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/rwsem.h>
28 #include <linux/sched.h>
29 #include <linux/slab.h>
30 #include <linux/stat.h>
31 #include <linux/string.h>
32 #include <linux/uaccess.h>
33 #include <linux/vfio.h>
34 #include <linux/wait.h>
36 #define DRIVER_VERSION "0.3"
37 #define DRIVER_AUTHOR "Alex Williamson <alex.williamson@redhat.com>"
38 #define DRIVER_DESC "VFIO - User Level meta-driver"
42 struct list_head iommu_drivers_list;
43 struct mutex iommu_drivers_lock;
44 struct list_head group_list;
46 struct mutex group_lock;
47 struct cdev group_cdev;
51 wait_queue_head_t release_q;
54 struct vfio_iommu_driver {
55 const struct vfio_iommu_driver_ops *ops;
56 struct list_head vfio_next;
59 struct vfio_container {
61 struct list_head group_list;
62 struct rw_semaphore group_lock;
63 struct vfio_iommu_driver *iommu_driver;
70 atomic_t container_users;
71 struct iommu_group *iommu_group;
72 struct vfio_container *container;
73 struct list_head device_list;
74 struct mutex device_lock;
76 struct notifier_block nb;
77 struct list_head vfio_next;
78 struct list_head container_next;
85 const struct vfio_device_ops *ops;
86 struct vfio_group *group;
87 struct list_head group_next;
92 * IOMMU driver registration
94 int vfio_register_iommu_driver(const struct vfio_iommu_driver_ops *ops)
96 struct vfio_iommu_driver *driver, *tmp;
98 driver = kzalloc(sizeof(*driver), GFP_KERNEL);
104 mutex_lock(&vfio.iommu_drivers_lock);
106 /* Check for duplicates */
107 list_for_each_entry(tmp, &vfio.iommu_drivers_list, vfio_next) {
108 if (tmp->ops == ops) {
109 mutex_unlock(&vfio.iommu_drivers_lock);
115 list_add(&driver->vfio_next, &vfio.iommu_drivers_list);
117 mutex_unlock(&vfio.iommu_drivers_lock);
121 EXPORT_SYMBOL_GPL(vfio_register_iommu_driver);
123 void vfio_unregister_iommu_driver(const struct vfio_iommu_driver_ops *ops)
125 struct vfio_iommu_driver *driver;
127 mutex_lock(&vfio.iommu_drivers_lock);
128 list_for_each_entry(driver, &vfio.iommu_drivers_list, vfio_next) {
129 if (driver->ops == ops) {
130 list_del(&driver->vfio_next);
131 mutex_unlock(&vfio.iommu_drivers_lock);
136 mutex_unlock(&vfio.iommu_drivers_lock);
138 EXPORT_SYMBOL_GPL(vfio_unregister_iommu_driver);
141 * Group minor allocation/free - both called with vfio.group_lock held
143 static int vfio_alloc_group_minor(struct vfio_group *group)
145 /* index 0 is used by /dev/vfio/vfio */
146 return idr_alloc(&vfio.group_idr, group, 1, MINORMASK + 1, GFP_KERNEL);
149 static void vfio_free_group_minor(int minor)
151 idr_remove(&vfio.group_idr, minor);
154 static int vfio_iommu_group_notifier(struct notifier_block *nb,
155 unsigned long action, void *data);
156 static void vfio_group_get(struct vfio_group *group);
159 * Container objects - containers are created when /dev/vfio/vfio is
160 * opened, but their lifecycle extends until the last user is done, so
161 * it's freed via kref. Must support container/group/device being
162 * closed in any order.
