xref: /openbmc/linux/drivers/vhost/vhost.c (revision 0c6dfa75)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright (C) 2009 Red Hat, Inc.
3  * Copyright (C) 2006 Rusty Russell IBM Corporation
4  *
5  * Author: Michael S. Tsirkin <mst@redhat.com>
6  *
7  * Inspiration, some code, and most witty comments come from
8  * Documentation/virtual/lguest/lguest.c, by Rusty Russell
9  *
10  * Generic code for virtio server in host kernel.
11  */
12 
13 #include <linux/eventfd.h>
14 #include <linux/vhost.h>
15 #include <linux/uio.h>
16 #include <linux/mm.h>
17 #include <linux/miscdevice.h>
18 #include <linux/mutex.h>
19 #include <linux/poll.h>
20 #include <linux/file.h>
21 #include <linux/highmem.h>
22 #include <linux/slab.h>
23 #include <linux/vmalloc.h>
24 #include <linux/kthread.h>
25 #include <linux/module.h>
26 #include <linux/sort.h>
27 #include <linux/sched/mm.h>
28 #include <linux/sched/signal.h>
29 #include <linux/sched/vhost_task.h>
30 #include <linux/interval_tree_generic.h>
31 #include <linux/nospec.h>
32 #include <linux/kcov.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 #ifdef CONFIG_VHOST_CROSS_ENDIAN_LEGACY
53 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
54 {
55 	vq->user_be = !virtio_legacy_is_little_endian();
56 }
57 
58 static void vhost_enable_cross_endian_big(struct vhost_virtqueue *vq)
59 {
60 	vq->user_be = true;
61 }
62 
63 static void vhost_enable_cross_endian_little(struct vhost_virtqueue *vq)
64 {
65 	vq->user_be = false;
66 }
67 
68 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
69 {
70 	struct vhost_vring_state s;
71 
72 	if (vq->private_data)
73 		return -EBUSY;
74 
75 	if (copy_from_user(&s, argp, sizeof(s)))
76 		return -EFAULT;
77 
78 	if (s.num != VHOST_VRING_LITTLE_ENDIAN &&
79 	    s.num != VHOST_VRING_BIG_ENDIAN)
80 		return -EINVAL;
81 
82 	if (s.num == VHOST_VRING_BIG_ENDIAN)
83 		vhost_enable_cross_endian_big(vq);
84 	else
85 		vhost_enable_cross_endian_little(vq);
86 
87 	return 0;
88 }
89 
90 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
91 				   int __user *argp)
92 {
93 	struct vhost_vring_state s = {
94 		.index = idx,
95 		.num = vq->user_be
96 	};
97 
98 	if (copy_to_user(argp, &s, sizeof(s)))
99 		return -EFAULT;
100 
101 	return 0;
102 }
103 
104 static void vhost_init_is_le(struct vhost_virtqueue *vq)
105 {
106 	/* Note for legacy virtio: user_be is initialized at reset time
107 	 * according to the host endianness. If userspace does not set an
108 	 * explicit endianness, the default behavior is native endian, as
109 	 * expected by legacy virtio.
110 	 */
111 	vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1) || !vq->user_be;
112 }
113 #else
114 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
115 {
116 }
117 
118 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
119 {
120 	return -ENOIOCTLCMD;
121 }
122 
123 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
124 				   int __user *argp)
125 {
126 	return -ENOIOCTLCMD;
127 }
128 
129 static void vhost_init_is_le(struct vhost_virtqueue *vq)
130 {
131 	vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1)
132 		|| virtio_legacy_is_little_endian();
133 }
134 #endif /* CONFIG_VHOST_CROSS_ENDIAN_LEGACY */
135 
136 static void vhost_reset_is_le(struct vhost_virtqueue *vq)
137 {
138 	vhost_init_is_le(vq);
139 }
140 
141 struct vhost_flush_struct {
142 	struct vhost_work work;
143 	struct completion wait_event;
144 };
145 
146 static void vhost_flush_work(struct vhost_work *work)
147 {
148 	struct vhost_flush_struct *s;
149 
150 	s = container_of(work, struct vhost_flush_struct, work);
151 	complete(&s->wait_event);
152 }
153 
154 static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
155 			    poll_table *pt)
156 {
157 	struct vhost_poll *poll;
158 
159 	poll = container_of(pt, struct vhost_poll, table);
160 	poll->wqh = wqh;
161 	add_wait_queue(wqh, &poll->wait);
162 }
163 
164 static int vhost_poll_wakeup(wait_queue_entry_t *wait, unsigned mode, int sync,
165 			     void *key)
166 {
167 	struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
168 	struct vhost_work *work = &poll->work;
169 
170 	if (!(key_to_poll(key) & poll->mask))
171 		return 0;
172 
173 	if (!poll->dev->use_worker)
174 		work->fn(work);
175 	else
176 		vhost_poll_queue(poll);
177 
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 }
186 EXPORT_SYMBOL_GPL(vhost_work_init);
187 
188 /* Init poll structure */
189 void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
190 		     __poll_t mask, struct vhost_dev *dev,
191 		     struct vhost_virtqueue *vq)
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 	poll->vq = vq;
199 
200 	vhost_work_init(&poll->work, fn);
201 }
202 EXPORT_SYMBOL_GPL(vhost_poll_init);
203 
204 /* Start polling a file. We add ourselves to file's wait queue. The caller must
205  * keep a reference to a file until after vhost_poll_stop is called. */
206 int vhost_poll_start(struct vhost_poll *poll, struct file *file)
207 {
208 	__poll_t mask;
209 
210 	if (poll->wqh)
211 		return 0;
212 
213 	mask = vfs_poll(file, &poll->table);
214 	if (mask)
215 		vhost_poll_wakeup(&poll->wait, 0, 0, poll_to_key(mask));
216 	if (mask & EPOLLERR) {
217 		vhost_poll_stop(poll);
218 		return -EINVAL;
219 	}
220 
221 	return 0;
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 static void vhost_worker_queue(struct vhost_worker *worker,
237 			       struct vhost_work *work)
238 {
239 	if (!test_and_set_bit(VHOST_WORK_QUEUED, &work->flags)) {
240 		/* We can only add the work to the list after we're
241 		 * sure it was not in the list.
242 		 * test_and_set_bit() implies a memory barrier.
243 		 */
244 		llist_add(&work->node, &worker->work_list);
245 		vhost_task_wake(worker->vtsk);
246 	}
247 }
248 
249 bool vhost_vq_work_queue(struct vhost_virtqueue *vq, struct vhost_work *work)
250 {
251 	struct vhost_worker *worker;
252 	bool queued = false;
253 
254 	rcu_read_lock();
255 	worker = rcu_dereference(vq->worker);
256 	if (worker) {
257 		queued = true;
258 		vhost_worker_queue(worker, work);
259 	}
260 	rcu_read_unlock();
261 
262 	return queued;
263 }
264 EXPORT_SYMBOL_GPL(vhost_vq_work_queue);
265 
266 void vhost_vq_flush(struct vhost_virtqueue *vq)
267 {
268 	struct vhost_flush_struct flush;
269 
270 	init_completion(&flush.wait_event);
271 	vhost_work_init(&flush.work, vhost_flush_work);
272 
273 	if (vhost_vq_work_queue(vq, &flush.work))
274 		wait_for_completion(&flush.wait_event);
275 }
276 EXPORT_SYMBOL_GPL(vhost_vq_flush);
277 
278 /**
279  * vhost_worker_flush - flush a worker
280  * @worker: worker to flush
281  *
282  * This does not use RCU to protect the worker, so the device or worker
283  * mutex must be held.
284  */
285 static void vhost_worker_flush(struct vhost_worker *worker)
286 {
287 	struct vhost_flush_struct flush;
288 
289 	init_completion(&flush.wait_event);
290 	vhost_work_init(&flush.work, vhost_flush_work);
291 
292 	vhost_worker_queue(worker, &flush.work);
293 	wait_for_completion(&flush.wait_event);
294 }
295 
296 void vhost_dev_flush(struct vhost_dev *dev)
297 {
298 	struct vhost_worker *worker;
299 	unsigned long i;
300 
301 	xa_for_each(&dev->worker_xa, i, worker) {
302 		mutex_lock(&worker->mutex);
303 		if (!worker->attachment_cnt) {
304 			mutex_unlock(&worker->mutex);
305 			continue;
306 		}
307 		vhost_worker_flush(worker);
308 		mutex_unlock(&worker->mutex);
309 	}
310 }
311 EXPORT_SYMBOL_GPL(vhost_dev_flush);
312 
313 /* A lockless hint for busy polling code to exit the loop */
314 bool vhost_vq_has_work(struct vhost_virtqueue *vq)
315 {
316 	struct vhost_worker *worker;
317 	bool has_work = false;
318 
319 	rcu_read_lock();
320 	worker = rcu_dereference(vq->worker);
321 	if (worker && !llist_empty(&worker->work_list))
322 		has_work = true;
323 	rcu_read_unlock();
324 
325 	return has_work;
326 }
327 EXPORT_SYMBOL_GPL(vhost_vq_has_work);
328 
329 void vhost_poll_queue(struct vhost_poll *poll)
330 {
331 	vhost_vq_work_queue(poll->vq, &poll->work);
332 }
333 EXPORT_SYMBOL_GPL(vhost_poll_queue);
334 
335 static void __vhost_vq_meta_reset(struct vhost_virtqueue *vq)
336 {
337 	int j;
338 
339 	for (j = 0; j < VHOST_NUM_ADDRS; j++)
340 		vq->meta_iotlb[j] = NULL;
341 }
342 
343 static void vhost_vq_meta_reset(struct vhost_dev *d)
344 {
345 	int i;
346 
347 	for (i = 0; i < d->nvqs; ++i)
348 		__vhost_vq_meta_reset(d->vqs[i]);
349 }
350 
351 static void vhost_vring_call_reset(struct vhost_vring_call *call_ctx)
352 {
353 	call_ctx->ctx = NULL;
354 	memset(&call_ctx->producer, 0x0, sizeof(struct irq_bypass_producer));
355 }
356 
357 bool vhost_vq_is_setup(struct vhost_virtqueue *vq)
358 {
359 	return vq->avail && vq->desc && vq->used && vhost_vq_access_ok(vq);
360 }
361 EXPORT_SYMBOL_GPL(vhost_vq_is_setup);
362 
363 static void vhost_vq_reset(struct vhost_dev *dev,
364 			   struct vhost_virtqueue *vq)
365 {
366 	vq->num = 1;
367 	vq->desc = NULL;
368 	vq->avail = NULL;
369 	vq->used = NULL;
370 	vq->last_avail_idx = 0;
371 	vq->avail_idx = 0;
372 	vq->last_used_idx = 0;
373 	vq->signalled_used = 0;
374 	vq->signalled_used_valid = false;
375 	vq->used_flags = 0;
