xref: /openbmc/linux/drivers/vfio/vfio_main.c (revision a48acad7)
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 unsigned int vfio_device_set_open_count(struct vfio_device_set *dev_set)
129 {
130 	struct vfio_device *cur;
131 	unsigned int open_count = 0;
132 
133 	lockdep_assert_held(&dev_set->lock);
134 
135 	list_for_each_entry(cur, &dev_set->device_list, dev_set_list)
136 		open_count += cur->open_count;
137 	return open_count;
138 }
139 EXPORT_SYMBOL_GPL(vfio_device_set_open_count);
140 
141 /*
142  * Group objects - create, release, get, put, search
143  */
144 static struct vfio_group *
145 __vfio_group_get_from_iommu(struct iommu_group *iommu_group)
146 {
147 	struct vfio_group *group;
148 
149 	/*
150 	 * group->iommu_group from the vfio.group_list cannot be NULL
151 	 * under the vfio.group_lock.
152 	 */
153 	list_for_each_entry(group, &vfio.group_list, vfio_next) {
154 		if (group->iommu_group == iommu_group) {
155 			refcount_inc(&group->drivers);
156 			return group;
157 		}
158 	}
159 	return NULL;
160 }
161 
162 static struct vfio_group *
163 vfio_group_get_from_iommu(struct iommu_group *iommu_group)
164 {
165 	struct vfio_group *group;
166 
167 	mutex_lock(&vfio.group_lock);
168 	group = __vfio_group_get_from_iommu(iommu_group);
169 	mutex_unlock(&vfio.group_lock);
170 	return group;
171 }
172 
173 static void vfio_group_release(struct device *dev)
174 {
175 	struct vfio_group *group = container_of(dev, struct vfio_group, dev);
176 
177 	mutex_destroy(&group->device_lock);
178 	mutex_destroy(&group->group_lock);
179 	WARN_ON(group->iommu_group);
180 	ida_free(&vfio.group_ida, MINOR(group->dev.devt));
181 	kfree(group);
182 }
183 
184 static struct vfio_group *vfio_group_alloc(struct iommu_group *iommu_group,
185 					   enum vfio_group_type type)
186 {
187 	struct vfio_group *group;
188 	int minor;
189 
190 	group = kzalloc(sizeof(*group), GFP_KERNEL);
191 	if (!group)
192 		return ERR_PTR(-ENOMEM);
193 
194 	minor = ida_alloc_max(&vfio.group_ida, MINORMASK, GFP_KERNEL);
195 	if (minor < 0) {
196 		kfree(group);
197 		return ERR_PTR(minor);
198 	}
199 
200 	device_initialize(&group->dev);
201 	group->dev.devt = MKDEV(MAJOR(vfio.group_devt), minor);
202 	group->dev.class = vfio.class;
203 	group->dev.release = vfio_group_release;
204 	cdev_init(&group->cdev, &vfio_group_fops);
205 	group->cdev.owner = THIS_MODULE;
206 
207 	refcount_set(&group->drivers, 1);
208 	mutex_init(&group->group_lock);
209 	INIT_LIST_HEAD(&group->device_list);
210 	mutex_init(&group->device_lock);
211 	group->iommu_group = iommu_group;
212 	/* put in vfio_group_release() */
213 	iommu_group_ref_get(iommu_group);
214 	group->type = type;
215 	BLOCKING_INIT_NOTIFIER_HEAD(&group->notifier);
216 
217 	return group;
218 }
219 
220 static struct vfio_group *vfio_create_group(struct iommu_group *iommu_group,
221 		enum vfio_group_type type)
222 {
223 	struct vfio_group *group;
224 	struct vfio_group *ret;
225 	int err;
226 
227 	group = vfio_group_alloc(iommu_group, type);
228 	if (IS_ERR(group))
229 		return group;
230 
231 	err = dev_set_name(&group->dev, "%s%d",
232 			   group->type == VFIO_NO_IOMMU ? "noiommu-" : "",
233 			   iommu_group_id(iommu_group));
234 	if (err) {
235 		ret = ERR_PTR(err);
236 		goto err_put;
237 	}
238 
239 	mutex_lock(&vfio.group_lock);
240 
241 	/* Did we race creating this group? */
242 	ret = __vfio_group_get_from_iommu(iommu_group);
243 	if (ret)
244 		goto err_unlock;
245 
246 	err = cdev_device_add(&group->cdev, &group->dev);
247 	if (err) {
248 		ret = ERR_PTR(err);
249 		goto err_unlock;
250 	}
251 
252 	list_add(&group->vfio_next, &vfio.group_list);
253 
254 	mutex_unlock(&vfio.group_lock);
255 	return group;
256 
257 err_unlock:
258 	mutex_unlock(&vfio.group_lock);
259 err_put:
260 	put_device(&group->dev);
261 	return ret;
262 }
263 
264 static void vfio_device_remove_group(struct vfio_device *device)
265 {
266 	struct vfio_group *group = device->group;
267 	struct iommu_group *iommu_group;
268 
269 	if (group->type == VFIO_NO_IOMMU || group->type == VFIO_EMULATED_IOMMU)
270 		iommu_group_remove_device(device->dev);
271 
272 	/* Pairs with vfio_create_group() / vfio_group_get_from_iommu() */
273 	if (!refcount_dec_and_mutex_lock(&group->drivers, &vfio.group_lock))
274 		return;
275 	list_del(&group->vfio_next);
276 
277 	/*
278 	 * We could concurrently probe another driver in the group that might
279 	 * race vfio_device_remove_group() with vfio_get_group(), so we have to
280 	 * ensure that the sysfs is all cleaned up under lock otherwise the
281 	 * cdev_device_add() will fail due to the name aready existing.
282 	 */
283 	cdev_device_del(&group->cdev, &group->dev);
284 
285 	mutex_lock(&group->group_lock);
286 	/*
287 	 * These data structures all have paired operations that can only be
288 	 * undone when the caller holds a live reference on the device. Since
289 	 * all pairs must be undone these WARN_ON's indicate some caller did not
290 	 * properly hold the group reference.
291 	 */
292 	WARN_ON(!list_empty(&group->device_list));
293 	WARN_ON(group->notifier.head);
294 
295 	/*
296 	 * Revoke all users of group->iommu_group. At this point we know there
297 	 * are no devices active because we are unplugging the last one. Setting
298 	 * iommu_group to NULL blocks all new users.
299 	 */
300 	if (group->container)
301 		vfio_group_detach_container(group);
302 	iommu_group = group->iommu_group;
303 	group->iommu_group = NULL;
304 	mutex_unlock(&group->group_lock);
305 	mutex_unlock(&vfio.group_lock);
306 
307 	iommu_group_put(iommu_group);
308 	put_device(&group->dev);
309 }
310 
311 /*
312  * Device objects - create, release, get, put, search
313  */
314 /* Device reference always implies a group reference */
315 static void vfio_device_put_registration(struct vfio_device *device)
316 {
317 	if (refcount_dec_and_test(&device->refcount))
318 		complete(&device->comp);
319 }
320 
321 static bool vfio_device_try_get_registration(struct vfio_device *device)
322 {
323 	return refcount_inc_not_zero(&device->refcount);
324 }
325 
326 static struct vfio_device *vfio_group_get_device(struct vfio_group *group,
327 						 struct device *dev)
328 {
329 	struct vfio_device *device;
330 
331 	mutex_lock(&group->device_lock);
332 	list_for_each_entry(device, &group->device_list, group_next) {
333 		if (device->dev == dev &&
334 		    vfio_device_try_get_registration(device)) {
335 			mutex_unlock(&group->device_lock);
336 			return device;
337 		}
338 	}
339 	mutex_unlock(&group->device_lock);
340 	return NULL;
341 }
342 
343 /*
344  * VFIO driver API
345  */
346 /* Release helper called by vfio_put_device() */
347 static void vfio_device_release(struct device *dev)
348 {
349 	struct vfio_device *device =
350 			container_of(dev, struct vfio_device, device);
351 
352 	vfio_release_device_set(device);
353 	ida_free(&vfio.device_ida, device->index);
354 
355 	/*
356 	 * kvfree() cannot be done here due to a life cycle mess in
357 	 * vfio-ccw. Before the ccw part is fixed all drivers are
358 	 * required to support @release and call vfio_free_device()
359 	 * from there.
360 	 */
361 	device->ops->release(device);
362 }
363 
364 /*
365  * Allocate and initialize vfio_device so it can be registered to vfio
366  * core.
367  *
368  * Drivers should use the wrapper vfio_alloc_device() for allocation.