164 static void vfio_container_get(struct vfio_container *container)
166 kref_get(&container->kref);
169 static void vfio_container_release(struct kref *kref)
171 struct vfio_container *container;
172 container = container_of(kref, struct vfio_container, kref);
177 static void vfio_container_put(struct vfio_container *container)
179 kref_put(&container->kref, vfio_container_release);
182 static void vfio_group_unlock_and_free(struct vfio_group *group)
184 mutex_unlock(&vfio.group_lock);
186 * Unregister outside of lock. A spurious callback is harmless now
187 * that the group is no longer in vfio.group_list.
189 iommu_group_unregister_notifier(group->iommu_group, &group->nb);
194 * Group objects - create, release, get, put, search
196 static struct vfio_group *vfio_create_group(struct iommu_group *iommu_group)
198 struct vfio_group *group, *tmp;
202 group = kzalloc(sizeof(*group), GFP_KERNEL);
204 return ERR_PTR(-ENOMEM);
206 kref_init(&group->kref);
207 INIT_LIST_HEAD(&group->device_list);
208 mutex_init(&group->device_lock);
209 atomic_set(&group->container_users, 0);
210 atomic_set(&group->opened, 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 vfio_group_unlock_and_free(group);
233 return ERR_PTR(minor);
236 /* Did we race creating this group? */
237 list_for_each_entry(tmp, &vfio.group_list, vfio_next) {
238 if (tmp->iommu_group == iommu_group) {
240 vfio_free_group_minor(minor);
241 vfio_group_unlock_and_free(group);
246 dev = device_create(vfio.class, NULL, MKDEV(MAJOR(vfio.devt), minor),
247 group, "%d", iommu_group_id(iommu_group));
249 vfio_free_group_minor(minor);
250 vfio_group_unlock_and_free(group);
251 return (struct vfio_group *)dev; /* ERR_PTR */
254 group->minor = minor;
257 list_add(&group->vfio_next, &vfio.group_list);
259 mutex_unlock(&vfio.group_lock);
264 /* called with vfio.group_lock held */
265 static void vfio_group_release(struct kref *kref)
267 struct vfio_group *group = container_of(kref, struct vfio_group, kref);
269 WARN_ON(!list_empty(&group->device_list));
271 device_destroy(vfio.class, MKDEV(MAJOR(vfio.devt), group->minor));
272 list_del(&group->vfio_next);
273 vfio_free_group_minor(group->minor);
274 vfio_group_unlock_and_free(group);
277 static void vfio_group_put(struct vfio_group *group)
279 kref_put_mutex(&group->kref, vfio_group_release, &vfio.group_lock);
282 /* Assume group_lock or group reference is held */
283 static void vfio_group_get(struct vfio_group *group)
285 kref_get(&group->kref);
289 * Not really a try as we will sleep for mutex, but we need to make
290 * sure the group pointer is valid under lock and get a reference.
292 static struct vfio_group *vfio_group_try_get(struct vfio_group *group)
294 struct vfio_group *target = group;
296 mutex_lock(&vfio.group_lock);
297 list_for_each_entry(group, &vfio.group_list, vfio_next) {
298 if (group == target) {
299 vfio_group_get(group);
300 mutex_unlock(&vfio.group_lock);
304 mutex_unlock(&vfio.group_lock);
310 struct vfio_group *vfio_group_get_from_iommu(struct iommu_group *iommu_group)
312 struct vfio_group *group;
314 mutex_lock(&vfio.group_lock);
315 list_for_each_entry(group, &vfio.group_list, vfio_next) {
316 if (group->iommu_group == iommu_group) {
317 vfio_group_get(group);
318 mutex_unlock(&vfio.group_lock);
322 mutex_unlock(&vfio.group_lock);
327 static struct vfio_group *vfio_group_get_from_minor(int minor)
329 struct vfio_group *group;
331 mutex_lock(&vfio.group_lock);
332 group = idr_find(&vfio.group_idr, minor);
334 mutex_unlock(&vfio.group_lock);
337 vfio_group_get(group);
338 mutex_unlock(&vfio.group_lock);
344 * Device objects - create, release, get, put, search
