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/module.h>
26 #include <linux/mutex.h>
27 #include <linux/sched.h>
28 #include <linux/slab.h>
29 #include <linux/string.h>
30 #include <linux/uaccess.h>
31 #include <linux/vfio.h>
32 #include <linux/wait.h>
34 #define DRIVER_VERSION "0.3"
35 #define DRIVER_AUTHOR "Alex Williamson <alex.williamson@redhat.com>"
36 #define DRIVER_DESC "VFIO - User Level meta-driver"
40 struct list_head iommu_drivers_list;
41 struct mutex iommu_drivers_lock;
42 struct list_head group_list;
44 struct mutex group_lock;
45 struct cdev group_cdev;
49 wait_queue_head_t release_q;
52 struct vfio_iommu_driver {
53 const struct vfio_iommu_driver_ops *ops;
54 struct list_head vfio_next;
57 struct vfio_container {
59 struct list_head group_list;
60 struct mutex group_lock;
61 struct vfio_iommu_driver *iommu_driver;
68 atomic_t container_users;
69 struct iommu_group *iommu_group;
70 struct vfio_container *container;
71 struct list_head device_list;
72 struct mutex device_lock;
74 struct notifier_block nb;
75 struct list_head vfio_next;
76 struct list_head container_next;
82 const struct vfio_device_ops *ops;
83 struct vfio_group *group;
84 struct list_head group_next;
89 * IOMMU driver registration
91 int vfio_register_iommu_driver(const struct vfio_iommu_driver_ops *ops)
93 struct vfio_iommu_driver *driver, *tmp;
95 driver = kzalloc(sizeof(*driver), GFP_KERNEL);
101 mutex_lock(&vfio.iommu_drivers_lock);
103 /* Check for duplicates */
104 list_for_each_entry(tmp, &vfio.iommu_drivers_list, vfio_next) {
105 if (tmp->ops == ops) {
106 mutex_unlock(&vfio.iommu_drivers_lock);
112 list_add(&driver->vfio_next, &vfio.iommu_drivers_list);
114 mutex_unlock(&vfio.iommu_drivers_lock);
118 EXPORT_SYMBOL_GPL(vfio_register_iommu_driver);
120 void vfio_unregister_iommu_driver(const struct vfio_iommu_driver_ops *ops)
122 struct vfio_iommu_driver *driver;
124 mutex_lock(&vfio.iommu_drivers_lock);
125 list_for_each_entry(driver, &vfio.iommu_drivers_list, vfio_next) {
126 if (driver->ops == ops) {
127 list_del(&driver->vfio_next);
128 mutex_unlock(&vfio.iommu_drivers_lock);
133 mutex_unlock(&vfio.iommu_drivers_lock);
135 EXPORT_SYMBOL_GPL(vfio_unregister_iommu_driver);
138 * Group minor allocation/free - both called with vfio.group_lock held
140 static int vfio_alloc_group_minor(struct vfio_group *group)
145 if (unlikely(idr_pre_get(&vfio.group_idr, GFP_KERNEL) == 0))
148 /* index 0 is used by /dev/vfio/vfio */
149 ret = idr_get_new_above(&vfio.group_idr, group, 1, &minor);
152 if (ret || minor > MINORMASK) {
153 if (minor > MINORMASK)
154 idr_remove(&vfio.group_idr, minor);
161 static void vfio_free_group_minor(int minor)
163 idr_remove(&vfio.group_idr, minor);
166 static int vfio_iommu_group_notifier(struct notifier_block *nb,
167 unsigned long action, void *data);
168 static void vfio_group_get(struct vfio_group *group);
171 * Container objects - containers are created when /dev/vfio/vfio is
172 * opened, but their lifecycle extends until the last user is done, so
173 * it's freed via kref. Must support container/group/device being
174 * closed in any order.