369  * @size is the size of the structure to be allocated, including any
370  * private data used by the driver.
371  *
372  * Driver may provide an @init callback to cover device private data.
373  *
374  * Use vfio_put_device() to release the structure after success return.
375  */
376 struct vfio_device *_vfio_alloc_device(size_t size, struct device *dev,
377 				       const struct vfio_device_ops *ops)
378 {
379 	struct vfio_device *device;
380 	int ret;
381 
382 	if (WARN_ON(size < sizeof(struct vfio_device)))
383 		return ERR_PTR(-EINVAL);
384 
385 	device = kvzalloc(size, GFP_KERNEL);
386 	if (!device)
387 		return ERR_PTR(-ENOMEM);
388 
389 	ret = vfio_init_device(device, dev, ops);
390 	if (ret)
391 		goto out_free;
392 	return device;
393 
394 out_free:
395 	kvfree(device);
396 	return ERR_PTR(ret);
397 }
398 EXPORT_SYMBOL_GPL(_vfio_alloc_device);
399 
400 /*
401  * Initialize a vfio_device so it can be registered to vfio core.
402  *
403  * Only vfio-ccw driver should call this interface.
404  */
405 int vfio_init_device(struct vfio_device *device, struct device *dev,
406 		     const struct vfio_device_ops *ops)
407 {
408 	int ret;
409 
410 	ret = ida_alloc_max(&vfio.device_ida, MINORMASK, GFP_KERNEL);
411 	if (ret < 0) {
412 		dev_dbg(dev, "Error to alloc index\n");
413 		return ret;
414 	}
415 
416 	device->index = ret;
417 	init_completion(&device->comp);
418 	device->dev = dev;
419 	device->ops = ops;
420 
421 	if (ops->init) {
422 		ret = ops->init(device);
423 		if (ret)
424 			goto out_uninit;
425 	}
426 
427 	device_initialize(&device->device);
428 	device->device.release = vfio_device_release;
429 	device->device.class = vfio.device_class;
430 	device->device.parent = device->dev;
431 	return 0;
432 
433 out_uninit:
434 	vfio_release_device_set(device);
435 	ida_free(&vfio.device_ida, device->index);
436 	return ret;
437 }
438 EXPORT_SYMBOL_GPL(vfio_init_device);
439 
440 /*
441  * The helper called by driver @release callback to free the device
442  * structure. Drivers which don't have private data to clean can
443  * simply use this helper as its @release.
444  */
445 void vfio_free_device(struct vfio_device *device)
446 {
447 	kvfree(device);
448 }
449 EXPORT_SYMBOL_GPL(vfio_free_device);
450 
451 static struct vfio_group *vfio_noiommu_group_alloc(struct device *dev,
452 		enum vfio_group_type type)
453 {
454 	struct iommu_group *iommu_group;
455 	struct vfio_group *group;
456 	int ret;
457 
458 	iommu_group = iommu_group_alloc();
459 	if (IS_ERR(iommu_group))
460 		return ERR_CAST(iommu_group);
461 
462 	ret = iommu_group_set_name(iommu_group, "vfio-noiommu");
463 	if (ret)
464 		goto out_put_group;
465 	ret = iommu_group_add_device(iommu_group, dev);
466 	if (ret)
467 		goto out_put_group;
468 
469 	group = vfio_create_group(iommu_group, type);
470 	if (IS_ERR(group)) {
471 		ret = PTR_ERR(group);
472 		goto out_remove_device;
473 	}
474 	iommu_group_put(iommu_group);
475 	return group;
476 
477 out_remove_device:
478 	iommu_group_remove_device(dev);
479 out_put_group:
480 	iommu_group_put(iommu_group);
481 	return ERR_PTR(ret);
482 }
483 
484 static struct vfio_group *vfio_group_find_or_alloc(struct device *dev)
485 {
486 	struct iommu_group *iommu_group;
487 	struct vfio_group *group;
488 
489 	iommu_group = iommu_group_get(dev);
490 	if (!iommu_group && vfio_noiommu) {
491 		/*
492 		 * With noiommu enabled, create an IOMMU group for devices that
493 		 * don't already have one, implying no IOMMU hardware/driver
494 		 * exists.  Taint the kernel because we're about to give a DMA
495 		 * capable device to a user without IOMMU protection.
496 		 */
497 		group = vfio_noiommu_group_alloc(dev, VFIO_NO_IOMMU);
498 		if (!IS_ERR(group)) {
499 			add_taint(TAINT_USER, LOCKDEP_STILL_OK);
500 			dev_warn(dev, "Adding kernel taint for vfio-noiommu group on device\n");
501 		}
502 		return group;
503 	}
504 
505 	if (!iommu_group)
506 		return ERR_PTR(-EINVAL);
507 
508 	/*
509 	 * VFIO always sets IOMMU_CACHE because we offer no way for userspace to
510 	 * restore cache coherency. It has to be checked here because it is only
511 	 * valid for cases where we are using iommu groups.
512 	 */
513 	if (!device_iommu_capable(dev, IOMMU_CAP_CACHE_COHERENCY)) {
514 		iommu_group_put(iommu_group);
515 		return ERR_PTR(-EINVAL);
516 	}
517 
518 	group = vfio_group_get_from_iommu(iommu_group);
519 	if (!group)
520 		group = vfio_create_group(iommu_group, VFIO_IOMMU);
521 
522 	/* The vfio_group holds a reference to the iommu_group */
523 	iommu_group_put(iommu_group);
524 	return group;
525 }
526 
527 static int __vfio_register_dev(struct vfio_device *device,
528 		struct vfio_group *group)
529 {
530 	struct vfio_device *existing_device;
531 	int ret;
532 
533 	/*
534 	 * In all cases group is the output of one of the group allocation
535 	 * functions and we have group->drivers incremented for us.
536 	 */
537 	if (IS_ERR(group))
538 		return PTR_ERR(group);
539 
540 	/*
541 	 * If the driver doesn't specify a set then the device is added to a
542 	 * singleton set just for itself.
543 	 */
544 	if (!device->dev_set)
545 		vfio_assign_device_set(device, device);
546 
547 	existing_device = vfio_group_get_device(group, device->dev);
548 	if (existing_device) {
549 		/*
550 		 * group->iommu_group is non-NULL because we hold the drivers
551 		 * refcount.
552 		 */
553 		dev_WARN(device->dev, "Device already exists on group %d\n",
554 			 iommu_group_id(group->iommu_group));
555 		vfio_device_put_registration(existing_device);
556 		ret = -EBUSY;
557 		goto err_out;
558 	}
559 
560 	/* Our reference on group is moved to the device */
561 	device->group = group;
562 
563 	ret = dev_set_name(&device->device, "vfio%d", device->index);
564 	if (ret)
565 		goto err_out;
566 
567 	ret = device_add(&device->device);
568 	if (ret)
569 		goto err_out;
570 
571 	/* Refcounting can't start until the driver calls register */
572 	refcount_set(&device->refcount, 1);
573 
574 	mutex_lock(&group->device_lock);
575 	list_add(&device->group_next, &group->device_list);
576 	mutex_unlock(&group->device_lock);
577 
578 	return 0;
579 err_out:
580 	vfio_device_remove_group(device);
581 	return ret;
582 }
583 
584 int vfio_register_group_dev(struct vfio_device *device)
585 {
586 	return __vfio_register_dev(device,
587 		vfio_group_find_or_alloc(device->dev));
588 }
589 EXPORT_SYMBOL_GPL(vfio_register_group_dev);
590 
591 /*
592  * Register a virtual device without IOMMU backing.  The user of this
593  * device must not be able to directly trigger unmediated DMA.
