xref: /linux/drivers/vfio/vfio_main.c (revision 6863aaa88516292b885fdce5dd91925a00c3a3de)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * VFIO core
4  *
5  * Copyright (C) 2012 Red Hat, Inc.  All rights reserved.
6  *     Author: Alex Williamson <alex.williamson@redhat.com>
7  *
8  * Derived from original vfio:
9  * Copyright 2010 Cisco Systems, Inc.  All rights reserved.
10  * Author: Tom Lyon, pugs@cisco.com
11  */
12 
13 #include <linux/cdev.h>
14 #include <linux/compat.h>
15 #include <linux/device.h>
16 #include <linux/file.h>
17 #include <linux/anon_inodes.h>
18 #include <linux/fs.h>
19 #include <linux/idr.h>
20 #include <linux/iommu.h>
21 #include <linux/list.h>
22 #include <linux/miscdevice.h>
23 #include <linux/module.h>
24 #include <linux/mutex.h>
25 #include <linux/pci.h>
26 #include <linux/rwsem.h>
27 #include <linux/sched.h>
28 #include <linux/slab.h>
29 #include <linux/stat.h>
30 #include <linux/string.h>
31 #include <linux/uaccess.h>
32 #include <linux/vfio.h>
33 #include <linux/wait.h>
34 #include <linux/sched/signal.h>
35 #include <linux/pm_runtime.h>
36 #include <linux/interval_tree.h>
37 #include <linux/iova_bitmap.h>
38 #include "vfio.h"
39 
40 #define DRIVER_VERSION	"0.3"
41 #define DRIVER_AUTHOR	"Alex Williamson <alex.williamson@redhat.com>"
42 #define DRIVER_DESC	"VFIO - User Level meta-driver"
43 
44 static struct vfio {
45 	struct class			*class;
46 	struct list_head		group_list;
47 	struct mutex			group_lock; /* locks group_list */
48 	struct ida			group_ida;
49 	dev_t				group_devt;
50 	struct class			*device_class;
51 	struct ida			device_ida;
52 } vfio;
53 
54 static DEFINE_XARRAY(vfio_device_set_xa);
55 static const struct file_operations vfio_group_fops;
56 
57 int vfio_assign_device_set(struct vfio_device *device, void *set_id)
58 {
59 	unsigned long idx = (unsigned long)set_id;
60 	struct vfio_device_set *new_dev_set;
61 	struct vfio_device_set *dev_set;
62 
63 	if (WARN_ON(!set_id))
64 		return -EINVAL;
65 
66 	/*
67 	 * Atomically acquire a singleton object in the xarray for this set_id
68 	 */
69 	xa_lock(&vfio_device_set_xa);
70 	dev_set = xa_load(&vfio_device_set_xa, idx);
71 	if (dev_set)
72 		goto found_get_ref;
73 	xa_unlock(&vfio_device_set_xa);
74 
75 	new_dev_set = kzalloc(sizeof(*new_dev_set), GFP_KERNEL);
76 	if (!new_dev_set)
77 		return -ENOMEM;
78 	mutex_init(&new_dev_set->lock);
79 	INIT_LIST_HEAD(&new_dev_set->device_list);
80 	new_dev_set->set_id = set_id;
81 
82 	xa_lock(&vfio_device_set_xa);
83 	dev_set = __xa_cmpxchg(&vfio_device_set_xa, idx, NULL, new_dev_set,
84 			       GFP_KERNEL);
85 	if (!dev_set) {
86 		dev_set = new_dev_set;
87 		goto found_get_ref;
88 	}
89 
90 	kfree(new_dev_set);
91 	if (xa_is_err(dev_set)) {
92 		xa_unlock(&vfio_device_set_xa);
93 		return xa_err(dev_set);
94 	}
95 
96 found_get_ref:
97 	dev_set->device_count++;
98 	xa_unlock(&vfio_device_set_xa);
99 	mutex_lock(&dev_set->lock);
100 	device->dev_set = dev_set;
101 	list_add_tail(&device->dev_set_list, &dev_set->device_list);
102 	mutex_unlock(&dev_set->lock);
103 	return 0;
104 }
105 EXPORT_SYMBOL_GPL(vfio_assign_device_set);
106 
107 static void vfio_release_device_set(struct vfio_device *device)
108 {
109 	struct vfio_device_set *dev_set = device->dev_set;
110 
111 	if (!dev_set)
112 		return;
113 
114 	mutex_lock(&dev_set->lock);
115 	list_del(&device->dev_set_list);
116 	mutex_unlock(&dev_set->lock);
117 
118 	xa_lock(&vfio_device_set_xa);
119 	if (!--dev_set->device_count) {
120 		__xa_erase(&vfio_device_set_xa,
121 			   (unsigned long)dev_set->set_id);
122 		mutex_destroy(&dev_set->lock);
123 		kfree(dev_set);
124 	}
125 	xa_unlock(&vfio_device_set_xa);
126 }
127 
128 /*
129  * Group objects - create, release, get, put, search
130  */
131 static struct vfio_group *
132 __vfio_group_get_from_iommu(struct iommu_group *iommu_group)
133 {
134 	struct vfio_group *group;
135 
136 	/*
137 	 * group->iommu_group from the vfio.group_list cannot be NULL
138 	 * under the vfio.group_lock.
139 	 */
140 	list_for_each_entry(group, &vfio.group_list, vfio_next) {
141 		if (group->iommu_group == iommu_group) {
142 			refcount_inc(&group->drivers);
143 			return group;
144 		}
145 	}
146 	return NULL;
147 }
148 
149 static struct vfio_group *
150 vfio_group_get_from_iommu(struct iommu_group *iommu_group)
151 {
152 	struct vfio_group *group;
153 
154 	mutex_lock(&vfio.group_lock);
155 	group = __vfio_group_get_from_iommu(iommu_group);
156 	mutex_unlock(&vfio.group_lock);
157 	return group;
158 }
159 
160 static void vfio_group_release(struct device *dev)
161 {
162 	struct vfio_group *group = container_of(dev, struct vfio_group, dev);
163 
164 	mutex_destroy(&group->device_lock);
165 	mutex_destroy(&group->group_lock);
166 	WARN_ON(group->iommu_group);
167 	ida_free(&vfio.group_ida, MINOR(group->dev.devt));
168 	kfree(group);
169 }
170 
171 static struct vfio_group *vfio_group_alloc(struct iommu_group *iommu_group,
172 					   enum vfio_group_type type)
173 {
174 	struct vfio_group *group;
175 	int minor;
176 
177 	group = kzalloc(sizeof(*group), GFP_KERNEL);
178 	if (!group)
179 		return ERR_PTR(-ENOMEM);
180 
181 	minor = ida_alloc_max(&vfio.group_ida, MINORMASK, GFP_KERNEL);
182 	if (minor < 0) {
183 		kfree(group);
184 		return ERR_PTR(minor);
185 	}
186 
187 	device_initialize(&group->dev);
188 	group->dev.devt = MKDEV(MAJOR(vfio.group_devt), minor);
189 	group->dev.class = vfio.class;
190 	group->dev.release = vfio_group_release;
191 	cdev_init(&group->cdev, &vfio_group_fops);
192 	group->cdev.owner = THIS_MODULE;
193 
194 	refcount_set(&group->drivers, 1);
195 	mutex_init(&group->group_lock);
196 	INIT_LIST_HEAD(&group->device_list);
197 	mutex_init(&group->device_lock);
198 	group->iommu_group = iommu_group;
199 	/* put in vfio_group_release() */
200 	iommu_group_ref_get(iommu_group);
201 	group->type = type;
202 	BLOCKING_INIT_NOTIFIER_HEAD(&group->notifier);
203 
204 	return group;
205 }
206 
207 static struct vfio_group *vfio_create_group(struct iommu_group *iommu_group,
208 		enum vfio_group_type type)
209 {
210 	struct vfio_group *group;
211 	struct vfio_group *ret;
212 	int err;
213 
214 	group = vfio_group_alloc(iommu_group, type);
215 	if (IS_ERR(group))
216 		return group;
217 
218 	err = dev_set_name(&group->dev, "%s%d",
219 			   group->type == VFIO_NO_IOMMU ? "noiommu-" : "",
220 			   iommu_group_id(iommu_group));
221 	if (err) {
222 		ret = ERR_PTR(err);
223 		goto err_put;
224 	}
225 
226 	mutex_lock(&vfio.group_lock);
227 
228 	/* Did we race creating this group? */
229 	ret = __vfio_group_get_from_iommu(iommu_group);
230 	if (ret)
231 		goto err_unlock;
232 
233 	err = cdev_device_add(&group->cdev, &group->dev);
234 	if (err) {
235 		ret = ERR_PTR(err);
236 		goto err_unlock;
237 	}
238 
239 	list_add(&group->vfio_next, &vfio.group_list);
240 
241 	mutex_unlock(&vfio.group_lock);
242 	return group;
243 
244 err_unlock:
245 	mutex_unlock(&vfio.group_lock);
246 err_put:
247 	put_device(&group->dev);
248 	return ret;
249 }
250 
251 static void vfio_device_remove_group(struct vfio_device *device)
252 {
253 	struct vfio_group *group = device->group;
254 	struct iommu_group *iommu_group;
255 
256 	if (group->type == VFIO_NO_IOMMU || group->type == VFIO_EMULATED_IOMMU)
257 		iommu_group_remove_device(device->dev);
258 
259 	/* Pairs with vfio_create_group() / vfio_group_get_from_iommu() */
260 	if (!refcount_dec_and_mutex_lock(&group->drivers, &vfio.group_lock))
261 		return;
262 	list_del(&group->vfio_next);
263 
264 	/*
265 	 * We could concurrently probe another driver in the group that might
266 	 * race vfio_device_remove_group() with vfio_get_group(), so we have to
267 	 * ensure that the sysfs is all cleaned up under lock otherwise the
268 	 * cdev_device_add() will fail due to the name aready existing.