347 struct vfio_device *vfio_group_create_device(struct vfio_group *group,
349 const struct vfio_device_ops *ops,
352 struct vfio_device *device;
355 device = kzalloc(sizeof(*device), GFP_KERNEL);
357 return ERR_PTR(-ENOMEM);
359 kref_init(&device->kref);
361 device->group = group;
363 device->device_data = device_data;
365 ret = dev_set_drvdata(dev, device);
371 /* No need to get group_lock, caller has group reference */
372 vfio_group_get(group);
374 mutex_lock(&group->device_lock);
375 list_add(&device->group_next, &group->device_list);
376 mutex_unlock(&group->device_lock);
381 static void vfio_device_release(struct kref *kref)
383 struct vfio_device *device = container_of(kref,
384 struct vfio_device, kref);
385 struct vfio_group *group = device->group;
387 list_del(&device->group_next);
388 mutex_unlock(&group->device_lock);
390 dev_set_drvdata(device->dev, NULL);
394 /* vfio_del_group_dev may be waiting for this device */
395 wake_up(&vfio.release_q);
398 /* Device reference always implies a group reference */
399 void vfio_device_put(struct vfio_device *device)
401 struct vfio_group *group = device->group;
402 kref_put_mutex(&device->kref, vfio_device_release, &group->device_lock);
403 vfio_group_put(group);
405 EXPORT_SYMBOL_GPL(vfio_device_put);
407 static void vfio_device_get(struct vfio_device *device)
409 vfio_group_get(device->group);
410 kref_get(&device->kref);
413 static struct vfio_device *vfio_group_get_device(struct vfio_group *group,
416 struct vfio_device *device;
418 mutex_lock(&group->device_lock);
419 list_for_each_entry(device, &group->device_list, group_next) {
420 if (device->dev == dev) {
421 vfio_device_get(device);
422 mutex_unlock(&group->device_lock);
426 mutex_unlock(&group->device_lock);
431 * Whitelist some drivers that we know are safe (no dma) or just sit on
432 * a device. It's not always practical to leave a device within a group
433 * driverless as it could get re-bound to something unsafe.
435 static const char * const vfio_driver_whitelist[] = { "pci-stub", "pcieport" };
437 static bool vfio_whitelisted_driver(struct device_driver *drv)
441 for (i = 0; i < ARRAY_SIZE(vfio_driver_whitelist); i++) {
442 if (!strcmp(drv->name, vfio_driver_whitelist[i]))
450 * A vfio group is viable for use by userspace if all devices are either
451 * driver-less or bound to a vfio or whitelisted driver. We test the
452 * latter by the existence of a struct vfio_device matching the dev.
454 static int vfio_dev_viable(struct device *dev, void *data)
456 struct vfio_group *group = data;
457 struct vfio_device *device;
458 struct device_driver *drv = ACCESS_ONCE(dev->driver);
460 if (!drv || vfio_whitelisted_driver(drv))
463 device = vfio_group_get_device(group, dev);
465 vfio_device_put(device);
473 * Async device support
475 static int vfio_group_nb_add_dev(struct vfio_group *group, struct device *dev)
477 struct vfio_device *device;
479 /* Do we already know about it? We shouldn't */
480 device = vfio_group_get_device(group, dev);
481 if (WARN_ON_ONCE(device)) {
482 vfio_device_put(device);
486 /* Nothing to do for idle groups */
487 if (!atomic_read(&group->container_users))
490 /* TODO Prevent device auto probing */
491 WARN("Device %s added to live group %d!\n", dev_name(dev),
492 iommu_group_id(group->iommu_group));
497 static int vfio_group_nb_del_dev(struct vfio_group *group, struct device *dev)
499 struct vfio_device *device;
502 * Expect to fall out here. If a device was in use, it would
503 * have been bound to a vfio sub-driver, which would have blocked
504 * in .remove at vfio_del_group_dev. Sanity check that we no
505 * longer track the device, so it's safe to remove.
