4 * Copyright (C) 2012 Red Hat, Inc. All rights reserved.
5 * Author: Alex Williamson <alex.williamson@redhat.com>
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.
11 * Derived from original vfio:
12 * Copyright 2010 Cisco Systems, Inc. All rights reserved.
13 * Author: Tom Lyon, pugs@cisco.com
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>
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/rwsem.h>
29 #include <linux/sched.h>
30 #include <linux/slab.h>
31 #include <linux/stat.h>
32 #include <linux/string.h>
33 #include <linux/uaccess.h>
34 #include <linux/vfio.h>
35 #include <linux/wait.h>
37 #define DRIVER_VERSION "0.3"
38 #define DRIVER_AUTHOR "Alex Williamson <alex.williamson@redhat.com>"
39 #define DRIVER_DESC "VFIO - User Level meta-driver"
43 struct list_head iommu_drivers_list;
44 struct mutex iommu_drivers_lock;
45 struct list_head group_list;
47 struct mutex group_lock;
48 struct cdev group_cdev;
50 wait_queue_head_t release_q;
53 struct vfio_iommu_driver {
54 const struct vfio_iommu_driver_ops *ops;
55 struct list_head vfio_next;
58 struct vfio_container {
60 struct list_head group_list;
61 struct rw_semaphore group_lock;
62 struct vfio_iommu_driver *iommu_driver;
66 struct vfio_unbound_dev {
68 struct list_head unbound_next;
74 atomic_t container_users;
75 struct iommu_group *iommu_group;
76 struct vfio_container *container;
77 struct list_head device_list;
78 struct mutex device_lock;
80 struct notifier_block nb;
81 struct list_head vfio_next;
82 struct list_head container_next;
83 struct list_head unbound_list;
84 struct mutex unbound_lock;
91 const struct vfio_device_ops *ops;
92 struct vfio_group *group;
93 struct list_head group_next;
98 * IOMMU driver registration
100 int vfio_register_iommu_driver(const struct vfio_iommu_driver_ops *ops)
102 struct vfio_iommu_driver *driver, *tmp;
104 driver = kzalloc(sizeof(*driver), GFP_KERNEL);
110 mutex_lock(&vfio.iommu_drivers_lock);
112 /* Check for duplicates */
113 list_for_each_entry(tmp, &vfio.iommu_drivers_list, vfio_next) {
114 if (tmp->ops == ops) {
115 mutex_unlock(&vfio.iommu_drivers_lock);
121 list_add(&driver->vfio_next, &vfio.iommu_drivers_list);
123 mutex_unlock(&vfio.iommu_drivers_lock);
127 EXPORT_SYMBOL_GPL(vfio_register_iommu_driver);
129 void vfio_unregister_iommu_driver(const struct vfio_iommu_driver_ops *ops)
131 struct vfio_iommu_driver *driver;
133 mutex_lock(&vfio.iommu_drivers_lock);
134 list_for_each_entry(driver, &vfio.iommu_drivers_list, vfio_next) {
135 if (driver->ops == ops) {
136 list_del(&driver->vfio_next);
137 mutex_unlock(&vfio.iommu_drivers_lock);
142 mutex_unlock(&vfio.iommu_drivers_lock);
144 EXPORT_SYMBOL_GPL(vfio_unregister_iommu_driver);
147 * Group minor allocation/free - both called with vfio.group_lock held
149 static int vfio_alloc_group_minor(struct vfio_group *group)
151 return idr_alloc(&vfio.group_idr, group, 0, MINORMASK + 1, GFP_KERNEL);
154 static void vfio_free_group_minor(int minor)
156 idr_remove(&vfio.group_idr, minor);
159 static int vfio_iommu_group_notifier(struct notifier_block *nb,
160 unsigned long action, void *data);
161 static void vfio_group_get(struct vfio_group *group);
164 * Container objects - containers are created when /dev/vfio/vfio is
165 * opened, but their lifecycle extends until the last user is done, so
166 * it's freed via kref. Must support container/group/device being
167 * closed in any order.
169 static void vfio_container_get(struct vfio_container *container)
171 kref_get(&container->kref);
174 static void vfio_container_release(struct kref *kref)
176 struct vfio_container *container;
177 container = container_of(kref, struct vfio_container, kref);
182 static void vfio_container_put(struct vfio_container *container)
184 kref_put(&container->kref, vfio_container_release);
187 static void vfio_group_unlock_and_free(struct vfio_group *group)
189 mutex_unlock(&vfio.group_lock);
191 * Unregister outside of lock. A spurious callback is harmless now
192 * that the group is no longer in vfio.group_list.
194 iommu_group_unregister_notifier(group->iommu_group, &group->nb);
199 * Group objects - create, release, get, put, search
201 static struct vfio_group *vfio_create_group(struct iommu_group *iommu_group)
203 struct vfio_group *group, *tmp;
207 group = kzalloc(sizeof(*group), GFP_KERNEL);
209 return ERR_PTR(-ENOMEM);
211 kref_init(&group->kref);
212 INIT_LIST_HEAD(&group->device_list);
213 mutex_init(&group->device_lock);
214 INIT_LIST_HEAD(&group->unbound_list);
215 mutex_init(&group->unbound_lock);
216 atomic_set(&group->container_users, 0);
217 atomic_set(&group->opened, 0);
218 group->iommu_group = iommu_group;
220 group->nb.notifier_call = vfio_iommu_group_notifier;
223 * blocking notifiers acquire a rwsem around registering and hold
224 * it around callback. Therefore, need to register outside of
225 * vfio.group_lock to avoid A-B/B-A contention. Our callback won't
226 * do anything unless it can find the group in vfio.group_list, so
227 * no harm in registering early.