376 	vq->log_used = false;
377 	vq->log_addr = -1ull;
378 	vq->private_data = NULL;
379 	vq->acked_features = 0;
380 	vq->acked_backend_features = 0;
381 	vq->log_base = NULL;
382 	vq->error_ctx = NULL;
383 	vq->kick = NULL;
384 	vq->log_ctx = NULL;
385 	vhost_disable_cross_endian(vq);
386 	vhost_reset_is_le(vq);
387 	vq->busyloop_timeout = 0;
388 	vq->umem = NULL;
389 	vq->iotlb = NULL;
390 	rcu_assign_pointer(vq->worker, NULL);
391 	vhost_vring_call_reset(&vq->call_ctx);
392 	__vhost_vq_meta_reset(vq);
393 }
394 
395 static bool vhost_worker(void *data)
396 {
397 	struct vhost_worker *worker = data;
398 	struct vhost_work *work, *work_next;
399 	struct llist_node *node;
400 
401 	node = llist_del_all(&worker->work_list);
402 	if (node) {
403 		__set_current_state(TASK_RUNNING);
404 
405 		node = llist_reverse_order(node);
406 		/* make sure flag is seen after deletion */
407 		smp_wmb();
408 		llist_for_each_entry_safe(work, work_next, node, node) {
409 			clear_bit(VHOST_WORK_QUEUED, &work->flags);
410 			kcov_remote_start_common(worker->kcov_handle);
411 			work->fn(work);
412 			kcov_remote_stop();
413 			cond_resched();
414 		}
415 	}
416 
417 	return !!node;
418 }
419 
420 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
421 {
422 	kfree(vq->indirect);
423 	vq->indirect = NULL;
424 	kfree(vq->log);
425 	vq->log = NULL;
426 	kfree(vq->heads);
427 	vq->heads = NULL;
428 }
429 
430 /* Helper to allocate iovec buffers for all vqs. */
431 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
432 {
433 	struct vhost_virtqueue *vq;
434 	int i;
435 
436 	for (i = 0; i < dev->nvqs; ++i) {
437 		vq = dev->vqs[i];
438 		vq->indirect = kmalloc_array(UIO_MAXIOV,
439 					     sizeof(*vq->indirect),
440 					     GFP_KERNEL);
441 		vq->log = kmalloc_array(dev->iov_limit, sizeof(*vq->log),
442 					GFP_KERNEL);
443 		vq->heads = kmalloc_array(dev->iov_limit, sizeof(*vq->heads),
444 					  GFP_KERNEL);
445 		if (!vq->indirect || !vq->log || !vq->heads)
446 			goto err_nomem;
447 	}
448 	return 0;
449 
450 err_nomem:
451 	for (; i >= 0; --i)
452 		vhost_vq_free_iovecs(dev->vqs[i]);
453 	return -ENOMEM;
454 }
455 
456 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
457 {
458 	int i;
459 
460 	for (i = 0; i < dev->nvqs; ++i)
461 		vhost_vq_free_iovecs(dev->vqs[i]);
462 }
463 
464 bool vhost_exceeds_weight(struct vhost_virtqueue *vq,
465 			  int pkts, int total_len)
466 {
467 	struct vhost_dev *dev = vq->dev;
468 
469 	if ((dev->byte_weight && total_len >= dev->byte_weight) ||
470 	    pkts >= dev->weight) {
471 		vhost_poll_queue(&vq->poll);
472 		return true;
473 	}
474 
475 	return false;
476 }
477 EXPORT_SYMBOL_GPL(vhost_exceeds_weight);
478 
479 static size_t vhost_get_avail_size(struct vhost_virtqueue *vq,
480 				   unsigned int num)
481 {
482 	size_t event __maybe_unused =
483 	       vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
484 
485 	return size_add(struct_size(vq->avail, ring, num), event);
486 }
487 
488 static size_t vhost_get_used_size(struct vhost_virtqueue *vq,
489 				  unsigned int num)
490 {
491 	size_t event __maybe_unused =
492 	       vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
493 
494 	return size_add(struct_size(vq->used, ring, num), event);
495 }
496 
497 static size_t vhost_get_desc_size(struct vhost_virtqueue *vq,
498 				  unsigned int num)
499 {
500 	return sizeof(*vq->desc) * num;
501 }
502 
503 void vhost_dev_init(struct vhost_dev *dev,
504 		    struct vhost_virtqueue **vqs, int nvqs,
505 		    int iov_limit, int weight, int byte_weight,
506 		    bool use_worker,
507 		    int (*msg_handler)(struct vhost_dev *dev, u32 asid,
508 				       struct vhost_iotlb_msg *msg))
509 {
510 	struct vhost_virtqueue *vq;
511 	int i;
512 
513 	dev->vqs = vqs;
514 	dev->nvqs = nvqs;
515 	mutex_init(&dev->mutex);
516 	dev->log_ctx = NULL;
517 	dev->umem = NULL;
518 	dev->iotlb = NULL;
519 	dev->mm = NULL;
520 	dev->iov_limit = iov_limit;
521 	dev->weight = weight;
522 	dev->byte_weight = byte_weight;
523 	dev->use_worker = use_worker;
524 	dev->msg_handler = msg_handler;
525 	init_waitqueue_head(&dev->wait);
526 	INIT_LIST_HEAD(&dev->read_list);
527 	INIT_LIST_HEAD(&dev->pending_list);
528 	spin_lock_init(&dev->iotlb_lock);
529 	xa_init_flags(&dev->worker_xa, XA_FLAGS_ALLOC);
530 
531 	for (i = 0; i < dev->nvqs; ++i) {
532 		vq = dev->vqs[i];
533 		vq->log = NULL;
534 		vq->indirect = NULL;
535 		vq->heads = NULL;
536 		vq->dev = dev;
537 		mutex_init(&vq->mutex);
538 		vhost_vq_reset(dev, vq);
539 		if (vq->handle_kick)
540 			vhost_poll_init(&vq->poll, vq->handle_kick,
541 					EPOLLIN, dev, vq);
542 	}
543 }
544 EXPORT_SYMBOL_GPL(vhost_dev_init);
545 
546 /* Caller should have device mutex */
547 long vhost_dev_check_owner(struct vhost_dev *dev)
548 {
549 	/* Are you the owner? If not, I don't think you mean to do that */
550 	return dev->mm == current->mm ? 0 : -EPERM;
551 }
552 EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
553 
554 /* Caller should have device mutex */
555 bool vhost_dev_has_owner(struct vhost_dev *dev)
556 {
557 	return dev->mm;
558 }
559 EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
560 
561 static void vhost_attach_mm(struct vhost_dev *dev)
562 {
563 	/* No owner, become one */
564 	if (dev->use_worker) {
565 		dev->mm = get_task_mm(current);
566 	} else {
567 		/* vDPA device does not use worker thead, so there's
568 		 * no need to hold the address space for mm. This help
569 		 * to avoid deadlock in the case of mmap() which may
570 		 * held the refcnt of the file and depends on release
571 		 * method to remove vma.
572 		 */
573 		dev->mm = current->mm;
574 		mmgrab(dev->mm);
575 	}
576 }
577 
578 static void vhost_detach_mm(struct vhost_dev *dev)
579 {
580 	if (!dev->mm)
581 		return;
582 
583 	if (dev->use_worker)
584 		mmput(dev->mm);
585 	else
586 		mmdrop(dev->mm);
587 
588 	dev->mm = NULL;
589 }
590 
591 static void vhost_worker_destroy(struct vhost_dev *dev,
592 				 struct vhost_worker *worker)
593 {
594 	if (!worker)
595 		return;
596 
597 	WARN_ON(!llist_empty(&worker->work_list));
598 	xa_erase(&dev->worker_xa, worker->id);
599 	vhost_task_stop(worker->vtsk);
600 	kfree(worker);
601 }
602 
603 static void vhost_workers_free(struct vhost_dev *dev)
604 {
605 	struct vhost_worker *worker;
606 	unsigned long i;
607 
608 	if (!dev->use_worker)
609 		return;
610 
611 	for (i = 0; i < dev->nvqs; i++)
612 		rcu_assign_pointer(dev->vqs[i]->worker, NULL);
613 	/*
614 	 * Free the default worker we created and cleanup workers userspace
615 	 * created but couldn't clean up (it forgot or crashed).
616 	 */
617 	xa_for_each(&dev->worker_xa, i, worker)
618 		vhost_worker_destroy(dev, worker);
619 	xa_destroy(&dev->worker_xa);
620 }
621 
622 static struct vhost_worker *vhost_worker_create(struct vhost_dev *dev)
623 {
624 	struct vhost_worker *worker;
625 	struct vhost_task *vtsk;
626 	char name[TASK_COMM_LEN];
627 	int ret;
628 	u32 id;
629 
630 	worker = kzalloc(sizeof(*worker), GFP_KERNEL_ACCOUNT);
631 	if (!worker)
632 		return NULL;
633 
634 	snprintf(name, sizeof(name), "vhost-%d", current->pid);
635 
636 	vtsk = vhost_task_create(vhost_worker, worker, name);
637 	if (!vtsk)
638 		goto free_worker;
639 
640 	mutex_init(&worker->mutex);
641 	init_llist_head(&worker->work_list);
642 	worker->kcov_handle = kcov_common_handle();
643 	worker->vtsk = vtsk;
644 
645 	vhost_task_start(vtsk);
646 
647 	ret = xa_alloc(&dev->worker_xa, &id, worker, xa_limit_32b, GFP_KERNEL);
648 	if (ret < 0)
649 		goto stop_worker;
650 	worker->id = id;
651 
652 	return worker;
653 
654 stop_worker:
655 	vhost_task_stop(vtsk);
656 free_worker:
657 	kfree(worker);
658 	return NULL;
659 }
660 
661 /* Caller must have device mutex */
662 static void __vhost_vq_attach_worker(struct vhost_virtqueue *vq,
663 				     struct vhost_worker *worker)
664 {
665 	struct vhost_worker *old_worker;
666 
667 	old_worker = rcu_dereference_check(vq->worker,
668 					   lockdep_is_held(&vq->dev->mutex));
669 
670 	mutex_lock(&worker->mutex);
671 	worker->attachment_cnt++;
672 	mutex_unlock(&worker->mutex);
673 	rcu_assign_pointer(vq->worker, worker);
674 
675 	if (!old_worker)
676 		return;
677 	/*
678 	 * Take the worker mutex to make sure we see the work queued from
679 	 * device wide flushes which doesn't use RCU for execution.
680 	 */
681 	mutex_lock(&old_worker->mutex);
682 	old_worker->attachment_cnt--;
683 	/*
684 	 * We don't want to call synchronize_rcu for every vq during setup
685 	 * because it will slow down VM startup. If we haven't done
686 	 * VHOST_SET_VRING_KICK and not done the driver specific
687 	 * SET_ENDPOINT/RUNNUNG then we can skip the sync since there will
688 	 * not be any works queued for scsi and net.
689 	 */
690 	mutex_lock(&vq->mutex);
691 	if (!vhost_vq_get_backend(vq) && !vq->kick) {
692 		mutex_unlock(&vq->mutex);
693 		mutex_unlock(&old_worker->mutex);
694 		/*
695 		 * vsock can queue anytime after VHOST_VSOCK_SET_GUEST_CID.
696 		 * Warn if it adds support for multiple workers but forgets to
697 		 * handle the early queueing case.