507 device = vfio_group_get_device(group, dev);
511 WARN("Device %s removed from live group %d!\n", dev_name(dev),
512 iommu_group_id(group->iommu_group));
514 vfio_device_put(device);
518 static int vfio_group_nb_verify(struct vfio_group *group, struct device *dev)
520 /* We don't care what happens when the group isn't in use */
521 if (!atomic_read(&group->container_users))
524 return vfio_dev_viable(dev, group);
527 static int vfio_iommu_group_notifier(struct notifier_block *nb,
528 unsigned long action, void *data)
530 struct vfio_group *group = container_of(nb, struct vfio_group, nb);
531 struct device *dev = data;
534 * Need to go through a group_lock lookup to get a reference or
535 * we risk racing a group being removed. Leave a WARN_ON for
536 * debuging, but if the group no longer exists, a spurious notify
539 group = vfio_group_try_get(group);
544 case IOMMU_GROUP_NOTIFY_ADD_DEVICE:
545 vfio_group_nb_add_dev(group, dev);
547 case IOMMU_GROUP_NOTIFY_DEL_DEVICE:
548 vfio_group_nb_del_dev(group, dev);
550 case IOMMU_GROUP_NOTIFY_BIND_DRIVER:
551 pr_debug("%s: Device %s, group %d binding to driver\n",
552 __func__, dev_name(dev),
553 iommu_group_id(group->iommu_group));
555 case IOMMU_GROUP_NOTIFY_BOUND_DRIVER:
556 pr_debug("%s: Device %s, group %d bound to driver %s\n",
557 __func__, dev_name(dev),
558 iommu_group_id(group->iommu_group), dev->driver->name);
559 BUG_ON(vfio_group_nb_verify(group, dev));
561 case IOMMU_GROUP_NOTIFY_UNBIND_DRIVER:
562 pr_debug("%s: Device %s, group %d unbinding from driver %s\n",
563 __func__, dev_name(dev),
564 iommu_group_id(group->iommu_group), dev->driver->name);
566 case IOMMU_GROUP_NOTIFY_UNBOUND_DRIVER:
567 pr_debug("%s: Device %s, group %d unbound from driver\n",
568 __func__, dev_name(dev),
569 iommu_group_id(group->iommu_group));
571 * XXX An unbound device in a live group is ok, but we'd
572 * really like to avoid the above BUG_ON by preventing other
573 * drivers from binding to it. Once that occurs, we have to
574 * stop the system to maintain isolation. At a minimum, we'd
575 * want a toggle to disable driver auto probe for this device.
580 vfio_group_put(group);
587 int vfio_add_group_dev(struct device *dev,
588 const struct vfio_device_ops *ops, void *device_data)
590 struct iommu_group *iommu_group;
591 struct vfio_group *group;
592 struct vfio_device *device;
594 iommu_group = iommu_group_get(dev);
598 group = vfio_group_get_from_iommu(iommu_group);
600 group = vfio_create_group(iommu_group);
602 iommu_group_put(iommu_group);
603 return PTR_ERR(group);
607 device = vfio_group_get_device(group, dev);
609 WARN(1, "Device %s already exists on group %d\n",
610 dev_name(dev), iommu_group_id(iommu_group));
611 vfio_device_put(device);
612 vfio_group_put(group);
613 iommu_group_put(iommu_group);
617 device = vfio_group_create_device(group, dev, ops, device_data);
618 if (IS_ERR(device)) {
619 vfio_group_put(group);
620 iommu_group_put(iommu_group);
621 return PTR_ERR(device);
625 * Added device holds reference to iommu_group and vfio_device
626 * (which in turn holds reference to vfio_group). Drop extra
627 * group reference used while acquiring device.
