1/*
2 * VFIO core
3 *
4 * Copyright (C) 2012 Red Hat, Inc.  All rights reserved.
5 *     Author: Alex Williamson <alex.williamson@redhat.com>
6 *
7 * This program is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License version 2 as
9 * published by the Free Software Foundation.
10 *
11 * Derived from original vfio:
12 * Copyright 2010 Cisco Systems, Inc.  All rights reserved.
13 * Author: Tom Lyon, pugs@cisco.com
14 */
15
16#include <linux/cdev.h>
17#include <linux/compat.h>
18#include <linux/device.h>
19#include <linux/file.h>
20#include <linux/anon_inodes.h>
21#include <linux/fs.h>
22#include <linux/idr.h>
23#include <linux/iommu.h>
24#include <linux/list.h>
25#include <linux/miscdevice.h>
26#include <linux/module.h>
27#include <linux/mutex.h>
28#include <linux/pci.h>
29#include <linux/rwsem.h>
30#include <linux/sched.h>
31#include <linux/slab.h>
32#include <linux/stat.h>
33#include <linux/string.h>
34#include <linux/uaccess.h>
35#include <linux/vfio.h>
36#include <linux/wait.h>
37
38#define DRIVER_VERSION	"0.3"
39#define DRIVER_AUTHOR	"Alex Williamson <alex.williamson@redhat.com>"
40#define DRIVER_DESC	"VFIO - User Level meta-driver"
41
42static struct vfio {
43	struct class			*class;
44	struct list_head		iommu_drivers_list;
45	struct mutex			iommu_drivers_lock;
46	struct list_head		group_list;
47	struct idr			group_idr;
48	struct mutex			group_lock;
49	struct cdev			group_cdev;
50	dev_t				group_devt;
51	wait_queue_head_t		release_q;
52} vfio;
53
54struct vfio_iommu_driver {
55	const struct vfio_iommu_driver_ops	*ops;
56	struct list_head			vfio_next;
57};
58
59struct vfio_container {
60	struct kref			kref;
61	struct list_head		group_list;
62	struct rw_semaphore		group_lock;
63	struct vfio_iommu_driver	*iommu_driver;
64	void				*iommu_data;
65};
66
67struct vfio_unbound_dev {
68	struct device			*dev;
69	struct list_head		unbound_next;
70};
71
72struct vfio_group {
73	struct kref			kref;
74	int				minor;
75	atomic_t			container_users;
76	struct iommu_group		*iommu_group;
77	struct vfio_container		*container;
78	struct list_head		device_list;
79	struct mutex			device_lock;
80	struct device			*dev;
81	struct notifier_block		nb;
82	struct list_head		vfio_next;
83	struct list_head		container_next;
84	struct list_head		unbound_list;
85	struct mutex			unbound_lock;
86	atomic_t			opened;
87};
88
89struct vfio_device {
90	struct kref			kref;
91	struct device			*dev;
92	const struct vfio_device_ops	*ops;
93	struct vfio_group		*group;
94	struct list_head		group_next;
95	void				*device_data;
96};
97
98/**
99 * IOMMU driver registration
100 */
101int vfio_register_iommu_driver(const struct vfio_iommu_driver_ops *ops)
102{
103	struct vfio_iommu_driver *driver, *tmp;
104
105	driver = kzalloc(sizeof(*driver), GFP_KERNEL);
106	if (!driver)
107		return -ENOMEM;
108
109	driver->ops = ops;
110
111	mutex_lock(&vfio.iommu_drivers_lock);
112
113	/* Check for duplicates */
114	list_for_each_entry(tmp, &vfio.iommu_drivers_list, vfio_next) {
115		if (tmp->ops == ops) {
116			mutex_unlock(&vfio.iommu_drivers_lock);
117			kfree(driver);
118			return -EINVAL;
119		}
120	}
121
122	list_add(&driver->vfio_next, &vfio.iommu_drivers_list);
123
124	mutex_unlock(&vfio.iommu_drivers_lock);
125
126	return 0;
127}
128EXPORT_SYMBOL_GPL(vfio_register_iommu_driver);
129
130void vfio_unregister_iommu_driver(const struct vfio_iommu_driver_ops *ops)
131{
132	struct vfio_iommu_driver *driver;
133
134	mutex_lock(&vfio.iommu_drivers_lock);
135	list_for_each_entry(driver, &vfio.iommu_drivers_list, vfio_next) {
136		if (driver->ops == ops) {
137			list_del(&driver->vfio_next);
138			mutex_unlock(&vfio.iommu_drivers_lock);
139			kfree(driver);
140			return;
141		}
142	}
143	mutex_unlock(&vfio.iommu_drivers_lock);
144}
145EXPORT_SYMBOL_GPL(vfio_unregister_iommu_driver);
146
147/**
148 * Group minor allocation/free - both called with vfio.group_lock held
149 */
150static int vfio_alloc_group_minor(struct vfio_group *group)
151{
152	return idr_alloc(&vfio.group_idr, group, 0, MINORMASK + 1, GFP_KERNEL);
153}
154
155static void vfio_free_group_minor(int minor)
156{
157	idr_remove(&vfio.group_idr, minor);
158}
159
160static int vfio_iommu_group_notifier(struct notifier_block *nb,
161				     unsigned long action, void *data);
162static void vfio_group_get(struct vfio_group *group);
163
164/**
165 * Container objects - containers are created when /dev/vfio/vfio is
166 * opened, but their lifecycle extends until the last user is done, so
167 * it's freed via kref.  Must support container/group/device being
168 * closed in any order.
169 */
170static void vfio_container_get(struct vfio_container *container)
171{
172	kref_get(&container->kref);
173}
174
175static void vfio_container_release(struct kref *kref)
176{
177	struct vfio_container *container;
178	container = container_of(kref, struct vfio_container, kref);
179
180	kfree(container);
181}
182
183static void vfio_container_put(struct vfio_container *container)
184{
185	kref_put(&container->kref, vfio_container_release);
186}
187
188static void vfio_group_unlock_and_free(struct vfio_group *group)
189{
190	mutex_unlock(&vfio.group_lock);
191	/*
192	 * Unregister outside of lock.  A spurious callback is harmless now
193	 * that the group is no longer in vfio.group_list.