594  */
595 int vfio_register_emulated_iommu_dev(struct vfio_device *device)
596 {
597 	return __vfio_register_dev(device,
598 		vfio_noiommu_group_alloc(device->dev, VFIO_EMULATED_IOMMU));
599 }
600 EXPORT_SYMBOL_GPL(vfio_register_emulated_iommu_dev);
601 
602 static struct vfio_device *vfio_device_get_from_name(struct vfio_group *group,
603 						     char *buf)
604 {
605 	struct vfio_device *it, *device = ERR_PTR(-ENODEV);
606 
607 	mutex_lock(&group->device_lock);
608 	list_for_each_entry(it, &group->device_list, group_next) {
609 		int ret;
610 
611 		if (it->ops->match) {
612 			ret = it->ops->match(it, buf);
613 			if (ret < 0) {
614 				device = ERR_PTR(ret);
615 				break;
616 			}
617 		} else {
618 			ret = !strcmp(dev_name(it->dev), buf);
619 		}
620 
621 		if (ret && vfio_device_try_get_registration(it)) {
622 			device = it;
623 			break;
624 		}
625 	}
626 	mutex_unlock(&group->device_lock);
627 
628 	return device;
629 }
630 
631 /*
632  * Decrement the device reference count and wait for the device to be
633  * removed.  Open file descriptors for the device... */
634 void vfio_unregister_group_dev(struct vfio_device *device)
635 {
636 	struct vfio_group *group = device->group;
637 	unsigned int i = 0;
638 	bool interrupted = false;
639 	long rc;
640 
641 	vfio_device_put_registration(device);
642 	rc = try_wait_for_completion(&device->comp);
643 	while (rc <= 0) {
644 		if (device->ops->request)
645 			device->ops->request(device, i++);
646 
647 		if (interrupted) {
648 			rc = wait_for_completion_timeout(&device->comp,
649 							 HZ * 10);
650 		} else {
651 			rc = wait_for_completion_interruptible_timeout(
652 				&device->comp, HZ * 10);
653 			if (rc < 0) {
654 				interrupted = true;
655 				dev_warn(device->dev,
656 					 "Device is currently in use, task"
657 					 " \"%s\" (%d) "
658 					 "blocked until device is released",
659 					 current->comm, task_pid_nr(current));
660 			}
661 		}
662 	}
663 
664 	mutex_lock(&group->device_lock);
665 	list_del(&device->group_next);
666 	mutex_unlock(&group->device_lock);
667 
668 	/* Balances device_add in register path */
669 	device_del(&device->device);
670 
671 	vfio_device_remove_group(device);
672 }
673 EXPORT_SYMBOL_GPL(vfio_unregister_group_dev);
674 
675 /*
676  * VFIO Group fd, /dev/vfio/$GROUP
677  */
678 /*
679  * VFIO_GROUP_UNSET_CONTAINER should fail if there are other users or
680  * if there was no container to unset.  Since the ioctl is called on
681  * the group, we know that still exists, therefore the only valid
682  * transition here is 1->0.
683  */
684 static int vfio_group_ioctl_unset_container(struct vfio_group *group)
685 {
686 	int ret = 0;
687 
688 	mutex_lock(&group->group_lock);
689 	if (!group->container) {
690 		ret = -EINVAL;
691 		goto out_unlock;
692 	}
693 	if (group->container_users != 1) {
694 		ret = -EBUSY;
695 		goto out_unlock;
696 	}
697 	vfio_group_detach_container(group);
698 
699 out_unlock:
700 	mutex_unlock(&group->group_lock);
701 	return ret;
702 }
703 
704 static int vfio_group_ioctl_set_container(struct vfio_group *group,
705 					  int __user *arg)
706 {
707 	struct vfio_container *container;
708 	struct fd f;
709 	int ret;
710 	int fd;
711 
712 	if (get_user(fd, arg))
713 		return -EFAULT;
714 
715 	f = fdget(fd);
716 	if (!f.file)
717 		return -EBADF;
718 
719 	mutex_lock(&group->group_lock);
720 	if (group->container || WARN_ON(group->container_users)) {
721 		ret = -EINVAL;
722 		goto out_unlock;
723 	}
724 	if (!group->iommu_group) {
725 		ret = -ENODEV;
726 		goto out_unlock;
727 	}
728 
729 	container = vfio_container_from_file(f.file);
730 	ret = -EINVAL;
731 	if (container) {
732 		ret = vfio_container_attach_group(container, group);
733 		goto out_unlock;
734 	}
735 
736 out_unlock:
737 	mutex_unlock(&group->group_lock);
738 	fdput(f);
739 	return ret;
740 }
741 
742 static const struct file_operations vfio_device_fops;
743 
744 /* true if the vfio_device has open_device() called but not close_device() */
745 bool vfio_assert_device_open(struct vfio_device *device)
746 {
747 	return !WARN_ON_ONCE(!READ_ONCE(device->open_count));
748 }
749 
750 static struct file *vfio_device_open(struct vfio_device *device)
751 {
752 	struct file *filep;
753 	int ret;
754 
755 	mutex_lock(&device->group->group_lock);
756 	ret = vfio_device_assign_container(device);
757 	mutex_unlock(&device->group->group_lock);
758 	if (ret)
759 		return ERR_PTR(ret);
760 
761 	if (!try_module_get(device->dev->driver->owner)) {
762 		ret = -ENODEV;
763 		goto err_unassign_container;
764 	}
765 
766 	mutex_lock(&device->dev_set->lock);
767 	device->open_count++;
768 	if (device->open_count == 1) {
769 		/*
770 		 * Here we pass the KVM pointer with the group under the read
771 		 * lock.  If the device driver will use it, it must obtain a
772 		 * reference and release it during close_device.
773 		 */
774 		mutex_lock(&device->group->group_lock);
775 		device->kvm = device->group->kvm;
776 
777 		if (device->ops->open_device) {
778 			ret = device->ops->open_device(device);
779 			if (ret)
780 				goto err_undo_count;
781 		}
782 		vfio_device_container_register(device);
783 		mutex_unlock(&device->group->group_lock);
784 	}
785 	mutex_unlock(&device->dev_set->lock);
786 
787 	/*
788 	 * We can't use anon_inode_getfd() because we need to modify
789 	 * the f_mode flags directly to allow more than just ioctls
790 	 */
791 	filep = anon_inode_getfile("[vfio-device]", &vfio_device_fops,
792 				   device, O_RDWR);
793 	if (IS_ERR(filep)) {
794 		ret = PTR_ERR(filep);
795 		goto err_close_device;
796 	}
797 
798 	/*
799 	 * TODO: add an anon_inode interface to do this.
800 	 * Appears to be missing by lack of need rather than
801 	 * explicitly prevented.  Now there's need.
802 	 */
803 	filep->f_mode |= (FMODE_PREAD | FMODE_PWRITE);
804 
805 	if (device->group->type == VFIO_NO_IOMMU)
806 		dev_warn(device->dev, "vfio-noiommu device opened by user "
807 			 "(%s:%d)\n", current->comm, task_pid_nr(current));
808 	/*
809 	 * On success the ref of device is moved to the file and
810 	 * put in vfio_device_fops_release()
811 	 */
812 	return filep;
813 
814 err_close_device:
815 	mutex_lock(&device->dev_set->lock);
816 	mutex_lock(&device->group->group_lock);
817 	if (device->open_count == 1) {
818 		if (device->ops->close_device)
819 			device->ops->close_device(device);
820 
821 		vfio_device_container_unregister(device);
822 	}
823 err_undo_count:
824 	mutex_unlock(&device->group->group_lock);
825 	device->open_count--;
826 	if (device->open_count == 0 && device->kvm)
827 		device->kvm = NULL;
828 	mutex_unlock(&device->dev_set->lock);
829 	module_put(device->dev->driver->owner);
830 err_unassign_container:
831 	vfio_device_unassign_container(device);
832 	return ERR_PTR(ret);
833 }
834 
835 static int vfio_group_ioctl_get_device_fd(struct vfio_group *group,
836 					  char __user *arg)
837 {
838 	struct vfio_device *device;
839 	struct file *filep;
840 	char *buf;
841 	int fdno;
842 	int ret;
843 
844 	buf = strndup_user(arg, PAGE_SIZE);
845 	if (IS_ERR(buf))
846 		return PTR_ERR(buf);
847 
848 	device = vfio_device_get_from_name(group, buf);
849 	kfree(buf);
850 	if (IS_ERR(device))
851 		return PTR_ERR(device);
852 
853 	fdno = get_unused_fd_flags(O_CLOEXEC);
854 	if (fdno < 0) {
855 		ret = fdno;
856 		goto err_put_device;
857 	}
858 
859 	filep = vfio_device_open(device);
860 	if (IS_ERR(filep)) {
861 		ret = PTR_ERR(filep);
862 		goto err_put_fdno;
863 	}
864 
865 	fd_install(fdno, filep);
866 	return fdno;
867 
868 err_put_fdno:
869 	put_unused_fd(fdno);
870 err_put_device:
871 	vfio_device_put_registration(device);
872 	return ret;
873 }
874 
875 static int vfio_group_ioctl_get_status(struct vfio_group *group,
876 				       struct vfio_group_status __user *arg)
877 {
878 	unsigned long minsz = offsetofend(struct vfio_group_status, flags);
879 	struct vfio_group_status status;
880 
881 	if (copy_from_user(&status, arg, minsz))
882 		return -EFAULT;
883 
884 	if (status.argsz < minsz)
885 		return -EINVAL;
886 
887 	status.flags = 0;
888 
889 	mutex_lock(&group->group_lock);
890 	if (!group->iommu_group) {
891 		mutex_unlock(&group->group_lock);
892 		return -ENODEV;
893 	}
894 
895 	if (group->container)
896 		status.flags |= VFIO_GROUP_FLAGS_CONTAINER_SET |
897 				VFIO_GROUP_FLAGS_VIABLE;
898 	else if (!iommu_group_dma_owner_claimed(group->iommu_group))
899 		status.flags |= VFIO_GROUP_FLAGS_VIABLE;
900 	mutex_unlock(&group->group_lock);
901 
902 	if (copy_to_user(arg, &status, minsz))
903 		return -EFAULT;
904 	return 0;
905 }
906 
907 static long vfio_group_fops_unl_ioctl(struct file *filep,
908 				      unsigned int cmd, unsigned long arg)
909 {
910 	struct vfio_group *group = filep->private_data;
911 	void __user *uarg = (void __user *)arg;
912 
913 	switch (cmd) {
914 	case VFIO_GROUP_GET_DEVICE_FD:
915 		return vfio_group_ioctl_get_device_fd(group, uarg);
916 	case VFIO_GROUP_GET_STATUS:
917 		return vfio_group_ioctl_get_status(group, uarg);
918 	case VFIO_GROUP_SET_CONTAINER:
919 		return vfio_group_ioctl_set_container(group, uarg);
920 	case VFIO_GROUP_UNSET_CONTAINER:
921 		return vfio_group_ioctl_unset_container(group);
922 	default:
923 		return -ENOTTY;
924 	}
925 }
926 
927 static int vfio_group_fops_open(struct inode *inode, struct file *filep)
928 {
929 	struct vfio_group *group =
930 		container_of(inode->i_cdev, struct vfio_group, cdev);
931 	int ret;
932 
933 	mutex_lock(&group->group_lock);
934 
935 	/*
936 	 * drivers can be zero if this races with vfio_device_remove_group(), it
937 	 * will be stable at 0 under the group rwsem
938 	 */
939 	if (refcount_read(&group->drivers) == 0) {
940 		ret = -ENODEV;
941 		goto out_unlock;
942 	}
943 
944 	if (group->type == VFIO_NO_IOMMU && !capable(CAP_SYS_RAWIO)) {
945 		ret = -EPERM;
946 		goto out_unlock;
947 	}
948 
949 	/*
950 	 * Do we need multiple instances of the group open?  Seems not.