229 ret = iommu_group_register_notifier(iommu_group, &group->nb);
235 mutex_lock(&vfio.group_lock);
237 /* Did we race creating this group? */
238 list_for_each_entry(tmp, &vfio.group_list, vfio_next) {
239 if (tmp->iommu_group == iommu_group) {
241 vfio_group_unlock_and_free(group);
246 minor = vfio_alloc_group_minor(group);
248 vfio_group_unlock_and_free(group);
249 return ERR_PTR(minor);
252 dev = device_create(vfio.class, NULL,
253 MKDEV(MAJOR(vfio.group_devt), minor),
254 group, "%d", iommu_group_id(iommu_group));
256 vfio_free_group_minor(minor);
257 vfio_group_unlock_and_free(group);
258 return (struct vfio_group *)dev; /* ERR_PTR */
261 group->minor = minor;
264 list_add(&group->vfio_next, &vfio.group_list);
266 mutex_unlock(&vfio.group_lock);
271 /* called with vfio.group_lock held */
272 static void vfio_group_release(struct kref *kref)
274 struct vfio_group *group = container_of(kref, struct vfio_group, kref);
275 struct vfio_unbound_dev *unbound, *tmp;
276 struct iommu_group *iommu_group = group->iommu_group;
278 WARN_ON(!list_empty(&group->device_list));
280 list_for_each_entry_safe(unbound, tmp,
281 &group->unbound_list, unbound_next) {
282 list_del(&unbound->unbound_next);
286 device_destroy(vfio.class, MKDEV(MAJOR(vfio.group_devt), group->minor));
287 list_del(&group->vfio_next);
288 vfio_free_group_minor(group->minor);
289 vfio_group_unlock_and_free(group);
290 iommu_group_put(iommu_group);
293 static void vfio_group_put(struct vfio_group *group)
295 kref_put_mutex(&group->kref, vfio_group_release, &vfio.group_lock);
298 /* Assume group_lock or group reference is held */
299 static void vfio_group_get(struct vfio_group *group)
301 kref_get(&group->kref);
305 * Not really a try as we will sleep for mutex, but we need to make
306 * sure the group pointer is valid under lock and get a reference.
308 static struct vfio_group *vfio_group_try_get(struct vfio_group *group)
310 struct vfio_group *target = group;
312 mutex_lock(&vfio.group_lock);
313 list_for_each_entry(group, &vfio.group_list, vfio_next) {
314 if (group == target) {
315 vfio_group_get(group);
316 mutex_unlock(&vfio.group_lock);
320 mutex_unlock(&vfio.group_lock);
326 struct vfio_group *vfio_group_get_from_iommu(struct iommu_group *iommu_group)
328 struct vfio_group *group;
330 mutex_lock(&vfio.group_lock);
331 list_for_each_entry(group, &vfio.group_list, vfio_next) {
332 if (group->iommu_group == iommu_group) {
333 vfio_group_get(group);
334 mutex_unlock(&vfio.group_lock);
338 mutex_unlock(&vfio.group_lock);
343 static struct vfio_group *vfio_group_get_from_minor(int minor)
345 struct vfio_group *group;
347 mutex_lock(&vfio.group_lock);
348 group = idr_find(&vfio.group_idr, minor);
350 mutex_unlock(&vfio.group_lock);
353 vfio_group_get(group);
354 mutex_unlock(&vfio.group_lock);
360 * Device objects - create, release, get, put, search
363 struct vfio_device *vfio_group_create_device(struct vfio_group *group,
365 const struct vfio_device_ops *ops,
368 struct vfio_device *device;
370 device = kzalloc(sizeof(*device), GFP_KERNEL);
372 return ERR_PTR(-ENOMEM);
374 kref_init(&device->kref);
376 device->group = group;
378 device->device_data = device_data;
379 dev_set_drvdata(dev, device);
381 /* No need to get group_lock, caller has group reference */
382 vfio_group_get(group);
384 mutex_lock(&group->device_lock);
385 list_add(&device->group_next, &group->device_list);
386 mutex_unlock(&group->device_lock);
391 static void vfio_device_release(struct kref *kref)
393 struct vfio_device *device = container_of(kref,
394 struct vfio_device, kref);
395 struct vfio_group *group = device->group;
397 list_del(&device->group_next);
398 mutex_unlock(&group->device_lock);
400 dev_set_drvdata(device->dev, NULL);
404 /* vfio_del_group_dev may be waiting for this device */
405 wake_up(&vfio.release_q);
408 /* Device reference always implies a group reference */
409 void vfio_device_put(struct vfio_device *device)
411 struct vfio_group *group = device->group;
412 kref_put_mutex(&device->kref, vfio_device_release, &group->device_lock);
413 vfio_group_put(group);
415 EXPORT_SYMBOL_GPL(vfio_device_put);
417 static void vfio_device_get(struct vfio_device *device)
419 vfio_group_get(device->group);
420 kref_get(&device->kref);
423 static struct vfio_device *vfio_group_get_device(struct vfio_group *group,
426 struct vfio_device *device;
428 mutex_lock(&group->device_lock);
429 list_for_each_entry(device, &group->device_list, group_next) {
430 if (device->dev == dev) {
431 vfio_device_get(device);
432 mutex_unlock(&group->device_lock);
436 mutex_unlock(&group->device_lock);
441 * Whitelist some drivers that we know are safe (no dma) or just sit on
442 * a device. It's not always practical to leave a device within a group
443 * driverless as it could get re-bound to something unsafe.