194	 */
195	iommu_group_unregister_notifier(group->iommu_group, &group->nb);
196	kfree(group);
197}
198
199/**
200 * Group objects - create, release, get, put, search
201 */
202static struct vfio_group *vfio_create_group(struct iommu_group *iommu_group)
203{
204	struct vfio_group *group, *tmp;
205	struct device *dev;
206	int ret, minor;
207
208	group = kzalloc(sizeof(*group), GFP_KERNEL);
209	if (!group)
210		return ERR_PTR(-ENOMEM);
211
212	kref_init(&group->kref);
213	INIT_LIST_HEAD(&group->device_list);
214	mutex_init(&group->device_lock);
215	INIT_LIST_HEAD(&group->unbound_list);
216	mutex_init(&group->unbound_lock);
217	atomic_set(&group->container_users, 0);
218	atomic_set(&group->opened, 0);
219	group->iommu_group = iommu_group;
220
221	group->nb.notifier_call = vfio_iommu_group_notifier;
222
223	/*
224	 * blocking notifiers acquire a rwsem around registering and hold
225	 * it around callback.  Therefore, need to register outside of
226	 * vfio.group_lock to avoid A-B/B-A contention.  Our callback won't
227	 * do anything unless it can find the group in vfio.group_list, so
228	 * no harm in registering early.
229	 */
230	ret = iommu_group_register_notifier(iommu_group, &group->nb);
231	if (ret) {
232		kfree(group);
233		return ERR_PTR(ret);
234	}
235
236	mutex_lock(&vfio.group_lock);
237
238	/* Did we race creating this group? */
239	list_for_each_entry(tmp, &vfio.group_list, vfio_next) {
240		if (tmp->iommu_group == iommu_group) {
241			vfio_group_get(tmp);
242			vfio_group_unlock_and_free(group);
243			return tmp;
244		}
245	}
246
247	minor = vfio_alloc_group_minor(group);
248	if (minor < 0) {
249		vfio_group_unlock_and_free(group);
250		return ERR_PTR(minor);
251	}
252
253	dev = device_create(vfio.class, NULL,
254			    MKDEV(MAJOR(vfio.group_devt), minor),
255			    group, "%d", iommu_group_id(iommu_group));
256	if (IS_ERR(dev)) {
257		vfio_free_group_minor(minor);
258		vfio_group_unlock_and_free(group);
259		return (struct vfio_group *)dev; /* ERR_PTR */
260	}
261
262	group->minor = minor;
263	group->dev = dev;
264
265	list_add(&group->vfio_next, &vfio.group_list);
266
267	mutex_unlock(&vfio.group_lock);
268
269	return group;
270}
271
272/* called with vfio.group_lock held */
273static void vfio_group_release(struct kref *kref)
274{
275	struct vfio_group *group = container_of(kref, struct vfio_group, kref);
276	struct vfio_unbound_dev *unbound, *tmp;
277	struct iommu_group *iommu_group = group->iommu_group;
278
279	WARN_ON(!list_empty(&group->device_list));
280
281	list_for_each_entry_safe(unbound, tmp,
282				 &group->unbound_list, unbound_next) {
283		list_del(&unbound->unbound_next);
284		kfree(unbound);
285	}
286
287	device_destroy(vfio.class, MKDEV(MAJOR(vfio.group_devt), group->minor));
288	list_del(&group->vfio_next);
289	vfio_free_group_minor(group->minor);
290	vfio_group_unlock_and_free(group);
291	iommu_group_put(iommu_group);
292}
293
294static void vfio_group_put(struct vfio_group *group)
295{
296	kref_put_mutex(&group->kref, vfio_group_release, &vfio.group_lock);
297}
298
299/* Assume group_lock or group reference is held */
300static void vfio_group_get(struct vfio_group *group)
301{
302	kref_get(&group->kref);
303}
304
305/*
306 * Not really a try as we will sleep for mutex, but we need to make
307 * sure the group pointer is valid under lock and get a reference.
308 */
309static struct vfio_group *vfio_group_try_get(struct vfio_group *group)
310{
311	struct vfio_group *target = group;
312
313	mutex_lock(&vfio.group_lock);
314	list_for_each_entry(group, &vfio.group_list, vfio_next) {
315		if (group == target) {
316			vfio_group_get(group);
317			mutex_unlock(&vfio.group_lock);
318			return group;
319		}
320	}
321	mutex_unlock(&vfio.group_lock);
322
323	return NULL;
324}
325
326static
327struct vfio_group *vfio_group_get_from_iommu(struct iommu_group *iommu_group)
328{
329	struct vfio_group *group;
330
331	mutex_lock(&vfio.group_lock);
332	list_for_each_entry(group, &vfio.group_list, vfio_next) {
333		if (group->iommu_group == iommu_group) {
334			vfio_group_get(group);
335			mutex_unlock(&vfio.group_lock);
336			return group;
337		}
338	}
339	mutex_unlock(&vfio.group_lock);
340
341	return NULL;
342}
343
344static struct vfio_group *vfio_group_get_from_minor(int minor)
345{
346	struct vfio_group *group;
347
348	mutex_lock(&vfio.group_lock);
349	group = idr_find(&vfio.group_idr, minor);
350	if (!group) {
351		mutex_unlock(&vfio.group_lock);
352		return NULL;
353	}
354	vfio_group_get(group);
355	mutex_unlock(&vfio.group_lock);
356
357	return group;
358}
359
360/**
361 * Device objects - create, release, get, put, search
362 */
363static
364struct vfio_device *vfio_group_create_device(struct vfio_group *group,
365					     struct device *dev,
366					     const struct vfio_device_ops *ops,
367					     void *device_data)
368{
369	struct vfio_device *device;
370
371	device = kzalloc(sizeof(*device), GFP_KERNEL);
372	if (!device)
373		return ERR_PTR(-ENOMEM);
374
375	kref_init(&device->kref);
376	device->dev = dev;
377	device->group = group;
378	device->ops = ops;
379	device->device_data = device_data;
380	dev_set_drvdata(dev, device);
381
382	/* No need to get group_lock, caller has group reference */
383	vfio_group_get(group);
384
385	mutex_lock(&group->device_lock);
386	list_add(&device->group_next, &group->device_list);
387	mutex_unlock(&group->device_lock);
388
389	return device;
390}
391
392static void vfio_device_release(struct kref *kref)
393{
394	struct vfio_device *device = container_of(kref,
395						  struct vfio_device, kref);
396	struct vfio_group *group = device->group;
397
398	list_del(&device->group_next);
399	mutex_unlock(&group->device_lock);
400
401	dev_set_drvdata(device->dev, NULL);
402
403	kfree(device);
404
405	/* vfio_del_group_dev may be waiting for this device */
406	wake_up(&vfio.release_q);
407}
408
409/* Device reference always implies a group reference */
410void vfio_device_put(struct vfio_device *device)
411{
412	struct vfio_group *group = device->group;
413	kref_put_mutex(&device->kref, vfio_device_release, &group->device_lock);
414	vfio_group_put(group);
415}
416EXPORT_SYMBOL_GPL(vfio_device_put);
417
418static void vfio_device_get(struct vfio_device *device)
419{
420	vfio_group_get(device->group);
421	kref_get(&device->kref);
422}
423
424static struct vfio_device *vfio_group_get_device(struct vfio_group *group,
425						 struct device *dev)
426{
427	struct vfio_device *device;
428
429	mutex_lock(&group->device_lock);
430	list_for_each_entry(device, &group->device_list, group_next) {
431		if (device->dev == dev) {
432			vfio_device_get(device);
433			mutex_unlock(&group->device_lock);
434			return device;
435		}
436	}
437	mutex_unlock(&group->device_lock);
438	return NULL;
439}
440
441/*
442 * Some drivers, like pci-stub, are only used to prevent other drivers from
443 * claiming a device and are therefore perfectly legitimate for a user owned
444 * group.  The pci-stub driver has no dependencies on DMA or the IOVA mapping
445 * of the device, but it does prevent the user from having direct access to
446 * the device, which is useful in some circumstances.