176 static void vfio_container_get(struct vfio_container *container)
178 kref_get(&container->kref);
181 static void vfio_container_release(struct kref *kref)
183 struct vfio_container *container;
184 container = container_of(kref, struct vfio_container, kref);
189 static void vfio_container_put(struct vfio_container *container)
191 kref_put(&container->kref, vfio_container_release);
195 * Group objects - create, release, get, put, search
197 static struct vfio_group *vfio_create_group(struct iommu_group *iommu_group)
199 struct vfio_group *group, *tmp;
203 group = kzalloc(sizeof(*group), GFP_KERNEL);
205 return ERR_PTR(-ENOMEM);
207 kref_init(&group->kref);
208 INIT_LIST_HEAD(&group->device_list);
209 mutex_init(&group->device_lock);
210 atomic_set(&group->container_users, 0);
211 group->iommu_group = iommu_group;
213 group->nb.notifier_call = vfio_iommu_group_notifier;
216 * blocking notifiers acquire a rwsem around registering and hold
217 * it around callback. Therefore, need to register outside of
218 * vfio.group_lock to avoid A-B/B-A contention. Our callback won't
219 * do anything unless it can find the group in vfio.group_list, so
220 * no harm in registering early.
222 ret = iommu_group_register_notifier(iommu_group, &group->nb);
228 mutex_lock(&vfio.group_lock);
230 minor = vfio_alloc_group_minor(group);
232 mutex_unlock(&vfio.group_lock);
234 return ERR_PTR(minor);
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_free_group_minor(minor);
242 mutex_unlock(&vfio.group_lock);
248 dev = device_create(vfio.class, NULL, MKDEV(MAJOR(vfio.devt), minor),
249 group, "%d", iommu_group_id(iommu_group));
251 vfio_free_group_minor(minor);
252 mutex_unlock(&vfio.group_lock);
254 return (struct vfio_group *)dev; /* ERR_PTR */
257 group->minor = minor;
260 list_add(&group->vfio_next, &vfio.group_list);
262 mutex_unlock(&vfio.group_lock);
267 static void vfio_group_release(struct kref *kref)
269 struct vfio_group *group = container_of(kref, struct vfio_group, kref);
271 WARN_ON(!list_empty(&group->device_list));
273 device_destroy(vfio.class, MKDEV(MAJOR(vfio.devt), group->minor));
274 list_del(&group->vfio_next);
275 vfio_free_group_minor(group->minor);
277 mutex_unlock(&vfio.group_lock);
280 * Unregister outside of lock. A spurious callback is harmless now
281 * that the group is no longer in vfio.group_list.
283 iommu_group_unregister_notifier(group->iommu_group, &group->nb);
288 static void vfio_group_put(struct vfio_group *group)
290 mutex_lock(&vfio.group_lock);
292 * Release needs to unlock to unregister the notifier, so only
293 * unlock if not released.
295 if (!kref_put(&group->kref, vfio_group_release))
296 mutex_unlock(&vfio.group_lock);
299 /* Assume group_lock or group reference is held */
300 static void vfio_group_get(struct vfio_group *group)
302 kref_get(&group->kref);
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.
309 static struct vfio_group *vfio_group_try_get(struct vfio_group *group)
311 struct vfio_group *target = group;
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);
321 mutex_unlock(&vfio.group_lock);
327 struct vfio_group *vfio_group_get_from_iommu(struct iommu_group *iommu_group)
329 struct vfio_group *group;
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);
339 mutex_unlock(&vfio.group_lock);
344 static struct vfio_group *vfio_group_get_from_minor(int minor)
346 struct vfio_group *group;
348 mutex_lock(&vfio.group_lock);
349 group = idr_find(&vfio.group_idr, minor);
351 mutex_unlock(&vfio.group_lock);
354 vfio_group_get(group);
355 mutex_unlock(&vfio.group_lock);
361 * Device objects - create, release, get, put, search
364 struct vfio_device *vfio_group_create_device(struct vfio_group *group,
366 const struct vfio_device_ops *ops,
369 struct vfio_device *device;
372 device = kzalloc(sizeof(*device), GFP_KERNEL);
374 return ERR_PTR(-ENOMEM);
376 kref_init(&device->kref);
378 device->group = group;
380 device->device_data = device_data;
382 ret = dev_set_drvdata(dev, device);
388 /* No need to get group_lock, caller has group reference */
389 vfio_group_get(group);
391 mutex_lock(&group->device_lock);
392 list_add(&device->group_next, &group->device_list);
393 mutex_unlock(&group->device_lock);
398 static void vfio_device_release(struct kref *kref)
400 struct vfio_device *device = container_of(kref,
401 struct vfio_device, kref);
402 struct vfio_group *group = device->group;
404 mutex_lock(&group->device_lock);
405 list_del(&device->group_next);
406 mutex_unlock(&group->device_lock);
408 dev_set_drvdata(device->dev, NULL);
412 /* vfio_del_group_dev may be waiting for this device */
413 wake_up(&vfio.release_q);
416 /* Device reference always implies a group reference */
417 static void vfio_device_put(struct vfio_device *device)
419 struct vfio_group *group = device->group;
420 kref_put(&device->kref, vfio_device_release);
421 vfio_group_put(group);
424 static void vfio_device_get(struct vfio_device *device)
426 vfio_group_get(device->group);
427 kref_get(&device->kref);
430 static struct vfio_device *vfio_group_get_device(struct vfio_group *group,
433 struct vfio_device *device;
435 mutex_lock(&group->device_lock);
436 list_for_each_entry(device, &group->device_list, group_next) {
437 if (device->dev == dev) {
438 vfio_device_get(device);
439 mutex_unlock(&group->device_lock);
443 mutex_unlock(&group->device_lock);
448 * Whitelist some drivers that we know are safe (no dma) or just sit on
449 * a device. It's not always practical to leave a device within a group
450 * driverless as it could get re-bound to something unsafe.