629 vfio_group_put(group);
633 EXPORT_SYMBOL_GPL(vfio_add_group_dev);
636 * Get a reference to the vfio_device for a device that is known to
637 * be bound to a vfio driver. The driver implicitly holds a
638 * vfio_device reference between vfio_add_group_dev and
639 * vfio_del_group_dev. We can therefore use drvdata to increment
640 * that reference from the struct device. This additional
641 * reference must be released by calling vfio_device_put.
643 struct vfio_device *vfio_device_get_from_dev(struct device *dev)
645 struct vfio_device *device = dev_get_drvdata(dev);
647 vfio_device_get(device);
651 EXPORT_SYMBOL_GPL(vfio_device_get_from_dev);
654 * Caller must hold a reference to the vfio_device
656 void *vfio_device_data(struct vfio_device *device)
658 return device->device_data;
660 EXPORT_SYMBOL_GPL(vfio_device_data);
662 /* Given a referenced group, check if it contains the device */
663 static bool vfio_dev_present(struct vfio_group *group, struct device *dev)
665 struct vfio_device *device;
667 device = vfio_group_get_device(group, dev);
671 vfio_device_put(device);
676 * Decrement the device reference count and wait for the device to be
677 * removed. Open file descriptors for the device... */
678 void *vfio_del_group_dev(struct device *dev)
680 struct vfio_device *device = dev_get_drvdata(dev);
681 struct vfio_group *group = device->group;
682 struct iommu_group *iommu_group = group->iommu_group;
683 void *device_data = device->device_data;
686 * The group exists so long as we have a device reference. Get
687 * a group reference and use it to scan for the device going away.
689 vfio_group_get(group);
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(group, dev));
696 vfio_group_put(group);
698 iommu_group_put(iommu_group);
702 EXPORT_SYMBOL_GPL(vfio_del_group_dev);
705 * VFIO base fd, /dev/vfio/vfio
707 static long vfio_ioctl_check_extension(struct vfio_container *container,
710 struct vfio_iommu_driver *driver;
713 down_read(&container->group_lock);
715 driver = container->iommu_driver;
718 /* No base extensions yet */
721 * If no driver is set, poll all registered drivers for
722 * extensions and return the first positive result. If
723 * a driver is already set, further queries will be passed
724 * only to that driver.
727 mutex_lock(&vfio.iommu_drivers_lock);
728 list_for_each_entry(driver, &vfio.iommu_drivers_list,
730 if (!try_module_get(driver->ops->owner))
733 ret = driver->ops->ioctl(NULL,
734 VFIO_CHECK_EXTENSION,
736 module_put(driver->ops->owner);
740 mutex_unlock(&vfio.iommu_drivers_lock);
742 ret = driver->ops->ioctl(container->iommu_data,
743 VFIO_CHECK_EXTENSION, arg);
746 up_read(&container->group_lock);
751 /* hold write lock on container->group_lock */
752 static int __vfio_container_attach_groups(struct vfio_container *container,
753 struct vfio_iommu_driver *driver,
756 struct vfio_group *group;
759 list_for_each_entry(group, &container->group_list, container_next) {
760 ret = driver->ops->attach_group(data, group->iommu_group);
768 list_for_each_entry_continue_reverse(group, &container->group_list,
770 driver->ops->detach_group(data, group->iommu_group);
776 static long vfio_ioctl_set_iommu(struct vfio_container *container,
779 struct vfio_iommu_driver *driver;
782 down_write(&container->group_lock);
785 * The container is designed to be an unprivileged interface while
786 * the group can be assigned to specific users. Therefore, only by
787 * adding a group to a container does the user get the privilege of
788 * enabling the iommu, which may allocate finite resources. There
789 * is no unset_iommu, but by removing all the groups from a container,
790 * the container is deprivileged and returns to an unset state.