447 *
448 * We also assume that we can include PCI interconnect devices, ie. bridges.
449 * IOMMU grouping on PCI necessitates that if we lack isolation on a bridge
450 * then all of the downstream devices will be part of the same IOMMU group as
451 * the bridge.  Thus, if placing the bridge into the user owned IOVA space
452 * breaks anything, it only does so for user owned devices downstream.  Note
453 * that error notification via MSI can be affected for platforms that handle
454 * MSI within the same IOVA space as DMA.
455 */
456static const char * const vfio_driver_whitelist[] = { "pci-stub" };
457
458static bool vfio_dev_whitelisted(struct device *dev, struct device_driver *drv)
459{
460	int i;
461
462	if (dev_is_pci(dev)) {
463		struct pci_dev *pdev = to_pci_dev(dev);
464
465		if (pdev->hdr_type != PCI_HEADER_TYPE_NORMAL)
466			return true;
467	}
468
469	for (i = 0; i < ARRAY_SIZE(vfio_driver_whitelist); i++) {
470		if (!strcmp(drv->name, vfio_driver_whitelist[i]))
471			return true;
472	}
473
474	return false;
475}
476
477/*
478 * A vfio group is viable for use by userspace if all devices are in
479 * one of the following states:
480 *  - driver-less
481 *  - bound to a vfio driver
482 *  - bound to a whitelisted driver
483 *  - a PCI interconnect device
484 *
485 * We use two methods to determine whether a device is bound to a vfio
486 * driver.  The first is to test whether the device exists in the vfio
487 * group.  The second is to test if the device exists on the group
488 * unbound_list, indicating it's in the middle of transitioning from
489 * a vfio driver to driver-less.
490 */
491static int vfio_dev_viable(struct device *dev, void *data)
492{
493	struct vfio_group *group = data;
494	struct vfio_device *device;
495	struct device_driver *drv = ACCESS_ONCE(dev->driver);
496	struct vfio_unbound_dev *unbound;
497	int ret = -EINVAL;
498
499	mutex_lock(&group->unbound_lock);
500	list_for_each_entry(unbound, &group->unbound_list, unbound_next) {
501		if (dev == unbound->dev) {
502			ret = 0;
503			break;
504		}
505	}
506	mutex_unlock(&group->unbound_lock);
507
508	if (!ret || !drv || vfio_dev_whitelisted(dev, drv))
509		return 0;
510
511	device = vfio_group_get_device(group, dev);
512	if (device) {
513		vfio_device_put(device);
514		return 0;
515	}
516
517	return ret;
518}
519
520/**
521 * Async device support
522 */
523static int vfio_group_nb_add_dev(struct vfio_group *group, struct device *dev)
524{
525	struct vfio_device *device;
526
527	/* Do we already know about it?  We shouldn't */
528	device = vfio_group_get_device(group, dev);
529	if (WARN_ON_ONCE(device)) {
530		vfio_device_put(device);
531		return 0;
532	}
533
534	/* Nothing to do for idle groups */
535	if (!atomic_read(&group->container_users))
536		return 0;
537
538	/* TODO Prevent device auto probing */
539	WARN(1, "Device %s added to live group %d!\n", dev_name(dev),
540	     iommu_group_id(group->iommu_group));
541
542	return 0;
543}
544
545static int vfio_group_nb_verify(struct vfio_group *group, struct device *dev)
546{
547	/* We don't care what happens when the group isn't in use */
548	if (!atomic_read(&group->container_users))
549		return 0;
550
551	return vfio_dev_viable(dev, group);
552}
553
554static int vfio_iommu_group_notifier(struct notifier_block *nb,
555				     unsigned long action, void *data)
556{
557	struct vfio_group *group = container_of(nb, struct vfio_group, nb);
558	struct device *dev = data;
559	struct vfio_unbound_dev *unbound;
560
561	/*
562	 * Need to go through a group_lock lookup to get a reference or we
563	 * risk racing a group being removed.  Ignore spurious notifies.
564	 */
565	group = vfio_group_try_get(group);
566	if (!group)
567		return NOTIFY_OK;
568
569	switch (action) {
570	case IOMMU_GROUP_NOTIFY_ADD_DEVICE:
571		vfio_group_nb_add_dev(group, dev);
572		break;
573	case IOMMU_GROUP_NOTIFY_DEL_DEVICE:
574		/*
575		 * Nothing to do here.  If the device is in use, then the
576		 * vfio sub-driver should block the remove callback until
577		 * it is unused.  If the device is unused or attached to a
578		 * stub driver, then it should be released and we don't
579		 * care that it will be going away.
580		 */
581		break;
582	case IOMMU_GROUP_NOTIFY_BIND_DRIVER:
583		pr_debug("%s: Device %s, group %d binding to driver\n",
584			 __func__, dev_name(dev),
585			 iommu_group_id(group->iommu_group));
586		break;
587	case IOMMU_GROUP_NOTIFY_BOUND_DRIVER:
588		pr_debug("%s: Device %s, group %d bound to driver %s\n",
589			 __func__, dev_name(dev),
590			 iommu_group_id(group->iommu_group), dev->driver->name);
591		BUG_ON(vfio_group_nb_verify(group, dev));
592		break;
593	case IOMMU_GROUP_NOTIFY_UNBIND_DRIVER:
594		pr_debug("%s: Device %s, group %d unbinding from driver %s\n",
595			 __func__, dev_name(dev),
596			 iommu_group_id(group->iommu_group), dev->driver->name);
597		break;
598	case IOMMU_GROUP_NOTIFY_UNBOUND_DRIVER:
599		pr_debug("%s: Device %s, group %d unbound from driver\n",
600			 __func__, dev_name(dev),
601			 iommu_group_id(group->iommu_group));
602		/*
603		 * XXX An unbound device in a live group is ok, but we'd
604		 * really like to avoid the above BUG_ON by preventing other
605		 * drivers from binding to it.  Once that occurs, we have to
606		 * stop the system to maintain isolation.  At a minimum, we'd
607		 * want a toggle to disable driver auto probe for this device.