452 static const char * const vfio_driver_whitelist[] = { "pci-stub" };
454 static bool vfio_whitelisted_driver(struct device_driver *drv)
458 for (i = 0; i < ARRAY_SIZE(vfio_driver_whitelist); i++) {
459 if (!strcmp(drv->name, vfio_driver_whitelist[i]))
467 * A vfio group is viable for use by userspace if all devices are either
468 * driver-less or bound to a vfio or whitelisted driver. We test the
469 * latter by the existence of a struct vfio_device matching the dev.
471 static int vfio_dev_viable(struct device *dev, void *data)
473 struct vfio_group *group = data;
474 struct vfio_device *device;
476 if (!dev->driver || vfio_whitelisted_driver(dev->driver))
479 device = vfio_group_get_device(group, dev);
481 vfio_device_put(device);
489 * Async device support
491 static int vfio_group_nb_add_dev(struct vfio_group *group, struct device *dev)
493 struct vfio_device *device;
495 /* Do we already know about it? We shouldn't */
496 device = vfio_group_get_device(group, dev);
497 if (WARN_ON_ONCE(device)) {
498 vfio_device_put(device);
502 /* Nothing to do for idle groups */
503 if (!atomic_read(&group->container_users))
506 /* TODO Prevent device auto probing */
507 WARN("Device %s added to live group %d!\n", dev_name(dev),
508 iommu_group_id(group->iommu_group));
513 static int vfio_group_nb_del_dev(struct vfio_group *group, struct device *dev)
515 struct vfio_device *device;
518 * Expect to fall out here. If a device was in use, it would
519 * have been bound to a vfio sub-driver, which would have blocked
520 * in .remove at vfio_del_group_dev. Sanity check that we no
521 * longer track the device, so it's safe to remove.
523 device = vfio_group_get_device(group, dev);
527 WARN("Device %s removed from live group %d!\n", dev_name(dev),
528 iommu_group_id(group->iommu_group));
530 vfio_device_put(device);
534 static int vfio_group_nb_verify(struct vfio_group *group, struct device *dev)
536 /* We don't care what happens when the group isn't in use */
537 if (!atomic_read(&group->container_users))
540 return vfio_dev_viable(dev, group);
543 static int vfio_iommu_group_notifier(struct notifier_block *nb,
544 unsigned long action, void *data)
546 struct vfio_group *group = container_of(nb, struct vfio_group, nb);
547 struct device *dev = data;
550 * Need to go through a group_lock lookup to get a reference or
551 * we risk racing a group being removed. Leave a WARN_ON for
552 * debuging, but if the group no longer exists, a spurious notify
555 group = vfio_group_try_get(group);
560 case IOMMU_GROUP_NOTIFY_ADD_DEVICE:
561 vfio_group_nb_add_dev(group, dev);
563 case IOMMU_GROUP_NOTIFY_DEL_DEVICE:
564 vfio_group_nb_del_dev(group, dev);
566 case IOMMU_GROUP_NOTIFY_BIND_DRIVER:
567 pr_debug("%s: Device %s, group %d binding to driver\n",
568 __func__, dev_name(dev),
569 iommu_group_id(group->iommu_group));
571 case IOMMU_GROUP_NOTIFY_BOUND_DRIVER:
572 pr_debug("%s: Device %s, group %d bound to driver %s\n",
573 __func__, dev_name(dev),
574 iommu_group_id(group->iommu_group), dev->driver->name);
575 BUG_ON(vfio_group_nb_verify(group, dev));
577 case IOMMU_GROUP_NOTIFY_UNBIND_DRIVER:
578 pr_debug("%s: Device %s, group %d unbinding from driver %s\n",
579 __func__, dev_name(dev),
580 iommu_group_id(group->iommu_group), dev->driver->name);
582 case IOMMU_GROUP_NOTIFY_UNBOUND_DRIVER:
583 pr_debug("%s: Device %s, group %d unbound from driver\n",
584 __func__, dev_name(dev),
585 iommu_group_id(group->iommu_group));
587 * XXX An unbound device in a live group is ok, but we'd
588 * really like to avoid the above BUG_ON by preventing other
589 * drivers from binding to it. Once that occurs, we have to
590 * stop the system to maintain isolation. At a minimum, we'd
591 * want a toggle to disable driver auto probe for this device.