269 	 */
270 	cdev_device_del(&group->cdev, &group->dev);
271 
272 	mutex_lock(&group->group_lock);
273 	/*
274 	 * These data structures all have paired operations that can only be
275 	 * undone when the caller holds a live reference on the device. Since
276 	 * all pairs must be undone these WARN_ON's indicate some caller did not
277 	 * properly hold the group reference.
278 	 */
279 	WARN_ON(!list_empty(&group->device_list));
280 	WARN_ON(group->notifier.head);
281 
282 	/*
283 	 * Revoke all users of group->iommu_group. At this point we know there
284 	 * are no devices active because we are unplugging the last one. Setting
285 	 * iommu_group to NULL blocks all new users.
286 	 */
287 	if (group->container)
288 		vfio_group_detach_container(group);
289 	iommu_group = group->iommu_group;
290 	group->iommu_group = NULL;
291 	mutex_unlock(&group->group_lock);
292 	mutex_unlock(&vfio.group_lock);
293 
294 	iommu_group_put(iommu_group);
295 	put_device(&group->dev);
296 }
297 
298 /*
299  * Device objects - create, release, get, put, search
300  */
301 /* Device reference always implies a group reference */
302 static void vfio_device_put_registration(struct vfio_device *device)
303 {
304 	if (refcount_dec_and_test(&device->refcount))
305 		complete(&device->comp);
306 }
307 
308 static bool vfio_device_try_get_registration(struct vfio_device *device)
309 {
310 	return refcount_inc_not_zero(&device->refcount);
311 }
312 
313 static struct vfio_device *vfio_group_get_device(struct vfio_group *group,
314 						 struct device *dev)
315 {
316 	struct vfio_device *device;
317 
318 	mutex_lock(&group->device_lock);
319 	list_for_each_entry(device, &group->device_list, group_next) {
320 		if (device->dev == dev &&
321 		    vfio_device_try_get_registration(device)) {
322 			mutex_unlock(&group->device_lock);
323 			return device;
324 		}
325 	}
326 	mutex_unlock(&group->device_lock);
327 	return NULL;
328 }
329 
330 /*
331  * VFIO driver API
332  */
333 /* Release helper called by vfio_put_device() */
334 static void vfio_device_release(struct device *dev)
335 {
336 	struct vfio_device *device =
337 			container_of(dev, struct vfio_device, device);
338 
339 	vfio_release_device_set(device);
340 	ida_free(&vfio.device_ida, device->index);
341 
342 	/*
343 	 * kvfree() cannot be done here due to a life cycle mess in
344 	 * vfio-ccw. Before the ccw part is fixed all drivers are
345 	 * required to support @release and call vfio_free_device()
346 	 * from there.
347 	 */
348 	device->ops->release(device);
349 }
350 
351 /*
352  * Allocate and initialize vfio_device so it can be registered to vfio
353  * core.
354  *
355  * Drivers should use the wrapper vfio_alloc_device() for allocation.
356  * @size is the size of the structure to be allocated, including any
357  * private data used by the driver.
358  *
359  * Driver may provide an @init callback to cover device private data.
360  *
361  * Use vfio_put_device() to release the structure after success return.
362  */
363 struct vfio_device *_vfio_alloc_device(size_t size, struct device *dev,
364 				       const struct vfio_device_ops *ops)
365 {
366 	struct vfio_device *device;
367 	int ret;
368 
369 	if (WARN_ON(size < sizeof(struct vfio_device)))
370 		return ERR_PTR(-EINVAL);
371 
372 	device = kvzalloc(size, GFP_KERNEL);
373 	if (!device)
374 		return ERR_PTR(-ENOMEM);
375 
376 	ret = vfio_init_device(device, dev, ops);
377 	if (ret)
378 		goto out_free;
379 	return device;
380 
381 out_free:
382 	kvfree(device);
383 	return ERR_PTR(ret);
384 }
385 EXPORT_SYMBOL_GPL(_vfio_alloc_device);
386 
387 /*
388  * Initialize a vfio_device so it can be registered to vfio core.
389  *
390  * Only vfio-ccw driver should call this interface.
391  */
392 int vfio_init_device(struct vfio_device *device, struct device *dev,
393 		     const struct vfio_device_ops *ops)
394 {
395 	int ret;
396 
397 	ret = ida_alloc_max(&vfio.device_ida, MINORMASK, GFP_KERNEL);
398 	if (ret < 0) {
399 		dev_dbg(dev, "Error to alloc index\n");
400 		return ret;
401 	}
402 
403 	device->index = ret;
404 	init_completion(&device->comp);
405 	device->dev = dev;
406 	device->ops = ops;
407 
408 	if (ops->init) {
409 		ret = ops->init(device);
410 		if (ret)
411 			goto out_uninit;
412 	}
413 
414 	device_initialize(&device->device);
415 	device->device.release = vfio_device_release;
416 	device->device.class = vfio.device_class;
417 	device->device.parent = device->dev;
418 	return 0;
419 
420 out_uninit:
421 	vfio_release_device_set(device);
422 	ida_free(&vfio.device_ida, device->index);
423 	return ret;
424 }
425 EXPORT_SYMBOL_GPL(vfio_init_device);
426 
427 /*
428  * The helper called by driver @release callback to free the device
429  * structure. Drivers which don't have private data to clean can
430  * simply use this helper as its @release.
431  */
432 void vfio_free_device(struct vfio_device *device)
433 {
434 	kvfree(device);
435 }
436 EXPORT_SYMBOL_GPL(vfio_free_device);
437 
438 static struct vfio_group *vfio_noiommu_group_alloc(struct device *dev,
439 		enum vfio_group_type type)
440 {
441 	struct iommu_group *iommu_group;
442 	struct vfio_group *group;
443 	int ret;
444 
445 	iommu_group = iommu_group_alloc();
446 	if (IS_ERR(iommu_group))
447 		return ERR_CAST(iommu_group);
448 
449 	ret = iommu_group_set_name(iommu_group, "vfio-noiommu");
450 	if (ret)
451 		goto out_put_group;
452 	ret = iommu_group_add_device(iommu_group, dev);
453 	if (ret)
454 		goto out_put_group;
455 
456 	group = vfio_create_group(iommu_group, type);
457 	if (IS_ERR(group)) {
458 		ret = PTR_ERR(group);
459 		goto out_remove_device;
460 	}
461 	iommu_group_put(iommu_group);
462 	return group;
463 
464 out_remove_device:
465 	iommu_group_remove_device(dev);
466 out_put_group:
467 	iommu_group_put(iommu_group);
468 	return ERR_PTR(ret);
469 }
470 
471 static struct vfio_group *vfio_group_find_or_alloc(struct device *dev)
472 {
473 	struct iommu_group *iommu_group;
474 	struct vfio_group *group;
475 
476 	iommu_group = iommu_group_get(dev);
477 	if (!iommu_group && vfio_noiommu) {
478 		/*
479 		 * With noiommu enabled, create an IOMMU group for devices that
480 		 * don't already have one, implying no IOMMU hardware/driver
481 		 * exists.  Taint the kernel because we're about to give a DMA
482 		 * capable device to a user without IOMMU protection.
483 		 */
484 		group = vfio_noiommu_group_alloc(dev, VFIO_NO_IOMMU);
485 		if (!IS_ERR(group)) {
486 			add_taint(TAINT_USER, LOCKDEP_STILL_OK);
487 			dev_warn(dev, "Adding kernel taint for vfio-noiommu group on device\n");
488 		}
489 		return group;
490 	}
491 
492 	if (!iommu_group)
493 		return ERR_PTR(-EINVAL);
494 
495 	/*
496 	 * VFIO always sets IOMMU_CACHE because we offer no way for userspace to
497 	 * restore cache coherency. It has to be checked here because it is only
498 	 * valid for cases where we are using iommu groups.