608		 */
609
610		mutex_lock(&group->unbound_lock);
611		list_for_each_entry(unbound,
612				    &group->unbound_list, unbound_next) {
613			if (dev == unbound->dev) {
614				list_del(&unbound->unbound_next);
615				kfree(unbound);
616				break;
617			}
618		}
619		mutex_unlock(&group->unbound_lock);
620		break;
621	}
622
623	vfio_group_put(group);
624	return NOTIFY_OK;
625}
626
627/**
628 * VFIO driver API
629 */
630int vfio_add_group_dev(struct device *dev,
631		       const struct vfio_device_ops *ops, void *device_data)
632{
633	struct iommu_group *iommu_group;
634	struct vfio_group *group;
635	struct vfio_device *device;
636
637	iommu_group = iommu_group_get(dev);
638	if (!iommu_group)
639		return -EINVAL;
640
641	group = vfio_group_get_from_iommu(iommu_group);
642	if (!group) {
643		group = vfio_create_group(iommu_group);
644		if (IS_ERR(group)) {
645			iommu_group_put(iommu_group);
646			return PTR_ERR(group);
647		}
648	} else {
649		/*
650		 * A found vfio_group already holds a reference to the
651		 * iommu_group.  A created vfio_group keeps the reference.
652		 */
653		iommu_group_put(iommu_group);
654	}
655
656	device = vfio_group_get_device(group, dev);
657	if (device) {
658		WARN(1, "Device %s already exists on group %d\n",
659		     dev_name(dev), iommu_group_id(iommu_group));
660		vfio_device_put(device);
661		vfio_group_put(group);
662		return -EBUSY;
663	}
664
665	device = vfio_group_create_device(group, dev, ops, device_data);
666	if (IS_ERR(device)) {
667		vfio_group_put(group);
668		return PTR_ERR(device);
669	}
670
671	/*
672	 * Drop all but the vfio_device reference.  The vfio_device holds
673	 * a reference to the vfio_group, which holds a reference to the
674	 * iommu_group.
675	 */
676	vfio_group_put(group);
677
678	return 0;
679}
680EXPORT_SYMBOL_GPL(vfio_add_group_dev);
681
682/**
683 * Get a reference to the vfio_device for a device.  Even if the
684 * caller thinks they own the device, they could be racing with a
685 * release call path, so we can't trust drvdata for the shortcut.
686 * Go the long way around, from the iommu_group to the vfio_group
687 * to the vfio_device.
688 */
689struct vfio_device *vfio_device_get_from_dev(struct device *dev)
690{
691	struct iommu_group *iommu_group;
692	struct vfio_group *group;
693	struct vfio_device *device;
694
695	iommu_group = iommu_group_get(dev);
696	if (!iommu_group)
697		return NULL;
698
699	group = vfio_group_get_from_iommu(iommu_group);
700	iommu_group_put(iommu_group);
701	if (!group)
702		return NULL;
703
704	device = vfio_group_get_device(group, dev);
705	vfio_group_put(group);
706
707	return device;
708}
709EXPORT_SYMBOL_GPL(vfio_device_get_from_dev);
710
711static struct vfio_device *vfio_device_get_from_name(struct vfio_group *group,
712						     char *buf)
713{
714	struct vfio_device *it, *device = NULL;
715
716	mutex_lock(&group->device_lock);
717	list_for_each_entry(it, &group->device_list, group_next) {
718		if (!strcmp(dev_name(it->dev), buf)) {
719			device = it;
720			vfio_device_get(device);
721			break;
722		}
723	}
724	mutex_unlock(&group->device_lock);
725
726	return device;
727}
728
729/*
730 * Caller must hold a reference to the vfio_device
731 */
732void *vfio_device_data(struct vfio_device *device)
733{
734	return device->device_data;
735}
736EXPORT_SYMBOL_GPL(vfio_device_data);
737
738/* Given a referenced group, check if it contains the device */
739static bool vfio_dev_present(struct vfio_group *group, struct device *dev)
740{
741	struct vfio_device *device;
742
743	device = vfio_group_get_device(group, dev);
744	if (!device)
745		return false;
746
747	vfio_device_put(device);
748	return true;
749}
750
751/*
752 * Decrement the device reference count and wait for the device to be
753 * removed.  Open file descriptors for the device... */
754void *vfio_del_group_dev(struct device *dev)
755{
756	struct vfio_device *device = dev_get_drvdata(dev);
757	struct vfio_group *group = device->group;
758	void *device_data = device->device_data;
759	struct vfio_unbound_dev *unbound;
760	unsigned int i = 0;
761	long ret;
762	bool interrupted = false;
763
764	/*
765	 * The group exists so long as we have a device reference.  Get
766	 * a group reference and use it to scan for the device going away.
767	 */
768	vfio_group_get(group);
769
770	/*
771	 * When the device is removed from the group, the group suddenly
772	 * becomes non-viable; the device has a driver (until the unbind
773	 * completes), but it's not present in the group.  This is bad news
774	 * for any external users that need to re-acquire a group reference
775	 * in order to match and release their existing reference.  To
776	 * solve this, we track such devices on the unbound_list to bridge
777	 * the gap until they're fully unbound.
778	 */
779	unbound = kzalloc(sizeof(*unbound), GFP_KERNEL);
780	if (unbound) {
781		unbound->dev = dev;
782		mutex_lock(&group->unbound_lock);
783		list_add(&unbound->unbound_next, &group->unbound_list);
784		mutex_unlock(&group->unbound_lock);
785	}
786	WARN_ON(!unbound);
787
788	vfio_device_put(device);
789
790	/*
791	 * If the device is still present in the group after the above
792	 * 'put', then it is in use and we need to request it from the
793	 * bus driver.  The driver may in turn need to request the
794	 * device from the user.  We send the request on an arbitrary
795	 * interval with counter to allow the driver to take escalating
796	 * measures to release the device if it has the ability to do so.
797	 */
798	do {
799		device = vfio_group_get_device(group, dev);
800		if (!device)
801			break;
802
803		if (device->ops->request)
804			device->ops->request(device_data, i++);
805
806		vfio_device_put(device);
807
808		if (interrupted) {
809			ret = wait_event_timeout(vfio.release_q,
810					!vfio_dev_present(group, dev), HZ * 10);
811		} else {
812			ret = wait_event_interruptible_timeout(vfio.release_q,
813					!vfio_dev_present(group, dev), HZ * 10);
814			if (ret == -ERESTARTSYS) {
815				interrupted = true;
816				dev_warn(dev,
817					 "Device is currently in use, task"
818					 " \"%s\" (%d) "
819					 "blocked until device is released",
820					 current->comm, task_pid_nr(current));
821			}
822		}
823	} while (ret <= 0);
824
825	vfio_group_put(group);
826
827	return device_data;
828}
829EXPORT_SYMBOL_GPL(vfio_del_group_dev);
830
831/**
832 * VFIO base fd, /dev/vfio/vfio
833 */
834static long vfio_ioctl_check_extension(struct vfio_container *container,
835				       unsigned long arg)
836{
837	struct vfio_iommu_driver *driver;
838	long ret = 0;
839
840	down_read(&container->group_lock);
841
842	driver = container->iommu_driver;
843
844	switch (arg) {
845		/* No base extensions yet */
846	default:
847		/*
848		 * If no driver is set, poll all registered drivers for
849		 * extensions and return the first positive result.  If
850		 * a driver is already set, further queries will be passed
851		 * only to that driver.