596 vfio_group_put(group);
603 int vfio_add_group_dev(struct device *dev,
604 const struct vfio_device_ops *ops, void *device_data)
606 struct iommu_group *iommu_group;
607 struct vfio_group *group;
608 struct vfio_device *device;
610 iommu_group = iommu_group_get(dev);
614 group = vfio_group_get_from_iommu(iommu_group);
616 group = vfio_create_group(iommu_group);
618 iommu_group_put(iommu_group);
619 return PTR_ERR(group);
623 device = vfio_group_get_device(group, dev);
625 WARN(1, "Device %s already exists on group %d\n",
626 dev_name(dev), iommu_group_id(iommu_group));
627 vfio_device_put(device);
628 vfio_group_put(group);
629 iommu_group_put(iommu_group);
633 device = vfio_group_create_device(group, dev, ops, device_data);
634 if (IS_ERR(device)) {
635 vfio_group_put(group);
636 iommu_group_put(iommu_group);
637 return PTR_ERR(device);
641 * Added device holds reference to iommu_group and vfio_device
642 * (which in turn holds reference to vfio_group). Drop extra
643 * group reference used while acquiring device.
645 vfio_group_put(group);
649 EXPORT_SYMBOL_GPL(vfio_add_group_dev);
651 /* Test whether a struct device is present in our tracking */
652 static bool vfio_dev_present(struct device *dev)
654 struct iommu_group *iommu_group;
655 struct vfio_group *group;
656 struct vfio_device *device;
658 iommu_group = iommu_group_get(dev);
662 group = vfio_group_get_from_iommu(iommu_group);
664 iommu_group_put(iommu_group);
668 device = vfio_group_get_device(group, dev);
670 vfio_group_put(group);
671 iommu_group_put(iommu_group);
675 vfio_device_put(device);
676 vfio_group_put(group);
677 iommu_group_put(iommu_group);
682 * Decrement the device reference count and wait for the device to be
683 * removed. Open file descriptors for the device... */
684 void *vfio_del_group_dev(struct device *dev)
686 struct vfio_device *device = dev_get_drvdata(dev);
687 struct vfio_group *group = device->group;
688 struct iommu_group *iommu_group = group->iommu_group;
689 void *device_data = device->device_data;
691 vfio_device_put(device);
693 /* TODO send a signal to encourage this to be released */
694 wait_event(vfio.release_q, !vfio_dev_present(dev));
696 iommu_group_put(iommu_group);
700 EXPORT_SYMBOL_GPL(vfio_del_group_dev);
703 * VFIO base fd, /dev/vfio/vfio
705 static long vfio_ioctl_check_extension(struct vfio_container *container,
708 struct vfio_iommu_driver *driver = container->iommu_driver;
712 /* No base extensions yet */
715 * If no driver is set, poll all registered drivers for
716 * extensions and return the first positive result. If
717 * a driver is already set, further queries will be passed
718 * only to that driver.