445 static const char * const vfio_driver_whitelist[] = { "pci-stub", "pcieport" };
447 static bool vfio_whitelisted_driver(struct device_driver *drv)
451 for (i = 0; i < ARRAY_SIZE(vfio_driver_whitelist); i++) {
452 if (!strcmp(drv->name, vfio_driver_whitelist[i]))
460 * A vfio group is viable for use by userspace if all devices are in
461 * one of the following states:
463 * - bound to a vfio driver
464 * - bound to a whitelisted driver
466 * We use two methods to determine whether a device is bound to a vfio
467 * driver. The first is to test whether the device exists in the vfio
468 * group. The second is to test if the device exists on the group
469 * unbound_list, indicating it's in the middle of transitioning from
470 * a vfio driver to driver-less.
472 static int vfio_dev_viable(struct device *dev, void *data)
474 struct vfio_group *group = data;
475 struct vfio_device *device;
476 struct device_driver *drv = ACCESS_ONCE(dev->driver);
477 struct vfio_unbound_dev *unbound;
480 mutex_lock(&group->unbound_lock);
481 list_for_each_entry(unbound, &group->unbound_list, unbound_next) {
482 if (dev == unbound->dev) {
487 mutex_unlock(&group->unbound_lock);
489 if (!ret || !drv || vfio_whitelisted_driver(drv))
492 device = vfio_group_get_device(group, dev);
494 vfio_device_put(device);
502 * Async device support
504 static int vfio_group_nb_add_dev(struct vfio_group *group, struct device *dev)
506 struct vfio_device *device;
508 /* Do we already know about it? We shouldn't */
509 device = vfio_group_get_device(group, dev);
510 if (WARN_ON_ONCE(device)) {
511 vfio_device_put(device);
515 /* Nothing to do for idle groups */
516 if (!atomic_read(&group->container_users))
519 /* TODO Prevent device auto probing */
520 WARN("Device %s added to live group %d!\n", dev_name(dev),
521 iommu_group_id(group->iommu_group));
526 static int vfio_group_nb_verify(struct vfio_group *group, struct device *dev)
528 /* We don't care what happens when the group isn't in use */
529 if (!atomic_read(&group->container_users))
532 return vfio_dev_viable(dev, group);
535 static int vfio_iommu_group_notifier(struct notifier_block *nb,
536 unsigned long action, void *data)
538 struct vfio_group *group = container_of(nb, struct vfio_group, nb);
539 struct device *dev = data;
540 struct vfio_unbound_dev *unbound;
543 * Need to go through a group_lock lookup to get a reference or we
544 * risk racing a group being removed. Ignore spurious notifies.
546 group = vfio_group_try_get(group);
551 case IOMMU_GROUP_NOTIFY_ADD_DEVICE:
552 vfio_group_nb_add_dev(group, dev);
554 case IOMMU_GROUP_NOTIFY_DEL_DEVICE:
556 * Nothing to do here. If the device is in use, then the
557 * vfio sub-driver should block the remove callback until
558 * it is unused. If the device is unused or attached to a
559 * stub driver, then it should be released and we don't
560 * care that it will be going away.
563 case IOMMU_GROUP_NOTIFY_BIND_DRIVER:
564 pr_debug("%s: Device %s, group %d binding to driver\n",
565 __func__, dev_name(dev),
566 iommu_group_id(group->iommu_group));
568 case IOMMU_GROUP_NOTIFY_BOUND_DRIVER:
569 pr_debug("%s: Device %s, group %d bound to driver %s\n",
570 __func__, dev_name(dev),
571 iommu_group_id(group->iommu_group), dev->driver->name);
572 BUG_ON(vfio_group_nb_verify(group, dev));
574 case IOMMU_GROUP_NOTIFY_UNBIND_DRIVER:
575 pr_debug("%s: Device %s, group %d unbinding from driver %s\n",
576 __func__, dev_name(dev),
577 iommu_group_id(group->iommu_group), dev->driver->name);
579 case IOMMU_GROUP_NOTIFY_UNBOUND_DRIVER:
580 pr_debug("%s: Device %s, group %d unbound from driver\n",
581 __func__, dev_name(dev),
582 iommu_group_id(group->iommu_group));
584 * XXX An unbound device in a live group is ok, but we'd
585 * really like to avoid the above BUG_ON by preventing other
586 * drivers from binding to it. Once that occurs, we have to
587 * stop the system to maintain isolation. At a minimum, we'd
588 * want a toggle to disable driver auto probe for this device.