499 	 */
500 	if (!device_iommu_capable(dev, IOMMU_CAP_CACHE_COHERENCY)) {
501 		iommu_group_put(iommu_group);
502 		return ERR_PTR(-EINVAL);
503 	}
504 
505 	group = vfio_group_get_from_iommu(iommu_group);
506 	if (!group)
507 		group = vfio_create_group(iommu_group, VFIO_IOMMU);
508 
509 	/* The vfio_group holds a reference to the iommu_group */
510 	iommu_group_put(iommu_group);
511 	return group;
512 }
513 
514 static int __vfio_register_dev(struct vfio_device *device,
515 		struct vfio_group *group)
516 {
517 	struct vfio_device *existing_device;
518 	int ret;
519 
520 	/*
521 	 * In all cases group is the output of one of the group allocation
522 	 * functions and we have group->drivers incremented for us.
523 	 */
524 	if (IS_ERR(group))
525 		return PTR_ERR(group);
526 
527 	/*
528 	 * If the driver doesn't specify a set then the device is added to a
529 	 * singleton set just for itself.
530 	 */
531 	if (!device->dev_set)
532 		vfio_assign_device_set(device, device);
533 
534 	existing_device = vfio_group_get_device(group, device->dev);
535 	if (existing_device) {
536 		/*
537 		 * group->iommu_group is non-NULL because we hold the drivers
538 		 * refcount.
539 		 */
540 		dev_WARN(device->dev, "Device already exists on group %d\n",
541 			 iommu_group_id(group->iommu_group));
542 		vfio_device_put_registration(existing_device);
543 		ret = -EBUSY;
544 		goto err_out;
545 	}
546 
547 	/* Our reference on group is moved to the device */
548 	device->group = group;
549 
550 	ret = dev_set_name(&device->device, "vfio%d", device->index);
551 	if (ret)
552 		goto err_out;
553 
554 	ret = device_add(&device->device);
555 	if (ret)
556 		goto err_out;
557 
558 	/* Refcounting can't start until the driver calls register */
559 	refcount_set(&device->refcount, 1);
560 
561 	mutex_lock(&group->device_lock);
562 	list_add(&device->group_next, &group->device_list);
563 	mutex_unlock(&group->device_lock);
564 
565 	return 0;
566 err_out:
567 	vfio_device_remove_group(device);
568 	return ret;
569 }
570 
571 int vfio_register_group_dev(struct vfio_device *device)
572 {
573 	return __vfio_register_dev(device,
574 		vfio_group_find_or_alloc(device->dev));
575 }
576 EXPORT_SYMBOL_GPL(vfio_register_group_dev);
577 
578 /*
579  * Register a virtual device without IOMMU backing.  The user of this
580  * device must not be able to directly trigger unmediated DMA.
581  */
582 int vfio_register_emulated_iommu_dev(struct vfio_device *device)
583 {
584 	return __vfio_register_dev(device,
585 		vfio_noiommu_group_alloc(device->dev, VFIO_EMULATED_IOMMU));
586 }
587 EXPORT_SYMBOL_GPL(vfio_register_emulated_iommu_dev);
588 
589 static struct vfio_device *vfio_device_get_from_name(struct vfio_group *group,
590 						     char *buf)
591 {
592 	struct vfio_device *it, *device = ERR_PTR(-ENODEV);
593 
594 	mutex_lock(&group->device_lock);
595 	list_for_each_entry(it, &group->device_list, group_next) {
596 		int ret;
597 
598 		if (it->ops->match) {
599 			ret = it->ops->match(it, buf);
600 			if (ret < 0) {
601 				device = ERR_PTR(ret);
602 				break;
603 			}
604 		} else {
605 			ret = !strcmp(dev_name(it->dev), buf);
606 		}
607 
608 		if (ret && vfio_device_try_get_registration(it)) {
609 			device = it;
610 			break;
611 		}
612 	}
613 	mutex_unlock(&group->device_lock);
614 
615 	return device;
616 }
617 
618 /*
619  * Decrement the device reference count and wait for the device to be
620  * removed.  Open file descriptors for the device... */
621 void vfio_unregister_group_dev(struct vfio_device *device)
622 {
623 	struct vfio_group *group = device->group;
624 	unsigned int i = 0;
625 	bool interrupted = false;
626 	long rc;
627 
628 	vfio_device_put_registration(device);
629 	rc = try_wait_for_completion(&device->comp);
630 	while (rc <= 0) {
631 		if (device->ops->request)
632 			device->ops->request(device, i++);
633 
634 		if (interrupted) {
635 			rc = wait_for_completion_timeout(&device->comp,
636 							 HZ * 10);
637 		} else {
638 			rc = wait_for_completion_interruptible_timeout(
639 				&device->comp, HZ * 10);
640 			if (rc < 0) {
641 				interrupted = true;
642 				dev_warn(device->dev,
643 					 "Device is currently in use, task"
644 					 " \"%s\" (%d) "
645 					 "blocked until device is released",
646 					 current->comm, task_pid_nr(current));
647 			}
648 		}
649 	}
650 
651 	mutex_lock(&group->device_lock);
652 	list_del(&device->group_next);
653 	mutex_unlock(&group->device_lock);
654 
655 	/* Balances device_add in register path */
656 	device_del(&device->device);
657 
658 	vfio_device_remove_group(device);
659 }
660 EXPORT_SYMBOL_GPL(vfio_unregister_group_dev);
661 
662 /*
663  * VFIO Group fd, /dev/vfio/$GROUP
664  */
665 /*
666  * VFIO_GROUP_UNSET_CONTAINER should fail if there are other users or
667  * if there was no container to unset.  Since the ioctl is called on
668  * the group, we know that still exists, therefore the only valid
669  * transition here is 1->0.
670  */
671 static int vfio_group_ioctl_unset_container(struct vfio_group *group)
672 {
673 	int ret = 0;
674 
675 	mutex_lock(&group->group_lock);
676 	if (!group->container) {
677 		ret = -EINVAL;
678 		goto out_unlock;
679 	}
680 	if (group->container_users != 1) {
681 		ret = -EBUSY;
682 		goto out_unlock;
683 	}
684 	vfio_group_detach_container(group);
685 
686 out_unlock:
687 	mutex_unlock(&group->group_lock);
688 	return ret;
689 }
690 
691 static int vfio_group_ioctl_set_container(struct vfio_group *group,
692 					  int __user *arg)
693 {
694 	struct vfio_container *container;
695 	struct fd f;
696 	int ret;
697 	int fd;
698 
699 	if (get_user(fd, arg))
700 		return -EFAULT;
701 
702 	f = fdget(fd);
703 	if (!f.file)
704 		return -EBADF;
705 
706 	mutex_lock(&group->group_lock);
707 	if (group->container || WARN_ON(group->container_users)) {
708 		ret = -EINVAL;
709 		goto out_unlock;
710 	}
711 	if (!group->iommu_group) {
712 		ret = -ENODEV;
713 		goto out_unlock;
714 	}
715 
716 	container = vfio_container_from_file(f.file);
717 	ret = -EINVAL;
718 	if (container) {
719 		ret = vfio_container_attach_group(container, group);
720 		goto out_unlock;
721 	}
722 
723 out_unlock:
724 	mutex_unlock(&group->group_lock);
725 	fdput(f);
726 	return ret;
727 }
728 
729 static const struct file_operations vfio_device_fops;
730 
731 /* true if the vfio_device has open_device() called but not close_device() */
732 bool vfio_assert_device_open(struct vfio_device *device)
733 {
734 	return !WARN_ON_ONCE(!READ_ONCE(device->open_count));
735 }
736 
737 static struct file *vfio_device_open(struct vfio_device *device)
738 {
739 	struct file *filep;
740 	int ret;
741 
742 	mutex_lock(&device->group->group_lock);
743 	ret = vfio_device_assign_container(device);
744 	mutex_unlock(&device->group->group_lock);
745 	if (ret)
746 		return ERR_PTR(ret);
747 
748 	if (!try_module_get(device->dev->driver->owner)) {
749 		ret = -ENODEV;
750 		goto err_unassign_container;
751 	}
752 
753 	mutex_lock(&device->dev_set->lock);
754 	device->open_count++;
755 	if (device->open_count == 1) {
756 		/*
757 		 * Here we pass the KVM pointer with the group under the read
758 		 * lock.  If the device driver will use it, it must obtain a
759 		 * reference and release it during close_device.
760 		 */
761 		mutex_lock(&device->group->group_lock);
762 		device->kvm = device->group->kvm;
763 
764 		if (device->ops->open_device) {
765 			ret = device->ops->open_device(device);
766 			if (ret)
767 				goto err_undo_count;
768 		}
769 		vfio_device_container_register(device);
770 		mutex_unlock(&device->group->group_lock);
771 	}
772 	mutex_unlock(&device->dev_set->lock);
773 
774 	/*
775 	 * We can't use anon_inode_getfd() because we need to modify
776 	 * the f_mode flags directly to allow more than just ioctls
777 	 */
778 	filep = anon_inode_getfile("[vfio-device]", &vfio_device_fops,
779 				   device, O_RDWR);
780 	if (IS_ERR(filep)) {
781 		ret = PTR_ERR(filep);
782 		goto err_close_device;
783 	}
784 
785 	/*
786 	 * TODO: add an anon_inode interface to do this.
787 	 * Appears to be missing by lack of need rather than
788 	 * explicitly prevented.  Now there's need.