721 mutex_lock(&vfio.iommu_drivers_lock);
722 list_for_each_entry(driver, &vfio.iommu_drivers_list,
724 if (!try_module_get(driver->ops->owner))
727 ret = driver->ops->ioctl(NULL,
728 VFIO_CHECK_EXTENSION,
730 module_put(driver->ops->owner);
734 mutex_unlock(&vfio.iommu_drivers_lock);
736 ret = driver->ops->ioctl(container->iommu_data,
737 VFIO_CHECK_EXTENSION, arg);
743 /* hold container->group_lock */
744 static int __vfio_container_attach_groups(struct vfio_container *container,
745 struct vfio_iommu_driver *driver,
748 struct vfio_group *group;
751 list_for_each_entry(group, &container->group_list, container_next) {
752 ret = driver->ops->attach_group(data, group->iommu_group);
760 list_for_each_entry_continue_reverse(group, &container->group_list,
762 driver->ops->detach_group(data, group->iommu_group);
768 static long vfio_ioctl_set_iommu(struct vfio_container *container,
771 struct vfio_iommu_driver *driver;
774 mutex_lock(&container->group_lock);
777 * The container is designed to be an unprivileged interface while
778 * the group can be assigned to specific users. Therefore, only by
779 * adding a group to a container does the user get the privilege of
780 * enabling the iommu, which may allocate finite resources. There
781 * is no unset_iommu, but by removing all the groups from a container,
782 * the container is deprivileged and returns to an unset state.
784 if (list_empty(&container->group_list) || container->iommu_driver) {
785 mutex_unlock(&container->group_lock);
789 mutex_lock(&vfio.iommu_drivers_lock);
790 list_for_each_entry(driver, &vfio.iommu_drivers_list, vfio_next) {
793 if (!try_module_get(driver->ops->owner))
797 * The arg magic for SET_IOMMU is the same as CHECK_EXTENSION,
798 * so test which iommu driver reported support for this
799 * extension and call open on them. We also pass them the
800 * magic, allowing a single driver to support multiple
801 * interfaces if they'd like.
803 if (driver->ops->ioctl(NULL, VFIO_CHECK_EXTENSION, arg) <= 0) {
804 module_put(driver->ops->owner);
808 /* module reference holds the driver we're working on */
809 mutex_unlock(&vfio.iommu_drivers_lock);
811 data = driver->ops->open(arg);
814 module_put(driver->ops->owner);
815 goto skip_drivers_unlock;
818 ret = __vfio_container_attach_groups(container, driver, data);
820 container->iommu_driver = driver;
821 container->iommu_data = data;
823 driver->ops->release(data);
824 module_put(driver->ops->owner);
827 goto skip_drivers_unlock;
830 mutex_unlock(&vfio.iommu_drivers_lock);
832 mutex_unlock(&container->group_lock);
837 static long vfio_fops_unl_ioctl(struct file *filep,
838 unsigned int cmd, unsigned long arg)
840 struct vfio_container *container = filep->private_data;
841 struct vfio_iommu_driver *driver;
848 driver = container->iommu_driver;
849 data = container->iommu_data;
852 case VFIO_GET_API_VERSION:
853 ret = VFIO_API_VERSION;
855 case VFIO_CHECK_EXTENSION:
856 ret = vfio_ioctl_check_extension(container, arg);
859 ret = vfio_ioctl_set_iommu(container, arg);
862 if (driver) /* passthrough all unrecognized ioctls */
863 ret = driver->ops->ioctl(data, cmd, arg);
870 static long vfio_fops_compat_ioctl(struct file *filep,
871 unsigned int cmd, unsigned long arg)
873 arg = (unsigned long)compat_ptr(arg);
874 return vfio_fops_unl_ioctl(filep, cmd, arg);
876 #endif /* CONFIG_COMPAT */
878 static int vfio_fops_open(struct inode *inode, struct file *filep)
880 struct vfio_container *container;
882 container = kzalloc(sizeof(*container), GFP_KERNEL);
886 INIT_LIST_HEAD(&container->group_list);
887 mutex_init(&container->group_lock);
888 kref_init(&container->kref);
890 filep->private_data = container;
895 static int vfio_fops_release(struct inode *inode, struct file *filep)
897 struct vfio_container *container = filep->private_data;
899 filep->private_data = NULL;
901 vfio_container_put(container);
907 * Once an iommu driver is set, we optionally pass read/write/mmap
908 * on to the driver, allowing management interfaces beyond ioctl.