792 if (list_empty(&container->group_list) || container->iommu_driver) {
793 up_write(&container->group_lock);
797 mutex_lock(&vfio.iommu_drivers_lock);
798 list_for_each_entry(driver, &vfio.iommu_drivers_list, vfio_next) {
801 if (!try_module_get(driver->ops->owner))
805 * The arg magic for SET_IOMMU is the same as CHECK_EXTENSION,
806 * so test which iommu driver reported support for this
807 * extension and call open on them. We also pass them the
808 * magic, allowing a single driver to support multiple
809 * interfaces if they'd like.
811 if (driver->ops->ioctl(NULL, VFIO_CHECK_EXTENSION, arg) <= 0) {
812 module_put(driver->ops->owner);
816 /* module reference holds the driver we're working on */
817 mutex_unlock(&vfio.iommu_drivers_lock);
819 data = driver->ops->open(arg);
822 module_put(driver->ops->owner);
823 goto skip_drivers_unlock;
826 ret = __vfio_container_attach_groups(container, driver, data);
828 container->iommu_driver = driver;
829 container->iommu_data = data;
831 driver->ops->release(data);
832 module_put(driver->ops->owner);
835 goto skip_drivers_unlock;
838 mutex_unlock(&vfio.iommu_drivers_lock);
840 up_write(&container->group_lock);
845 static long vfio_fops_unl_ioctl(struct file *filep,
846 unsigned int cmd, unsigned long arg)
848 struct vfio_container *container = filep->private_data;
849 struct vfio_iommu_driver *driver;
857 case VFIO_GET_API_VERSION:
858 ret = VFIO_API_VERSION;
860 case VFIO_CHECK_EXTENSION:
861 ret = vfio_ioctl_check_extension(container, arg);
864 ret = vfio_ioctl_set_iommu(container, arg);
867 down_read(&container->group_lock);
869 driver = container->iommu_driver;
870 data = container->iommu_data;
872 if (driver) /* passthrough all unrecognized ioctls */
873 ret = driver->ops->ioctl(data, cmd, arg);
875 up_read(&container->group_lock);
882 static long vfio_fops_compat_ioctl(struct file *filep,
883 unsigned int cmd, unsigned long arg)
885 arg = (unsigned long)compat_ptr(arg);
886 return vfio_fops_unl_ioctl(filep, cmd, arg);
888 #endif /* CONFIG_COMPAT */
890 static int vfio_fops_open(struct inode *inode, struct file *filep)
892 struct vfio_container *container;
894 container = kzalloc(sizeof(*container), GFP_KERNEL);
898 INIT_LIST_HEAD(&container->group_list);
899 init_rwsem(&container->group_lock);
900 kref_init(&container->kref);
902 filep->private_data = container;
907 static int vfio_fops_release(struct inode *inode, struct file *filep)
909 struct vfio_container *container = filep->private_data;
911 filep->private_data = NULL;
913 vfio_container_put(container);
919 * Once an iommu driver is set, we optionally pass read/write/mmap
920 * on to the driver, allowing management interfaces beyond ioctl.