698 		 */
699 		WARN_ON(!old_worker->attachment_cnt &&
700 			!llist_empty(&old_worker->work_list));
701 		return;
702 	}
703 	mutex_unlock(&vq->mutex);
704 
705 	/* Make sure new vq queue/flush/poll calls see the new worker */
706 	synchronize_rcu();
707 	/* Make sure whatever was queued gets run */
708 	vhost_worker_flush(old_worker);
709 	mutex_unlock(&old_worker->mutex);
710 }
711 
712  /* Caller must have device mutex */
713 static int vhost_vq_attach_worker(struct vhost_virtqueue *vq,
714 				  struct vhost_vring_worker *info)
715 {
716 	unsigned long index = info->worker_id;
717 	struct vhost_dev *dev = vq->dev;
718 	struct vhost_worker *worker;
719 
720 	if (!dev->use_worker)
721 		return -EINVAL;
722 
723 	worker = xa_find(&dev->worker_xa, &index, UINT_MAX, XA_PRESENT);
724 	if (!worker || worker->id != info->worker_id)
725 		return -ENODEV;
726 
727 	__vhost_vq_attach_worker(vq, worker);
728 	return 0;
729 }
730 
731 /* Caller must have device mutex */
732 static int vhost_new_worker(struct vhost_dev *dev,
733 			    struct vhost_worker_state *info)
734 {
735 	struct vhost_worker *worker;
736 
737 	worker = vhost_worker_create(dev);
738 	if (!worker)
739 		return -ENOMEM;
740 
741 	info->worker_id = worker->id;
742 	return 0;
743 }
744 
745 /* Caller must have device mutex */
746 static int vhost_free_worker(struct vhost_dev *dev,
747 			     struct vhost_worker_state *info)
748 {
749 	unsigned long index = info->worker_id;
750 	struct vhost_worker *worker;
751 
752 	worker = xa_find(&dev->worker_xa, &index, UINT_MAX, XA_PRESENT);
753 	if (!worker || worker->id != info->worker_id)
754 		return -ENODEV;
755 
756 	mutex_lock(&worker->mutex);
757 	if (worker->attachment_cnt) {
758 		mutex_unlock(&worker->mutex);
759 		return -EBUSY;
760 	}
761 	mutex_unlock(&worker->mutex);
762 
763 	vhost_worker_destroy(dev, worker);
764 	return 0;
765 }
766 
767 static int vhost_get_vq_from_user(struct vhost_dev *dev, void __user *argp,
768 				  struct vhost_virtqueue **vq, u32 *id)
769 {
770 	u32 __user *idxp = argp;
771 	u32 idx;
772 	long r;
773 
774 	r = get_user(idx, idxp);
775 	if (r < 0)
776 		return r;
777 
778 	if (idx >= dev->nvqs)
779 		return -ENOBUFS;
780 
781 	idx = array_index_nospec(idx, dev->nvqs);
782 
783 	*vq = dev->vqs[idx];
784 	*id = idx;
785 	return 0;
786 }
787 
788 /* Caller must have device mutex */
789 long vhost_worker_ioctl(struct vhost_dev *dev, unsigned int ioctl,
790 			void __user *argp)
791 {
792 	struct vhost_vring_worker ring_worker;
793 	struct vhost_worker_state state;
794 	struct vhost_worker *worker;
795 	struct vhost_virtqueue *vq;
796 	long ret;
797 	u32 idx;
798 
799 	if (!dev->use_worker)
800 		return -EINVAL;
801 
802 	if (!vhost_dev_has_owner(dev))
803 		return -EINVAL;
804 
805 	ret = vhost_dev_check_owner(dev);
806 	if (ret)
807 		return ret;
808 
809 	switch (ioctl) {
810 	/* dev worker ioctls */
811 	case VHOST_NEW_WORKER:
812 		ret = vhost_new_worker(dev, &state);
813 		if (!ret && copy_to_user(argp, &state, sizeof(state)))
814 			ret = -EFAULT;
815 		return ret;
816 	case VHOST_FREE_WORKER:
817 		if (copy_from_user(&state, argp, sizeof(state)))
818 			return -EFAULT;
819 		return vhost_free_worker(dev, &state);
820 	/* vring worker ioctls */
821 	case VHOST_ATTACH_VRING_WORKER:
822 	case VHOST_GET_VRING_WORKER:
823 		break;
824 	default:
825 		return -ENOIOCTLCMD;
826 	}
827 
828 	ret = vhost_get_vq_from_user(dev, argp, &vq, &idx);
829 	if (ret)
830 		return ret;
831 
832 	switch (ioctl) {
833 	case VHOST_ATTACH_VRING_WORKER:
834 		if (copy_from_user(&ring_worker, argp, sizeof(ring_worker))) {
835 			ret = -EFAULT;
836 			break;
837 		}
838 
839 		ret = vhost_vq_attach_worker(vq, &ring_worker);
840 		break;
841 	case VHOST_GET_VRING_WORKER:
842 		worker = rcu_dereference_check(vq->worker,
843 					       lockdep_is_held(&dev->mutex));
844 		if (!worker) {
845 			ret = -EINVAL;
846 			break;
847 		}
848 
849 		ring_worker.index = idx;
850 		ring_worker.worker_id = worker->id;
851 
852 		if (copy_to_user(argp, &ring_worker, sizeof(ring_worker)))
853 			ret = -EFAULT;
854 		break;
855 	default:
856 		ret = -ENOIOCTLCMD;
857 		break;
858 	}
859 
860 	return ret;
861 }
862 EXPORT_SYMBOL_GPL(vhost_worker_ioctl);
863 
864 /* Caller should have device mutex */
865 long vhost_dev_set_owner(struct vhost_dev *dev)
866 {
867 	struct vhost_worker *worker;
868 	int err, i;
869 
870 	/* Is there an owner already? */
871 	if (vhost_dev_has_owner(dev)) {
872 		err = -EBUSY;
873 		goto err_mm;
874 	}
875 
876 	vhost_attach_mm(dev);
877 
878 	err = vhost_dev_alloc_iovecs(dev);
879 	if (err)
880 		goto err_iovecs;
881 
882 	if (dev->use_worker) {
883 		/*
884 		 * This should be done last, because vsock can queue work
885 		 * before VHOST_SET_OWNER so it simplifies the failure path
886 		 * below since we don't have to worry about vsock queueing
887 		 * while we free the worker.
888 		 */
889 		worker = vhost_worker_create(dev);
890 		if (!worker) {
891 			err = -ENOMEM;
892 			goto err_worker;
893 		}
894 
895 		for (i = 0; i < dev->nvqs; i++)
896 			__vhost_vq_attach_worker(dev->vqs[i], worker);
897 	}
898 
899 	return 0;
900 
901 err_worker:
902 	vhost_dev_free_iovecs(dev);
903 err_iovecs:
904 	vhost_detach_mm(dev);
905 err_mm:
906 	return err;
907 }
908 EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
909 
910 static struct vhost_iotlb *iotlb_alloc(void)
911 {
912 	return vhost_iotlb_alloc(max_iotlb_entries,
913 				 VHOST_IOTLB_FLAG_RETIRE);
914 }
915 
916 struct vhost_iotlb *vhost_dev_reset_owner_prepare(void)
917 {
918 	return iotlb_alloc();
919 }
920 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
921 
922 /* Caller should have device mutex */
923 void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_iotlb *umem)
924 {
925 	int i;
926 
927 	vhost_dev_cleanup(dev);
928 
929 	dev->umem = umem;
930 	/* We don't need VQ locks below since vhost_dev_cleanup makes sure
931 	 * VQs aren't running.
932 	 */
933 	for (i = 0; i < dev->nvqs; ++i)
934 		dev->vqs[i]->umem = umem;
935 }
936 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
937 
938 void vhost_dev_stop(struct vhost_dev *dev)
939 {
940 	int i;
941 
942 	for (i = 0; i < dev->nvqs; ++i) {
943 		if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick)
944 			vhost_poll_stop(&dev->vqs[i]->poll);
945 	}
946 
947 	vhost_dev_flush(dev);
948 }
949 EXPORT_SYMBOL_GPL(vhost_dev_stop);
950 
951 void vhost_clear_msg(struct vhost_dev *dev)
952 {
953 	struct vhost_msg_node *node, *n;
954 
955 	spin_lock(&dev->iotlb_lock);
956 
957 	list_for_each_entry_safe(node, n, &dev->read_list, node) {
958 		list_del(&node->node);
959 		kfree(node);
960 	}
961 
962 	list_for_each_entry_safe(node, n, &dev->pending_list, node) {
963 		list_del(&node->node);
964 		kfree(node);
965 	}
966 
967 	spin_unlock(&dev->iotlb_lock);
968 }
969 EXPORT_SYMBOL_GPL(vhost_clear_msg);
970 
971 void vhost_dev_cleanup(struct vhost_dev *dev)
972 {
973 	int i;
974 
975 	for (i = 0; i < dev->nvqs; ++i) {
976 		if (dev->vqs[i]->error_ctx)
977 			eventfd_ctx_put(dev->vqs[i]->error_ctx);
978 		if (dev->vqs[i]->kick)
979 			fput(dev->vqs[i]->kick);
980 		if (dev->vqs[i]->call_ctx.ctx)
981 			eventfd_ctx_put(dev->vqs[i]->call_ctx.ctx);
982 		vhost_vq_reset(dev, dev->vqs[i]);
983 	}
984 	vhost_dev_free_iovecs(dev);
985 	if (dev->log_ctx)
986 		eventfd_ctx_put(dev->log_ctx);
987 	dev->log_ctx = NULL;
988 	/* No one will access memory at this point */
989 	vhost_iotlb_free(dev->umem);
990 	dev->umem = NULL;
991 	vhost_iotlb_free(dev->iotlb);
992 	dev->iotlb = NULL;
993 	vhost_clear_msg(dev);
994 	wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
995 	vhost_workers_free(dev);
996 	vhost_detach_mm(dev);
997 }
998 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
999 
1000 static bool log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
1001 {
1002 	u64 a = addr / VHOST_PAGE_SIZE / 8;
1003 
1004 	/* Make sure 64 bit math will not overflow. */
1005 	if (a > ULONG_MAX - (unsigned long)log_base ||
1006 	    a + (unsigned long)log_base > ULONG_MAX)
1007 		return false;
1008 
1009 	return access_ok(log_base + a,
1010 			 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
1011 }
1012 
1013 /* Make sure 64 bit math will not overflow. */
1014 static bool vhost_overflow(u64 uaddr, u64 size)
1015 {
1016 	if (uaddr > ULONG_MAX || size > ULONG_MAX)
1017 		return true;
1018 
1019 	if (!size)
1020 		return false;
1021 
1022 	return uaddr > ULONG_MAX - size + 1;
1023 }
1024 
1025 /* Caller should have vq mutex and device mutex. */
1026 static bool vq_memory_access_ok(void __user *log_base, struct vhost_iotlb *umem,
1027 				int log_all)
1028 {
1029 	struct vhost_iotlb_map *map;
1030 
1031 	if (!umem)
1032 		return false;
1033 
1034 	list_for_each_entry(map, &umem->list, link) {
1035 		unsigned long a = map->addr;
1036 
1037 		if (vhost_overflow(map->addr, map->size))
1038 			return false;
1039 
1040 
1041 		if (!access_ok((void __user *)a, map->size))
1042 			return false;
1043 		else if (log_all && !log_access_ok(log_base,
1044 						   map->start,
1045 						   map->size))
1046 			return false;
1047 	}
1048 	return true;
1049 }
1050 
1051 static inline void __user *vhost_vq_meta_fetch(struct vhost_virtqueue *vq,
1052 					       u64 addr, unsigned int size,
1053 					       int type)
1054 {
1055 	const struct vhost_iotlb_map *map = vq->meta_iotlb[type];
1056 
1057 	if (!map)
1058 		return NULL;
1059 
1060 	return (void __user *)(uintptr_t)(map->addr + addr - map->start);
1061 }
1062 
1063 /* Can we switch to this memory table? */
1064 /* Caller should have device mutex but not vq mutex */
1065 static bool memory_access_ok(struct vhost_dev *d, struct vhost_iotlb *umem,
1066 			     int log_all)
1067 {
1068 	int i;
1069 
1070 	for (i = 0; i < d->nvqs; ++i) {
1071 		bool ok;
1072 		bool log;
1073 
1074 		mutex_lock(&d->vqs[i]->mutex);
1075 		log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
1076 		/* If ring is inactive, will check when it's enabled. */
1077 		if (d->vqs[i]->private_data)
1078 			ok = vq_memory_access_ok(d->vqs[i]->log_base,
1079 						 umem, log);
1080 		else
1081 			ok = true;
1082 		mutex_unlock(&d->vqs[i]->mutex);
1083 		if (!ok)
1084 			return false;
1085 	}
1086 	return true;
1087 }
1088 
1089 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
1090 			  struct iovec iov[], int iov_size, int access);
1091 
1092 static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
1093 			      const void *from, unsigned size)
1094 {
1095 	int ret;
1096 
1097 	if (!vq->iotlb)
1098 		return __copy_to_user(to, from, size);
1099 	else {
1100 		/* This function should be called after iotlb
1101 		 * prefetch, which means we're sure that all vq
1102 		 * could be access through iotlb. So -EAGAIN should
1103 		 * not happen in this case.
1104 		 */
1105 		struct iov_iter t;
1106 		void __user *uaddr = vhost_vq_meta_fetch(vq,
1107 				     (u64)(uintptr_t)to, size,
1108 				     VHOST_ADDR_USED);
1109 
1110 		if (uaddr)
1111 			return __copy_to_user(uaddr, from, size);
1112 
1113 		ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
1114 				     ARRAY_SIZE(vq->iotlb_iov),
1115 				     VHOST_ACCESS_WO);
1116 		if (ret < 0)
1117 			goto out;
1118 		iov_iter_init(&t, ITER_DEST, vq->iotlb_iov, ret, size);
1119 		ret = copy_to_iter(from, size, &t);
1120 		if (ret == size)
1121 			ret = 0;
1122 	}
1123 out:
1124 	return ret;
1125 }
1126 
1127 static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
1128 				void __user *from, unsigned size)
1129 {
1130 	int ret;
1131 
1132 	if (!vq->iotlb)
1133 		return __copy_from_user(to, from, size);
1134 	else {
1135 		/* This function should be called after iotlb
1136 		 * prefetch, which means we're sure that vq
1137 		 * could be access through iotlb. So -EAGAIN should
1138 		 * not happen in this case.