852		 */
853		if (!driver) {
854			mutex_lock(&vfio.iommu_drivers_lock);
855			list_for_each_entry(driver, &vfio.iommu_drivers_list,
856					    vfio_next) {
857				if (!try_module_get(driver->ops->owner))
858					continue;
859
860				ret = driver->ops->ioctl(NULL,
861							 VFIO_CHECK_EXTENSION,
862							 arg);
863				module_put(driver->ops->owner);
864				if (ret > 0)
865					break;
866			}
867			mutex_unlock(&vfio.iommu_drivers_lock);
868		} else
869			ret = driver->ops->ioctl(container->iommu_data,
870						 VFIO_CHECK_EXTENSION, arg);
871	}
872
873	up_read(&container->group_lock);
874
875	return ret;
876}
877
878/* hold write lock on container->group_lock */
879static int __vfio_container_attach_groups(struct vfio_container *container,
880					  struct vfio_iommu_driver *driver,
881					  void *data)
882{
883	struct vfio_group *group;
884	int ret = -ENODEV;
885
886	list_for_each_entry(group, &container->group_list, container_next) {
887		ret = driver->ops->attach_group(data, group->iommu_group);
888		if (ret)
889			goto unwind;
890	}
891
892	return ret;
893
894unwind:
895	list_for_each_entry_continue_reverse(group, &container->group_list,
896					     container_next) {
897		driver->ops->detach_group(data, group->iommu_group);
898	}
899
900	return ret;
901}
902
903static long vfio_ioctl_set_iommu(struct vfio_container *container,
904				 unsigned long arg)
905{
906	struct vfio_iommu_driver *driver;
907	long ret = -ENODEV;
908
909	down_write(&container->group_lock);
910
911	/*
912	 * The container is designed to be an unprivileged interface while
913	 * the group can be assigned to specific users.  Therefore, only by
914	 * adding a group to a container does the user get the privilege of
915	 * enabling the iommu, which may allocate finite resources.  There
916	 * is no unset_iommu, but by removing all the groups from a container,
917	 * the container is deprivileged and returns to an unset state.
918	 */
919	if (list_empty(&container->group_list) || container->iommu_driver) {
920		up_write(&container->group_lock);
921		return -EINVAL;
922	}
923
924	mutex_lock(&vfio.iommu_drivers_lock);
925	list_for_each_entry(driver, &vfio.iommu_drivers_list, vfio_next) {
926		void *data;
927
928		if (!try_module_get(driver->ops->owner))
929			continue;
930
931		/*
932		 * The arg magic for SET_IOMMU is the same as CHECK_EXTENSION,
933		 * so test which iommu driver reported support for this
934		 * extension and call open on them.  We also pass them the
935		 * magic, allowing a single driver to support multiple
936		 * interfaces if they'd like.
937		 */
938		if (driver->ops->ioctl(NULL, VFIO_CHECK_EXTENSION, arg) <= 0) {
939			module_put(driver->ops->owner);
940			continue;
941		}
942
943		/* module reference holds the driver we're working on */
944		mutex_unlock(&vfio.iommu_drivers_lock);
945
946		data = driver->ops->open(arg);
947		if (IS_ERR(data)) {
948			ret = PTR_ERR(data);
949			module_put(driver->ops->owner);
950			goto skip_drivers_unlock;
951		}
952
953		ret = __vfio_container_attach_groups(container, driver, data);
954		if (!ret) {
955			container->iommu_driver = driver;
956			container->iommu_data = data;
957		} else {
958			driver->ops->release(data);
959			module_put(driver->ops->owner);
960		}
961
962		goto skip_drivers_unlock;
963	}
964
965	mutex_unlock(&vfio.iommu_drivers_lock);
966skip_drivers_unlock:
967	up_write(&container->group_lock);
968
969	return ret;
970}
971
972static long vfio_fops_unl_ioctl(struct file *filep,
973				unsigned int cmd, unsigned long arg)
974{
975	struct vfio_container *container = filep->private_data;
976	struct vfio_iommu_driver *driver;
977	void *data;
978	long ret = -EINVAL;
979
980	if (!container)
981		return ret;
982
983	switch (cmd) {
984	case VFIO_GET_API_VERSION:
985		ret = VFIO_API_VERSION;
986		break;
987	case VFIO_CHECK_EXTENSION:
988		ret = vfio_ioctl_check_extension(container, arg);
989		break;
990	case VFIO_SET_IOMMU:
991		ret = vfio_ioctl_set_iommu(container, arg);
992		break;
993	default:
994		down_read(&container->group_lock);
995
996		driver = container->iommu_driver;
997		data = container->iommu_data;
998
999		if (driver) /* passthrough all unrecognized ioctls */
1000			ret = driver->ops->ioctl(data, cmd, arg);
1001
1002		up_read(&container->group_lock);
1003	}
1004
1005	return ret;
1006}
1007
1008#ifdef CONFIG_COMPAT
1009static long vfio_fops_compat_ioctl(struct file *filep,
1010				   unsigned int cmd, unsigned long arg)
1011{
1012	arg = (unsigned long)compat_ptr(arg);
1013	return vfio_fops_unl_ioctl(filep, cmd, arg);
1014}
1015#endif	/* CONFIG_COMPAT */
1016
1017static int vfio_fops_open(struct inode *inode, struct file *filep)
1018{
1019	struct vfio_container *container;
1020
1021	container = kzalloc(sizeof(*container), GFP_KERNEL);
1022	if (!container)
1023		return -ENOMEM;
1024
1025	INIT_LIST_HEAD(&container->group_list);
1026	init_rwsem(&container->group_lock);
1027	kref_init(&container->kref);
1028
1029	filep->private_data = container;
1030
1031	return 0;
1032}
1033
1034static int vfio_fops_release(struct inode *inode, struct file *filep)
1035{
1036	struct vfio_container *container = filep->private_data;
1037
1038	filep->private_data = NULL;
1039
1040	vfio_container_put(container);
1041
1042	return 0;
1043}
1044
1045/*
1046 * Once an iommu driver is set, we optionally pass read/write/mmap
1047 * on to the driver, allowing management interfaces beyond ioctl.
