291 }
2292 EXPORT_SYMBOL_GPL(vhost_enable_notify);
2293 
2294 /* We don't need to be notified again. */
2295 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2296 {
2297 	int r;
2298 
2299 	if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2300 		return;
2301 	vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2302 	if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2303 		r = vhost_update_used_flags(vq);
2304 		if (r)
2305 			vq_err(vq, "Failed to enable notification at %p: %d\n",
2306 			       &vq->used->flags, r);
2307 	}
2308 }
2309 EXPORT_SYMBOL_GPL(vhost_disable_notify);
2310 
2311 /* Create a new message. */
2312 struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2313 {
2314 	struct vhost_msg_node *node = kmalloc(sizeof *node, GFP_KERNEL);
2315 	if (!node)
2316 		return NULL;
2317 	node->vq = vq;
2318 	node->msg.type = type;
2319 	return node;
2320 }
2321 EXPORT_SYMBOL_GPL(vhost_new_msg);
2322 
2323 void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2324 		       struct vhost_msg_node *node)
2325 {
2326 	spin_lock(&dev->iotlb_lock);
2327 	list_add_tail(&node->node, head);
2328 	spin_unlock(&dev->iotlb_lock);
2329 
2330 	wake_up_interruptible_poll(&dev->wait, POLLIN | POLLRDNORM);
2331 }
2332 EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2333 
2334 struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2335 					 struct list_head *head)
2336 {
2337 	struct vhost_msg_node *node = NULL;
2338 
2339 	spin_lock(&dev->iotlb_lock);
2340 	if (!list_empty(head)) {
2341 		node = list_first_entry(head, struct vhost_msg_node,
2342 					node);
2343 		list_del(&node->node);
2344 	}
2345 	spin_unlock(&dev->iotlb_lock);
2346 
2347 	return node;
2348 }
2349 EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2350 
2351 
2352 static int __init vhost_init(void)
2353 {
2354 	return 0;
2355 }
2356 
2357 static void __exit vhost_exit(void)
2358 {
2359 }
2360 
2361 module_init(vhost_init);
2362 module_exit(vhost_exit);
2363 
2364 MODULE_VERSION("0.0.1");
2365 MODULE_LICENSE("GPL v2");
2366 MODULE_AUTHOR("Michael S. Tsirkin");
2367 MODULE_DESCRIPTION("Host kernel accelerator for virtio");
2368