1139 		 */
1140 		void __user *uaddr = vhost_vq_meta_fetch(vq,
1141 				     (u64)(uintptr_t)from, size,
1142 				     VHOST_ADDR_DESC);
1143 		struct iov_iter f;
1144 
1145 		if (uaddr)
1146 			return __copy_from_user(to, uaddr, size);
1147 
1148 		ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
1149 				     ARRAY_SIZE(vq->iotlb_iov),
1150 				     VHOST_ACCESS_RO);
1151 		if (ret < 0) {
1152 			vq_err(vq, "IOTLB translation failure: uaddr "
1153 			       "%p size 0x%llx\n", from,
1154 			       (unsigned long long) size);
1155 			goto out;
1156 		}
1157 		iov_iter_init(&f, ITER_SOURCE, vq->iotlb_iov, ret, size);
1158 		ret = copy_from_iter(to, size, &f);
1159 		if (ret == size)
1160 			ret = 0;
1161 	}
1162 
1163 out:
1164 	return ret;
1165 }
1166 
1167 static void __user *__vhost_get_user_slow(struct vhost_virtqueue *vq,
1168 					  void __user *addr, unsigned int size,
1169 					  int type)
1170 {
1171 	int ret;
1172 
1173 	ret = translate_desc(vq, (u64)(uintptr_t)addr, size, vq->iotlb_iov,
1174 			     ARRAY_SIZE(vq->iotlb_iov),
1175 			     VHOST_ACCESS_RO);
1176 	if (ret < 0) {
1177 		vq_err(vq, "IOTLB translation failure: uaddr "
1178 			"%p size 0x%llx\n", addr,
1179 			(unsigned long long) size);
1180 		return NULL;
1181 	}
1182 
1183 	if (ret != 1 || vq->iotlb_iov[0].iov_len != size) {
1184 		vq_err(vq, "Non atomic userspace memory access: uaddr "
1185 			"%p size 0x%llx\n", addr,
1186 			(unsigned long long) size);
1187 		return NULL;
1188 	}
1189 
1190 	return vq->iotlb_iov[0].iov_base;
1191 }
1192 
1193 /* This function should be called after iotlb
1194  * prefetch, which means we're sure that vq
1195  * could be access through iotlb. So -EAGAIN should
1196  * not happen in this case.
1197  */
1198 static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
1199 					    void __user *addr, unsigned int size,
1200 					    int type)
1201 {
1202 	void __user *uaddr = vhost_vq_meta_fetch(vq,
1203 			     (u64)(uintptr_t)addr, size, type);
1204 	if (uaddr)
1205 		return uaddr;
1206 
1207 	return __vhost_get_user_slow(vq, addr, size, type);
1208 }
1209 
1210 #define vhost_put_user(vq, x, ptr)		\
1211 ({ \
1212 	int ret; \
1213 	if (!vq->iotlb) { \
1214 		ret = __put_user(x, ptr); \
1215 	} else { \
1216 		__typeof__(ptr) to = \
1217 			(__typeof__(ptr)) __vhost_get_user(vq, ptr,	\
1218 					  sizeof(*ptr), VHOST_ADDR_USED); \
1219 		if (to != NULL) \
1220 			ret = __put_user(x, to); \
1221 		else \
1222 			ret = -EFAULT;	\
1223 	} \
1224 	ret; \
1225 })
1226 
1227 static inline int vhost_put_avail_event(struct vhost_virtqueue *vq)
1228 {
1229 	return vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
1230 			      vhost_avail_event(vq));
1231 }
1232 
1233 static inline int vhost_put_used(struct vhost_virtqueue *vq,
1234 				 struct vring_used_elem *head, int idx,
1235 				 int count)
1236 {
1237 	return vhost_copy_to_user(vq, vq->used->ring + idx, head,
1238 				  count * sizeof(*head));
1239 }
1240 
1241 static inline int vhost_put_used_flags(struct vhost_virtqueue *vq)
1242 
1243 {
1244 	return vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
1245 			      &vq->used->flags);
1246 }
1247 
1248 static inline int vhost_put_used_idx(struct vhost_virtqueue *vq)
1249 
1250 {
1251 	return vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
1252 			      &vq->used->idx);
1253 }
1254 
1255 #define vhost_get_user(vq, x, ptr, type)		\
1256 ({ \
1257 	int ret; \
1258 	if (!vq->iotlb) { \
1259 		ret = __get_user(x, ptr); \
1260 	} else { \
1261 		__typeof__(ptr) from = \
1262 			(__typeof__(ptr)) __vhost_get_user(vq, ptr, \
1263 							   sizeof(*ptr), \
1264 							   type); \
1265 		if (from != NULL) \
1266 			ret = __get_user(x, from); \
1267 		else \
1268 			ret = -EFAULT; \
1269 	} \
1270 	ret; \
1271 })
1272 
1273 #define vhost_get_avail(vq, x, ptr) \
1274 	vhost_get_user(vq, x, ptr, VHOST_ADDR_AVAIL)
1275 
1276 #define vhost_get_used(vq, x, ptr) \
1277 	vhost_get_user(vq, x, ptr, VHOST_ADDR_USED)
1278 
1279 static void vhost_dev_lock_vqs(struct vhost_dev *d)
1280 {
1281 	int i = 0;
1282 	for (i = 0; i < d->nvqs; ++i)
1283 		mutex_lock_nested(&d->vqs[i]->mutex, i);
1284 }
1285 
1286 static void vhost_dev_unlock_vqs(struct vhost_dev *d)
1287 {
1288 	int i = 0;
1289 	for (i = 0; i < d->nvqs; ++i)
1290 		mutex_unlock(&d->vqs[i]->mutex);
1291 }
1292 
1293 static inline int vhost_get_avail_idx(struct vhost_virtqueue *vq,
1294 				      __virtio16 *idx)
1295 {
1296 	return vhost_get_avail(vq, *idx, &vq->avail->idx);
1297 }
1298 
1299 static inline int vhost_get_avail_head(struct vhost_virtqueue *vq,
1300 				       __virtio16 *head, int idx)
1301 {
1302 	return vhost_get_avail(vq, *head,
1303 			       &vq->avail->ring[idx & (vq->num - 1)]);
1304 }
1305 
1306 static inline int vhost_get_avail_flags(struct vhost_virtqueue *vq,
1307 					__virtio16 *flags)
1308 {
1309 	return vhost_get_avail(vq, *flags, &vq->avail->flags);
1310 }
1311 
1312 static inline int vhost_get_used_event(struct vhost_virtqueue *vq,
1313 				       __virtio16 *event)
1314 {
1315 	return vhost_get_avail(vq, *event, vhost_used_event(vq));
1316 }
1317 
1318 static inline int vhost_get_used_idx(struct vhost_virtqueue *vq,
1319 				     __virtio16 *idx)
1320 {
1321 	return vhost_get_used(vq, *idx, &vq->used->idx);
1322 }
1323 
1324 static inline int vhost_get_desc(struct vhost_virtqueue *vq,
1325 				 struct vring_desc *desc, int idx)
1326 {
1327 	return vhost_copy_from_user(vq, desc, vq->desc + idx, sizeof(*desc));
1328 }
1329 
1330 static void vhost_iotlb_notify_vq(struct vhost_dev *d,
1331 				  struct vhost_iotlb_msg *msg)
1332 {
1333 	struct vhost_msg_node *node, *n;
1334 
1335 	spin_lock(&d->iotlb_lock);
1336 
1337 	list_for_each_entry_safe(node, n, &d->pending_list, node) {
1338 		struct vhost_iotlb_msg *vq_msg = &node->msg.iotlb;
1339 		if (msg->iova <= vq_msg->iova &&
1340 		    msg->iova + msg->size - 1 >= vq_msg->iova &&
1341 		    vq_msg->type == VHOST_IOTLB_MISS) {
1342 			vhost_poll_queue(&node->vq->poll);
1343 			list_del(&node->node);
1344 			kfree(node);
1345 		}
1346 	}
1347 
1348 	spin_unlock(&d->iotlb_lock);
1349 }
1350 
1351 static bool umem_access_ok(u64 uaddr, u64 size, int access)
1352 {
1353 	unsigned long a = uaddr;
1354 
1355 	/* Make sure 64 bit math will not overflow. */
1356 	if (vhost_overflow(uaddr, size))
1357 		return false;
1358 
1359 	if ((access & VHOST_ACCESS_RO) &&
1360 	    !access_ok((void __user *)a, size))
1361 		return false;
1362 	if ((access & VHOST_ACCESS_WO) &&
1363 	    !access_ok((void __user *)a, size))
1364 		return false;
1365 	return true;
1366 }
1367 
1368 static int vhost_process_iotlb_msg(struct vhost_dev *dev, u32 asid,
1369 				   struct vhost_iotlb_msg *msg)
1370 {
1371 	int ret = 0;
1372 
1373 	if (asid != 0)
1374 		return -EINVAL;
1375 
1376 	mutex_lock(&dev->mutex);
1377 	vhost_dev_lock_vqs(dev);
1378 	switch (msg->type) {
1379 	case VHOST_IOTLB_UPDATE:
1380 		if (!dev->iotlb) {
1381 			ret = -EFAULT;
1382 			break;
1383 		}
1384 		if (!umem_access_ok(msg->uaddr, msg->size, msg->perm)) {
1385 			ret = -EFAULT;
1386 			break;
1387 		}
1388 		vhost_vq_meta_reset(dev);
1389 		if (vhost_iotlb_add_range(dev->iotlb, msg->iova,
1390 					  msg->iova + msg->size - 1,
1391 					  msg->uaddr, msg->perm)) {
1392 			ret = -ENOMEM;
1393 			break;
1394 		}
1395 		vhost_iotlb_notify_vq(dev, msg);
1396 		break;
1397 	case VHOST_IOTLB_INVALIDATE:
1398 		if (!dev->iotlb) {
1399 			ret = -EFAULT;
1400 			break;
1401 		}
1402 		vhost_vq_meta_reset(dev);
1403 		vhost_iotlb_del_range(dev->iotlb, msg->iova,
1404 				      msg->iova + msg->size - 1);
1405 		break;
1406 	default:
1407 		ret = -EINVAL;
1408 		break;
1409 	}
1410 
1411 	vhost_dev_unlock_vqs(dev);
1412 	mutex_unlock(&dev->mutex);
1413 
1414 	return ret;
1415 }
1416 ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
1417 			     struct iov_iter *from)
1418 {
1419 	struct vhost_iotlb_msg msg;
1420 	size_t offset;
1421 	int type, ret;
1422 	u32 asid = 0;
1423 
1424 	ret = copy_from_iter(&type, sizeof(type), from);
1425 	if (ret != sizeof(type)) {
1426 		ret = -EINVAL;
1427 		goto done;
1428 	}
1429 
1430 	switch (type) {
1431 	case VHOST_IOTLB_MSG:
1432 		/* There maybe a hole after type for V1 message type,
1433 		 * so skip it here.