1048 */
1049static ssize_t vfio_fops_read(struct file *filep, char __user *buf,
1050			      size_t count, loff_t *ppos)
1051{
1052	struct vfio_container *container = filep->private_data;
1053	struct vfio_iommu_driver *driver;
1054	ssize_t ret = -EINVAL;
1055
1056	down_read(&container->group_lock);
1057
1058	driver = container->iommu_driver;
1059	if (likely(driver && driver->ops->read))
1060		ret = driver->ops->read(container->iommu_data,
1061					buf, count, ppos);
1062
1063	up_read(&container->group_lock);
1064
1065	return ret;
1066}
1067
1068static ssize_t vfio_fops_write(struct file *filep, const char __user *buf,
1069			       size_t count, loff_t *ppos)
1070{
1071	struct vfio_container *container = filep->private_data;
1072	struct vfio_iommu_driver *driver;
1073	ssize_t ret = -EINVAL;
1074
1075	down_read(&container->group_lock);
1076
1077	driver = container->iommu_driver;
1078	if (likely(driver && driver->ops->write))
1079		ret = driver->ops->write(container->iommu_data,
1080					 buf, count, ppos);
1081
1082	up_read(&container->group_lock);
1083
1084	return ret;
1085}
1086
1087static int vfio_fops_mmap(struct file *filep, struct vm_area_struct *vma)
1088{
1089	struct vfio_container *container = filep->private_data;
1090	struct vfio_iommu_driver *driver;
1091	int ret = -EINVAL;
1092
1093	down_read(&container->group_lock);
1094
1095	driver = container->iommu_driver;
1096	if (likely(driver && driver->ops->mmap))
1097		ret = driver->ops->mmap(container->iommu_data, vma);
1098
1099	up_read(&container->group_lock);
1100
1101	return ret;
1102}
1103
1104static const struct file_operations vfio_fops = {
1105	.owner		= THIS_MODULE,
1106	.open		= vfio_fops_open,
1107	.release	= vfio_fops_release,
1108	.read		= vfio_fops_read,
1109	.write		= vfio_fops_write,
1110	.unlocked_ioctl	= vfio_fops_unl_ioctl,
1111#ifdef CONFIG_COMPAT
1112	.compat_ioctl	= vfio_fops_compat_ioctl,
1113#endif
1114	.mmap		= vfio_fops_mmap,
1115};
1116
1117/**
1118 * VFIO Group fd, /dev/vfio/$GROUP
1119 */
1120static void __vfio_group_unset_container(struct vfio_group *group)
1121{
1122	struct vfio_container *container = group->container;
1123	struct vfio_iommu_driver *driver;
1124
1125	down_write(&container->group_lock);
1126
1127	driver = container->iommu_driver;
1128	if (driver)
1129		driver->ops->detach_group(container->iommu_data,
1130					  group->iommu_group);
1131
1132	group->container = NULL;
1133	list_del(&group->container_next);
1134
1135	/* Detaching the last group deprivileges a container, remove iommu */
1136	if (driver && list_empty(&container->group_list)) {
1137		driver->ops->release(container->iommu_data);
1138		module_put(driver->ops->owner);
1139		container->iommu_driver = NULL;
1140		container->iommu_data = NULL;
1141	}
1142
1143	up_write(&container->group_lock);
1144
1145	vfio_container_put(container);
1146}
1147
1148/*
1149 * VFIO_GROUP_UNSET_CONTAINER should fail if there are other users or
1150 * if there was no container to unset.  Since the ioctl is called on
1151 * the group, we know that still exists, therefore the only valid
1152 * transition here is 1->0.
1153 */
1154static int vfio_group_unset_container(struct vfio_group *group)
1155{
1156	int users = atomic_cmpxchg(&group->container_users, 1, 0);
1157
1158	if (!users)
1159		return -EINVAL;
1160	if (users != 1)
1161		return -EBUSY;
1162
1163	__vfio_group_unset_container(group);
1164
1165	return 0;
1166}
1167
1168/*
1169 * When removing container users, anything that removes the last user
1170 * implicitly removes the group from the container.  That is, if the
1171 * group file descriptor is closed, as well as any device file descriptors,
1172 * the group is free.
1173 */
1174static void vfio_group_try_dissolve_container(struct vfio_group *group)
1175{
1176	if (0 == atomic_dec_if_positive(&group->container_users))
1177		__vfio_group_unset_container(group);
1178}
1179
1180static int vfio_group_set_container(struct vfio_group *group, int container_fd)
1181{
1182	struct fd f;
1183	struct vfio_container *container;
1184	struct vfio_iommu_driver *driver;
1185	int ret = 0;
1186
1187	if (atomic_read(&group->container_users))
1188		return -EINVAL;
1189
1190	f = fdget(container_fd);
1191	if (!f.file)
1192		return -EBADF;
1193
1194	/* Sanity check, is this really our fd? */
1195	if (f.file->f_op != &vfio_fops) {
1196		fdput(f);
1197		return -EINVAL;
1198	}
1199
1200	container = f.file->private_data;
1201	WARN_ON(!container); /* fget ensures we don't race vfio_release */
1202
1203	down_write(&container->group_lock);
1204
1205	driver = container->iommu_driver;
1206	if (driver) {
1207		ret = driver->ops->attach_group(container->iommu_data,
1208						group->iommu_group);
1209		if (ret)
1210			goto unlock_out;
1211	}
1212
1213	group->container = container;
1214	list_add(&group->container_next, &container->group_list);
1215
1216	/* Get a reference on the container and mark a user within the group */
1217	vfio_container_get(container);
1218	atomic_inc(&group->container_users);
1219
1220unlock_out:
1221	up_write(&container->group_lock);
1222	fdput(f);
1223	return ret;
1224}
1225
1226static bool vfio_group_viable(struct vfio_group *group)
1227{
1228	return (iommu_group_for_each_dev(group->iommu_group,
1229					 group, vfio_dev_viable) == 0);
1230}
1231
1232static const struct file_operations vfio_device_fops;
1233
1234static int vfio_group_get_device_fd(struct vfio_group *group, char *buf)
1235{
1236	struct vfio_device *device;
1237	struct file *filep;
1238	int ret;
1239
1240	if (0 == atomic_read(&group->container_users) ||
1241	    !group->container->iommu_driver || !vfio_group_viable(group))
1242		return -EINVAL;
1243
1244	device = vfio_device_get_from_name(group, buf);
1245	if (!device)
1246		return -ENODEV;
1247
1248	ret = device->ops->open(device->device_data);
1249	if (ret) {
1250		vfio_device_put(device);
1251		return ret;
1252	}
1253
1254	/*
1255	 * We can't use anon_inode_getfd() because we need to modify
1256	 * the f_mode flags directly to allow more than just ioctls
1257	 */
1258	ret = get_unused_fd_flags(O_CLOEXEC);
1259	if (ret < 0) {
1260		device->ops->release(device->device_data);
1261		vfio_device_put(device);
1262		return ret;
1263	}
1264
1265	filep = anon_inode_getfile("[vfio-device]", &vfio_device_fops,
1266				   device, O_RDWR);
1267	if (IS_ERR(filep)) {
1268		put_unused_fd(ret);
1269		ret = PTR_ERR(filep);
1270		device->ops->release(device->device_data);
1271		vfio_device_put(device);
1272		return ret;
1273	}
1274
1275	/*
1276	 * TODO: add an anon_inode interface to do this.