591 mutex_lock(&group->unbound_lock);
592 list_for_each_entry(unbound,
593 &group->unbound_list, unbound_next) {
594 if (dev == unbound->dev) {
595 list_del(&unbound->unbound_next);
600 mutex_unlock(&group->unbound_lock);
604 vfio_group_put(group);
611 int vfio_add_group_dev(struct device *dev,
612 const struct vfio_device_ops *ops, void *device_data)
614 struct iommu_group *iommu_group;
615 struct vfio_group *group;
616 struct vfio_device *device;
618 iommu_group = iommu_group_get(dev);
622 group = vfio_group_get_from_iommu(iommu_group);
624 group = vfio_create_group(iommu_group);
626 iommu_group_put(iommu_group);
627 return PTR_ERR(group);
631 * A found vfio_group already holds a reference to the
632 * iommu_group. A created vfio_group keeps the reference.
634 iommu_group_put(iommu_group);
637 device = vfio_group_get_device(group, dev);
639 WARN(1, "Device %s already exists on group %d\n",
640 dev_name(dev), iommu_group_id(iommu_group));
641 vfio_device_put(device);
642 vfio_group_put(group);
646 device = vfio_group_create_device(group, dev, ops, device_data);
647 if (IS_ERR(device)) {
648 vfio_group_put(group);
649 return PTR_ERR(device);
653 * Drop all but the vfio_device reference. The vfio_device holds
654 * a reference to the vfio_group, which holds a reference to the
657 vfio_group_put(group);
661 EXPORT_SYMBOL_GPL(vfio_add_group_dev);
664 * Get a reference to the vfio_device for a device. Even if the
665 * caller thinks they own the device, they could be racing with a
666 * release call path, so we can't trust drvdata for the shortcut.
667 * Go the long way around, from the iommu_group to the vfio_group
668 * to the vfio_device.
670 struct vfio_device *vfio_device_get_from_dev(struct device *dev)
672 struct iommu_group *iommu_group;
673 struct vfio_group *group;
674 struct vfio_device *device;
676 iommu_group = iommu_group_get(dev);
680 group = vfio_group_get_from_iommu(iommu_group);
681 iommu_group_put(iommu_group);
685 device = vfio_group_get_device(group, dev);
686 vfio_group_put(group);
690 EXPORT_SYMBOL_GPL(vfio_device_get_from_dev);
693 * Caller must hold a reference to the vfio_device
695 void *vfio_device_data(struct vfio_device *device)
697 return device->device_data;
699 EXPORT_SYMBOL_GPL(vfio_device_data);
701 /* Given a referenced group, check if it contains the device */
702 static bool vfio_dev_present(struct vfio_group *group, struct device *dev)
704 struct vfio_device *device;
706 device = vfio_group_get_device(group, dev);
710 vfio_device_put(device);
715 * Decrement the device reference count and wait for the device to be
716 * removed. Open file descriptors for the device... */
717 void *vfio_del_group_dev(struct device *dev)
719 struct vfio_device *device = dev_get_drvdata(dev);
720 struct vfio_group *group = device->group;
721 void *device_data = device->device_data;
722 struct vfio_unbound_dev *unbound;
725 bool interrupted = false;
728 * The group exists so long as we have a device reference. Get
729 * a group reference and use it to scan for the device going away.
731 vfio_group_get(group);
734 * When the device is removed from the group, the group suddenly
735 * becomes non-viable; the device has a driver (until the unbind
736 * completes), but it's not present in the group. This is bad news
737 * for any external users that need to re-acquire a group reference
738 * in order to match and release their existing reference. To
739 * solve this, we track such devices on the unbound_list to bridge
740 * the gap until they're fully unbound.
742 unbound = kzalloc(sizeof(*unbound), GFP_KERNEL);
745 mutex_lock(&group->unbound_lock);
746 list_add(&unbound->unbound_next, &group->unbound_list);
747 mutex_unlock(&group->unbound_lock);
751 vfio_device_put(device);
754 * If the device is still present in the group after the above
755 * 'put', then it is in use and we need to request it from the
756 * bus driver. The driver may in turn need to request the
757 * device from the user. We send the request on an arbitrary
758 * interval with counter to allow the driver to take escalating
759 * measures to release the device if it has the ability to do so.
762 device = vfio_group_get_device(group, dev);
766 if (device->ops->request)
767 device->ops->request(device_data, i++);
769 vfio_device_put(device);
772 ret = wait_event_timeout(vfio.release_q,
773 !vfio_dev_present(group, dev), HZ * 10);
775 ret = wait_event_interruptible_timeout(vfio.release_q,
776 !vfio_dev_present(group, dev), HZ * 10);
777 if (ret == -ERESTARTSYS) {
780 "Device is currently in use, task"
782 "blocked until device is released",
783 current->comm, task_pid_nr(current));
788 vfio_group_put(group);
792 EXPORT_SYMBOL_GPL(vfio_del_group_dev);
795 * VFIO base fd, /dev/vfio/vfio
797 static long vfio_ioctl_check_extension(struct vfio_container *container,
800 struct vfio_iommu_driver *driver;
803 down_read(&container->group_lock);
805 driver = container->iommu_driver;
808 /* No base extensions yet */
811 * If no driver is set, poll all registered drivers for
812 * extensions and return the first positive result. If
813 * a driver is already set, further queries will be passed
814 * only to that driver.