789 	 */
790 	filep->f_mode |= (FMODE_PREAD | FMODE_PWRITE);
791 
792 	if (device->group->type == VFIO_NO_IOMMU)
793 		dev_warn(device->dev, "vfio-noiommu device opened by user "
794 			 "(%s:%d)\n", current->comm, task_pid_nr(current));
795 	/*
796 	 * On success the ref of device is moved to the file and
797 	 * put in vfio_device_fops_release()
798 	 */
799 	return filep;
800 
801 err_close_device:
802 	mutex_lock(&device->dev_set->lock);
803 	mutex_lock(&device->group->group_lock);
804 	if (device->open_count == 1 && device->ops->close_device) {
805 		device->ops->close_device(device);
806 
807 		vfio_device_container_unregister(device);
808 	}
809 err_undo_count:
810 	mutex_unlock(&device->group->group_lock);
811 	device->open_count--;
812 	if (device->open_count == 0 && device->kvm)
813 		device->kvm = NULL;
814 	mutex_unlock(&device->dev_set->lock);
815 	module_put(device->dev->driver->owner);
816 err_unassign_container:
817 	vfio_device_unassign_container(device);
818 	return ERR_PTR(ret);
819 }
820 
821 static int vfio_group_ioctl_get_device_fd(struct vfio_group *group,
822 					  char __user *arg)
823 {
824 	struct vfio_device *device;
825 	struct file *filep;
826 	char *buf;
827 	int fdno;
828 	int ret;
829 
830 	buf = strndup_user(arg, PAGE_SIZE);
831 	if (IS_ERR(buf))
832 		return PTR_ERR(buf);
833 
834 	device = vfio_device_get_from_name(group, buf);
835 	kfree(buf);
836 	if (IS_ERR(device))
837 		return PTR_ERR(device);
838 
839 	fdno = get_unused_fd_flags(O_CLOEXEC);
840 	if (fdno < 0) {
841 		ret = fdno;
842 		goto err_put_device;
843 	}
844 
845 	filep = vfio_device_open(device);
846 	if (IS_ERR(filep)) {
847 		ret = PTR_ERR(filep);
848 		goto err_put_fdno;
849 	}
850 
851 	fd_install(fdno, filep);
852 	return fdno;
853 
854 err_put_fdno:
855 	put_unused_fd(fdno);
856 err_put_device:
857 	vfio_device_put_registration(device);
858 	return ret;
859 }
860 
861 static int vfio_group_ioctl_get_status(struct vfio_group *group,
862 				       struct vfio_group_status __user *arg)
863 {
864 	unsigned long minsz = offsetofend(struct vfio_group_status, flags);
865 	struct vfio_group_status status;
866 
867 	if (copy_from_user(&status, arg, minsz))
868 		return -EFAULT;
869 
870 	if (status.argsz < minsz)
871 		return -EINVAL;
872 
873 	status.flags = 0;
874 
875 	mutex_lock(&group->group_lock);
876 	if (!group->iommu_group) {
877 		mutex_unlock(&group->group_lock);
878 		return -ENODEV;
879 	}
880 
881 	if (group->container)
882 		status.flags |= VFIO_GROUP_FLAGS_CONTAINER_SET |
883 				VFIO_GROUP_FLAGS_VIABLE;
884 	else if (!iommu_group_dma_owner_claimed(group->iommu_group))
885 		status.flags |= VFIO_GROUP_FLAGS_VIABLE;
886 	mutex_unlock(&group->group_lock);
887 
888 	if (copy_to_user(arg, &status, minsz))
889 		return -EFAULT;
890 	return 0;
891 }
892 
893 static long vfio_group_fops_unl_ioctl(struct file *filep,
894 				      unsigned int cmd, unsigned long arg)
895 {
896 	struct vfio_group *group = filep->private_data;
897 	void __user *uarg = (void __user *)arg;
898 
899 	switch (cmd) {
900 	case VFIO_GROUP_GET_DEVICE_FD:
901 		return vfio_group_ioctl_get_device_fd(group, uarg);
902 	case VFIO_GROUP_GET_STATUS:
903 		return vfio_group_ioctl_get_status(group, uarg);
904 	case VFIO_GROUP_SET_CONTAINER:
905 		return vfio_group_ioctl_set_container(group, uarg);
906 	case VFIO_GROUP_UNSET_CONTAINER:
907 		return vfio_group_ioctl_unset_container(group);
908 	default:
909 		return -ENOTTY;
910 	}
911 }
912 
913 static int vfio_group_fops_open(struct inode *inode, struct file *filep)
914 {
915 	struct vfio_group *group =
916 		container_of(inode->i_cdev, struct vfio_group, cdev);
917 	int ret;
918 
919 	mutex_lock(&group->group_lock);
920 
921 	/*
922 	 * drivers can be zero if this races with vfio_device_remove_group(), it
923 	 * will be stable at 0 under the group rwsem
924 	 */
925 	if (refcount_read(&group->drivers) == 0) {
926 		ret = -ENODEV;
927 		goto out_unlock;
928 	}
929 
930 	if (group->type == VFIO_NO_IOMMU && !capable(CAP_SYS_RAWIO)) {
931 		ret = -EPERM;
932 		goto out_unlock;
933 	}
934 
935 	/*
936 	 * Do we need multiple instances of the group open?  Seems not.
937 	 */
938 	if (group->opened_file) {
939 		ret = -EBUSY;
940 		goto out_unlock;
941 	}
942 	group->opened_file = filep;
943 	filep->private_data = group;
944 	ret = 0;
945 out_unlock:
946 	mutex_unlock(&group->group_lock);
947 	return ret;
948 }
949 
950 static int vfio_group_fops_release(struct inode *inode, struct file *filep)
951 {
952 	struct vfio_group *group = filep->private_data;
953 
954 	filep->private_data = NULL;
955 
956 	mutex_lock(&group->group_lock);
957 	/*
958 	 * Device FDs hold a group file reference, therefore the group release
959 	 * is only called when there are no open devices.
960 	 */
961 	WARN_ON(group->notifier.head);
962 	if (group->container)
963 		vfio_group_detach_container(group);
964 	group->opened_file = NULL;
965 	mutex_unlock(&group->group_lock);
966 	return 0;
967 }
968 
969 static const struct file_operations vfio_group_fops = {
970 	.owner		= THIS_MODULE,
971 	.unlocked_ioctl	= vfio_group_fops_unl_ioctl,
972 	.compat_ioctl	= compat_ptr_ioctl,
973 	.open		= vfio_group_fops_open,
974 	.release	= vfio_group_fops_release,
975 };
976 
977 /*
978  * Wrapper around pm_runtime_resume_and_get().
979  * Return error code on failure or 0 on success.
980  */
981 static inline int vfio_device_pm_runtime_get(struct vfio_device *device)
982 {
983 	struct device *dev = device->dev;
984 
985 	if (dev->driver && dev->driver->pm) {
986 		int ret;
987 
988 		ret = pm_runtime_resume_and_get(dev);
989 		if (ret) {
990 			dev_info_ratelimited(dev,
991 				"vfio: runtime resume failed %d\n", ret);
992 			return -EIO;
993 		}
994 	}
995 
996 	return 0;
997 }
998 
999 /*
1000  * Wrapper around pm_runtime_put().
1001  */
1002 static inline void vfio_device_pm_runtime_put(struct vfio_device *device)
1003 {
1004 	struct device *dev = device->dev;
1005 
1006 	if (dev->driver && dev->driver->pm)
1007 		pm_runtime_put(dev);
1008 }
1009 
1010 /*
1011  * VFIO Device fd
1012  */
1013 static int vfio_device_fops_release(struct inode *inode, struct file *filep)
1014 {
1015 	struct vfio_device *device = filep->private_data;
1016 
1017 	mutex_lock(&device->dev_set->lock);
1018 	vfio_assert_device_open(device);
1019 	mutex_lock(&device->group->group_lock);
1020 	if (device->open_count == 1 && device->ops->close_device)
1021 		device->ops->close_device(device);
1022 
1023 	vfio_device_container_unregister(device);
1024 	mutex_unlock(&device->group->group_lock);
1025 	device->open_count--;
1026 	if (device->open_count == 0)
1027 		device->kvm = NULL;
1028 	mutex_unlock(&device->dev_set->lock);
1029 
1030 	module_put(device->dev->driver->owner);
1031 
1032 	vfio_device_unassign_container(device);
1033 
1034 	vfio_device_put_registration(device);
1035 
1036 	return 0;
1037 }
1038 
1039 /*
1040  * vfio_mig_get_next_state - Compute the next step in the FSM
1041  * @cur_fsm - The current state the device is in
1042  * @new_fsm - The target state to reach
1043  * @next_fsm - Pointer to the next step to get to new_fsm
1044  *
1045  * Return 0 upon success, otherwise -errno
1046  * Upon success the next step in the state progression between cur_fsm and
1047  * new_fsm will be set in next_fsm.
1048  *
1049  * This breaks down requests for combination transitions into smaller steps and
1050  * returns the next step to get to new_fsm. The function may need to be called
1051  * multiple times before reaching new_fsm.