1277	 * Appears to be missing by lack of need rather than
1278	 * explicitly prevented.  Now there's need.
1279	 */
1280	filep->f_mode |= (FMODE_LSEEK | FMODE_PREAD | FMODE_PWRITE);
1281
1282	atomic_inc(&group->container_users);
1283
1284	fd_install(ret, filep);
1285
1286	return ret;
1287}
1288
1289static long vfio_group_fops_unl_ioctl(struct file *filep,
1290				      unsigned int cmd, unsigned long arg)
1291{
1292	struct vfio_group *group = filep->private_data;
1293	long ret = -ENOTTY;
1294
1295	switch (cmd) {
1296	case VFIO_GROUP_GET_STATUS:
1297	{
1298		struct vfio_group_status status;
1299		unsigned long minsz;
1300
1301		minsz = offsetofend(struct vfio_group_status, flags);
1302
1303		if (copy_from_user(&status, (void __user *)arg, minsz))
1304			return -EFAULT;
1305
1306		if (status.argsz < minsz)
1307			return -EINVAL;
1308
1309		status.flags = 0;
1310
1311		if (vfio_group_viable(group))
1312			status.flags |= VFIO_GROUP_FLAGS_VIABLE;
1313
1314		if (group->container)
1315			status.flags |= VFIO_GROUP_FLAGS_CONTAINER_SET;
1316
1317		if (copy_to_user((void __user *)arg, &status, minsz))
1318			return -EFAULT;
1319
1320		ret = 0;
1321		break;
1322	}
1323	case VFIO_GROUP_SET_CONTAINER:
1324	{
1325		int fd;
1326
1327		if (get_user(fd, (int __user *)arg))
1328			return -EFAULT;
1329
1330		if (fd < 0)
1331			return -EINVAL;
1332
1333		ret = vfio_group_set_container(group, fd);
1334		break;
1335	}
1336	case VFIO_GROUP_UNSET_CONTAINER:
1337		ret = vfio_group_unset_container(group);
1338		break;
1339	case VFIO_GROUP_GET_DEVICE_FD:
1340	{
1341		char *buf;
1342
1343		buf = strndup_user((const char __user *)arg, PAGE_SIZE);
1344		if (IS_ERR(buf))
1345			return PTR_ERR(buf);
1346
1347		ret = vfio_group_get_device_fd(group, buf);
1348		kfree(buf);
1349		break;
1350	}
1351	}
1352
1353	return ret;
1354}
1355
1356#ifdef CONFIG_COMPAT
1357static long vfio_group_fops_compat_ioctl(struct file *filep,
1358					 unsigned int cmd, unsigned long arg)
1359{
1360	arg = (unsigned long)compat_ptr(arg);
1361	return vfio_group_fops_unl_ioctl(filep, cmd, arg);
1362}
1363#endif	/* CONFIG_COMPAT */
1364
1365static int vfio_group_fops_open(struct inode *inode, struct file *filep)
1366{
1367	struct vfio_group *group;
1368	int opened;
1369
1370	group = vfio_group_get_from_minor(iminor(inode));
1371	if (!group)
1372		return -ENODEV;
1373
1374	/* Do we need multiple instances of the group open?  Seems not. */
1375	opened = atomic_cmpxchg(&group->opened, 0, 1);
1376	if (opened) {
1377		vfio_group_put(group);
1378		return -EBUSY;
1379	}
1380
1381	/* Is something still in use from a previous open? */
1382	if (group->container) {
1383		atomic_dec(&group->opened);
1384		vfio_group_put(group);
1385		return -EBUSY;
1386	}
1387
1388	filep->private_data = group;
1389
1390	return 0;
1391}
1392
1393static int vfio_group_fops_release(struct inode *inode, struct file *filep)
1394{
1395	struct vfio_group *group = filep->private_data;
1396
1397	filep->private_data = NULL;
1398
1399	vfio_group_try_dissolve_container(group);
1400
1401	atomic_dec(&group->opened);
1402
1403	vfio_group_put(group);
1404
1405	return 0;
1406}
1407
1408static const struct file_operations vfio_group_fops = {
1409	.owner		= THIS_MODULE,
1410	.unlocked_ioctl	= vfio_group_fops_unl_ioctl,
1411#ifdef CONFIG_COMPAT
1412	.compat_ioctl	= vfio_group_fops_compat_ioctl,
1413#endif
1414	.open		= vfio_group_fops_open,
1415	.release	= vfio_group_fops_release,
1416};
1417
1418/**
1419 * VFIO Device fd
1420 */
1421static int vfio_device_fops_release(struct inode *inode, struct file *filep)
1422{
1423	struct vfio_device *device = filep->private_data;
1424
1425	device->ops->release(device->device_data);
1426
1427	vfio_group_try_dissolve_container(device->group);
1428
1429	vfio_device_put(device);
1430
1431	return 0;
1432}
1433
1434static long vfio_device_fops_unl_ioctl(struct file *filep,
1435				       unsigned int cmd, unsigned long arg)
1436{
1437	struct vfio_device *device = filep->private_data;
1438
1439	if (unlikely(!device->ops->ioctl))
1440		return -EINVAL;
1441
1442	return device->ops->ioctl(device->device_data, cmd, arg);
1443}
1444
1445static ssize_t vfio_device_fops_read(struct file *filep, char __user *buf,
1446				     size_t count, loff_t *ppos)
1447{
1448	struct vfio_device *device = filep->private_data;
1449
1450	if (unlikely(!device->ops->read))
1451		return -EINVAL;
1452
1453	return device->ops->read(device->device_data, buf, count, ppos);
1454}
1455
1456static ssize_t vfio_device_fops_write(struct file *filep,
1457				      const char __user *buf,
1458				      size_t count, loff_t *ppos)
1459{
1460	struct vfio_device *device = filep->private_data;
1461
1462	if (unlikely(!device->ops->write))
1463		return -EINVAL;
1464
1465	return device->ops->write(device->device_data, buf, count, ppos);
1466}
1467
1468static int vfio_device_fops_mmap(struct file *filep, struct vm_area_struct *vma)
1469{
1470	struct vfio_device *device = filep->private_data;
1471
1472	if (unlikely(!device->ops->mmap))
1473		return -EINVAL;
1474
1475	return device->ops->mmap(device->device_data, vma);
1476}
1477
1478#ifdef CONFIG_COMPAT
1479static long vfio_device_fops_compat_ioctl(struct file *filep,
1480					  unsigned int cmd, unsigned long arg)
1481{
1482	arg = (unsigned long)compat_ptr(arg);
1483	return vfio_device_fops_unl_ioctl(filep, cmd, arg);
1484}
1485#endif	/* CONFIG_COMPAT */
1486
1487static const struct file_operations vfio_device_fops = {
1488	.owner		= THIS_MODULE,
1489	.release	= vfio_device_fops_release,
1490	.read		= vfio_device_fops_read,
1491	.write		= vfio_device_fops_write,
1492	.unlocked_ioctl	= vfio_device_fops_unl_ioctl,
1493#ifdef CONFIG_COMPAT
1494	.compat_ioctl	= vfio_device_fops_compat_ioctl,
1495#endif
1496	.mmap		= vfio_device_fops_mmap,
1497};
1498
1499/**
1500 * External user API, exported by symbols to be linked dynamically.