817 mutex_lock(&vfio.iommu_drivers_lock);
818 list_for_each_entry(driver, &vfio.iommu_drivers_list,
820 if (!try_module_get(driver->ops->owner))
823 ret = driver->ops->ioctl(NULL,
824 VFIO_CHECK_EXTENSION,
826 module_put(driver->ops->owner);
830 mutex_unlock(&vfio.iommu_drivers_lock);
832 ret = driver->ops->ioctl(container->iommu_data,
833 VFIO_CHECK_EXTENSION, arg);
836 up_read(&container->group_lock);
841 /* hold write lock on container->group_lock */
842 static int __vfio_container_attach_groups(struct vfio_container *container,
843 struct vfio_iommu_driver *driver,
846 struct vfio_group *group;
849 list_for_each_entry(group, &container->group_list, container_next) {
850 ret = driver->ops->attach_group(data, group->iommu_group);
858 list_for_each_entry_continue_reverse(group, &container->group_list,
860 driver->ops->detach_group(data, group->iommu_group);
866 static long vfio_ioctl_set_iommu(struct vfio_container *container,
869 struct vfio_iommu_driver *driver;
872 down_write(&container->group_lock);
875 * The container is designed to be an unprivileged interface while
876 * the group can be assigned to specific users. Therefore, only by
877 * adding a group to a container does the user get the privilege of
878 * enabling the iommu, which may allocate finite resources. There
879 * is no unset_iommu, but by removing all the groups from a container,
880 * the container is deprivileged and returns to an unset state.
882 if (list_empty(&container->group_list) || container->iommu_driver) {
883 up_write(&container->group_lock);
887 mutex_lock(&vfio.iommu_drivers_lock);
888 list_for_each_entry(driver, &vfio.iommu_drivers_list, vfio_next) {
891 if (!try_module_get(driver->ops->owner))
895 * The arg magic for SET_IOMMU is the same as CHECK_EXTENSION,
896 * so test which iommu driver reported support for this
897 * extension and call open on them. We also pass them the
898 * magic, allowing a single driver to support multiple
899 * interfaces if they'd like.
901 if (driver->ops->ioctl(NULL, VFIO_CHECK_EXTENSION, arg) <= 0) {
902 module_put(driver->ops->owner);
906 /* module reference holds the driver we're working on */
907 mutex_unlock(&vfio.iommu_drivers_lock);
909 data = driver->ops->open(arg);
912 module_put(driver->ops->owner);
913 goto skip_drivers_unlock;
916 ret = __vfio_container_attach_groups(container, driver, data);
918 container->iommu_driver = driver;
919 container->iommu_data = data;
921 driver->ops->release(data);
922 module_put(driver->ops->owner);
925 goto skip_drivers_unlock;
928 mutex_unlock(&vfio.iommu_drivers_lock);
930 up_write(&container->group_lock);
935 static long vfio_fops_unl_ioctl(struct file *filep,
936 unsigned int cmd, unsigned long arg)
938 struct vfio_container *container = filep->private_data;
939 struct vfio_iommu_driver *driver;
947 case VFIO_GET_API_VERSION:
948 ret = VFIO_API_VERSION;
950 case VFIO_CHECK_EXTENSION:
951 ret = vfio_ioctl_check_extension(container, arg);
954 ret = vfio_ioctl_set_iommu(container, arg);
957 down_read(&container->group_lock);
959 driver = container->iommu_driver;
960 data = container->iommu_data;
962 if (driver) /* passthrough all unrecognized ioctls */
963 ret = driver->ops->ioctl(data, cmd, arg);
965 up_read(&container->group_lock);
972 static long vfio_fops_compat_ioctl(struct file *filep,
973 unsigned int cmd, unsigned long arg)
975 arg = (unsigned long)compat_ptr(arg);
976 return vfio_fops_unl_ioctl(filep, cmd, arg);
978 #endif /* CONFIG_COMPAT */
980 static int vfio_fops_open(struct inode *inode, struct file *filep)
982 struct vfio_container *container;
984 container = kzalloc(sizeof(*container), GFP_KERNEL);
988 INIT_LIST_HEAD(&container->group_list);
989 init_rwsem(&container->group_lock);
990 kref_init(&container->kref);
992 filep->private_data = container;
997 static int vfio_fops_release(struct inode *inode, struct file *filep)
999 struct vfio_container *container = filep->private_data;
1001 filep->private_data = NULL;
1003 vfio_container_put(container);
1009 * Once an iommu driver is set, we optionally pass read/write/mmap
1010 * on to the driver, allowing management interfaces beyond ioctl.