922 static ssize_t vfio_fops_read(struct file *filep, char __user *buf,
923 size_t count, loff_t *ppos)
925 struct vfio_container *container = filep->private_data;
926 struct vfio_iommu_driver *driver;
927 ssize_t ret = -EINVAL;
929 down_read(&container->group_lock);
931 driver = container->iommu_driver;
932 if (likely(driver && driver->ops->read))
933 ret = driver->ops->read(container->iommu_data,
936 up_read(&container->group_lock);
941 static ssize_t vfio_fops_write(struct file *filep, const char __user *buf,
942 size_t count, loff_t *ppos)
944 struct vfio_container *container = filep->private_data;
945 struct vfio_iommu_driver *driver;
946 ssize_t ret = -EINVAL;
948 down_read(&container->group_lock);
950 driver = container->iommu_driver;
951 if (likely(driver && driver->ops->write))
952 ret = driver->ops->write(container->iommu_data,
955 up_read(&container->group_lock);
960 static int vfio_fops_mmap(struct file *filep, struct vm_area_struct *vma)
962 struct vfio_container *container = filep->private_data;
963 struct vfio_iommu_driver *driver;
966 down_read(&container->group_lock);
968 driver = container->iommu_driver;
969 if (likely(driver && driver->ops->mmap))
970 ret = driver->ops->mmap(container->iommu_data, vma);
972 up_read(&container->group_lock);
977 static const struct file_operations vfio_fops = {
978 .owner = THIS_MODULE,
979 .open = vfio_fops_open,
980 .release = vfio_fops_release,
981 .read = vfio_fops_read,
982 .write = vfio_fops_write,
983 .unlocked_ioctl = vfio_fops_unl_ioctl,
985 .compat_ioctl = vfio_fops_compat_ioctl,
987 .mmap = vfio_fops_mmap,
991 * VFIO Group fd, /dev/vfio/$GROUP
993 static void __vfio_group_unset_container(struct vfio_group *group)
995 struct vfio_container *container = group->container;
996 struct vfio_iommu_driver *driver;
998 down_write(&container->group_lock);
1000 driver = container->iommu_driver;
1002 driver->ops->detach_group(container->iommu_data,
1003 group->iommu_group);
1005 group->container = NULL;
1006 list_del(&group->container_next);
1008 /* Detaching the last group deprivileges a container, remove iommu */
1009 if (driver && list_empty(&container->group_list)) {
1010 driver->ops->release(container->iommu_data);
1011 module_put(driver->ops->owner);
1012 container->iommu_driver = NULL;
1013 container->iommu_data = NULL;
1016 up_write(&container->group_lock);
1018 vfio_container_put(container);
1022 * VFIO_GROUP_UNSET_CONTAINER should fail if there are other users or
1023 * if there was no container to unset. Since the ioctl is called on
1024 * the group, we know that still exists, therefore the only valid
1025 * transition here is 1->0.
1027 static int vfio_group_unset_container(struct vfio_group *group)
1029 int users = atomic_cmpxchg(&group->container_users, 1, 0);
1036 __vfio_group_unset_container(group);
1042 * When removing container users, anything that removes the last user
1043 * implicitly removes the group from the container. That is, if the
1044 * group file descriptor is closed, as well as any device file descriptors,
1045 * the group is free.
1047 static void vfio_group_try_dissolve_container(struct vfio_group *group)
1049 if (0 == atomic_dec_if_positive(&group->container_users))
1050 __vfio_group_unset_container(group);
1053 static int vfio_group_set_container(struct vfio_group *group, int container_fd)
1056 struct vfio_container *container;
1057 struct vfio_iommu_driver *driver;
1060 if (atomic_read(&group->container_users))
1063 f = fdget(container_fd);
1067 /* Sanity check, is this really our fd? */
1068 if (f.file->f_op != &vfio_fops) {
1073 container = f.file->private_data;
1074 WARN_ON(!container); /* fget ensures we don't race vfio_release */
1076 down_write(&container->group_lock);
1078 driver = container->iommu_driver;
1080 ret = driver->ops->attach_group(container->iommu_data,
1081 group->iommu_group);
1086 group->container = container;
1087 list_add(&group->container_next, &container->group_list);
1089 /* Get a reference on the container and mark a user within the group */
1090 vfio_container_get(container);
1091 atomic_inc(&group->container_users);
1094 up_write(&container->group_lock);
1099 static bool vfio_group_viable(struct vfio_group *group)
1101 return (iommu_group_for_each_dev(group->iommu_group,
1102 group, vfio_dev_viable) == 0);
1105 static const struct file_operations vfio_device_fops;
1107 static int vfio_group_get_device_fd(struct vfio_group *group, char *buf)
1109 struct vfio_device *device;
1113 if (0 == atomic_read(&group->container_users) ||
1114 !group->container->iommu_driver || !vfio_group_viable(group))
1117 mutex_lock(&group->device_lock);
1118 list_for_each_entry(device, &group->device_list, group_next) {
1119 if (strcmp(dev_name(device->dev), buf))
1122 ret = device->ops->open(device->device_data);
1126 * We can't use anon_inode_getfd() because we need to modify
1127 * the f_mode flags directly to allow more than just ioctls
1129 ret = get_unused_fd();
1131 device->ops->release(device->device_data);
1135 filep = anon_inode_getfile("[vfio-device]", &vfio_device_fops,
1137 if (IS_ERR(filep)) {
1139 ret = PTR_ERR(filep);
1140 device->ops->release(device->device_data);
1145 * TODO: add an anon_inode interface to do this.