1052  *
1053  */
1054 int vfio_mig_get_next_state(struct vfio_device *device,
1055 			    enum vfio_device_mig_state cur_fsm,
1056 			    enum vfio_device_mig_state new_fsm,
1057 			    enum vfio_device_mig_state *next_fsm)
1058 {
1059 	enum { VFIO_DEVICE_NUM_STATES = VFIO_DEVICE_STATE_RUNNING_P2P + 1 };
1060 	/*
1061 	 * The coding in this table requires the driver to implement the
1062 	 * following FSM arcs:
1063 	 *         RESUMING -> STOP
1064 	 *         STOP -> RESUMING
1065 	 *         STOP -> STOP_COPY
1066 	 *         STOP_COPY -> STOP
1067 	 *
1068 	 * If P2P is supported then the driver must also implement these FSM
1069 	 * arcs:
1070 	 *         RUNNING -> RUNNING_P2P
1071 	 *         RUNNING_P2P -> RUNNING
1072 	 *         RUNNING_P2P -> STOP
1073 	 *         STOP -> RUNNING_P2P
1074 	 * Without P2P the driver must implement:
1075 	 *         RUNNING -> STOP
1076 	 *         STOP -> RUNNING
1077 	 *
1078 	 * The coding will step through multiple states for some combination
1079 	 * transitions; if all optional features are supported, this means the
1080 	 * following ones:
1081 	 *         RESUMING -> STOP -> RUNNING_P2P
1082 	 *         RESUMING -> STOP -> RUNNING_P2P -> RUNNING
1083 	 *         RESUMING -> STOP -> STOP_COPY
1084 	 *         RUNNING -> RUNNING_P2P -> STOP
1085 	 *         RUNNING -> RUNNING_P2P -> STOP -> RESUMING
1086 	 *         RUNNING -> RUNNING_P2P -> STOP -> STOP_COPY
1087 	 *         RUNNING_P2P -> STOP -> RESUMING
1088 	 *         RUNNING_P2P -> STOP -> STOP_COPY
1089 	 *         STOP -> RUNNING_P2P -> RUNNING
1090 	 *         STOP_COPY -> STOP -> RESUMING
1091 	 *         STOP_COPY -> STOP -> RUNNING_P2P
1092 	 *         STOP_COPY -> STOP -> RUNNING_P2P -> RUNNING
1093 	 */
1094 	static const u8 vfio_from_fsm_table[VFIO_DEVICE_NUM_STATES][VFIO_DEVICE_NUM_STATES] = {
1095 		[VFIO_DEVICE_STATE_STOP] = {
1096 			[VFIO_DEVICE_STATE_STOP] = VFIO_DEVICE_STATE_STOP,
1097 			[VFIO_DEVICE_STATE_RUNNING] = VFIO_DEVICE_STATE_RUNNING_P2P,
1098 			[VFIO_DEVICE_STATE_STOP_COPY] = VFIO_DEVICE_STATE_STOP_COPY,
1099 			[VFIO_DEVICE_STATE_RESUMING] = VFIO_DEVICE_STATE_RESUMING,
1100 			[VFIO_DEVICE_STATE_RUNNING_P2P] = VFIO_DEVICE_STATE_RUNNING_P2P,
1101 			[VFIO_DEVICE_STATE_ERROR] = VFIO_DEVICE_STATE_ERROR,
1102 		},
1103 		[VFIO_DEVICE_STATE_RUNNING] = {
1104 			[VFIO_DEVICE_STATE_STOP] = VFIO_DEVICE_STATE_RUNNING_P2P,
1105 			[VFIO_DEVICE_STATE_RUNNING] = VFIO_DEVICE_STATE_RUNNING,
1106 			[VFIO_DEVICE_STATE_STOP_COPY] = VFIO_DEVICE_STATE_RUNNING_P2P,
1107 			[VFIO_DEVICE_STATE_RESUMING] = VFIO_DEVICE_STATE_RUNNING_P2P,
1108 			[VFIO_DEVICE_STATE_RUNNING_P2P] = VFIO_DEVICE_STATE_RUNNING_P2P,
1109 			[VFIO_DEVICE_STATE_ERROR] = VFIO_DEVICE_STATE_ERROR,
1110 		},
1111 		[VFIO_DEVICE_STATE_STOP_COPY] = {
1112 			[VFIO_DEVICE_STATE_STOP] = VFIO_DEVICE_STATE_STOP,
1113 			[VFIO_DEVICE_STATE_RUNNING] = VFIO_DEVICE_STATE_STOP,
1114 			[VFIO_DEVICE_STATE_STOP_COPY] = VFIO_DEVICE_STATE_STOP_COPY,
1115 			[VFIO_DEVICE_STATE_RESUMING] = VFIO_DEVICE_STATE_STOP,
1116 			[VFIO_DEVICE_STATE_RUNNING_P2P] = VFIO_DEVICE_STATE_STOP,
1117 			[VFIO_DEVICE_STATE_ERROR] = VFIO_DEVICE_STATE_ERROR,
1118 		},
1119 		[VFIO_DEVICE_STATE_RESUMING] = {
1120 			[VFIO_DEVICE_STATE_STOP] = VFIO_DEVICE_STATE_STOP,
1121 			[VFIO_DEVICE_STATE_RUNNING] = VFIO_DEVICE_STATE_STOP,
1122 			[VFIO_DEVICE_STATE_STOP_COPY] = VFIO_DEVICE_STATE_STOP,
1123 			[VFIO_DEVICE_STATE_RESUMING] = VFIO_DEVICE_STATE_RESUMING,
1124 			[VFIO_DEVICE_STATE_RUNNING_P2P] = VFIO_DEVICE_STATE_STOP,
1125 			[VFIO_DEVICE_STATE_ERROR] = VFIO_DEVICE_STATE_ERROR,
1126 		},
1127 		[VFIO_DEVICE_STATE_RUNNING_P2P] = {
1128 			[VFIO_DEVICE_STATE_STOP] = VFIO_DEVICE_STATE_STOP,
1129 			[VFIO_DEVICE_STATE_RUNNING] = VFIO_DEVICE_STATE_RUNNING,
1130 			[VFIO_DEVICE_STATE_STOP_COPY] = VFIO_DEVICE_STATE_STOP,
1131 			[VFIO_DEVICE_STATE_RESUMING] = VFIO_DEVICE_STATE_STOP,
1132 			[VFIO_DEVICE_STATE_RUNNING_P2P] = VFIO_DEVICE_STATE_RUNNING_P2P,
1133 			[VFIO_DEVICE_STATE_ERROR] = VFIO_DEVICE_STATE_ERROR,
1134 		},
1135 		[VFIO_DEVICE_STATE_ERROR] = {
1136 			[VFIO_DEVICE_STATE_STOP] = VFIO_DEVICE_STATE_ERROR,
1137 			[VFIO_DEVICE_STATE_RUNNING] = VFIO_DEVICE_STATE_ERROR,
1138 			[VFIO_DEVICE_STATE_STOP_COPY] = VFIO_DEVICE_STATE_ERROR,
1139 			[VFIO_DEVICE_STATE_RESUMING] = VFIO_DEVICE_STATE_ERROR,
1140 			[VFIO_DEVICE_STATE_RUNNING_P2P] = VFIO_DEVICE_STATE_ERROR,
1141 			[VFIO_DEVICE_STATE_ERROR] = VFIO_DEVICE_STATE_ERROR,
1142 		},
1143 	};
1144 
1145 	static const unsigned int state_flags_table[VFIO_DEVICE_NUM_STATES] = {
1146 		[VFIO_DEVICE_STATE_STOP] = VFIO_MIGRATION_STOP_COPY,
1147 		[VFIO_DEVICE_STATE_RUNNING] = VFIO_MIGRATION_STOP_COPY,
1148 		[VFIO_DEVICE_STATE_STOP_COPY] = VFIO_MIGRATION_STOP_COPY,
1149 		[VFIO_DEVICE_STATE_RESUMING] = VFIO_MIGRATION_STOP_COPY,
1150 		[VFIO_DEVICE_STATE_RUNNING_P2P] =
1151 			VFIO_MIGRATION_STOP_COPY | VFIO_MIGRATION_P2P,
1152 		[VFIO_DEVICE_STATE_ERROR] = ~0U,
1153 	};
1154 
1155 	if (WARN_ON(cur_fsm >= ARRAY_SIZE(vfio_from_fsm_table) ||
1156 		    (state_flags_table[cur_fsm] & device->migration_flags) !=
1157 			state_flags_table[cur_fsm]))
1158 		return -EINVAL;
1159 
1160 	if (new_fsm >= ARRAY_SIZE(vfio_from_fsm_table) ||
1161 	   (state_flags_table[new_fsm] & device->migration_flags) !=
1162 			state_flags_table[new_fsm])
1163 		return -EINVAL;
1164 
1165 	/*
1166 	 * Arcs touching optional and unsupported states are skipped over. The
1167 	 * driver will instead see an arc from the original state to the next
1168 	 * logical state, as per the above comment.