1012 static ssize_t vfio_fops_read(struct file *filep, char __user *buf,
1013 size_t count, loff_t *ppos)
1015 struct vfio_container *container = filep->private_data;
1016 struct vfio_iommu_driver *driver;
1017 ssize_t ret = -EINVAL;
1019 down_read(&container->group_lock);
1021 driver = container->iommu_driver;
1022 if (likely(driver && driver->ops->read))
1023 ret = driver->ops->read(container->iommu_data,
1026 up_read(&container->group_lock);
1031 static ssize_t vfio_fops_write(struct file *filep, const char __user *buf,
1032 size_t count, loff_t *ppos)
1034 struct vfio_container *container = filep->private_data;
1035 struct vfio_iommu_driver *driver;
1036 ssize_t ret = -EINVAL;
1038 down_read(&container->group_lock);
1040 driver = container->iommu_driver;
1041 if (likely(driver && driver->ops->write))
1042 ret = driver->ops->write(container->iommu_data,
1045 up_read(&container->group_lock);
1050 static int vfio_fops_mmap(struct file *filep, struct vm_area_struct *vma)
1052 struct vfio_container *container = filep->private_data;
1053 struct vfio_iommu_driver *driver;
1056 down_read(&container->group_lock);
1058 driver = container->iommu_driver;
1059 if (likely(driver && driver->ops->mmap))
1060 ret = driver->ops->mmap(container->iommu_data, vma);
1062 up_read(&container->group_lock);
1067 static const struct file_operations vfio_fops = {
1068 .owner = THIS_MODULE,
1069 .open = vfio_fops_open,
1070 .release = vfio_fops_release,
1071 .read = vfio_fops_read,
1072 .write = vfio_fops_write,
1073 .unlocked_ioctl = vfio_fops_unl_ioctl,
1074 #ifdef CONFIG_COMPAT
1075 .compat_ioctl = vfio_fops_compat_ioctl,
1077 .mmap = vfio_fops_mmap,
1081 * VFIO Group fd, /dev/vfio/$GROUP
1083 static void __vfio_group_unset_container(struct vfio_group *group)
1085 struct vfio_container *container = group->container;
1086 struct vfio_iommu_driver *driver;
1088 down_write(&container->group_lock);
1090 driver = container->iommu_driver;
1092 driver->ops->detach_group(container->iommu_data,
1093 group->iommu_group);
1095 group->container = NULL;
1096 list_del(&group->container_next);
1098 /* Detaching the last group deprivileges a container, remove iommu */
1099 if (driver && list_empty(&container->group_list)) {
1100 driver->ops->release(container->iommu_data);
1101 module_put(driver->ops->owner);
1102 container->iommu_driver = NULL;
1103 container->iommu_data = NULL;
1106 up_write(&container->group_lock);
1108 vfio_container_put(container);
1112 * VFIO_GROUP_UNSET_CONTAINER should fail if there are other users or
1113 * if there was no container to unset. Since the ioctl is called on
1114 * the group, we know that still exists, therefore the only valid
1115 * transition here is 1->0.
1117 static int vfio_group_unset_container(struct vfio_group *group)
1119 int users = atomic_cmpxchg(&group->container_users, 1, 0);
1126 __vfio_group_unset_container(group);
1132 * When removing container users, anything that removes the last user
1133 * implicitly removes the group from the container. That is, if the
1134 * group file descriptor is closed, as well as any device file descriptors,
1135 * the group is free.
1137 static void vfio_group_try_dissolve_container(struct vfio_group *group)
1139 if (0 == atomic_dec_if_positive(&group->container_users))
1140 __vfio_group_unset_container(group);
1143 static int vfio_group_set_container(struct vfio_group *group, int container_fd)
1146 struct vfio_container *container;
1147 struct vfio_iommu_driver *driver;
1150 if (atomic_read(&group->container_users))
1153 f = fdget(container_fd);
1157 /* Sanity check, is this really our fd? */
1158 if (f.file->f_op != &vfio_fops) {
1163 container = f.file->private_data;
1164 WARN_ON(!container); /* fget ensures we don't race vfio_release */
1166 down_write(&container->group_lock);
1168 driver = container->iommu_driver;
1170 ret = driver->ops->attach_group(container->iommu_data,
1171 group->iommu_group);
1176 group->container = container;
1177 list_add(&group->container_next, &container->group_list);
1179 /* Get a reference on the container and mark a user within the group */
1180 vfio_container_get(container);
1181 atomic_inc(&group->container_users);
1184 up_write(&container->group_lock);
1189 static bool vfio_group_viable(struct vfio_group *group)
1191 return (iommu_group_for_each_dev(group->iommu_group,
1192 group, vfio_dev_viable) == 0);
1195 static const struct file_operations vfio_device_fops;
1197 static int vfio_group_get_device_fd(struct vfio_group *group, char *buf)
1199 struct vfio_device *device;
1203 if (0 == atomic_read(&group->container_users) ||
1204 !group->container->iommu_driver || !vfio_group_viable(group))
1207 mutex_lock(&group->device_lock);
1208 list_for_each_entry(device, &group->device_list, group_next) {
1209 if (strcmp(dev_name(device->dev), buf))
1212 ret = device->ops->open(device->device_data);
1216 * We can't use anon_inode_getfd() because we need to modify
1217 * the f_mode flags directly to allow more than just ioctls
1219 ret = get_unused_fd_flags(O_CLOEXEC);
1221 device->ops->release(device->device_data);
1225 filep = anon_inode_getfile("[vfio-device]", &vfio_device_fops,
1227 if (IS_ERR(filep)) {
1229 ret = PTR_ERR(filep);
1230 device->ops->release(device->device_data);
1235 * TODO: add an anon_inode interface to do this.
1236 * Appears to be missing by lack of need rather than
1237 * explicitly prevented. Now there's need.