951 	 */
952 	if (group->opened_file) {
953 		ret = -EBUSY;
954 		goto out_unlock;
955 	}
956 	group->opened_file = filep;
957 	filep->private_data = group;
958 	ret = 0;
959 out_unlock:
960 	mutex_unlock(&group->group_lock);
961 	return ret;
962 }
963 
964 static int vfio_group_fops_release(struct inode *inode, struct file *filep)
965 {
966 	struct vfio_group *group = filep->private_data;
967 
968 	filep->private_data = NULL;
969 
970 	mutex_lock(&group->group_lock);
971 	/*
972 	 * Device FDs hold a group file reference, therefore the group release
973 	 * is only called when there are no open devices.
974 	 */
975 	WARN_ON(group->notifier.head);
976 	if (group->container)
977 		vfio_group_detach_container(group);
978 	group->opened_file = NULL;
979 	mutex_unlock(&group->group_lock);
980 	return 0;
981 }
982 
983 static const struct file_operations vfio_group_fops = {
984 	.owner		= THIS_MODULE,
985 	.unlocked_ioctl	= vfio_group_fops_unl_ioctl,
986 	.compat_ioctl	= compat_ptr_ioctl,
987 	.open		= vfio_group_fops_open,
988 	.release	= vfio_group_fops_release,
989 };
990 
991 /*
992  * Wrapper around pm_runtime_resume_and_get().
993  * Return error code on failure or 0 on success.
994  */
995 static inline int vfio_device_pm_runtime_get(struct vfio_device *device)
996 {
997 	struct device *dev = device->dev;
998 
999 	if (dev->driver && dev->driver->pm) {
1000 		int ret;
1001 
1002 		ret = pm_runtime_resume_and_get(dev);
1003 		if (ret) {
1004 			dev_info_ratelimited(dev,
1005 				"vfio: runtime resume failed %d\n", ret);
1006 			return -EIO;
1007 		}
1008 	}
1009 
1010 	return 0;
1011 }
1012 
1013 /*
1014  * Wrapper around pm_runtime_put().
1015  */
1016 static inline void vfio_device_pm_runtime_put(struct vfio_device *device)
1017 {
1018 	struct device *dev = device->dev;
1019 
1020 	if (dev->driver && dev->driver->pm)
1021 		pm_runtime_put(dev);
1022 }
1023 
1024 /*
1025  * VFIO Device fd
1026  */
1027 static int vfio_device_fops_release(struct inode *inode, struct file *filep)
1028 {
1029 	struct vfio_device *device = filep->private_data;
1030 
1031 	mutex_lock(&device->dev_set->lock);
1032 	vfio_assert_device_open(device);
1033 	mutex_lock(&device->group->group_lock);
1034 	if (device->open_count == 1) {
1035 		if (device->ops->close_device)
1036 			device->ops->close_device(device);
1037 
1038 		vfio_device_container_unregister(device);
1039 	}
1040 	mutex_unlock(&device->group->group_lock);
1041 	device->open_count--;
1042 	if (device->open_count == 0)
1043 		device->kvm = NULL;
1044 	mutex_unlock(&device->dev_set->lock);
1045 
1046 	module_put(device->dev->driver->owner);
1047 
1048 	vfio_device_unassign_container(device);
1049 
1050 	vfio_device_put_registration(device);
1051 
1052 	return 0;
1053 }
1054 
1055 /*
1056  * vfio_mig_get_next_state - Compute the next step in the FSM
1057  * @cur_fsm - The current state the device is in
1058  * @new_fsm - The target state to reach
1059  * @next_fsm - Pointer to the next step to get to new_fsm
1060  *
1061  * Return 0 upon success, otherwise -errno
1062  * Upon success the next step in the state progression between cur_fsm and
1063  * new_fsm will be set in next_fsm.
1064  *
1065  * This breaks down requests for combination transitions into smaller steps and
1066  * returns the next step to get to new_fsm. The function may need to be called
1067  * multiple times before reaching new_fsm.