1169 	 */
1170 	*next_fsm = vfio_from_fsm_table[cur_fsm][new_fsm];
1171 	while ((state_flags_table[*next_fsm] & device->migration_flags) !=
1172 			state_flags_table[*next_fsm])
1173 		*next_fsm = vfio_from_fsm_table[*next_fsm][new_fsm];
1174 
1175 	return (*next_fsm != VFIO_DEVICE_STATE_ERROR) ? 0 : -EINVAL;
1176 }
1177 EXPORT_SYMBOL_GPL(vfio_mig_get_next_state);
1178 
1179 /*
1180  * Convert the drivers's struct file into a FD number and return it to userspace
1181  */
1182 static int vfio_ioct_mig_return_fd(struct file *filp, void __user *arg,
1183 				   struct vfio_device_feature_mig_state *mig)
1184 {
1185 	int ret;
1186 	int fd;
1187 
1188 	fd = get_unused_fd_flags(O_CLOEXEC);
1189 	if (fd < 0) {
1190 		ret = fd;
1191 		goto out_fput;
1192 	}
1193 
1194 	mig->data_fd = fd;
1195 	if (copy_to_user(arg, mig, sizeof(*mig))) {
1196 		ret = -EFAULT;
1197 		goto out_put_unused;
1198 	}
1199 	fd_install(fd, filp);
1200 	return 0;
1201 
1202 out_put_unused:
1203 	put_unused_fd(fd);
1204 out_fput:
1205 	fput(filp);
1206 	return ret;
1207 }
1208 
1209 static int
1210 vfio_ioctl_device_feature_mig_device_state(struct vfio_device *device,
1211 					   u32 flags, void __user *arg,
1212 					   size_t argsz)
1213 {
1214 	size_t minsz =
1215 		offsetofend(struct vfio_device_feature_mig_state, data_fd);
1216 	struct vfio_device_feature_mig_state mig;
1217 	struct file *filp = NULL;
1218 	int ret;
1219 
1220 	if (!device->mig_ops)
1221 		return -ENOTTY;
1222 
1223 	ret = vfio_check_feature(flags, argsz,
1224 				 VFIO_DEVICE_FEATURE_SET |
1225 				 VFIO_DEVICE_FEATURE_GET,
1226 				 sizeof(mig));
1227 	if (ret != 1)
1228 		return ret;
1229 
1230 	if (copy_from_user(&mig, arg, minsz))
1231 		return -EFAULT;
1232 
1233 	if (flags & VFIO_DEVICE_FEATURE_GET) {
1234 		enum vfio_device_mig_state curr_state;
1235 
1236 		ret = device->mig_ops->migration_get_state(device,
1237 							   &curr_state);
1238 		if (ret)
1239 			return ret;
1240 		mig.device_state = curr_state;
1241 		goto out_copy;
1242 	}
1243 
1244 	/* Handle the VFIO_DEVICE_FEATURE_SET */
1245 	filp = device->mig_ops->migration_set_state(device, mig.device_state);
1246 	if (IS_ERR(filp) || !filp)
1247 		goto out_copy;
1248 
1249 	return vfio_ioct_mig_return_fd(filp, arg, &mig);
1250 out_copy:
1251 	mig.data_fd = -1;
1252 	if (copy_to_user(arg, &mig, sizeof(mig)))
1253 		return -EFAULT;
1254 	if (IS_ERR(filp))
1255 		return PTR_ERR(filp);
1256 	return 0;
1257 }
1258 
1259 static int vfio_ioctl_device_feature_migration(struct vfio_device *device,
1260 					       u32 flags, void __user *arg,
1261 					       size_t argsz)
1262 {
1263 	struct vfio_device_feature_migration mig = {
1264 		.flags = device->migration_flags,
1265 	};
1266 	int ret;
1267 
1268 	if (!device->mig_ops)
1269 		return -ENOTTY;
1270 
1271 	ret = vfio_check_feature(flags, argsz, VFIO_DEVICE_FEATURE_GET,
1272 				 sizeof(mig));
1273 	if (ret != 1)
1274 		return ret;
1275 	if (copy_to_user(arg, &mig, sizeof(mig)))
1276 		return -EFAULT;
1277 	return 0;
1278 }
1279 
1280 /* Ranges should fit into a single kernel page */
1281 #define LOG_MAX_RANGES \
1282 	(PAGE_SIZE / sizeof(struct vfio_device_feature_dma_logging_range))
1283 
1284 static int
1285 vfio_ioctl_device_feature_logging_start(struct vfio_device *device,
1286 					u32 flags, void __user *arg,
1287 					size_t argsz)
1288 {
1289 	size_t minsz =
1290 		offsetofend(struct vfio_device_feature_dma_logging_control,
1291 			    ranges);
1292 	struct vfio_device_feature_dma_logging_range __user *ranges;
1293 	struct vfio_device_feature_dma_logging_control control;
1294 	struct vfio_device_feature_dma_logging_range range;
1295 	struct rb_root_cached root = RB_ROOT_CACHED;
1296 	struct interval_tree_node *nodes;
1297 	u64 iova_end;
1298 	u32 nnodes;
1299 	int i, ret;
1300 
1301 	if (!device->log_ops)
1302 		return -ENOTTY;
1303 
1304 	ret = vfio_check_feature(flags, argsz,
1305 				 VFIO_DEVICE_FEATURE_SET,
1306 				 sizeof(control));
1307 	if (ret != 1)
1308 		return ret;
1309 
1310 	if (copy_from_user(&control, arg, minsz))
1311 		return -EFAULT;
1312 
1313 	nnodes = control.num_ranges;
1314 	if (!nnodes)
1315 		return -EINVAL;
1316 
1317 	if (nnodes > LOG_MAX_RANGES)
1318 		return -E2BIG;
1319 
1320 	ranges = u64_to_user_ptr(control.ranges);
1321 	nodes = kmalloc_array(nnodes, sizeof(struct interval_tree_node),
1322 			      GFP_KERNEL);
1323 	if (!nodes)
1324 		return -ENOMEM;
1325 
1326 	for (i = 0; i < nnodes; i++) {
1327 		if (copy_from_user(&range, &ranges[i], sizeof(range))) {
1328 			ret = -EFAULT;
1329 			goto end;
1330 		}
1331 		if (!IS_ALIGNED(range.iova, control.page_size) ||
1332 		    !IS_ALIGNED(range.length, control.page_size)) {
1333 			ret = -EINVAL;
1334 			goto end;
1335 		}
1336 
1337 		if (check_add_overflow(range.iova, range.length, &iova_end) ||
1338 		    iova_end > ULONG_MAX) {
1339 			ret = -EOVERFLOW;
1340 			goto end;
1341 		}
1342 
1343 		nodes[i].start = range.iova;
1344 		nodes[i].last = range.iova + range.length - 1;
1345 		if (interval_tree_iter_first(&root, nodes[i].start,
1346 					     nodes[i].last)) {
1347 			/* Range overlapping */
1348 			ret = -EINVAL;
1349 			goto end;
1350 		}
1351 		interval_tree_insert(nodes + i, &root);
1352 	}
1353 
1354 	ret = device->log_ops->log_start(device, &root, nnodes,
1355 					 &control.page_size);
1356 	if (ret)
1357 		goto end;
1358 
1359 	if (copy_to_user(arg, &control, sizeof(control))) {
1360 		ret = -EFAULT;
1361 		device->log_ops->log_stop(device);
1362 	}
1363 
1364 end:
1365 	kfree(nodes);
1366 	return ret;
1367 }
1368 
1369 static int
1370 vfio_ioctl_device_feature_logging_stop(struct vfio_device *device,
1371 				       u32 flags, void __user *arg,
1372 				       size_t argsz)
1373 {
1374 	int ret;
1375 
1376 	if (!device->log_ops)
1377 		return -ENOTTY;
1378 
1379 	ret = vfio_check_feature(flags, argsz,
1380 				 VFIO_DEVICE_FEATURE_SET, 0);
1381 	if (ret != 1)
1382 		return ret;
1383 
1384 	return device->log_ops->log_stop(device);
1385 }
1386 
1387 static int vfio_device_log_read_and_clear(struct iova_bitmap *iter,
1388 					  unsigned long iova, size_t length,
1389 					  void *opaque)
1390 {
1391 	struct vfio_device *device = opaque;
1392 
1393 	return device->log_ops->log_read_and_clear(device, iova, length, iter);
1394 }
1395 
1396 static int
1397 vfio_ioctl_device_feature_logging_report(struct vfio_device *device,
1398 					 u32 flags, void __user *arg,
1399 					 size_t argsz)
1400 {
1401 	size_t minsz =
1402 		offsetofend(struct vfio_device_feature_dma_logging_report,
1403 			    bitmap);
1404 	struct vfio_device_feature_dma_logging_report report;
1405 	struct iova_bitmap *iter;
1406 	u64 iova_end;
1407 	int ret;
1408 
1409 	if (!device->log_ops)
1410 		return -ENOTTY;
1411 
1412 	ret = vfio_check_feature(flags, argsz,
1413 				 VFIO_DEVICE_FEATURE_GET,
1414 				 sizeof(report));
1415 	if (ret != 1)
1416 		return ret;
1417 
1418 	if (copy_from_user(&report, arg, minsz))
1419 		return -EFAULT;
1420 
1421 	if (report.page_size < SZ_4K || !is_power_of_2(report.page_size))
1422 		return -EINVAL;
1423 
1424 	if (check_add_overflow(report.iova, report.length, &iova_end) ||