1501 *
1502 * The protocol includes:
1503 *  1. do normal VFIO init operation:
1504 *	- opening a new container;
1505 *	- attaching group(s) to it;
1506 *	- setting an IOMMU driver for a container.
1507 * When IOMMU is set for a container, all groups in it are
1508 * considered ready to use by an external user.
1509 *
1510 * 2. User space passes a group fd to an external user.
1511 * The external user calls vfio_group_get_external_user()
1512 * to verify that:
1513 *	- the group is initialized;
1514 *	- IOMMU is set for it.
1515 * If both checks passed, vfio_group_get_external_user()
1516 * increments the container user counter to prevent
1517 * the VFIO group from disposal before KVM exits.
1518 *
1519 * 3. The external user calls vfio_external_user_iommu_id()
1520 * to know an IOMMU ID.
1521 *
1522 * 4. When the external KVM finishes, it calls
1523 * vfio_group_put_external_user() to release the VFIO group.
1524 * This call decrements the container user counter.
1525 */
1526struct vfio_group *vfio_group_get_external_user(struct file *filep)
1527{
1528	struct vfio_group *group = filep->private_data;
1529
1530	if (filep->f_op != &vfio_group_fops)
1531		return ERR_PTR(-EINVAL);
1532
1533	if (!atomic_inc_not_zero(&group->container_users))
1534		return ERR_PTR(-EINVAL);
1535
1536	if (!group->container->iommu_driver ||
1537			!vfio_group_viable(group)) {
1538		atomic_dec(&group->container_users);
1539		return ERR_PTR(-EINVAL);
1540	}
1541
1542	vfio_group_get(group);
1543
1544	return group;
1545}
1546EXPORT_SYMBOL_GPL(vfio_group_get_external_user);
1547
1548void vfio_group_put_external_user(struct vfio_group *group)
1549{
1550	vfio_group_put(group);
1551	vfio_group_try_dissolve_container(group);
1552}
1553EXPORT_SYMBOL_GPL(vfio_group_put_external_user);
1554
1555int vfio_external_user_iommu_id(struct vfio_group *group)
1556{
1557	return iommu_group_id(group->iommu_group);
1558}
1559EXPORT_SYMBOL_GPL(vfio_external_user_iommu_id);
1560
1561long vfio_external_check_extension(struct vfio_group *group, unsigned long arg)
1562{
1563	return vfio_ioctl_check_extension(group->container, arg);
1564}
1565EXPORT_SYMBOL_GPL(vfio_external_check_extension);
1566
1567/**
1568 * Module/class support
1569 */
1570static char *vfio_devnode(struct device *dev, umode_t *mode)
1571{
1572	return kasprintf(GFP_KERNEL, "vfio/%s", dev_name(dev));
1573}
1574
1575static struct miscdevice vfio_dev = {
1576	.minor = VFIO_MINOR,
1577	.name = "vfio",
1578	.fops = &vfio_fops,
1579	.nodename = "vfio/vfio",
1580	.mode = S_IRUGO | S_IWUGO,
1581};
1582
1583static int __init vfio_init(void)
1584{
1585	int ret;
1586
1587	idr_init(&vfio.group_idr);
1588	mutex_init(&vfio.group_lock);
1589	mutex_init(&vfio.iommu_drivers_lock);
1590	INIT_LIST_HEAD(&vfio.group_list);
1591	INIT_LIST_HEAD(&vfio.iommu_drivers_list);
1592	init_waitqueue_head(&vfio.release_q);
1593
1594	ret = misc_register(&vfio_dev);
1595	if (ret) {
1596		pr_err("vfio: misc device register failed\n");
1597		return ret;
1598	}
1599
1600	/* /dev/vfio/$GROUP */
1601	vfio.class = class_create(THIS_MODULE, "vfio");
1602	if (IS_ERR(vfio.class)) {
1603		ret = PTR_ERR(vfio.class);
1604		goto err_class;
1605	}
1606
1607	vfio.class->devnode = vfio_devnode;
1608
1609	ret = alloc_chrdev_region(&vfio.group_devt, 0, MINORMASK, "vfio");
1610	if (ret)
1611		goto err_alloc_chrdev;
1612
1613	cdev_init(&vfio.group_cdev, &vfio_group_fops);
1614	ret = cdev_add(&vfio.group_cdev, vfio.group_devt, MINORMASK);
1615	if (ret)
1616		goto err_cdev_add;
1617
1618	pr_info(DRIVER_DESC " version: " DRIVER_VERSION "\n");
1619
1620	/*
1621	 * Attempt to load known iommu-drivers.  This gives us a working
1622	 * environment without the user needing to explicitly load iommu
1623	 * drivers.
1624	 */
1625	request_module_nowait("vfio_iommu_type1");
1626	request_module_nowait("vfio_iommu_spapr_tce");
1627
1628	return 0;
1629
1630err_cdev_add:
1631	unregister_chrdev_region(vfio.group_devt, MINORMASK);
1632err_alloc_chrdev:
1633	class_destroy(vfio.class);
1634	vfio.class = NULL;
1635err_class:
1636	misc_deregister(&vfio_dev);
1637	return ret;
1638}
1639
1640static void __exit vfio_cleanup(void)
1641{
1642	WARN_ON(!list_empty(&vfio.group_list));
1643
1644	idr_destroy(&vfio.group_idr);
1645	cdev_del(&vfio.group_cdev);
1646	unregister_chrdev_region(vfio.group_devt, MINORMASK);
1647	class_destroy(vfio.class);
1648	vfio.class = NULL;
1649	misc_deregister(&vfio_dev);
1650}
1651
1652module_init(vfio_init);
1653module_exit(vfio_cleanup);
1654
1655MODULE_VERSION(DRIVER_VERSION);
1656MODULE_LICENSE("GPL v2");
1657MODULE_AUTHOR(DRIVER_AUTHOR);
1658MODULE_DESCRIPTION(DRIVER_DESC);
1659MODULE_ALIAS_MISCDEV(VFIO_MINOR);
1660MODULE_ALIAS("devname:vfio/vfio");
1661