1434 		 */
1435 		offset = offsetof(struct vhost_msg, iotlb) - sizeof(int);
1436 		break;
1437 	case VHOST_IOTLB_MSG_V2:
1438 		if (vhost_backend_has_feature(dev->vqs[0],
1439 					      VHOST_BACKEND_F_IOTLB_ASID)) {
1440 			ret = copy_from_iter(&asid, sizeof(asid), from);
1441 			if (ret != sizeof(asid)) {
1442 				ret = -EINVAL;
1443 				goto done;
1444 			}
1445 			offset = 0;
1446 		} else
1447 			offset = sizeof(__u32);
1448 		break;
1449 	default:
1450 		ret = -EINVAL;
1451 		goto done;
1452 	}
1453 
1454 	iov_iter_advance(from, offset);
1455 	ret = copy_from_iter(&msg, sizeof(msg), from);
1456 	if (ret != sizeof(msg)) {
1457 		ret = -EINVAL;
1458 		goto done;
1459 	}
1460 
1461 	if (msg.type == VHOST_IOTLB_UPDATE && msg.size == 0) {
1462 		ret = -EINVAL;
1463 		goto done;
1464 	}
1465 
1466 	if (dev->msg_handler)
1467 		ret = dev->msg_handler(dev, asid, &msg);
1468 	else
1469 		ret = vhost_process_iotlb_msg(dev, asid, &msg);
1470 	if (ret) {
1471 		ret = -EFAULT;
1472 		goto done;
1473 	}
1474 
1475 	ret = (type == VHOST_IOTLB_MSG) ? sizeof(struct vhost_msg) :
1476 	      sizeof(struct vhost_msg_v2);
1477 done:
1478 	return ret;
1479 }
1480 EXPORT_SYMBOL(vhost_chr_write_iter);
1481 
1482 __poll_t vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1483 			    poll_table *wait)
1484 {
1485 	__poll_t mask = 0;
1486 
1487 	poll_wait(file, &dev->wait, wait);
1488 
1489 	if (!list_empty(&dev->read_list))
1490 		mask |= EPOLLIN | EPOLLRDNORM;
1491 
1492 	return mask;
1493 }
1494 EXPORT_SYMBOL(vhost_chr_poll);
1495 
1496 ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1497 			    int noblock)
1498 {
1499 	DEFINE_WAIT(wait);
1500 	struct vhost_msg_node *node;
1501 	ssize_t ret = 0;
1502 	unsigned size = sizeof(struct vhost_msg);
1503 
1504 	if (iov_iter_count(to) < size)
1505 		return 0;
1506 
1507 	while (1) {
1508 		if (!noblock)
1509 			prepare_to_wait(&dev->wait, &wait,
1510 					TASK_INTERRUPTIBLE);
1511 
1512 		node = vhost_dequeue_msg(dev, &dev->read_list);
1513 		if (node)
1514 			break;
1515 		if (noblock) {
1516 			ret = -EAGAIN;
1517 			break;
1518 		}
1519 		if (signal_pending(current)) {
1520 			ret = -ERESTARTSYS;
1521 			break;
1522 		}
1523 		if (!dev->iotlb) {
1524 			ret = -EBADFD;
1525 			break;
1526 		}
1527 
1528 		schedule();
1529 	}
1530 
1531 	if (!noblock)
1532 		finish_wait(&dev->wait, &wait);
1533 
1534 	if (node) {
1535 		struct vhost_iotlb_msg *msg;
1536 		void *start = &node->msg;
1537 
1538 		switch (node->msg.type) {
1539 		case VHOST_IOTLB_MSG:
1540 			size = sizeof(node->msg);
1541 			msg = &node->msg.iotlb;
1542 			break;
1543 		case VHOST_IOTLB_MSG_V2:
1544 			size = sizeof(node->msg_v2);
1545 			msg = &node->msg_v2.iotlb;
1546 			break;
1547 		default:
1548 			BUG();
1549 			break;
1550 		}
1551 
1552 		ret = copy_to_iter(start, size, to);
1553 		if (ret != size || msg->type != VHOST_IOTLB_MISS) {
1554 			kfree(node);
1555 			return ret;
1556 		}
1557 		vhost_enqueue_msg(dev, &dev->pending_list, node);
1558 	}
1559 
1560 	return ret;
1561 }
1562 EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1563 
1564 static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1565 {
1566 	struct vhost_dev *dev = vq->dev;
1567 	struct vhost_msg_node *node;
1568 	struct vhost_iotlb_msg *msg;
1569 	bool v2 = vhost_backend_has_feature(vq, VHOST_BACKEND_F_IOTLB_MSG_V2);
1570 
1571 	node = vhost_new_msg(vq, v2 ? VHOST_IOTLB_MSG_V2 : VHOST_IOTLB_MSG);
1572 	if (!node)
1573 		return -ENOMEM;
1574 
1575 	if (v2) {
1576 		node->msg_v2.type = VHOST_IOTLB_MSG_V2;
1577 		msg = &node->msg_v2.iotlb;
1578 	} else {
1579 		msg = &node->msg.iotlb;
1580 	}
1581 
1582 	msg->type = VHOST_IOTLB_MISS;
1583 	msg->iova = iova;
1584 	msg->perm = access;
1585 
1586 	vhost_enqueue_msg(dev, &dev->read_list, node);
1587 
1588 	return 0;
1589 }
1590 
1591 static bool vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1592 			 vring_desc_t __user *desc,
1593 			 vring_avail_t __user *avail,
1594 			 vring_used_t __user *used)
1595 
1596 {
1597 	/* If an IOTLB device is present, the vring addresses are
1598 	 * GIOVAs. Access validation occurs at prefetch time. */
1599 	if (vq->iotlb)
1600 		return true;
1601 
1602 	return access_ok(desc, vhost_get_desc_size(vq, num)) &&
1603 	       access_ok(avail, vhost_get_avail_size(vq, num)) &&
1604 	       access_ok(used, vhost_get_used_size(vq, num));
1605 }
1606 
1607 static void vhost_vq_meta_update(struct vhost_virtqueue *vq,
1608 				 const struct vhost_iotlb_map *map,
1609 				 int type)
1610 {
1611 	int access = (type == VHOST_ADDR_USED) ?
1612 		     VHOST_ACCESS_WO : VHOST_ACCESS_RO;
1613 
1614 	if (likely(map->perm & access))
1615 		vq->meta_iotlb[type] = map;
1616 }
1617 
1618 static bool iotlb_access_ok(struct vhost_virtqueue *vq,
1619 			    int access, u64 addr, u64 len, int type)
1620 {
1621 	const struct vhost_iotlb_map *map;
1622 	struct vhost_iotlb *umem = vq->iotlb;
1623 	u64 s = 0, size, orig_addr = addr, last = addr + len - 1;
1624 
1625 	if (vhost_vq_meta_fetch(vq, addr, len, type))
1626 		return true;
1627 
1628 	while (len > s) {
1629 		map = vhost_iotlb_itree_first(umem, addr, last);
1630 		if (map == NULL || map->start > addr) {
1631 			vhost_iotlb_miss(vq, addr, access);
1632 			return false;
1633 		} else if (!(map->perm & access)) {
1634 			/* Report the possible access violation by
1635 			 * request another translation from userspace.
1636 			 */
1637 			return false;
1638 		}
1639 
1640 		size = map->size - addr + map->start;
1641 
1642 		if (orig_addr == addr && size >= len)
1643 			vhost_vq_meta_update(vq, map, type);
1644 
1645 		s += size;
1646 		addr += size;
1647 	}
1648 
1649 	return true;
1650 }
1651 
1652 int vq_meta_prefetch(struct vhost_virtqueue *vq)
1653 {
1654 	unsigned int num = vq->num;
1655 
1656 	if (!vq->iotlb)
1657 		return 1;
1658 
1659 	return iotlb_access_ok(vq, VHOST_MAP_RO, (u64)(uintptr_t)vq->desc,
1660 			       vhost_get_desc_size(vq, num), VHOST_ADDR_DESC) &&
1661 	       iotlb_access_ok(vq, VHOST_MAP_RO, (u64)(uintptr_t)vq->avail,
1662 			       vhost_get_avail_size(vq, num),
1663 			       VHOST_ADDR_AVAIL) &&
1664 	       iotlb_access_ok(vq, VHOST_MAP_WO, (u64)(uintptr_t)vq->used,
1665 			       vhost_get_used_size(vq, num), VHOST_ADDR_USED);
1666 }
1667 EXPORT_SYMBOL_GPL(vq_meta_prefetch);
1668 
1669 /* Can we log writes? */
1670 /* Caller should have device mutex but not vq mutex */
1671 bool vhost_log_access_ok(struct vhost_dev *dev)
1672 {
1673 	return memory_access_ok(dev, dev->umem, 1);
1674 }
1675 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1676 
1677 static bool vq_log_used_access_ok(struct vhost_virtqueue *vq,
1678 				  void __user *log_base,
1679 				  bool log_used,
1680 				  u64 log_addr)
1681 {
1682 	/* If an IOTLB device is present, log_addr is a GIOVA that
1683 	 * will never be logged by log_used(). */
1684 	if (vq->iotlb)
1685 		return true;
1686 
1687 	return !log_used || log_access_ok(log_base, log_addr,
1688 					  vhost_get_used_size(vq, vq->num));
1689 }
1690 
1691 /* Verify access for write logging. */
1692 /* Caller should have vq mutex and device mutex */
1693 static bool vq_log_access_ok(struct vhost_virtqueue *vq,
1694 			     void __user *log_base)
1695 {
1696 	return vq_memory_access_ok(log_base, vq->umem,
1697 				   vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1698 		vq_log_used_access_ok(vq, log_base, vq->log_used, vq->log_addr);
1699 }
1700 
1701 /* Can we start vq? */
1702 /* Caller should have vq mutex and device mutex */
1703 bool vhost_vq_access_ok(struct vhost_virtqueue *vq)
1704 {
1705 	if (!vq_log_access_ok(vq, vq->log_base))
1706 		return false;
1707 
1708 	return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used);
1709 }
1710 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1711 
1712 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1713 {
1714 	struct vhost_memory mem, *newmem;
1715 	struct vhost_memory_region *region;
1716 	struct vhost_iotlb *newumem, *oldumem;
1717 	unsigned long size = offsetof(struct vhost_memory, regions);
1718 	int i;
1719 
1720 	if (copy_from_user(&mem, m, size))
1721 		return -EFAULT;
1722 	if (mem.padding)
1723 		return -EOPNOTSUPP;
1724 	if (mem.nregions > max_mem_regions)
1725 		return -E2BIG;
1726 	newmem = kvzalloc(struct_size(newmem, regions, mem.nregions),
1727 			GFP_KERNEL);
1728 	if (!newmem)
1729 		return -ENOMEM;
1730 
1731 	memcpy(newmem, &mem, size);
1732 	if (copy_from_user(newmem->regions, m->regions,
1733 			   flex_array_size(newmem, regions, mem.nregions))) {
1734 		kvfree(newmem);
1735 		return -EFAULT;
1736 	}
1737 
1738 	newumem = iotlb_alloc();
1739 	if (!newumem) {
1740 		kvfree(newmem);
1741 		return -ENOMEM;
1742 	}
1743 
1744 	for (region = newmem->regions;
1745 	     region < newmem->regions + mem.nregions;
1746 	     region++) {
1747 		if (vhost_iotlb_add_range(newumem,
1748 					  region->guest_phys_addr,
1749 					  region->guest_phys_addr +
1750 					  region->memory_size - 1,
1751 					  region->userspace_addr,
1752 					  VHOST_MAP_RW))
1753 			goto err;
1754 	}
1755 
1756 	if (!memory_access_ok(d, newumem, 0))
1757 		goto err;
1758 
1759 	oldumem = d->umem;
1760 	d->umem = newumem;
1761 
1762 	/* All memory accesses are done under some VQ mutex. */
1763 	for (i = 0; i < d->nvqs; ++i) {
1764 		mutex_lock(&d->vqs[i]->mutex);
1765 		d->vqs[i]->umem = newumem;
1766 		mutex_unlock(&d->vqs[i]->mutex);
1767 	}
1768 
1769 	kvfree(newmem);
1770 	vhost_iotlb_free(oldumem);
1771 	return 0;
1772 
1773 err:
1774 	vhost_iotlb_free(newumem);
1775 	kvfree(newmem);
1776 	return -EFAULT;
1777 }
1778 
1779 static long vhost_vring_set_num(struct vhost_dev *d,
1780 				struct vhost_virtqueue *vq,
1781 				void __user *argp)
1782 {
1783 	struct vhost_vring_state s;
1784 
1785 	/* Resizing ring with an active backend?
1786 	 * You don't want to do that. */
1787 	if (vq->private_data)
1788 		return -EBUSY;
1789 
1790 	if (copy_from_user(&s, argp, sizeof s))
1791 		return -EFAULT;
1792 
1793 	if (!s.num || s.num > 0xffff || (s.num & (s.num - 1)))
1794 		return -EINVAL;
1795 	vq->num = s.num;
1796 
1797 	return 0;
1798 }
1799 
1800 static long vhost_vring_set_addr(struct vhost_dev *d,
1801 				 struct vhost_virtqueue *vq,
1802 				 void __user *argp)
1803 {
1804 	struct vhost_vring_addr a;
1805 
1806 	if (copy_from_user(&a, argp, sizeof a))
1807 		return -EFAULT;
1808 	if (a.flags & ~(0x1 << VHOST_VRING_F_LOG))
1809 		return -EOPNOTSUPP;
1810 
1811 	/* For 32bit, verify that the top 32bits of the user
1812 	   data are set to zero. */
1813 	if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1814 	    (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1815 	    (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr)
1816 		return -EFAULT;
1817 
1818 	/* Make sure it's safe to cast pointers to vring types. */
1819 	BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1820 	BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1821 	if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1822 	    (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1823 	    (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1)))
1824 		return -EINVAL;
1825 
1826 	/* We only verify access here if backend is configured.
1827 	 * If it is not, we don't as size might not have been setup.