1425 	    iova_end > ULONG_MAX)
1426 		return -EOVERFLOW;
1427 
1428 	iter = iova_bitmap_alloc(report.iova, report.length,
1429 				 report.page_size,
1430 				 u64_to_user_ptr(report.bitmap));
1431 	if (IS_ERR(iter))
1432 		return PTR_ERR(iter);
1433 
1434 	ret = iova_bitmap_for_each(iter, device,
1435 				   vfio_device_log_read_and_clear);
1436 
1437 	iova_bitmap_free(iter);
1438 	return ret;
1439 }
1440 
1441 static int vfio_ioctl_device_feature(struct vfio_device *device,
1442 				     struct vfio_device_feature __user *arg)
1443 {
1444 	size_t minsz = offsetofend(struct vfio_device_feature, flags);
1445 	struct vfio_device_feature feature;
1446 
1447 	if (copy_from_user(&feature, arg, minsz))
1448 		return -EFAULT;
1449 
1450 	if (feature.argsz < minsz)
1451 		return -EINVAL;
1452 
1453 	/* Check unknown flags */
1454 	if (feature.flags &
1455 	    ~(VFIO_DEVICE_FEATURE_MASK | VFIO_DEVICE_FEATURE_SET |
1456 	      VFIO_DEVICE_FEATURE_GET | VFIO_DEVICE_FEATURE_PROBE))
1457 		return -EINVAL;
1458 
1459 	/* GET & SET are mutually exclusive except with PROBE */
1460 	if (!(feature.flags & VFIO_DEVICE_FEATURE_PROBE) &&
1461 	    (feature.flags & VFIO_DEVICE_FEATURE_SET) &&
1462 	    (feature.flags & VFIO_DEVICE_FEATURE_GET))
1463 		return -EINVAL;
1464 
1465 	switch (feature.flags & VFIO_DEVICE_FEATURE_MASK) {
1466 	case VFIO_DEVICE_FEATURE_MIGRATION:
1467 		return vfio_ioctl_device_feature_migration(
1468 			device, feature.flags, arg->data,
1469 			feature.argsz - minsz);
1470 	case VFIO_DEVICE_FEATURE_MIG_DEVICE_STATE:
1471 		return vfio_ioctl_device_feature_mig_device_state(
1472 			device, feature.flags, arg->data,
1473 			feature.argsz - minsz);
1474 	case VFIO_DEVICE_FEATURE_DMA_LOGGING_START:
1475 		return vfio_ioctl_device_feature_logging_start(
1476 			device, feature.flags, arg->data,
1477 			feature.argsz - minsz);
1478 	case VFIO_DEVICE_FEATURE_DMA_LOGGING_STOP:
1479 		return vfio_ioctl_device_feature_logging_stop(
1480 			device, feature.flags, arg->data,
1481 			feature.argsz - minsz);
1482 	case VFIO_DEVICE_FEATURE_DMA_LOGGING_REPORT:
1483 		return vfio_ioctl_device_feature_logging_report(
1484 			device, feature.flags, arg->data,
1485 			feature.argsz - minsz);
1486 	default:
1487 		if (unlikely(!device->ops->device_feature))
1488 			return -EINVAL;
1489 		return device->ops->device_feature(device, feature.flags,
1490 						   arg->data,
1491 						   feature.argsz - minsz);
1492 	}
1493 }
1494 
1495 static long vfio_device_fops_unl_ioctl(struct file *filep,
1496 				       unsigned int cmd, unsigned long arg)
1497 {
1498 	struct vfio_device *device = filep->private_data;
1499 	int ret;
1500 
1501 	ret = vfio_device_pm_runtime_get(device);
1502 	if (ret)
1503 		return ret;
1504 
1505 	switch (cmd) {
1506 	case VFIO_DEVICE_FEATURE:
1507 		ret = vfio_ioctl_device_feature(device, (void __user *)arg);
1508 		break;
1509 
1510 	default:
1511 		if (unlikely(!device->ops->ioctl))
1512 			ret = -EINVAL;
1513 		else
1514 			ret = device->ops->ioctl(device, cmd, arg);
1515 		break;
1516 	}
1517 
1518 	vfio_device_pm_runtime_put(device);
1519 	return ret;
1520 }
1521 
1522 static ssize_t vfio_device_fops_read(struct file *filep, char __user *buf,
1523 				     size_t count, loff_t *ppos)
1524 {
1525 	struct vfio_device *device = filep->private_data;
1526 
1527 	if (unlikely(!device->ops->read))
1528 		return -EINVAL;
1529 
1530 	return device->ops->read(device, buf, count, ppos);
1531 }
1532 
1533 static ssize_t vfio_device_fops_write(struct file *filep,
1534 				      const char __user *buf,
1535 				      size_t count, loff_t *ppos)
1536 {
1537 	struct vfio_device *device = filep->private_data;
1538 
1539 	if (unlikely(!device->ops->write))
1540 		return -EINVAL;
1541 
1542 	return device->ops->write(device, buf, count, ppos);
1543 }
1544 
1545 static int vfio_device_fops_mmap(struct file *filep, struct vm_area_struct *vma)
1546 {
1547 	struct vfio_device *device = filep->private_data;
1548 
1549 	if (unlikely(!device->ops->mmap))
1550 		return -EINVAL;
1551 
1552 	return device->ops->mmap(device, vma);
1553 }
1554 
1555 static const struct file_operations vfio_device_fops = {
1556 	.owner		= THIS_MODULE,
1557 	.release	= vfio_device_fops_release,
1558 	.read		= vfio_device_fops_read,
1559 	.write		= vfio_device_fops_write,
1560 	.unlocked_ioctl	= vfio_device_fops_unl_ioctl,
1561 	.compat_ioctl	= compat_ptr_ioctl,
1562 	.mmap		= vfio_device_fops_mmap,
1563 };
1564 
1565 /**
1566  * vfio_file_iommu_group - Return the struct iommu_group for the vfio group file
1567  * @file: VFIO group file
1568  *
1569  * The returned iommu_group is valid as long as a ref is held on the file. This
1570  * returns a reference on the group. This function is deprecated, only the SPAPR
1571  * path in kvm should call it.
1572  */
1573 struct iommu_group *vfio_file_iommu_group(struct file *file)
1574 {
1575 	struct vfio_group *group = file->private_data;
1576 	struct iommu_group *iommu_group = NULL;
1577 
1578 	if (!IS_ENABLED(CONFIG_SPAPR_TCE_IOMMU))
1579 		return NULL;
1580 
1581 	if (!vfio_file_is_group(file))
1582 		return NULL;
1583 
1584 	mutex_lock(&group->group_lock);
1585 	if (group->iommu_group) {
1586 		iommu_group = group->iommu_group;
1587 		iommu_group_ref_get(iommu_group);
1588 	}
1589 	mutex_unlock(&group->group_lock);
1590 	return iommu_group;
1591 }
1592 EXPORT_SYMBOL_GPL(vfio_file_iommu_group);
1593 
1594 /**
1595  * vfio_file_is_group - True if the file is usable with VFIO aPIS
1596  * @file: VFIO group file
1597  */
1598 bool vfio_file_is_group(struct file *file)
1599 {
1600 	return file->f_op == &vfio_group_fops;
1601 }
1602 EXPORT_SYMBOL_GPL(vfio_file_is_group);
1603 
1604 /**
1605  * vfio_file_enforced_coherent - True if the DMA associated with the VFIO file
1606  *        is always CPU cache coherent
1607  * @file: VFIO group file
1608  *
1609  * Enforced coherency means that the IOMMU ignores things like the PCIe no-snoop
1610  * bit in DMA transactions. A return of false indicates that the user has
1611  * rights to access additional instructions such as wbinvd on x86.
1612  */
1613 bool vfio_file_enforced_coherent(struct file *file)
1614 {
1615 	struct vfio_group *group = file->private_data;
1616 	bool ret;
1617 
1618 	if (!vfio_file_is_group(file))
1619 		return true;
1620 
1621 	mutex_lock(&group->group_lock);
1622 	if (group->container) {
1623 		ret = vfio_container_ioctl_check_extension(group->container,
1624 							   VFIO_DMA_CC_IOMMU);
1625 	} else {
1626 		/*
1627 		 * Since the coherency state is determined only once a container
1628 		 * is attached the user must do so before they can prove they
1629 		 * have permission.
1630 		 */
1631 		ret = true;
1632 	}
1633 	mutex_unlock(&group->group_lock);
1634 	return ret;
1635 }
1636 EXPORT_SYMBOL_GPL(vfio_file_enforced_coherent);
1637 
1638 /**
1639  * vfio_file_set_kvm - Link a kvm with VFIO drivers
1640  * @file: VFIO group file
1641  * @kvm: KVM to link
1642  *
1643  * When a VFIO device is first opened the KVM will be available in
1644  * device->kvm if one was associated with the group.