1239 filep->f_mode |= (FMODE_LSEEK | FMODE_PREAD | FMODE_PWRITE);
1241 vfio_device_get(device);
1242 atomic_inc(&group->container_users);
1244 fd_install(ret, filep);
1247 mutex_unlock(&group->device_lock);
1252 static long vfio_group_fops_unl_ioctl(struct file *filep,
1253 unsigned int cmd, unsigned long arg)
1255 struct vfio_group *group = filep->private_data;
1259 case VFIO_GROUP_GET_STATUS:
1261 struct vfio_group_status status;
1262 unsigned long minsz;
1264 minsz = offsetofend(struct vfio_group_status, flags);
1266 if (copy_from_user(&status, (void __user *)arg, minsz))
1269 if (status.argsz < minsz)
1274 if (vfio_group_viable(group))
1275 status.flags |= VFIO_GROUP_FLAGS_VIABLE;
1277 if (group->container)
1278 status.flags |= VFIO_GROUP_FLAGS_CONTAINER_SET;
1280 if (copy_to_user((void __user *)arg, &status, minsz))
1286 case VFIO_GROUP_SET_CONTAINER:
1290 if (get_user(fd, (int __user *)arg))
1296 ret = vfio_group_set_container(group, fd);
1299 case VFIO_GROUP_UNSET_CONTAINER:
1300 ret = vfio_group_unset_container(group);
1302 case VFIO_GROUP_GET_DEVICE_FD:
1306 buf = strndup_user((const char __user *)arg, PAGE_SIZE);
1308 return PTR_ERR(buf);
1310 ret = vfio_group_get_device_fd(group, buf);
1319 #ifdef CONFIG_COMPAT
1320 static long vfio_group_fops_compat_ioctl(struct file *filep,
1321 unsigned int cmd, unsigned long arg)
1323 arg = (unsigned long)compat_ptr(arg);
1324 return vfio_group_fops_unl_ioctl(filep, cmd, arg);
1326 #endif /* CONFIG_COMPAT */
1328 static int vfio_group_fops_open(struct inode *inode, struct file *filep)
1330 struct vfio_group *group;
1333 group = vfio_group_get_from_minor(iminor(inode));
1337 /* Do we need multiple instances of the group open? Seems not. */
1338 opened = atomic_cmpxchg(&group->opened, 0, 1);
1340 vfio_group_put(group);
1344 /* Is something still in use from a previous open? */
1345 if (group->container) {
1346 atomic_dec(&group->opened);
1347 vfio_group_put(group);
1351 filep->private_data = group;
1356 static int vfio_group_fops_release(struct inode *inode, struct file *filep)
1358 struct vfio_group *group = filep->private_data;
1360 filep->private_data = NULL;
1362 vfio_group_try_dissolve_container(group);
1364 atomic_dec(&group->opened);
1366 vfio_group_put(group);
1371 static const struct file_operations vfio_group_fops = {
1372 .owner = THIS_MODULE,
1373 .unlocked_ioctl = vfio_group_fops_unl_ioctl,
1374 #ifdef CONFIG_COMPAT
1375 .compat_ioctl = vfio_group_fops_compat_ioctl,
1377 .open = vfio_group_fops_open,
1378 .release = vfio_group_fops_release,
1384 static int vfio_device_fops_release(struct inode *inode, struct file *filep)
1386 struct vfio_device *device = filep->private_data;
1388 device->ops->release(device->device_data);
1390 vfio_group_try_dissolve_container(device->group);
1392 vfio_device_put(device);
1397 static long vfio_device_fops_unl_ioctl(struct file *filep,
1398 unsigned int cmd, unsigned long arg)
1400 struct vfio_device *device = filep->private_data;
1402 if (unlikely(!device->ops->ioctl))
1405 return device->ops->ioctl(device->device_data, cmd, arg);
1408 static ssize_t vfio_device_fops_read(struct file *filep, char __user *buf,
1409 size_t count, loff_t *ppos)
1411 struct vfio_device *device = filep->private_data;
1413 if (unlikely(!device->ops->read))
1416 return device->ops->read(device->device_data, buf, count, ppos);
1419 static ssize_t vfio_device_fops_write(struct file *filep,
1420 const char __user *buf,
1421 size_t count, loff_t *ppos)
1423 struct vfio_device *device = filep->private_data;
1425 if (unlikely(!device->ops->write))
1428 return device->ops->write(device->device_data, buf, count, ppos);
1431 static int vfio_device_fops_mmap(struct file *filep, struct vm_area_struct *vma)
1433 struct vfio_device *device = filep->private_data;
1435 if (unlikely(!device->ops->mmap))
1438 return device->ops->mmap(device->device_data, vma);
1441 #ifdef CONFIG_COMPAT
1442 static long vfio_device_fops_compat_ioctl(struct file *filep,
1443 unsigned int cmd, unsigned long arg)
1445 arg = (unsigned long)compat_ptr(arg);
1446 return vfio_device_fops_unl_ioctl(filep, cmd, arg);
1448 #endif /* CONFIG_COMPAT */
1450 static const struct file_operations vfio_device_fops = {
1451 .owner = THIS_MODULE,
1452 .release = vfio_device_fops_release,
1453 .read = vfio_device_fops_read,
1454 .write = vfio_device_fops_write,
1455 .unlocked_ioctl = vfio_device_fops_unl_ioctl,
1456 #ifdef CONFIG_COMPAT
1457 .compat_ioctl = vfio_device_fops_compat_ioctl,
1459 .mmap = vfio_device_fops_mmap,
1463 * External user API, exported by symbols to be linked dynamically.