1828 	 * We will verify when backend is configured. */
1829 	if (vq->private_data) {
1830 		if (!vq_access_ok(vq, vq->num,
1831 			(void __user *)(unsigned long)a.desc_user_addr,
1832 			(void __user *)(unsigned long)a.avail_user_addr,
1833 			(void __user *)(unsigned long)a.used_user_addr))
1834 			return -EINVAL;
1835 
1836 		/* Also validate log access for used ring if enabled. */
1837 		if (!vq_log_used_access_ok(vq, vq->log_base,
1838 				a.flags & (0x1 << VHOST_VRING_F_LOG),
1839 				a.log_guest_addr))
1840 			return -EINVAL;
1841 	}
1842 
1843 	vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1844 	vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1845 	vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1846 	vq->log_addr = a.log_guest_addr;
1847 	vq->used = (void __user *)(unsigned long)a.used_user_addr;
1848 
1849 	return 0;
1850 }
1851 
1852 static long vhost_vring_set_num_addr(struct vhost_dev *d,
1853 				     struct vhost_virtqueue *vq,
1854 				     unsigned int ioctl,
1855 				     void __user *argp)
1856 {
1857 	long r;
1858 
1859 	mutex_lock(&vq->mutex);
1860 
1861 	switch (ioctl) {
1862 	case VHOST_SET_VRING_NUM:
1863 		r = vhost_vring_set_num(d, vq, argp);
1864 		break;
1865 	case VHOST_SET_VRING_ADDR:
1866 		r = vhost_vring_set_addr(d, vq, argp);
1867 		break;
1868 	default:
1869 		BUG();
1870 	}
1871 
1872 	mutex_unlock(&vq->mutex);
1873 
1874 	return r;
1875 }
1876 long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1877 {
1878 	struct file *eventfp, *filep = NULL;
1879 	bool pollstart = false, pollstop = false;
1880 	struct eventfd_ctx *ctx = NULL;
1881 	struct vhost_virtqueue *vq;
1882 	struct vhost_vring_state s;
1883 	struct vhost_vring_file f;
1884 	u32 idx;
1885 	long r;
1886 
1887 	r = vhost_get_vq_from_user(d, argp, &vq, &idx);
1888 	if (r < 0)
1889 		return r;
1890 
1891 	if (ioctl == VHOST_SET_VRING_NUM ||
1892 	    ioctl == VHOST_SET_VRING_ADDR) {
1893 		return vhost_vring_set_num_addr(d, vq, ioctl, argp);
1894 	}
1895 
1896 	mutex_lock(&vq->mutex);
1897 
1898 	switch (ioctl) {
1899 	case VHOST_SET_VRING_BASE:
1900 		/* Moving base with an active backend?
1901 		 * You don't want to do that. */
1902 		if (vq->private_data) {
1903 			r = -EBUSY;
1904 			break;
1905 		}
1906 		if (copy_from_user(&s, argp, sizeof s)) {
1907 			r = -EFAULT;
1908 			break;
1909 		}
1910 		if (vhost_has_feature(vq, VIRTIO_F_RING_PACKED)) {
1911 			vq->last_avail_idx = s.num & 0xffff;
1912 			vq->last_used_idx = (s.num >> 16) & 0xffff;
1913 		} else {
1914 			if (s.num > 0xffff) {
1915 				r = -EINVAL;
1916 				break;
1917 			}
1918 			vq->last_avail_idx = s.num;
1919 		}
1920 		/* Forget the cached index value. */
1921 		vq->avail_idx = vq->last_avail_idx;
1922 		break;
1923 	case VHOST_GET_VRING_BASE:
1924 		s.index = idx;
1925 		if (vhost_has_feature(vq, VIRTIO_F_RING_PACKED))
1926 			s.num = (u32)vq->last_avail_idx | ((u32)vq->last_used_idx << 16);
1927 		else
1928 			s.num = vq->last_avail_idx;
1929 		if (copy_to_user(argp, &s, sizeof s))
1930 			r = -EFAULT;
1931 		break;
1932 	case VHOST_SET_VRING_KICK:
1933 		if (copy_from_user(&f, argp, sizeof f)) {
1934 			r = -EFAULT;
1935 			break;
1936 		}
1937 		eventfp = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_fget(f.fd);
1938 		if (IS_ERR(eventfp)) {
1939 			r = PTR_ERR(eventfp);
1940 			break;
1941 		}
1942 		if (eventfp != vq->kick) {
1943 			pollstop = (filep = vq->kick) != NULL;
1944 			pollstart = (vq->kick = eventfp) != NULL;
1945 		} else
1946 			filep = eventfp;
1947 		break;
1948 	case VHOST_SET_VRING_CALL:
1949 		if (copy_from_user(&f, argp, sizeof f)) {
1950 			r = -EFAULT;
1951 			break;
1952 		}
1953 		ctx = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(f.fd);
1954 		if (IS_ERR(ctx)) {
1955 			r = PTR_ERR(ctx);
1956 			break;
1957 		}
1958 
1959 		swap(ctx, vq->call_ctx.ctx);
1960 		break;
1961 	case VHOST_SET_VRING_ERR:
1962 		if (copy_from_user(&f, argp, sizeof f)) {
1963 			r = -EFAULT;
1964 			break;
1965 		}
1966 		ctx = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(f.fd);
1967 		if (IS_ERR(ctx)) {
1968 			r = PTR_ERR(ctx);
1969 			break;
1970 		}
1971 		swap(ctx, vq->error_ctx);
1972 		break;
1973 	case VHOST_SET_VRING_ENDIAN:
1974 		r = vhost_set_vring_endian(vq, argp);
1975 		break;
1976 	case VHOST_GET_VRING_ENDIAN:
1977 		r = vhost_get_vring_endian(vq, idx, argp);
1978 		break;
1979 	case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
1980 		if (copy_from_user(&s, argp, sizeof(s))) {
1981 			r = -EFAULT;
1982 			break;
1983 		}
1984 		vq->busyloop_timeout = s.num;
1985 		break;
1986 	case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
1987 		s.index = idx;
1988 		s.num = vq->busyloop_timeout;
1989 		if (copy_to_user(argp, &s, sizeof(s)))
1990 			r = -EFAULT;
1991 		break;
1992 	default:
1993 		r = -ENOIOCTLCMD;
1994 	}
1995 
1996 	if (pollstop && vq->handle_kick)
1997 		vhost_poll_stop(&vq->poll);
1998 
1999 	if (!IS_ERR_OR_NULL(ctx))
2000 		eventfd_ctx_put(ctx);
2001 	if (filep)
2002 		fput(filep);
2003 
2004 	if (pollstart && vq->handle_kick)
2005 		r = vhost_poll_start(&vq->poll, vq->kick);
2006 
2007 	mutex_unlock(&vq->mutex);
2008 
2009 	if (pollstop && vq->handle_kick)
2010 		vhost_dev_flush(vq->poll.dev);
2011 	return r;
2012 }
2013 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
2014 
2015 int vhost_init_device_iotlb(struct vhost_dev *d)
2016 {
2017 	struct vhost_iotlb *niotlb, *oiotlb;
2018 	int i;
2019 
2020 	niotlb = iotlb_alloc();
2021 	if (!niotlb)
2022 		return -ENOMEM;
2023 
2024 	oiotlb = d->iotlb;
2025 	d->iotlb = niotlb;
2026 
2027 	for (i = 0; i < d->nvqs; ++i) {
2028 		struct vhost_virtqueue *vq = d->vqs[i];
2029 
2030 		mutex_lock(&vq->mutex);
2031 		vq->iotlb = niotlb;
2032 		__vhost_vq_meta_reset(vq);
2033 		mutex_unlock(&vq->mutex);
2034 	}
2035 
2036 	vhost_iotlb_free(oiotlb);
2037 
2038 	return 0;
2039 }
2040 EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
2041 
2042 /* Caller must have device mutex */
2043 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
2044 {
2045 	struct eventfd_ctx *ctx;
2046 	u64 p;
2047 	long r;
2048 	int i, fd;
2049 
2050 	/* If you are not the owner, you can become one */
2051 	if (ioctl == VHOST_SET_OWNER) {
2052 		r = vhost_dev_set_owner(d);
2053 		goto done;
2054 	}
2055 
2056 	/* You must be the owner to do anything else */
2057 	r = vhost_dev_check_owner(d);
2058 	if (r)
2059 		goto done;
2060 
2061 	switch (ioctl) {
2062 	case VHOST_SET_MEM_TABLE:
2063 		r = vhost_set_memory(d, argp);
2064 		break;
2065 	case VHOST_SET_LOG_BASE:
2066 		if (copy_from_user(&p, argp, sizeof p)) {
2067 			r = -EFAULT;
2068 			break;
2069 		}
2070 		if ((u64)(unsigned long)p != p) {
2071 			r = -EFAULT;
2072 			break;
2073 		}
2074 		for (i = 0; i < d->nvqs; ++i) {
2075 			struct vhost_virtqueue *vq;
2076 			void __user *base = (void __user *)(unsigned long)p;
2077 			vq = d->vqs[i];
2078 			mutex_lock(&vq->mutex);
2079 			/* If ring is inactive, will check when it's enabled. */
2080 			if (vq->private_data && !vq_log_access_ok(vq, base))
2081 				r = -EFAULT;
2082 			else
2083 				vq->log_base = base;
2084 			mutex_unlock(&vq->mutex);
2085 		}
2086 		break;
2087 	case VHOST_SET_LOG_FD:
2088 		r = get_user(fd, (int __user *)argp);
2089 		if (r < 0)
2090 			break;
2091 		ctx = fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(fd);
2092 		if (IS_ERR(ctx)) {
2093 			r = PTR_ERR(ctx);
2094 			break;
2095 		}
2096 		swap(ctx, d->log_ctx);
2097 		for (i = 0; i < d->nvqs; ++i) {
2098 			mutex_lock(&d->vqs[i]->mutex);
2099 			d->vqs[i]->log_ctx = d->log_ctx;
2100 			mutex_unlock(&d->vqs[i]->mutex);
2101 		}
2102 		if (ctx)
2103 			eventfd_ctx_put(ctx);
2104 		break;
2105 	default:
2106 		r = -ENOIOCTLCMD;
2107 		break;
2108 	}
2109 done:
2110 	return r;
2111 }
2112 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
2113 
2114 /* TODO: This is really inefficient.  We need something like get_user()
2115  * (instruction directly accesses the data, with an exception table entry
2116  * returning -EFAULT). See Documentation/arch/x86/exception-tables.rst.
2117  */
2118 static int set_bit_to_user(int nr, void __user *addr)
2119 {
2120 	unsigned long log = (unsigned long)addr;
2121 	struct page *page;
2122 	void *base;
2123 	int bit = nr + (log % PAGE_SIZE) * 8;
2124 	int r;
2125 
2126 	r = pin_user_pages_fast(log, 1, FOLL_WRITE, &page);
2127 	if (r < 0)
2128 		return r;
2129 	BUG_ON(r != 1);
2130 	base = kmap_atomic(page);
2131 	set_bit(bit, base);
2132 	kunmap_atomic(base);
2133 	unpin_user_pages_dirty_lock(&page, 1, true);
2134 	return 0;
2135 }
2136 
2137 static int log_write(void __user *log_base,
2138 		     u64 write_address, u64 write_length)
2139 {
2140 	u64 write_page = write_address / VHOST_PAGE_SIZE;
2141 	int r;
2142 
2143 	if (!write_length)
2144 		return 0;
2145 	write_length += write_address % VHOST_PAGE_SIZE;
2146 	for (;;) {
2147 		u64 base = (u64)(unsigned long)log_base;
2148 		u64 log = base + write_page / 8;
2149 		int bit = write_page % 8;
2150 		if ((u64)(unsigned long)log != log)
2151 			return -EFAULT;
2152 		r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
2153 		if (r < 0)
2154 			return r;
2155 		if (write_length <= VHOST_PAGE_SIZE)
2156 			break;
2157 		write_length -= VHOST_PAGE_SIZE;
2158 		write_page += 1;
2159 	}
2160 	return r;
2161 }
2162 
2163 static int log_write_hva(struct vhost_virtqueue *vq, u64 hva, u64 len)
2164 {
2165 	struct vhost_iotlb *umem = vq->umem;
2166 	struct vhost_iotlb_map *u;
2167 	u64 start, end, l, min;
2168 	int r;
2169 	bool hit = false;
2170 
2171 	while (len) {
2172 		min = len;
2173 		/* More than one GPAs can be mapped into a single HVA. So
2174 		 * iterate all possible umems here to be safe.