910 static ssize_t vfio_fops_read(struct file *filep, char __user *buf,
911 size_t count, loff_t *ppos)
913 struct vfio_container *container = filep->private_data;
914 struct vfio_iommu_driver *driver = container->iommu_driver;
916 if (unlikely(!driver || !driver->ops->read))
919 return driver->ops->read(container->iommu_data, buf, count, ppos);
922 static ssize_t vfio_fops_write(struct file *filep, const char __user *buf,
923 size_t count, loff_t *ppos)
925 struct vfio_container *container = filep->private_data;
926 struct vfio_iommu_driver *driver = container->iommu_driver;
928 if (unlikely(!driver || !driver->ops->write))
931 return driver->ops->write(container->iommu_data, buf, count, ppos);
934 static int vfio_fops_mmap(struct file *filep, struct vm_area_struct *vma)
936 struct vfio_container *container = filep->private_data;
937 struct vfio_iommu_driver *driver = container->iommu_driver;
939 if (unlikely(!driver || !driver->ops->mmap))
942 return driver->ops->mmap(container->iommu_data, vma);
945 static const struct file_operations vfio_fops = {
946 .owner = THIS_MODULE,
947 .open = vfio_fops_open,
948 .release = vfio_fops_release,
949 .read = vfio_fops_read,
950 .write = vfio_fops_write,
951 .unlocked_ioctl = vfio_fops_unl_ioctl,
953 .compat_ioctl = vfio_fops_compat_ioctl,
955 .mmap = vfio_fops_mmap,
959 * VFIO Group fd, /dev/vfio/$GROUP
961 static void __vfio_group_unset_container(struct vfio_group *group)
963 struct vfio_container *container = group->container;
964 struct vfio_iommu_driver *driver;
966 mutex_lock(&container->group_lock);
968 driver = container->iommu_driver;
970 driver->ops->detach_group(container->iommu_data,
973 group->container = NULL;
974 list_del(&group->container_next);
976 /* Detaching the last group deprivileges a container, remove iommu */
977 if (driver && list_empty(&container->group_list)) {
978 driver->ops->release(container->iommu_data);
979 module_put(driver->ops->owner);
980 container->iommu_driver = NULL;
981 container->iommu_data = NULL;
984 mutex_unlock(&container->group_lock);
986 vfio_container_put(container);
990 * VFIO_GROUP_UNSET_CONTAINER should fail if there are other users or
991 * if there was no container to unset. Since the ioctl is called on
992 * the group, we know that still exists, therefore the only valid
993 * transition here is 1->0.
995 static int vfio_group_unset_container(struct vfio_group *group)
997 int users = atomic_cmpxchg(&group->container_users, 1, 0);
1004 __vfio_group_unset_container(group);
1010 * When removing container users, anything that removes the last user
1011 * implicitly removes the group from the container. That is, if the
1012 * group file descriptor is closed, as well as any device file descriptors,
1013 * the group is free.
1015 static void vfio_group_try_dissolve_container(struct vfio_group *group)
1017 if (0 == atomic_dec_if_positive(&group->container_users))
1018 __vfio_group_unset_container(group);
1021 static int vfio_group_set_container(struct vfio_group *group, int container_fd)
1024 struct vfio_container *container;
1025 struct vfio_iommu_driver *driver;
1028 if (atomic_read(&group->container_users))
1031 filep = fget(container_fd);
1035 /* Sanity check, is this really our fd? */
1036 if (filep->f_op != &vfio_fops) {
1041 container = filep->private_data;
1042 WARN_ON(!container); /* fget ensures we don't race vfio_release */
1044 mutex_lock(&container->group_lock);
1046 driver = container->iommu_driver;
1048 ret = driver->ops->attach_group(container->iommu_data,
1049 group->iommu_group);
1054 group->container = container;
1055 list_add(&group->container_next, &container->group_list);
1057 /* Get a reference on the container and mark a user within the group */
1058 vfio_container_get(container);
1059 atomic_inc(&group->container_users);
1062 mutex_unlock(&container->group_lock);
1068 static bool vfio_group_viable(struct vfio_group *group)
1070 return (iommu_group_for_each_dev(group->iommu_group,
1071 group, vfio_dev_viable) == 0);
1074 static const struct file_operations vfio_device_fops;
1076 static int vfio_group_get_device_fd(struct vfio_group *group, char *buf)
1078 struct vfio_device *device;
1082 if (0 == atomic_read(&group->container_users) ||
1083 !group->container->iommu_driver || !vfio_group_viable(group))
1086 mutex_lock(&group->device_lock);
1087 list_for_each_entry(device, &group->device_list, group_next) {
1088 if (strcmp(dev_name(device->dev), buf))
1091 ret = device->ops->open(device->device_data);
1095 * We can't use anon_inode_getfd() because we need to modify
1096 * the f_mode flags directly to allow more than just ioctls
1098 ret = get_unused_fd();
1100 device->ops->release(device->device_data);
1104 filep = anon_inode_getfile("[vfio-device]", &vfio_device_fops,
1106 if (IS_ERR(filep)) {
1108 ret = PTR_ERR(filep);
1109 device->ops->release(device->device_data);
1114 * TODO: add an anon_inode interface to do this.