1068  *
1069  */
1070 int vfio_mig_get_next_state(struct vfio_device *device,
1071 			    enum vfio_device_mig_state cur_fsm,
1072 			    enum vfio_device_mig_state new_fsm,
1073 			    enum vfio_device_mig_state *next_fsm)
1074 {
1075 	enum { VFIO_DEVICE_NUM_STATES = VFIO_DEVICE_STATE_RUNNING_P2P + 1 };
1076 	/*
1077 	 * The coding in this table requires the driver to implement the
1078 	 * following FSM arcs:
1079 	 *         RESUMING -> STOP
1080 	 *         STOP -> RESUMING
1081 	 *         STOP -> STOP_COPY
1082 	 *         STOP_COPY -> STOP
1083 	 *
1084 	 * If P2P is supported then the driver must also implement these FSM
1085 	 * arcs:
1086 	 *         RUNNING -> RUNNING_P2P
1087 	 *         RUNNING_P2P -> RUNNING
1088 	 *         RUNNING_P2P -> STOP
1089 	 *         STOP -> RUNNING_P2P
1090 	 * Without P2P the driver must implement:
1091 	 *         RUNNING -> STOP
1092 	 *         STOP -> RUNNING
1093 	 *
1094 	 * The coding will step through multiple states for some combination
1095 	 * transitions; if all optional features are supported, this means the
1096 	 * following ones:
1097 	 *         RESUMING -> STOP -> RUNNING_P2P
1098 	 *         RESUMING -> STOP -> RUNNING_P2P -> RUNNING
1099 	 *         RESUMING -> STOP -> STOP_COPY
1100 	 *         RUNNING -> RUNNING_P2P -> STOP
1101 	 *         RUNNING -> RUNNING_P2P -> STOP -> RESUMING
1102 	 *         RUNNING -> RUNNING_P2P -> STOP -> STOP_COPY
1103 	 *         RUNNING_P2P -> STOP -> RESUMING
1104 	 *         RUNNING_P2P -> STOP -> STOP_COPY
1105 	 *         STOP -> RUNNING_P2P -> RUNNING
1106 	 *         STOP_COPY -> STOP -> RESUMING
1107 	 *         STOP_COPY -> STOP -> RUNNING_P2P
1108 	 *         STOP_COPY -> STOP -> RUNNING_P2P -> RUNNING
1109 	 */
1110 	static const u8 vfio_from_fsm_table[VFIO_DEVICE_NUM_STATES][VFIO_DEVICE_NUM_STATES] = {
1111 		[VFIO_DEVICE_STATE_STOP] = {
1112 			[VFIO_DEVICE_STATE_STOP] = VFIO_DEVICE_STATE_STOP,
1113 			[VFIO_DEVICE_STATE_RUNNING] = VFIO_DEVICE_STATE_RUNNING_P2P,
1114 			[VFIO_DEVICE_STATE_STOP_COPY] = VFIO_DEVICE_STATE_STOP_COPY,
1115 			[VFIO_DEVICE_STATE_RESUMING] = VFIO_DEVICE_STATE_RESUMING,
1116 			[VFIO_DEVICE_STATE_RUNNING_P2P] = VFIO_DEVICE_STATE_RUNNING_P2P,
1117 			[VFIO_DEVICE_STATE_ERROR] = VFIO_DEVICE_STATE_ERROR,
1118 		},
1119 		[VFIO_DEVICE_STATE_RUNNING] = {
1120 			[VFIO_DEVICE_STATE_STOP] = VFIO_DEVICE_STATE_RUNNING_P2P,
1121 			[VFIO_DEVICE_STATE_RUNNING] = VFIO_DEVICE_STATE_RUNNING,
1122 			[VFIO_DEVICE_STATE_STOP_COPY] = VFIO_DEVICE_STATE_RUNNING_P2P,
1123 			[VFIO_DEVICE_STATE_RESUMING] = VFIO_DEVICE_STATE_RUNNING_P2P,
1124 			[VFIO_DEVICE_STATE_RUNNING_P2P] = VFIO_DEVICE_STATE_RUNNING_P2P,
1125 			[VFIO_DEVICE_STATE_ERROR] = VFIO_DEVICE_STATE_ERROR,
1126 		},
1127 		[VFIO_DEVICE_STATE_STOP_COPY] = {
1128 			[VFIO_DEVICE_STATE_STOP] = VFIO_DEVICE_STATE_STOP,
1129 			[VFIO_DEVICE_STATE_RUNNING] = VFIO_DEVICE_STATE_STOP,
1130 			[VFIO_DEVICE_STATE_STOP_COPY] = VFIO_DEVICE_STATE_STOP_COPY,
1131 			[VFIO_DEVICE_STATE_RESUMING] = VFIO_DEVICE_STATE_STOP,
1132 			[VFIO_DEVICE_STATE_RUNNING_P2P] = VFIO_DEVICE_STATE_STOP,
1133 			[VFIO_DEVICE_STATE_ERROR] = VFIO_DEVICE_STATE_ERROR,
1134 		},
1135 		[VFIO_DEVICE_STATE_RESUMING] = {
1136 			[VFIO_DEVICE_STATE_STOP] = VFIO_DEVICE_STATE_STOP,
1137 			[VFIO_DEVICE_STATE_RUNNING] = VFIO_DEVICE_STATE_STOP,
1138 			[VFIO_DEVICE_STATE_STOP_COPY] = VFIO_DEVICE_STATE_STOP,
1139 			[VFIO_DEVICE_STATE_RESUMING] = VFIO_DEVICE_STATE_RESUMING,
1140 			[VFIO_DEVICE_STATE_RUNNING_P2P] = VFIO_DEVICE_STATE_STOP,
1141 			[VFIO_DEVICE_STATE_ERROR] = VFIO_DEVICE_STATE_ERROR,
1142 		},
1143 		[VFIO_DEVICE_STATE_RUNNING_P2P] = {
1144 			[VFIO_DEVICE_STATE_STOP] = VFIO_DEVICE_STATE_STOP,
1145 			[VFIO_DEVICE_STATE_RUNNING] = VFIO_DEVICE_STATE_RUNNING,
1146 			[VFIO_DEVICE_STATE_STOP_COPY] = VFIO_DEVICE_STATE_STOP,
1147 			[VFIO_DEVICE_STATE_RESUMING] = VFIO_DEVICE_STATE_STOP,
1148 			[VFIO_DEVICE_STATE_RUNNING_P2P] = VFIO_DEVICE_STATE_RUNNING_P2P,
1149 			[VFIO_DEVICE_STATE_ERROR] = VFIO_DEVICE_STATE_ERROR,
1150 		},
1151 		[VFIO_DEVICE_STATE_ERROR] = {
1152 			[VFIO_DEVICE_STATE_STOP] = VFIO_DEVICE_STATE_ERROR,
1153 			[VFIO_DEVICE_STATE_RUNNING] = VFIO_DEVICE_STATE_ERROR,
1154 			[VFIO_DEVICE_STATE_STOP_COPY] = VFIO_DEVICE_STATE_ERROR,
1155 			[VFIO_DEVICE_STATE_RESUMING] = VFIO_DEVICE_STATE_ERROR,
1156 			[VFIO_DEVICE_STATE_RUNNING_P2P] = VFIO_DEVICE_STATE_ERROR,
1157 			[VFIO_DEVICE_STATE_ERROR] = VFIO_DEVICE_STATE_ERROR,
1158 		},
1159 	};
1160 
1161 	static const unsigned int state_flags_table[VFIO_DEVICE_NUM_STATES] = {
1162 		[VFIO_DEVICE_STATE_STOP] = VFIO_MIGRATION_STOP_COPY,
1163 		[VFIO_DEVICE_STATE_RUNNING] = VFIO_MIGRATION_STOP_COPY,
1164 		[VFIO_DEVICE_STATE_STOP_COPY] = VFIO_MIGRATION_STOP_COPY,
1165 		[VFIO_DEVICE_STATE_RESUMING] = VFIO_MIGRATION_STOP_COPY,
1166 		[VFIO_DEVICE_STATE_RUNNING_P2P] =
1167 			VFIO_MIGRATION_STOP_COPY | VFIO_MIGRATION_P2P,
1168 		[VFIO_DEVICE_STATE_ERROR] = ~0U,
1169 	};
1170 
1171 	if (WARN_ON(cur_fsm >= ARRAY_SIZE(vfio_from_fsm_table) ||
1172 		    (state_flags_table[cur_fsm] & device->migration_flags) !=
1173 			state_flags_table[cur_fsm]))
1174 		return -EINVAL;
1175 
1176 	if (new_fsm >= ARRAY_SIZE(vfio_from_fsm_table) ||
1177 	   (state_flags_table[new_fsm] & device->migration_flags) !=
1178 			state_flags_table[new_fsm])
1179 		return -EINVAL;
1180 
1181 	/*
1182 	 * Arcs touching optional and unsupported states are skipped over. The
1183 	 * driver will instead see an arc from the original state to the next
1184 	 * logical state, as per the above comment.