1645  */
1646 void vfio_file_set_kvm(struct file *file, struct kvm *kvm)
1647 {
1648 	struct vfio_group *group = file->private_data;
1649 
1650 	if (!vfio_file_is_group(file))
1651 		return;
1652 
1653 	mutex_lock(&group->group_lock);
1654 	group->kvm = kvm;
1655 	mutex_unlock(&group->group_lock);
1656 }
1657 EXPORT_SYMBOL_GPL(vfio_file_set_kvm);
1658 
1659 /**
1660  * vfio_file_has_dev - True if the VFIO file is a handle for device
1661  * @file: VFIO file to check
1662  * @device: Device that must be part of the file
1663  *
1664  * Returns true if given file has permission to manipulate the given device.
1665  */
1666 bool vfio_file_has_dev(struct file *file, struct vfio_device *device)
1667 {
1668 	struct vfio_group *group = file->private_data;
1669 
1670 	if (!vfio_file_is_group(file))
1671 		return false;
1672 
1673 	return group == device->group;
1674 }
1675 EXPORT_SYMBOL_GPL(vfio_file_has_dev);
1676 
1677 /*
1678  * Sub-module support
1679  */
1680 /*
1681  * Helper for managing a buffer of info chain capabilities, allocate or
1682  * reallocate a buffer with additional @size, filling in @id and @version
1683  * of the capability.  A pointer to the new capability is returned.
1684  *
1685  * NB. The chain is based at the head of the buffer, so new entries are
1686  * added to the tail, vfio_info_cap_shift() should be called to fixup the
1687  * next offsets prior to copying to the user buffer.
1688  */
1689 struct vfio_info_cap_header *vfio_info_cap_add(struct vfio_info_cap *caps,
1690 					       size_t size, u16 id, u16 version)
1691 {
1692 	void *buf;
1693 	struct vfio_info_cap_header *header, *tmp;
1694 
1695 	buf = krealloc(caps->buf, caps->size + size, GFP_KERNEL);
1696 	if (!buf) {
1697 		kfree(caps->buf);
1698 		caps->buf = NULL;
1699 		caps->size = 0;
1700 		return ERR_PTR(-ENOMEM);
1701 	}
1702 
1703 	caps->buf = buf;
1704 	header = buf + caps->size;
1705 
1706 	/* Eventually copied to user buffer, zero */
1707 	memset(header, 0, size);
1708 
1709 	header->id = id;
1710 	header->version = version;
1711 
1712 	/* Add to the end of the capability chain */
1713 	for (tmp = buf; tmp->next; tmp = buf + tmp->next)
1714 		; /* nothing */
1715 
1716 	tmp->next = caps->size;
1717 	caps->size += size;
1718 
1719 	return header;
1720 }
1721 EXPORT_SYMBOL_GPL(vfio_info_cap_add);
1722 
1723 void vfio_info_cap_shift(struct vfio_info_cap *caps, size_t offset)
1724 {
1725 	struct vfio_info_cap_header *tmp;
1726 	void *buf = (void *)caps->buf;
1727 
1728 	for (tmp = buf; tmp->next; tmp = buf + tmp->next - offset)
1729 		tmp->next += offset;
1730 }
1731 EXPORT_SYMBOL(vfio_info_cap_shift);
1732 
1733 int vfio_info_add_capability(struct vfio_info_cap *caps,
1734 			     struct vfio_info_cap_header *cap, size_t size)
1735 {
1736 	struct vfio_info_cap_header *header;
1737 
1738 	header = vfio_info_cap_add(caps, size, cap->id, cap->version);
1739 	if (IS_ERR(header))
1740 		return PTR_ERR(header);
1741 
1742 	memcpy(header + 1, cap + 1, size - sizeof(*header));
1743 
1744 	return 0;
1745 }
1746 EXPORT_SYMBOL(vfio_info_add_capability);
1747 
1748 int vfio_set_irqs_validate_and_prepare(struct vfio_irq_set *hdr, int num_irqs,
1749 				       int max_irq_type, size_t *data_size)
1750 {
1751 	unsigned long minsz;
1752 	size_t size;
1753 
1754 	minsz = offsetofend(struct vfio_irq_set, count);
1755 
1756 	if ((hdr->argsz < minsz) || (hdr->index >= max_irq_type) ||
1757 	    (hdr->count >= (U32_MAX - hdr->start)) ||
1758 	    (hdr->flags & ~(VFIO_IRQ_SET_DATA_TYPE_MASK |
1759 				VFIO_IRQ_SET_ACTION_TYPE_MASK)))
1760 		return -EINVAL;
1761 
1762 	if (data_size)
1763 		*data_size = 0;
1764 
1765 	if (hdr->start >= num_irqs || hdr->start + hdr->count > num_irqs)
1766 		return -EINVAL;
1767 
1768 	switch (hdr->flags & VFIO_IRQ_SET_DATA_TYPE_MASK) {
1769 	case VFIO_IRQ_SET_DATA_NONE:
1770 		size = 0;
1771 		break;
1772 	case VFIO_IRQ_SET_DATA_BOOL:
1773 		size = sizeof(uint8_t);
1774 		break;
1775 	case VFIO_IRQ_SET_DATA_EVENTFD:
1776 		size = sizeof(int32_t);
1777 		break;
1778 	default:
1779 		return -EINVAL;
1780 	}
1781 
1782 	if (size) {
1783 		if (hdr->argsz - minsz < hdr->count * size)
1784 			return -EINVAL;
1785 
1786 		if (!data_size)
1787 			return -EINVAL;
1788 
1789 		*data_size = hdr->count * size;
1790 	}
1791 
1792 	return 0;
1793 }
1794 EXPORT_SYMBOL(vfio_set_irqs_validate_and_prepare);
1795 
1796 /*
1797  * Module/class support
1798  */
1799 static char *vfio_devnode(struct device *dev, umode_t *mode)
1800 {
1801 	return kasprintf(GFP_KERNEL, "vfio/%s", dev_name(dev));
1802 }
1803 
1804 static int __init vfio_init(void)
1805 {
1806 	int ret;
1807 
1808 	ida_init(&vfio.group_ida);
1809 	ida_init(&vfio.device_ida);
1810 	mutex_init(&vfio.group_lock);
1811 	INIT_LIST_HEAD(&vfio.group_list);
1812 
1813 	ret = vfio_container_init();
1814 	if (ret)
1815 		return ret;
1816 
1817 	/* /dev/vfio/$GROUP */
1818 	vfio.class = class_create(THIS_MODULE, "vfio");
1819 	if (IS_ERR(vfio.class)) {
1820 		ret = PTR_ERR(vfio.class);
1821 		goto err_group_class;
1822 	}
1823 
1824 	vfio.class->devnode = vfio_devnode;
1825 
1826 	/* /sys/class/vfio-dev/vfioX */
1827 	vfio.device_class = class_create(THIS_MODULE, "vfio-dev");
1828 	if (IS_ERR(vfio.device_class)) {
1829 		ret = PTR_ERR(vfio.device_class);
1830 		goto err_dev_class;
1831 	}
1832 
1833 	ret = alloc_chrdev_region(&vfio.group_devt, 0, MINORMASK + 1, "vfio");
1834 	if (ret)
1835 		goto err_alloc_chrdev;
1836 
1837 	pr_info(DRIVER_DESC " version: " DRIVER_VERSION "\n");
1838 	return 0;
1839 
1840 err_alloc_chrdev:
1841 	class_destroy(vfio.device_class);
1842 	vfio.device_class = NULL;
1843 err_dev_class:
1844 	class_destroy(vfio.class);
1845 	vfio.class = NULL;
1846 err_group_class:
1847 	vfio_container_cleanup();
1848 	return ret;
1849 }
1850 
1851 static void __exit vfio_cleanup(void)
1852 {
1853 	WARN_ON(!list_empty(&vfio.group_list));
1854 
1855 	ida_destroy(&vfio.device_ida);
1856 	ida_destroy(&vfio.group_ida);
1857 	unregister_chrdev_region(vfio.group_devt, MINORMASK + 1);
1858 	class_destroy(vfio.device_class);
1859 	vfio.device_class = NULL;
1860 	class_destroy(vfio.class);
1861 	vfio_container_cleanup();
1862 	vfio.class = NULL;
1863 	xa_destroy(&vfio_device_set_xa);
1864 }
1865 
1866 module_init(vfio_init);
1867 module_exit(vfio_cleanup);
1868 
1869 MODULE_VERSION(DRIVER_VERSION);
1870 MODULE_LICENSE("GPL v2");
1871 MODULE_AUTHOR(DRIVER_AUTHOR);
1872 MODULE_DESCRIPTION(DRIVER_DESC);
1873 MODULE_ALIAS_MISCDEV(VFIO_MINOR);
1874 MODULE_ALIAS("devname:vfio/vfio");
1875 MODULE_SOFTDEP("post: vfio_iommu_type1 vfio_iommu_spapr_tce");
1876