2175 		 */
2176 		list_for_each_entry(u, &umem->list, link) {
2177 			if (u->addr > hva - 1 + len ||
2178 			    u->addr - 1 + u->size < hva)
2179 				continue;
2180 			start = max(u->addr, hva);
2181 			end = min(u->addr - 1 + u->size, hva - 1 + len);
2182 			l = end - start + 1;
2183 			r = log_write(vq->log_base,
2184 				      u->start + start - u->addr,
2185 				      l);
2186 			if (r < 0)
2187 				return r;
2188 			hit = true;
2189 			min = min(l, min);
2190 		}
2191 
2192 		if (!hit)
2193 			return -EFAULT;
2194 
2195 		len -= min;
2196 		hva += min;
2197 	}
2198 
2199 	return 0;
2200 }
2201 
2202 static int log_used(struct vhost_virtqueue *vq, u64 used_offset, u64 len)
2203 {
2204 	struct iovec *iov = vq->log_iov;
2205 	int i, ret;
2206 
2207 	if (!vq->iotlb)
2208 		return log_write(vq->log_base, vq->log_addr + used_offset, len);
2209 
2210 	ret = translate_desc(vq, (uintptr_t)vq->used + used_offset,
2211 			     len, iov, 64, VHOST_ACCESS_WO);
2212 	if (ret < 0)
2213 		return ret;
2214 
2215 	for (i = 0; i < ret; i++) {
2216 		ret = log_write_hva(vq,	(uintptr_t)iov[i].iov_base,
2217 				    iov[i].iov_len);
2218 		if (ret)
2219 			return ret;
2220 	}
2221 
2222 	return 0;
2223 }
2224 
2225 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
2226 		    unsigned int log_num, u64 len, struct iovec *iov, int count)
2227 {
2228 	int i, r;
2229 
2230 	/* Make sure data written is seen before log. */
2231 	smp_wmb();
2232 
2233 	if (vq->iotlb) {
2234 		for (i = 0; i < count; i++) {
2235 			r = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
2236 					  iov[i].iov_len);
2237 			if (r < 0)
2238 				return r;
2239 		}
2240 		return 0;
2241 	}
2242 
2243 	for (i = 0; i < log_num; ++i) {
2244 		u64 l = min(log[i].len, len);
2245 		r = log_write(vq->log_base, log[i].addr, l);
2246 		if (r < 0)
2247 			return r;
2248 		len -= l;
2249 		if (!len) {
2250 			if (vq->log_ctx)
2251 				eventfd_signal(vq->log_ctx, 1);
2252 			return 0;
2253 		}
2254 	}
2255 	/* Length written exceeds what we have stored. This is a bug. */
2256 	BUG();
2257 	return 0;
2258 }
2259 EXPORT_SYMBOL_GPL(vhost_log_write);
2260 
2261 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
2262 {
2263 	void __user *used;
2264 	if (vhost_put_used_flags(vq))
2265 		return -EFAULT;
2266 	if (unlikely(vq->log_used)) {
2267 		/* Make sure the flag is seen before log. */
2268 		smp_wmb();
2269 		/* Log used flag write. */
2270 		used = &vq->used->flags;
2271 		log_used(vq, (used - (void __user *)vq->used),
2272 			 sizeof vq->used->flags);
2273 		if (vq->log_ctx)
2274 			eventfd_signal(vq->log_ctx, 1);
2275 	}
2276 	return 0;
2277 }
2278 
2279 static int vhost_update_avail_event(struct vhost_virtqueue *vq)
2280 {
2281 	if (vhost_put_avail_event(vq))
2282 		return -EFAULT;
2283 	if (unlikely(vq->log_used)) {
2284 		void __user *used;
2285 		/* Make sure the event is seen before log. */
2286 		smp_wmb();
2287 		/* Log avail event write */
2288 		used = vhost_avail_event(vq);
2289 		log_used(vq, (used - (void __user *)vq->used),
2290 			 sizeof *vhost_avail_event(vq));
2291 		if (vq->log_ctx)
2292 			eventfd_signal(vq->log_ctx, 1);
2293 	}
2294 	return 0;
2295 }
2296 
2297 int vhost_vq_init_access(struct vhost_virtqueue *vq)
2298 {
2299 	__virtio16 last_used_idx;
2300 	int r;
2301 	bool is_le = vq->is_le;
2302 
2303 	if (!vq->private_data)
2304 		return 0;
2305 
2306 	vhost_init_is_le(vq);
2307 
2308 	r = vhost_update_used_flags(vq);
2309 	if (r)
2310 		goto err;
2311 	vq->signalled_used_valid = false;
2312 	if (!vq->iotlb &&
2313 	    !access_ok(&vq->used->idx, sizeof vq->used->idx)) {
2314 		r = -EFAULT;
2315 		goto err;
2316 	}
2317 	r = vhost_get_used_idx(vq, &last_used_idx);
2318 	if (r) {
2319 		vq_err(vq, "Can't access used idx at %p\n",
2320 		       &vq->used->idx);
2321 		goto err;
2322 	}
2323 	vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
2324 	return 0;
2325 
2326 err:
2327 	vq->is_le = is_le;
2328 	return r;
2329 }
2330 EXPORT_SYMBOL_GPL(vhost_vq_init_access);
2331 
2332 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
2333 			  struct iovec iov[], int iov_size, int access)
2334 {
2335 	const struct vhost_iotlb_map *map;
2336 	struct vhost_dev *dev = vq->dev;
2337 	struct vhost_iotlb *umem = dev->iotlb ? dev->iotlb : dev->umem;
2338 	struct iovec *_iov;
2339 	u64 s = 0, last = addr + len - 1;
2340 	int ret = 0;
2341 
2342 	while ((u64)len > s) {
2343 		u64 size;
2344 		if (unlikely(ret >= iov_size)) {
2345 			ret = -ENOBUFS;
2346 			break;
2347 		}
2348 
2349 		map = vhost_iotlb_itree_first(umem, addr, last);
2350 		if (map == NULL || map->start > addr) {
2351 			if (umem != dev->iotlb) {
2352 				ret = -EFAULT;
2353 				break;
2354 			}
2355 			ret = -EAGAIN;
2356 			break;
2357 		} else if (!(map->perm & access)) {
2358 			ret = -EPERM;
2359 			break;
2360 		}
2361 
2362 		_iov = iov + ret;
2363 		size = map->size - addr + map->start;
2364 		_iov->iov_len = min((u64)len - s, size);
2365 		_iov->iov_base = (void __user *)(unsigned long)
2366 				 (map->addr + addr - map->start);
2367 		s += size;
2368 		addr += size;
2369 		++ret;
2370 	}
2371 
2372 	if (ret == -EAGAIN)
2373 		vhost_iotlb_miss(vq, addr, access);
2374 	return ret;
2375 }
2376 
2377 /* Each buffer in the virtqueues is actually a chain of descriptors.  This
2378  * function returns the next descriptor in the chain,
2379  * or -1U if we're at the end. */
2380 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
2381 {
2382 	unsigned int next;
2383 
2384 	/* If this descriptor says it doesn't chain, we're done. */
2385 	if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
2386 		return -1U;
2387 
2388 	/* Check they're not leading us off end of descriptors. */
2389 	next = vhost16_to_cpu(vq, READ_ONCE(desc->next));
2390 	return next;
2391 }
2392 
2393 static int get_indirect(struct vhost_virtqueue *vq,
2394 			struct iovec iov[], unsigned int iov_size,
2395 			unsigned int *out_num, unsigned int *in_num,
2396 			struct vhost_log *log, unsigned int *log_num,
2397 			struct vring_desc *indirect)
2398 {
2399 	struct vring_desc desc;
2400 	unsigned int i = 0, count, found = 0;
2401 	u32 len = vhost32_to_cpu(vq, indirect->len);
2402 	struct iov_iter from;
2403 	int ret, access;
2404 
2405 	/* Sanity check */
2406 	if (unlikely(len % sizeof desc)) {
2407 		vq_err(vq, "Invalid length in indirect descriptor: "
2408 		       "len 0x%llx not multiple of 0x%zx\n",
2409 		       (unsigned long long)len,
2410 		       sizeof desc);
2411 		return -EINVAL;
2412 	}
2413 
2414 	ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
2415 			     UIO_MAXIOV, VHOST_ACCESS_RO);
2416 	if (unlikely(ret < 0)) {
2417 		if (ret != -EAGAIN)
2418 			vq_err(vq, "Translation failure %d in indirect.\n", ret);
2419 		return ret;
2420 	}
2421 	iov_iter_init(&from, ITER_SOURCE, vq->indirect, ret, len);
2422 	count = len / sizeof desc;
2423 	/* Buffers are chained via a 16 bit next field, so
2424 	 * we can have at most 2^16 of these. */
2425 	if (unlikely(count > USHRT_MAX + 1)) {
2426 		vq_err(vq, "Indirect buffer length too big: %d\n",
2427 		       indirect->len);
2428 		return -E2BIG;
2429 	}
2430 
2431 	do {
2432 		unsigned iov_count = *in_num + *out_num;
2433 		if (unlikely(++found > count)) {
2434 			vq_err(vq, "Loop detected: last one at %u "
2435 			       "indirect size %u\n",
2436 			       i, count);
2437 			return -EINVAL;
2438 		}
2439 		if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
2440 			vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
2441 			       i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2442 			return -EINVAL;
2443 		}
2444 		if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
2445 			vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
2446 			       i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2447 			return -EINVAL;
2448 		}
2449 
2450 		if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2451 			access = VHOST_ACCESS_WO;
2452 		else
2453 			access = VHOST_ACCESS_RO;
2454 
2455 		ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2456 				     vhost32_to_cpu(vq, desc.len), iov + iov_count,
2457 				     iov_size - iov_count, access);
2458 		if (unlikely(ret < 0)) {
2459 			if (ret != -EAGAIN)
2460 				vq_err(vq, "Translation failure %d indirect idx %d\n",
2461 					ret, i);
2462 			return ret;
2463 		}
2464 		/* If this is an input descriptor, increment that count. */
2465 		if (access == VHOST_ACCESS_WO) {
2466 			*in_num += ret;
2467 			if (unlikely(log && ret)) {
2468 				log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2469 				log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2470 				++*log_num;
2471 			}
2472 		} else {
2473 			/* If it's an output descriptor, they're all supposed
2474 			 * to come before any input descriptors. */
2475 			if (unlikely(*in_num)) {
2476 				vq_err(vq, "Indirect descriptor "
2477 				       "has out after in: idx %d\n", i);
2478 				return -EINVAL;
2479 			}
2480 			*out_num += ret;
2481 		}
2482 	} while ((i = next_desc(vq, &desc)) != -1);
2483 	return 0;
2484 }
2485 
2486 /* This looks in the virtqueue and for the first available buffer, and converts
2487  * it to an iovec for convenient access.  Since descriptors consist of some
2488  * number of output then some number of input descriptors, it's actually two
2489  * iovecs, but we pack them into one and note how many of each there were.
2490  *
2491  * This function returns the descriptor number found, or vq->num (which is
2492  * never a valid descriptor number) if none was found.  A negative code is
2493  * returned on error. */
2494 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
2495 		      struct iovec iov[], unsigned int iov_size,
2496 		      unsigned int *out_num, unsigned int *in_num,
2497 		      struct vhost_log *log, unsigned int *log_num)
2498 {
2499 	struct vring_desc desc;
2500 	unsigned int i, head, found = 0;
2501 	u16 last_avail_idx;
2502 	__virtio16 avail_idx;
2503 	__virtio16 ring_head;
2504 	int ret, access;
2505 
2506 	/* Check it isn't doing very strange things with descriptor numbers. */
2507 	last_avail_idx = vq->last_avail_idx;
2508 
2509 	if (vq->avail_idx == vq->last_avail_idx) {
2510 		if (unlikely(vhost_get_avail_idx(vq, &avail_idx))) {
2511 			vq_err(vq, "Failed to access avail idx at %p\n",
2512 				&vq->avail->idx);
2513 			return -EFAULT;
2514 		}
2515 		vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2516 
2517 		if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
2518 			vq_err(vq, "Guest moved used index from %u to %u",
2519 				last_avail_idx, vq->avail_idx);
2520 			return -EFAULT;
2521 		}
2522 
2523 		/* If there's nothing new since last we looked, return
2524 		 * invalid.
2525 		 */
2526 		if (vq->avail_idx == last_avail_idx)
2527 			return vq->num;
2528 
2529 		/* Only get avail ring entries after they have been
2530 		 * exposed by guest.