1185 	 */
1186 	*next_fsm = vfio_from_fsm_table[cur_fsm][new_fsm];
1187 	while ((state_flags_table[*next_fsm] & device->migration_flags) !=
1188 			state_flags_table[*next_fsm])
1189 		*next_fsm = vfio_from_fsm_table[*next_fsm][new_fsm];
1190 
1191 	return (*next_fsm != VFIO_DEVICE_STATE_ERROR) ? 0 : -EINVAL;
1192 }
1193 EXPORT_SYMBOL_GPL(vfio_mig_get_next_state);
1194 
1195 /*
1196  * Convert the drivers's struct file into a FD number and return it to userspace
1197  */
1198 static int vfio_ioct_mig_return_fd(struct file *filp, void __user *arg,
1199 				   struct vfio_device_feature_mig_state *mig)
1200 {
1201 	int ret;
1202 	int fd;
1203 
1204 	fd = get_unused_fd_flags(O_CLOEXEC);
1205 	if (fd < 0) {
1206 		ret = fd;
1207 		goto out_fput;
1208 	}
1209 
1210 	mig->data_fd = fd;
1211 	if (copy_to_user(arg, mig, sizeof(*mig))) {
1212 		ret = -EFAULT;
1213 		goto out_put_unused;
1214 	}
1215 	fd_install(fd, filp);
1216 	return 0;
1217 
1218 out_put_unused:
1219 	put_unused_fd(fd);
1220 out_fput:
1221 	fput(filp);
1222 	return ret;
1223 }
1224 
1225 static int
1226 vfio_ioctl_device_feature_mig_device_state(struct vfio_device *device,
1227 					   u32 flags, void __user *arg,
1228 					   size_t argsz)
1229 {
1230 	size_t minsz =
1231 		offsetofend(struct vfio_device_feature_mig_state, data_fd);
1232 	struct vfio_device_feature_mig_state mig;
1233 	struct file *filp = NULL;
1234 	int ret;
1235 
1236 	if (!device->mig_ops)
1237 		return -ENOTTY;
1238 
1239 	ret = vfio_check_feature(flags, argsz,
1240 				 VFIO_DEVICE_FEATURE_SET |
1241 				 VFIO_DEVICE_FEATURE_GET,
1242 				 sizeof(mig));
1243 	if (ret != 1)
1244 		return ret;
1245 
1246 	if (copy_from_user(&mig, arg, minsz))
1247 		return -EFAULT;
1248 
1249 	if (flags & VFIO_DEVICE_FEATURE_GET) {
1250 		enum vfio_device_mig_state curr_state;
1251 
1252 		ret = device->mig_ops->migration_get_state(device,
1253 							   &curr_state);
1254 		if (ret)
1255 			return ret;
1256 		mig.device_state = curr_state;
1257 		goto out_copy;
1258 	}
1259 
1260 	/* Handle the VFIO_DEVICE_FEATURE_SET */
1261 	filp = device->mig_ops->migration_set_state(device, mig.device_state);
1262 	if (IS_ERR(filp) || !filp)
1263 		goto out_copy;
1264 
1265 	return vfio_ioct_mig_return_fd(filp, arg, &mig);
1266 out_copy:
1267 	mig.data_fd = -1;
1268 	if (copy_to_user(arg, &mig, sizeof(mig)))
1269 		return -EFAULT;
1270 	if (IS_ERR(filp))
1271 		return PTR_ERR(filp);
1272 	return 0;
1273 }
1274 
1275 static int vfio_ioctl_device_feature_migration(struct vfio_device *device,
1276 					       u32 flags, void __user *arg,
1277 					       size_t argsz)
1278 {
1279 	struct vfio_device_feature_migration mig = {
1280 		.flags = device->migration_flags,
1281 	};
1282 	int ret;
1283 
1284 	if (!device->mig_ops)
1285 		return -ENOTTY;
1286 
1287 	ret = vfio_check_feature(flags, argsz, VFIO_DEVICE_FEATURE_GET,
1288 				 sizeof(mig));
1289 	if (ret != 1)
1290 		return ret;
1291 	if (copy_to_user(arg, &mig, sizeof(mig)))
1292 		return -EFAULT;
1293 	return 0;
1294 }
1295 
1296 /* Ranges should fit into a single kernel page */
1297 #define LOG_MAX_RANGES \
1298 	(PAGE_SIZE / sizeof(struct vfio_device_feature_dma_logging_range))
1299 
1300 static int
1301 vfio_ioctl_device_feature_logging_start(struct vfio_device *device,
1302 					u32 flags, void __user *arg,
1303 					size_t argsz)
1304 {
1305 	size_t minsz =
1306 		offsetofend(struct vfio_device_feature_dma_logging_control,
1307 			    ranges);
1308 	struct vfio_device_feature_dma_logging_range __user *ranges;
1309 	struct vfio_device_feature_dma_logging_control control;
1310 	struct vfio_device_feature_dma_logging_range range;
1311 	struct rb_root_cached root = RB_ROOT_CACHED;
1312 	struct interval_tree_node *nodes;
1313 	u64 iova_end;
1314 	u32 nnodes;
1315 	int i, ret;
1316 
1317 	if (!device->log_ops)
1318 		return -ENOTTY;
1319 
1320 	ret = vfio_check_feature(flags, argsz,
1321 				 VFIO_DEVICE_FEATURE_SET,
1322 				 sizeof(control));
1323 	if (ret != 1)
1324 		return ret;
1325 
1326 	if (copy_from_user(&control, arg, minsz))
1327 		return -EFAULT;
1328 
1329 	nnodes = control.num_ranges;
1330 	if (!nnodes)
1331 		return -EINVAL;
1332 
1333 	if (nnodes > LOG_MAX_RANGES)
1334 		return -E2BIG;
1335 
1336 	ranges = u64_to_user_ptr(control.ranges);
1337 	nodes = kmalloc_array(nnodes, sizeof(struct interval_tree_node),
1338 			      GFP_KERNEL);
1339 	if (!nodes)
1340 		return -ENOMEM;
1341 
1342 	for (i = 0; i < nnodes; i++) {
1343 		if (copy_from_user(&range, &ranges[i], sizeof(range))) {
1344 			ret = -EFAULT;
1345 			goto end;
1346 		}
1347 		if (!IS_ALIGNED(range.iova, control.page_size) ||
1348 		    !IS_ALIGNED(range.length, control.page_size)) {
1349 			ret = -EINVAL;
1350 			goto end;
1351 		}
1352 
1353 		if (check_add_overflow(range.iova, range.length, &iova_end) ||
1354 		    iova_end > ULONG_MAX) {
1355 			ret = -EOVERFLOW;
1356 			goto end;
1357 		}
1358 
1359 		nodes[i].start = range.iova;
1360 		nodes[i].last = range.iova + range.length - 1;
1361 		if (interval_tree_iter_first(&root, nodes[i].start,
1362 					     nodes[i].last)) {
1363 			/* Range overlapping */
1364 			ret = -EINVAL;
1365 			goto end;
1366 		}
1367 		interval_tree_insert(nodes + i, &root);
1368 	}
1369 
1370 	ret = device->log_ops->log_start(device, &root, nnodes,
1371 					 &control.page_size);
1372 	if (ret)
1373 		goto end;
1374 
1375 	if (copy_to_user(arg, &control, sizeof(control))) {
1376 		ret = -EFAULT;
1377 		device->log_ops->log_stop(device);
1378 	}
1379 
1380 end:
1381 	kfree(nodes);
1382 	return ret;
1383 }
1384 
1385 static int
1386 vfio_ioctl_device_feature_logging_stop(struct vfio_device *device,
1387 				       u32 flags, void __user *arg,
1388 				       size_t argsz)
1389 {
1390 	int ret;
1391 
1392 	if (!device->log_ops)
1393 		return -ENOTTY;
1394 
1395 	ret = vfio_check_feature(flags, argsz,
1396 				 VFIO_DEVICE_FEATURE_SET, 0);
1397 	if (ret != 1)
1398 		return ret;
1399 
1400 	return device->log_ops->log_stop(device);
1401 }
1402 
1403 static int vfio_device_log_read_and_clear(struct iova_bitmap *iter,
1404 					  unsigned long iova, size_t length,
1405 					  void *opaque)
1406 {
1407 	struct vfio_device *device = opaque;
1408 
1409 	return device->log_ops->log_read_and_clear(device, iova, length, iter);
1410 }
1411 
1412 static int
1413 vfio_ioctl_device_feature_logging_report(struct vfio_device *device,
1414 					 u32 flags, void __user *arg,
1415 					 size_t argsz)
1416 {
1417 	size_t minsz =
1418 		offsetofend(struct vfio_device_feature_dma_logging_report,
1419 			    bitmap);
1420 	struct vfio_device_feature_dma_logging_report report;
1421 	struct iova_bitmap *iter;
1422 	u64 iova_end;
1423 	int ret;
1424 
1425 	if (!device->log_ops)
1426 		return -ENOTTY;
1427 
1428 	ret = vfio_check_feature(flags, argsz,
1429 				 VFIO_DEVICE_FEATURE_GET,
1430 				 sizeof(report));
1431 	if (ret != 1)
1432 		return ret;
1433 
1434 	if (copy_from_user(&report, arg, minsz))
1435 		return -EFAULT;
1436 
1437 	if (report.page_size < SZ_4K || !is_power_of_2(report.page_size))
1438 		return -EINVAL;
1439 
1440 	if (check_add_overflow(report.iova, report.length, &iova_end) ||
1441 	    iova_end > ULONG_MAX)
1442 		return -EOVERFLOW;
1443 
1444 	iter = iova_bitmap_alloc(report.iova, report.length,
1445 				 report.page_size,
1446 				 u64_to_user_ptr(report.bitmap));
1447 	if (IS_ERR(iter))
1448 		return PTR_ERR(iter);
1449 