1465 * The protocol includes:
1466 * 1. do normal VFIO init operation:
1467 * - opening a new container;
1468 * - attaching group(s) to it;
1469 * - setting an IOMMU driver for a container.
1470 * When IOMMU is set for a container, all groups in it are
1471 * considered ready to use by an external user.
1473 * 2. User space passes a group fd to an external user.
1474 * The external user calls vfio_group_get_external_user()
1476 * - the group is initialized;
1477 * - IOMMU is set for it.
1478 * If both checks passed, vfio_group_get_external_user()
1479 * increments the container user counter to prevent
1480 * the VFIO group from disposal before KVM exits.
1482 * 3. The external user calls vfio_external_user_iommu_id()
1483 * to know an IOMMU ID.
1485 * 4. When the external KVM finishes, it calls
1486 * vfio_group_put_external_user() to release the VFIO group.
1487 * This call decrements the container user counter.
1489 struct vfio_group *vfio_group_get_external_user(struct file *filep)
1491 struct vfio_group *group = filep->private_data;
1493 if (filep->f_op != &vfio_group_fops)
1494 return ERR_PTR(-EINVAL);
1496 if (!atomic_inc_not_zero(&group->container_users))
1497 return ERR_PTR(-EINVAL);
1499 if (!group->container->iommu_driver ||
1500 !vfio_group_viable(group)) {
1501 atomic_dec(&group->container_users);
1502 return ERR_PTR(-EINVAL);
1505 vfio_group_get(group);
1509 EXPORT_SYMBOL_GPL(vfio_group_get_external_user);
1511 void vfio_group_put_external_user(struct vfio_group *group)
1513 vfio_group_put(group);
1514 vfio_group_try_dissolve_container(group);
1516 EXPORT_SYMBOL_GPL(vfio_group_put_external_user);
1518 int vfio_external_user_iommu_id(struct vfio_group *group)
1520 return iommu_group_id(group->iommu_group);
1522 EXPORT_SYMBOL_GPL(vfio_external_user_iommu_id);
1524 long vfio_external_check_extension(struct vfio_group *group, unsigned long arg)
1526 return vfio_ioctl_check_extension(group->container, arg);
1528 EXPORT_SYMBOL_GPL(vfio_external_check_extension);
1531 * Module/class support
1533 static char *vfio_devnode(struct device *dev, umode_t *mode)
1535 return kasprintf(GFP_KERNEL, "vfio/%s", dev_name(dev));
1538 static struct miscdevice vfio_dev = {
1539 .minor = VFIO_MINOR,
1542 .nodename = "vfio/vfio",
1543 .mode = S_IRUGO | S_IWUGO,
1546 static int __init vfio_init(void)
1550 idr_init(&vfio.group_idr);
1551 mutex_init(&vfio.group_lock);
1552 mutex_init(&vfio.iommu_drivers_lock);
1553 INIT_LIST_HEAD(&vfio.group_list);
1554 INIT_LIST_HEAD(&vfio.iommu_drivers_list);
1555 init_waitqueue_head(&vfio.release_q);
1557 ret = misc_register(&vfio_dev);
1559 pr_err("vfio: misc device register failed\n");
1563 /* /dev/vfio/$GROUP */
1564 vfio.class = class_create(THIS_MODULE, "vfio");
1565 if (IS_ERR(vfio.class)) {
1566 ret = PTR_ERR(vfio.class);
1570 vfio.class->devnode = vfio_devnode;
1572 ret = alloc_chrdev_region(&vfio.group_devt, 0, MINORMASK, "vfio");
1574 goto err_alloc_chrdev;
1576 cdev_init(&vfio.group_cdev, &vfio_group_fops);
1577 ret = cdev_add(&vfio.group_cdev, vfio.group_devt, MINORMASK);
1581 pr_info(DRIVER_DESC " version: " DRIVER_VERSION "\n");
1584 * Attempt to load known iommu-drivers. This gives us a working
1585 * environment without the user needing to explicitly load iommu
1588 request_module_nowait("vfio_iommu_type1");
1589 request_module_nowait("vfio_iommu_spapr_tce");
1594 unregister_chrdev_region(vfio.group_devt, MINORMASK);
1596 class_destroy(vfio.class);
1599 misc_deregister(&vfio_dev);
1603 static void __exit vfio_cleanup(void)
1605 WARN_ON(!list_empty(&vfio.group_list));
1607 idr_destroy(&vfio.group_idr);
1608 cdev_del(&vfio.group_cdev);
1609 unregister_chrdev_region(vfio.group_devt, MINORMASK);
1610 class_destroy(vfio.class);
1612 misc_deregister(&vfio_dev);
1615 module_init(vfio_init);
1616 module_exit(vfio_cleanup);
1618 MODULE_VERSION(DRIVER_VERSION);
1619 MODULE_LICENSE("GPL v2");
1620 MODULE_AUTHOR(DRIVER_AUTHOR);
1621 MODULE_DESCRIPTION(DRIVER_DESC);
1622 MODULE_ALIAS_MISCDEV(VFIO_MINOR);
1623 MODULE_ALIAS("devname:vfio/vfio");