1115 * Appears to be missing by lack of need rather than
1116 * explicitly prevented. Now there's need.
1118 filep->f_mode |= (FMODE_LSEEK | FMODE_PREAD | FMODE_PWRITE);
1120 fd_install(ret, filep);
1122 vfio_device_get(device);
1123 atomic_inc(&group->container_users);
1126 mutex_unlock(&group->device_lock);
1131 static long vfio_group_fops_unl_ioctl(struct file *filep,
1132 unsigned int cmd, unsigned long arg)
1134 struct vfio_group *group = filep->private_data;
1138 case VFIO_GROUP_GET_STATUS:
1140 struct vfio_group_status status;
1141 unsigned long minsz;
1143 minsz = offsetofend(struct vfio_group_status, flags);
1145 if (copy_from_user(&status, (void __user *)arg, minsz))
1148 if (status.argsz < minsz)
1153 if (vfio_group_viable(group))
1154 status.flags |= VFIO_GROUP_FLAGS_VIABLE;
1156 if (group->container)
1157 status.flags |= VFIO_GROUP_FLAGS_CONTAINER_SET;
1159 if (copy_to_user((void __user *)arg, &status, minsz))
1165 case VFIO_GROUP_SET_CONTAINER:
1169 if (get_user(fd, (int __user *)arg))
1175 ret = vfio_group_set_container(group, fd);
1178 case VFIO_GROUP_UNSET_CONTAINER:
1179 ret = vfio_group_unset_container(group);
1181 case VFIO_GROUP_GET_DEVICE_FD:
1185 buf = strndup_user((const char __user *)arg, PAGE_SIZE);
1187 return PTR_ERR(buf);
1189 ret = vfio_group_get_device_fd(group, buf);
1198 #ifdef CONFIG_COMPAT
1199 static long vfio_group_fops_compat_ioctl(struct file *filep,
1200 unsigned int cmd, unsigned long arg)
1202 arg = (unsigned long)compat_ptr(arg);
1203 return vfio_group_fops_unl_ioctl(filep, cmd, arg);
1205 #endif /* CONFIG_COMPAT */
1207 static int vfio_group_fops_open(struct inode *inode, struct file *filep)
1209 struct vfio_group *group;
1211 group = vfio_group_get_from_minor(iminor(inode));
1215 if (group->container) {
1216 vfio_group_put(group);
1220 filep->private_data = group;
1225 static int vfio_group_fops_release(struct inode *inode, struct file *filep)
1227 struct vfio_group *group = filep->private_data;
1229 filep->private_data = NULL;
1231 vfio_group_try_dissolve_container(group);
1233 vfio_group_put(group);
1238 static const struct file_operations vfio_group_fops = {
1239 .owner = THIS_MODULE,
1240 .unlocked_ioctl = vfio_group_fops_unl_ioctl,
1241 #ifdef CONFIG_COMPAT
1242 .compat_ioctl = vfio_group_fops_compat_ioctl,
1244 .open = vfio_group_fops_open,
1245 .release = vfio_group_fops_release,
1251 static int vfio_device_fops_release(struct inode *inode, struct file *filep)
1253 struct vfio_device *device = filep->private_data;
1255 device->ops->release(device->device_data);
1257 vfio_group_try_dissolve_container(device->group);
1259 vfio_device_put(device);
1264 static long vfio_device_fops_unl_ioctl(struct file *filep,
1265 unsigned int cmd, unsigned long arg)
1267 struct vfio_device *device = filep->private_data;
1269 if (unlikely(!device->ops->ioctl))
1272 return device->ops->ioctl(device->device_data, cmd, arg);
1275 static ssize_t vfio_device_fops_read(struct file *filep, char __user *buf,
1276 size_t count, loff_t *ppos)
1278 struct vfio_device *device = filep->private_data;
1280 if (unlikely(!device->ops->read))
1283 return device->ops->read(device->device_data, buf, count, ppos);