1450 	ret = iova_bitmap_for_each(iter, device,
1451 				   vfio_device_log_read_and_clear);
1452 
1453 	iova_bitmap_free(iter);
1454 	return ret;
1455 }
1456 
1457 static int vfio_ioctl_device_feature(struct vfio_device *device,
1458 				     struct vfio_device_feature __user *arg)
1459 {
1460 	size_t minsz = offsetofend(struct vfio_device_feature, flags);
1461 	struct vfio_device_feature feature;
1462 
1463 	if (copy_from_user(&feature, arg, minsz))
1464 		return -EFAULT;
1465 
1466 	if (feature.argsz < minsz)
1467 		return -EINVAL;
1468 
1469 	/* Check unknown flags */
1470 	if (feature.flags &
1471 	    ~(VFIO_DEVICE_FEATURE_MASK | VFIO_DEVICE_FEATURE_SET |
1472 	      VFIO_DEVICE_FEATURE_GET | VFIO_DEVICE_FEATURE_PROBE))
1473 		return -EINVAL;
1474 
1475 	/* GET & SET are mutually exclusive except with PROBE */
1476 	if (!(feature.flags & VFIO_DEVICE_FEATURE_PROBE) &&
1477 	    (feature.flags & VFIO_DEVICE_FEATURE_SET) &&
1478 	    (feature.flags & VFIO_DEVICE_FEATURE_GET))
1479 		return -EINVAL;
1480 
1481 	switch (feature.flags & VFIO_DEVICE_FEATURE_MASK) {
1482 	case VFIO_DEVICE_FEATURE_MIGRATION:
1483 		return vfio_ioctl_device_feature_migration(
1484 			device, feature.flags, arg->data,
1485 			feature.argsz - minsz);
1486 	case VFIO_DEVICE_FEATURE_MIG_DEVICE_STATE:
1487 		return vfio_ioctl_device_feature_mig_device_state(
1488 			device, feature.flags, arg->data,
1489 			feature.argsz - minsz);
1490 	case VFIO_DEVICE_FEATURE_DMA_LOGGING_START:
1491 		return vfio_ioctl_device_feature_logging_start(
1492 			device, feature.flags, arg->data,
1493 			feature.argsz - minsz);
1494 	case VFIO_DEVICE_FEATURE_DMA_LOGGING_STOP:
1495 		return vfio_ioctl_device_feature_logging_stop(
1496 			device, feature.flags, arg->data,
1497 			feature.argsz - minsz);
1498 	case VFIO_DEVICE_FEATURE_DMA_LOGGING_REPORT:
1499 		return vfio_ioctl_device_feature_logging_report(
1500 			device, feature.flags, arg->data,
1501 			feature.argsz - minsz);
1502 	default:
1503 		if (unlikely(!device->ops->device_feature))
1504 			return -EINVAL;
1505 		return device->ops->device_feature(device, feature.flags,
1506 						   arg->data,
1507 						   feature.argsz - minsz);
1508 	}
1509 }
1510 
1511 static long vfio_device_fops_unl_ioctl(struct file *filep,
1512 				       unsigned int cmd, unsigned long arg)
1513 {
1514 	struct vfio_device *device = filep->private_data;
1515 	int ret;
1516 
1517 	ret = vfio_device_pm_runtime_get(device);
1518 	if (ret)
1519 		return ret;
1520 
1521 	switch (cmd) {
1522 	case VFIO_DEVICE_FEATURE:
1523 		ret = vfio_ioctl_device_feature(device, (void __user *)arg);
1524 		break;
1525 
1526 	default:
1527 		if (unlikely(!device->ops->ioctl))
1528 			ret = -EINVAL;
1529 		else
1530 			ret = device->ops->ioctl(device, cmd, arg);
1531 		break;
1532 	}
1533 
1534 	vfio_device_pm_runtime_put(device);
1535 	return ret;
1536 }
1537 
1538 static ssize_t vfio_device_fops_read(struct file *filep, char __user *buf,
1539 				     size_t count, loff_t *ppos)
1540 {
1541 	struct vfio_device *device = filep->private_data;
1542 
1543 	if (unlikely(!device->ops->read))
1544 		return -EINVAL;
1545 
1546 	return device->ops->read(device, buf, count, ppos);
1547 }
1548 
1549 static ssize_t vfio_device_fops_write(struct file *filep,
1550 				      const char __user *buf,
1551 				      size_t count, loff_t *ppos)
1552 {
1553 	struct vfio_device *device = filep->private_data;
1554 
1555 	if (unlikely(!device->ops->write))
1556 		return -EINVAL;
1557 
1558 	return device->ops->write(device, buf, count, ppos);
1559 }
1560 
1561 static int vfio_device_fops_mmap(struct file *filep, struct vm_area_struct *vma)
1562 {
1563 	struct vfio_device *device = filep->private_data;
1564 
1565 	if (unlikely(!device->ops->mmap))
1566 		return -EINVAL;
1567 
1568 	return device->ops->mmap(device, vma);
1569 }
1570 
1571 static const struct file_operations vfio_device_fops = {
1572 	.owner		= THIS_MODULE,
1573 	.release	= vfio_device_fops_release,
1574 	.read		= vfio_device_fops_read,
1575 	.write		= vfio_device_fops_write,
1576 	.unlocked_ioctl	= vfio_device_fops_unl_ioctl,
1577 	.compat_ioctl	= compat_ptr_ioctl,
1578 	.mmap		= vfio_device_fops_mmap,
1579 };
1580 
1581 /**
1582  * vfio_file_iommu_group - Return the struct iommu_group for the vfio group file
1583  * @file: VFIO group file
1584  *
1585  * The returned iommu_group is valid as long as a ref is held on the file. This
1586  * returns a reference on the group. This function is deprecated, only the SPAPR
1587  * path in kvm should call it.
1588  */
1589 struct iommu_group *vfio_file_iommu_group(struct file *file)
1590 {
1591 	struct vfio_group *group = file->private_data;
1592 	struct iommu_group *iommu_group = NULL;
1593 
1594 	if (!IS_ENABLED(CONFIG_SPAPR_TCE_IOMMU))
1595 		return NULL;
1596 
1597 	if (!vfio_file_is_group(file))
1598 		return NULL;
1599 
1600 	mutex_lock(&group->group_lock);
1601 	if (group->iommu_group) {
1602 		iommu_group = group->iommu_group;
1603 		iommu_group_ref_get(iommu_group);
1604 	}
1605 	mutex_unlock(&group->group_lock);
1606 	return iommu_group;
1607 }
1608 EXPORT_SYMBOL_GPL(vfio_file_iommu_group);
1609 
1610 /**
1611  * vfio_file_is_group - True if the file is usable with VFIO aPIS
1612  * @file: VFIO group file
1613  */
1614 bool vfio_file_is_group(struct file *file)
1615 {
1616 	return file->f_op == &vfio_group_fops;
1617 }
1618 EXPORT_SYMBOL_GPL(vfio_file_is_group);
1619 
1620 /**
1621  * vfio_file_enforced_coherent - True if the DMA associated with the VFIO file
1622  *        is always CPU cache coherent
1623  * @file: VFIO group file
1624  *
1625  * Enforced coherency means that the IOMMU ignores things like the PCIe no-snoop
1626  * bit in DMA transactions. A return of false indicates that the user has
1627  * rights to access additional instructions such as wbinvd on x86.
1628  */
1629 bool vfio_file_enforced_coherent(struct file *file)
1630 {
1631 	struct vfio_group *group = file->private_data;
1632 	bool ret;
1633 
1634 	if (!vfio_file_is_group(file))
1635 		return true;
1636 
1637 	mutex_lock(&group->group_lock);
1638 	if (group->container) {
1639 		ret = vfio_container_ioctl_check_extension(group->container,
1640 							   VFIO_DMA_CC_IOMMU);
1641 	} else {
1642 		/*
1643 		 * Since the coherency state is determined only once a container
1644 		 * is attached the user must do so before they can prove they
1645 		 * have permission.
1646 		 */
1647 		ret = true;
1648 	}
1649 	mutex_unlock(&group->group_lock);
1650 	return ret;
1651 }
1652 EXPORT_SYMBOL_GPL(vfio_file_enforced_coherent);
1653 
1654 /**
1655  * vfio_file_set_kvm - Link a kvm with VFIO drivers
1656  * @file: VFIO group file
1657  * @kvm: KVM to link
1658  *
1659  * When a VFIO device is first opened the KVM will be available in
1660  * device->kvm if one was associated with the group.
1661  */
1662 void vfio_file_set_kvm(struct file *file, struct kvm *kvm)
1663 {
1664 	struct vfio_group *group = file->private_data;
1665 
1666 	if (!vfio_file_is_group(file))
1667 		return;
1668 
1669 	mutex_lock(&group->group_lock);
1670 	group->kvm = kvm;
1671 	mutex_unlock(&group->group_lock);
1672 }
1673 EXPORT_SYMBOL_GPL(vfio_file_set_kvm);
1674 
1675 /**
1676  * vfio_file_has_dev - True if the VFIO file is a handle for device
1677  * @file: VFIO file to check
1678  * @device: Device that must be part of the file
1679  *
1680  * Returns true if given file has permission to manipulate the given device.