2531 		 */
2532 		smp_rmb();
2533 	}
2534 
2535 	/* Grab the next descriptor number they're advertising, and increment
2536 	 * the index we've seen. */
2537 	if (unlikely(vhost_get_avail_head(vq, &ring_head, last_avail_idx))) {
2538 		vq_err(vq, "Failed to read head: idx %d address %p\n",
2539 		       last_avail_idx,
2540 		       &vq->avail->ring[last_avail_idx % vq->num]);
2541 		return -EFAULT;
2542 	}
2543 
2544 	head = vhost16_to_cpu(vq, ring_head);
2545 
2546 	/* If their number is silly, that's an error. */
2547 	if (unlikely(head >= vq->num)) {
2548 		vq_err(vq, "Guest says index %u > %u is available",
2549 		       head, vq->num);
2550 		return -EINVAL;
2551 	}
2552 
2553 	/* When we start there are none of either input nor output. */
2554 	*out_num = *in_num = 0;
2555 	if (unlikely(log))
2556 		*log_num = 0;
2557 
2558 	i = head;
2559 	do {
2560 		unsigned iov_count = *in_num + *out_num;
2561 		if (unlikely(i >= vq->num)) {
2562 			vq_err(vq, "Desc index is %u > %u, head = %u",
2563 			       i, vq->num, head);
2564 			return -EINVAL;
2565 		}
2566 		if (unlikely(++found > vq->num)) {
2567 			vq_err(vq, "Loop detected: last one at %u "
2568 			       "vq size %u head %u\n",
2569 			       i, vq->num, head);
2570 			return -EINVAL;
2571 		}
2572 		ret = vhost_get_desc(vq, &desc, i);
2573 		if (unlikely(ret)) {
2574 			vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
2575 			       i, vq->desc + i);
2576 			return -EFAULT;
2577 		}
2578 		if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2579 			ret = get_indirect(vq, iov, iov_size,
2580 					   out_num, in_num,
2581 					   log, log_num, &desc);
2582 			if (unlikely(ret < 0)) {
2583 				if (ret != -EAGAIN)
2584 					vq_err(vq, "Failure detected "
2585 						"in indirect descriptor at idx %d\n", i);
2586 				return ret;
2587 			}
2588 			continue;
2589 		}
2590 
2591 		if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2592 			access = VHOST_ACCESS_WO;
2593 		else
2594 			access = VHOST_ACCESS_RO;
2595 		ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2596 				     vhost32_to_cpu(vq, desc.len), iov + iov_count,
2597 				     iov_size - iov_count, access);
2598 		if (unlikely(ret < 0)) {
2599 			if (ret != -EAGAIN)
2600 				vq_err(vq, "Translation failure %d descriptor idx %d\n",
2601 					ret, i);
2602 			return ret;
2603 		}
2604 		if (access == VHOST_ACCESS_WO) {
2605 			/* If this is an input descriptor,
2606 			 * increment that count. */
2607 			*in_num += ret;
2608 			if (unlikely(log && ret)) {
2609 				log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2610 				log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2611 				++*log_num;
2612 			}
2613 		} else {
2614 			/* If it's an output descriptor, they're all supposed
2615 			 * to come before any input descriptors. */
2616 			if (unlikely(*in_num)) {
2617 				vq_err(vq, "Descriptor has out after in: "
2618 				       "idx %d\n", i);
2619 				return -EINVAL;
2620 			}
2621 			*out_num += ret;
2622 		}
2623 	} while ((i = next_desc(vq, &desc)) != -1);
2624 
2625 	/* On success, increment avail index. */
2626 	vq->last_avail_idx++;
2627 
2628 	/* Assume notifications from guest are disabled at this point,
2629 	 * if they aren't we would need to update avail_event index. */
2630 	BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2631 	return head;
2632 }
2633 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2634 
2635 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2636 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2637 {
2638 	vq->last_avail_idx -= n;
2639 }
2640 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2641 
2642 /* After we've used one of their buffers, we tell them about it.  We'll then
2643  * want to notify the guest, using eventfd. */
2644 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2645 {
2646 	struct vring_used_elem heads = {
2647 		cpu_to_vhost32(vq, head),
2648 		cpu_to_vhost32(vq, len)
2649 	};
2650 
2651 	return vhost_add_used_n(vq, &heads, 1);
2652 }
2653 EXPORT_SYMBOL_GPL(vhost_add_used);
2654 
2655 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2656 			    struct vring_used_elem *heads,
2657 			    unsigned count)
2658 {
2659 	vring_used_elem_t __user *used;
2660 	u16 old, new;
2661 	int start;
2662 
2663 	start = vq->last_used_idx & (vq->num - 1);
2664 	used = vq->used->ring + start;
2665 	if (vhost_put_used(vq, heads, start, count)) {
2666 		vq_err(vq, "Failed to write used");
2667 		return -EFAULT;
2668 	}
2669 	if (unlikely(vq->log_used)) {
2670 		/* Make sure data is seen before log. */
2671 		smp_wmb();
2672 		/* Log used ring entry write. */
2673 		log_used(vq, ((void __user *)used - (void __user *)vq->used),
2674 			 count * sizeof *used);
2675 	}
2676 	old = vq->last_used_idx;
2677 	new = (vq->last_used_idx += count);
2678 	/* If the driver never bothers to signal in a very long while,
2679 	 * used index might wrap around. If that happens, invalidate
2680 	 * signalled_used index we stored. TODO: make sure driver
2681 	 * signals at least once in 2^16 and remove this. */
2682 	if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2683 		vq->signalled_used_valid = false;
2684 	return 0;
2685 }
2686 
2687 /* After we've used one of their buffers, we tell them about it.  We'll then
2688  * want to notify the guest, using eventfd. */
2689 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2690 		     unsigned count)
2691 {
2692 	int start, n, r;
2693 
2694 	start = vq->last_used_idx & (vq->num - 1);
2695 	n = vq->num - start;
2696 	if (n < count) {
2697 		r = __vhost_add_used_n(vq, heads, n);
2698 		if (r < 0)
2699 			return r;
2700 		heads += n;
2701 		count -= n;
2702 	}
2703 	r = __vhost_add_used_n(vq, heads, count);
2704 
2705 	/* Make sure buffer is written before we update index. */
2706 	smp_wmb();
2707 	if (vhost_put_used_idx(vq)) {
2708 		vq_err(vq, "Failed to increment used idx");
2709 		return -EFAULT;
2710 	}
2711 	if (unlikely(vq->log_used)) {
2712 		/* Make sure used idx is seen before log. */
2713 		smp_wmb();
2714 		/* Log used index update. */
2715 		log_used(vq, offsetof(struct vring_used, idx),
2716 			 sizeof vq->used->idx);
2717 		if (vq->log_ctx)
2718 			eventfd_signal(vq->log_ctx, 1);
2719 	}
2720 	return r;
2721 }
2722 EXPORT_SYMBOL_GPL(vhost_add_used_n);
2723 
2724 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2725 {
2726 	__u16 old, new;
2727 	__virtio16 event;
2728 	bool v;
2729 	/* Flush out used index updates. This is paired
2730 	 * with the barrier that the Guest executes when enabling
2731 	 * interrupts. */
2732 	smp_mb();
2733 
2734 	if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2735 	    unlikely(vq->avail_idx == vq->last_avail_idx))
2736 		return true;
2737 
2738 	if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2739 		__virtio16 flags;
2740 		if (vhost_get_avail_flags(vq, &flags)) {
2741 			vq_err(vq, "Failed to get flags");
2742 			return true;
2743 		}
2744 		return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2745 	}
2746 	old = vq->signalled_used;
2747 	v = vq->signalled_used_valid;
2748 	new = vq->signalled_used = vq->last_used_idx;
2749 	vq->signalled_used_valid = true;
2750 
2751 	if (unlikely(!v))
2752 		return true;
2753 
2754 	if (vhost_get_used_event(vq, &event)) {
2755 		vq_err(vq, "Failed to get used event idx");
2756 		return true;
2757 	}
2758 	return vring_need_event(vhost16_to_cpu(vq, event), new, old);
2759 }
2760 
2761 /* This actually signals the guest, using eventfd. */
2762 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2763 {
2764 	/* Signal the Guest tell them we used something up. */
2765 	if (vq->call_ctx.ctx && vhost_notify(dev, vq))
2766 		eventfd_signal(vq->call_ctx.ctx, 1);
2767 }
2768 EXPORT_SYMBOL_GPL(vhost_signal);
2769 
2770 /* And here's the combo meal deal.  Supersize me! */
2771 void vhost_add_used_and_signal(struct vhost_dev *dev,
2772 			       struct vhost_virtqueue *vq,
2773 			       unsigned int head, int len)
2774 {
2775 	vhost_add_used(vq, head, len);
2776 	vhost_signal(dev, vq);
2777 }
2778 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2779 
2780 /* multi-buffer version of vhost_add_used_and_signal */
2781 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2782 				 struct vhost_virtqueue *vq,
2783 				 struct vring_used_elem *heads, unsigned count)
2784 {
2785 	vhost_add_used_n(vq, heads, count);
2786 	vhost_signal(dev, vq);
2787 }
2788 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2789 
2790 /* return true if we're sure that avaiable ring is empty */
2791 bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2792 {
2793 	__virtio16 avail_idx;
2794 	int r;
2795 
2796 	if (vq->avail_idx != vq->last_avail_idx)
2797 		return false;
2798 
2799 	r = vhost_get_avail_idx(vq, &avail_idx);
2800 	if (unlikely(r))
2801 		return false;
2802 	vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2803 
2804 	return vq->avail_idx == vq->last_avail_idx;
2805 }
2806 EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2807 
2808 /* OK, now we need to know about added descriptors. */
2809 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2810 {
2811 	__virtio16 avail_idx;
2812 	int r;
2813 
2814 	if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2815 		return false;
2816 	vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2817 	if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2818 		r = vhost_update_used_flags(vq);
2819 		if (r) {
2820 			vq_err(vq, "Failed to enable notification at %p: %d\n",
2821 			       &vq->used->flags, r);
2822 			return false;
2823 		}
2824 	} else {
2825 		r = vhost_update_avail_event(vq);
2826 		if (r) {
2827 			vq_err(vq, "Failed to update avail event index at %p: %d\n",
2828 			       vhost_avail_event(vq), r);
2829 			return false;
2830 		}
2831 	}
2832 	/* They could have slipped one in as we were doing that: make
2833 	 * sure it's written, then check again. */
2834 	smp_mb();
2835 	r = vhost_get_avail_idx(vq, &avail_idx);
2836 	if (r) {
2837 		vq_err(vq, "Failed to check avail idx at %p: %d\n",
2838 		       &vq->avail->idx, r);
2839 		return false;
2840 	}
2841 	vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2842 
2843 	return vq->avail_idx != vq->last_avail_idx;
2844 }
2845 EXPORT_SYMBOL_GPL(vhost_enable_notify);
2846 
2847 /* We don't need to be notified again. */
2848 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2849 {
2850 	int r;
2851 
2852 	if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2853 		return;
2854 	vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2855 	if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2856 		r = vhost_update_used_flags(vq);
2857 		if (r)
2858 			vq_err(vq, "Failed to disable notification at %p: %d\n",
2859 			       &vq->used->flags, r);
2860 	}
2861 }
2862 EXPORT_SYMBOL_GPL(vhost_disable_notify);
2863 
2864 /* Create a new message. */
2865 struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2866 {
2867 	/* Make sure all padding within the structure is initialized. */
2868 	struct vhost_msg_node *node = kzalloc(sizeof(*node), GFP_KERNEL);
2869 	if (!node)
2870 		return NULL;
2871 
2872 	node->vq = vq;
2873 	node->msg.type = type;
2874 	return node;
2875 }
2876 EXPORT_SYMBOL_GPL(vhost_new_msg);
2877 
2878 void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2879 		       struct vhost_msg_node *node)
2880 {
2881 	spin_lock(&dev->iotlb_lock);
2882 	list_add_tail(&node->node, head);
2883 	spin_unlock(&dev->iotlb_lock);
2884 
2885 	wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
2886 }
2887 EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2888 
2889 struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2890 					 struct list_head *head)
2891 {
2892 	struct vhost_msg_node *node = NULL;
2893 
2894 	spin_lock(&dev->iotlb_lock);
2895 	if (!list_empty(head)) {
2896 		node = list_first_entry(head, struct vhost_msg_node,
2897 					node);
2898 		list_del(&node->node);
2899 	}
2900 	spin_unlock(&dev->iotlb_lock);
2901 
2902 	return node;
2903 }
2904 EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2905 
2906 void vhost_set_backend_features(struct vhost_dev *dev, u64 features)
2907 {
2908 	struct vhost_virtqueue *vq;
2909 	int i;
2910 
2911 	mutex_lock(&dev->mutex);
2912 	for (i = 0; i < dev->nvqs; ++i) {
2913 		vq = dev->vqs[i];
2914 		mutex_lock(&vq->mutex);
2915 		vq->acked_backend_features = features;
2916 		mutex_unlock(&vq->mutex);
2917 	}
2918 	mutex_unlock(&dev->mutex);
2919 }
2920 EXPORT_SYMBOL_GPL(vhost_set_backend_features);
2921 
2922 static int __init vhost_init(void)
2923 {
2924 	return 0;
2925 }
2926 
2927 static void __exit vhost_exit(void)
2928 {
2929 }
2930 
2931 module_init(vhost_init);
2932 module_exit(vhost_exit);
2933 
2934 MODULE_VERSION("0.0.1");
2935 MODULE_LICENSE("GPL v2");
2936 MODULE_AUTHOR("Michael S. Tsirkin");
2937 MODULE_DESCRIPTION("Host kernel accelerator for virtio");
2938