1286 static ssize_t vfio_device_fops_write(struct file *filep,
1287 const char __user *buf,
1288 size_t count, loff_t *ppos)
1290 struct vfio_device *device = filep->private_data;
1292 if (unlikely(!device->ops->write))
1295 return device->ops->write(device->device_data, buf, count, ppos);
1298 static int vfio_device_fops_mmap(struct file *filep, struct vm_area_struct *vma)
1300 struct vfio_device *device = filep->private_data;
1302 if (unlikely(!device->ops->mmap))
1305 return device->ops->mmap(device->device_data, vma);
1308 #ifdef CONFIG_COMPAT
1309 static long vfio_device_fops_compat_ioctl(struct file *filep,
1310 unsigned int cmd, unsigned long arg)
1312 arg = (unsigned long)compat_ptr(arg);
1313 return vfio_device_fops_unl_ioctl(filep, cmd, arg);
1315 #endif /* CONFIG_COMPAT */
1317 static const struct file_operations vfio_device_fops = {
1318 .owner = THIS_MODULE,
1319 .release = vfio_device_fops_release,
1320 .read = vfio_device_fops_read,
1321 .write = vfio_device_fops_write,
1322 .unlocked_ioctl = vfio_device_fops_unl_ioctl,
1323 #ifdef CONFIG_COMPAT
1324 .compat_ioctl = vfio_device_fops_compat_ioctl,
1326 .mmap = vfio_device_fops_mmap,
1330 * Module/class support
1332 static char *vfio_devnode(struct device *dev, umode_t *mode)
1334 return kasprintf(GFP_KERNEL, "vfio/%s", dev_name(dev));
1337 static int __init vfio_init(void)
1341 idr_init(&vfio.group_idr);
1342 mutex_init(&vfio.group_lock);
1343 mutex_init(&vfio.iommu_drivers_lock);
1344 INIT_LIST_HEAD(&vfio.group_list);
1345 INIT_LIST_HEAD(&vfio.iommu_drivers_list);
1346 init_waitqueue_head(&vfio.release_q);
1348 vfio.class = class_create(THIS_MODULE, "vfio");
1349 if (IS_ERR(vfio.class)) {
1350 ret = PTR_ERR(vfio.class);
1354 vfio.class->devnode = vfio_devnode;
1356 ret = alloc_chrdev_region(&vfio.devt, 0, MINORMASK, "vfio");
1358 goto err_base_chrdev;
1360 cdev_init(&vfio.cdev, &vfio_fops);
1361 ret = cdev_add(&vfio.cdev, vfio.devt, 1);
1365 vfio.dev = device_create(vfio.class, NULL, vfio.devt, NULL, "vfio");
1366 if (IS_ERR(vfio.dev)) {
1367 ret = PTR_ERR(vfio.dev);
1371 /* /dev/vfio/$GROUP */
1372 cdev_init(&vfio.group_cdev, &vfio_group_fops);
1373 ret = cdev_add(&vfio.group_cdev,
1374 MKDEV(MAJOR(vfio.devt), 1), MINORMASK - 1);
1376 goto err_groups_cdev;
1378 pr_info(DRIVER_DESC " version: " DRIVER_VERSION "\n");
1381 * Attempt to load known iommu-drivers. This gives us a working
1382 * environment without the user needing to explicitly load iommu
1385 request_module_nowait("vfio_iommu_type1");
1390 device_destroy(vfio.class, vfio.devt);
1392 cdev_del(&vfio.cdev);
1394 unregister_chrdev_region(vfio.devt, MINORMASK);
1396 class_destroy(vfio.class);
1402 static void __exit vfio_cleanup(void)
1404 WARN_ON(!list_empty(&vfio.group_list));
1406 idr_destroy(&vfio.group_idr);
1407 cdev_del(&vfio.group_cdev);
1408 device_destroy(vfio.class, vfio.devt);
1409 cdev_del(&vfio.cdev);
1410 unregister_chrdev_region(vfio.devt, MINORMASK);
1411 class_destroy(vfio.class);
1415 module_init(vfio_init);
1416 module_exit(vfio_cleanup);
1418 MODULE_VERSION(DRIVER_VERSION);
1419 MODULE_LICENSE("GPL v2");
1420 MODULE_AUTHOR(DRIVER_AUTHOR);
1421 MODULE_DESCRIPTION(DRIVER_DESC);