1681  */
1682 bool vfio_file_has_dev(struct file *file, struct vfio_device *device)
1683 {
1684 	struct vfio_group *group = file->private_data;
1685 
1686 	if (!vfio_file_is_group(file))
1687 		return false;
1688 
1689 	return group == device->group;
1690 }
1691 EXPORT_SYMBOL_GPL(vfio_file_has_dev);
1692 
1693 /*
1694  * Sub-module support
1695  */
1696 /*
1697  * Helper for managing a buffer of info chain capabilities, allocate or
1698  * reallocate a buffer with additional @size, filling in @id and @version
1699  * of the capability.  A pointer to the new capability is returned.
1700  *
1701  * NB. The chain is based at the head of the buffer, so new entries are
1702  * added to the tail, vfio_info_cap_shift() should be called to fixup the
1703  * next offsets prior to copying to the user buffer.
1704  */
1705 struct vfio_info_cap_header *vfio_info_cap_add(struct vfio_info_cap *caps,
1706 					       size_t size, u16 id, u16 version)
1707 {
1708 	void *buf;
1709 	struct vfio_info_cap_header *header, *tmp;
1710 
1711 	buf = krealloc(caps->buf, caps->size + size, GFP_KERNEL);
1712 	if (!buf) {
1713 		kfree(caps->buf);
1714 		caps->buf = NULL;
1715 		caps->size = 0;
1716 		return ERR_PTR(-ENOMEM);
1717 	}
1718 
1719 	caps->buf = buf;
1720 	header = buf + caps->size;
1721 
1722 	/* Eventually copied to user buffer, zero */
1723 	memset(header, 0, size);
1724 
1725 	header->id = id;
1726 	header->version = version;
1727 
1728 	/* Add to the end of the capability chain */
1729 	for (tmp = buf; tmp->next; tmp = buf + tmp->next)
1730 		; /* nothing */
1731 
1732 	tmp->next = caps->size;
1733 	caps->size += size;
1734 
1735 	return header;
1736 }
1737 EXPORT_SYMBOL_GPL(vfio_info_cap_add);
1738 
1739 void vfio_info_cap_shift(struct vfio_info_cap *caps, size_t offset)
1740 {
1741 	struct vfio_info_cap_header *tmp;
1742 	void *buf = (void *)caps->buf;
1743 
1744 	for (tmp = buf; tmp->next; tmp = buf + tmp->next - offset)
1745 		tmp->next += offset;
1746 }
1747 EXPORT_SYMBOL(vfio_info_cap_shift);
1748 
1749 int vfio_info_add_capability(struct vfio_info_cap *caps,
1750 			     struct vfio_info_cap_header *cap, size_t size)
1751 {
1752 	struct vfio_info_cap_header *header;
1753 
1754 	header = vfio_info_cap_add(caps, size, cap->id, cap->version);
1755 	if (IS_ERR(header))
1756 		return PTR_ERR(header);
1757 
1758 	memcpy(header + 1, cap + 1, size - sizeof(*header));
1759 
1760 	return 0;
1761 }
1762 EXPORT_SYMBOL(vfio_info_add_capability);
1763 
1764 int vfio_set_irqs_validate_and_prepare(struct vfio_irq_set *hdr, int num_irqs,
1765 				       int max_irq_type, size_t *data_size)
1766 {
1767 	unsigned long minsz;
1768 	size_t size;
1769 
1770 	minsz = offsetofend(struct vfio_irq_set, count);
1771 
1772 	if ((hdr->argsz < minsz) || (hdr->index >= max_irq_type) ||
1773 	    (hdr->count >= (U32_MAX - hdr->start)) ||
1774 	    (hdr->flags & ~(VFIO_IRQ_SET_DATA_TYPE_MASK |
1775 				VFIO_IRQ_SET_ACTION_TYPE_MASK)))
1776 		return -EINVAL;
1777 
1778 	if (data_size)
1779 		*data_size = 0;
1780 
1781 	if (hdr->start >= num_irqs || hdr->start + hdr->count > num_irqs)
1782 		return -EINVAL;
1783 
1784 	switch (hdr->flags & VFIO_IRQ_SET_DATA_TYPE_MASK) {
1785 	case VFIO_IRQ_SET_DATA_NONE:
1786 		size = 0;
1787 		break;
1788 	case VFIO_IRQ_SET_DATA_BOOL:
1789 		size = sizeof(uint8_t);
1790 		break;
1791 	case VFIO_IRQ_SET_DATA_EVENTFD:
1792 		size = sizeof(int32_t);
1793 		break;
1794 	default:
1795 		return -EINVAL;
1796 	}
1797 
1798 	if (size) {
1799 		if (hdr->argsz - minsz < hdr->count * size)
1800 			return -EINVAL;
1801 
1802 		if (!data_size)
1803 			return -EINVAL;
1804 
1805 		*data_size = hdr->count * size;
1806 	}
1807 
1808 	return 0;
1809 }
1810 EXPORT_SYMBOL(vfio_set_irqs_validate_and_prepare);
1811 
1812 /*
1813  * Module/class support
1814  */
1815 static char *vfio_devnode(struct device *dev, umode_t *mode)
1816 {
1817 	return kasprintf(GFP_KERNEL, "vfio/%s", dev_name(dev));
1818 }
1819 
1820 static int __init vfio_init(void)
1821 {
1822 	int ret;
1823 
1824 	ida_init(&vfio.group_ida);
1825 	ida_init(&vfio.device_ida);
1826 	mutex_init(&vfio.group_lock);
1827 	INIT_LIST_HEAD(&vfio.group_list);
1828 
1829 	ret = vfio_container_init();
1830 	if (ret)
1831 		return ret;
1832 
1833 	/* /dev/vfio/$GROUP */
1834 	vfio.class = class_create(THIS_MODULE, "vfio");
1835 	if (IS_ERR(vfio.class)) {
1836 		ret = PTR_ERR(vfio.class);
1837 		goto err_group_class;
1838 	}
1839 
1840 	vfio.class->devnode = vfio_devnode;
1841 
1842 	/* /sys/class/vfio-dev/vfioX */
1843 	vfio.device_class = class_create(THIS_MODULE, "vfio-dev");
1844 	if (IS_ERR(vfio.device_class)) {
1845 		ret = PTR_ERR(vfio.device_class);
1846 		goto err_dev_class;
1847 	}
1848 
1849 	ret = alloc_chrdev_region(&vfio.group_devt, 0, MINORMASK + 1, "vfio");
1850 	if (ret)
1851 		goto err_alloc_chrdev;
1852 
1853 	pr_info(DRIVER_DESC " version: " DRIVER_VERSION "\n");
1854 	return 0;
1855 
1856 err_alloc_chrdev:
1857 	class_destroy(vfio.device_class);
1858 	vfio.device_class = NULL;
1859 err_dev_class:
1860 	class_destroy(vfio.class);
1861 	vfio.class = NULL;
1862 err_group_class:
1863 	vfio_container_cleanup();
1864 	return ret;
1865 }
1866 
1867 static void __exit vfio_cleanup(void)
1868 {
1869 	WARN_ON(!list_empty(&vfio.group_list));
1870 
1871 	ida_destroy(&vfio.device_ida);
1872 	ida_destroy(&vfio.group_ida);
1873 	unregister_chrdev_region(vfio.group_devt, MINORMASK + 1);
1874 	class_destroy(vfio.device_class);
1875 	vfio.device_class = NULL;
1876 	class_destroy(vfio.class);
1877 	vfio_container_cleanup();
1878 	vfio.class = NULL;
1879 	xa_destroy(&vfio_device_set_xa);
1880 }
1881 
1882 module_init(vfio_init);
1883 module_exit(vfio_cleanup);
1884 
1885 MODULE_VERSION(DRIVER_VERSION);
1886 MODULE_LICENSE("GPL v2");
1887 MODULE_AUTHOR(DRIVER_AUTHOR);
1888 MODULE_DESCRIPTION(DRIVER_DESC);
1889 MODULE_ALIAS_MISCDEV(VFIO_MINOR);
1890 MODULE_ALIAS("devname:vfio/vfio");
1891 MODULE_SOFTDEP("post: vfio_iommu_type1 vfio_iommu_spapr_tce");
1892