1146 * Appears to be missing by lack of need rather than
1147 * explicitly prevented. Now there's need.
1149 filep->f_mode |= (FMODE_LSEEK | FMODE_PREAD | FMODE_PWRITE);
1151 vfio_device_get(device);
1152 atomic_inc(&group->container_users);
1154 fd_install(ret, filep);
1157 mutex_unlock(&group->device_lock);
1162 static long vfio_group_fops_unl_ioctl(struct file *filep,
1163 unsigned int cmd, unsigned long arg)
1165 struct vfio_group *group = filep->private_data;
1169 case VFIO_GROUP_GET_STATUS:
1171 struct vfio_group_status status;
1172 unsigned long minsz;
1174 minsz = offsetofend(struct vfio_group_status, flags);
1176 if (copy_from_user(&status, (void __user *)arg, minsz))
1179 if (status.argsz < minsz)
1184 if (vfio_group_viable(group))
1185 status.flags |= VFIO_GROUP_FLAGS_VIABLE;
1187 if (group->container)
1188 status.flags |= VFIO_GROUP_FLAGS_CONTAINER_SET;
1190 if (copy_to_user((void __user *)arg, &status, minsz))
1196 case VFIO_GROUP_SET_CONTAINER:
1200 if (get_user(fd, (int __user *)arg))
1206 ret = vfio_group_set_container(group, fd);
1209 case VFIO_GROUP_UNSET_CONTAINER:
1210 ret = vfio_group_unset_container(group);
1212 case VFIO_GROUP_GET_DEVICE_FD:
1216 buf = strndup_user((const char __user *)arg, PAGE_SIZE);
1218 return PTR_ERR(buf);
1220 ret = vfio_group_get_device_fd(group, buf);
1229 #ifdef CONFIG_COMPAT
1230 static long vfio_group_fops_compat_ioctl(struct file *filep,
1231 unsigned int cmd, unsigned long arg)
1233 arg = (unsigned long)compat_ptr(arg);
1234 return vfio_group_fops_unl_ioctl(filep, cmd, arg);
1236 #endif /* CONFIG_COMPAT */
1238 static int vfio_group_fops_open(struct inode *inode, struct file *filep)
1240 struct vfio_group *group;
1243 group = vfio_group_get_from_minor(iminor(inode));
1247 /* Do we need multiple instances of the group open? Seems not. */
1248 opened = atomic_cmpxchg(&group->opened, 0, 1);
1250 vfio_group_put(group);
1254 /* Is something still in use from a previous open? */
1255 if (group->container) {
1256 atomic_dec(&group->opened);
1257 vfio_group_put(group);
1261 filep->private_data = group;
1266 static int vfio_group_fops_release(struct inode *inode, struct file *filep)
1268 struct vfio_group *group = filep->private_data;
1270 filep->private_data = NULL;
1272 vfio_group_try_dissolve_container(group);
1274 atomic_dec(&group->opened);
1276 vfio_group_put(group);
1281 static const struct file_operations vfio_group_fops = {
1282 .owner = THIS_MODULE,
1283 .unlocked_ioctl = vfio_group_fops_unl_ioctl,
1284 #ifdef CONFIG_COMPAT
1285 .compat_ioctl = vfio_group_fops_compat_ioctl,
1287 .open = vfio_group_fops_open,
1288 .release = vfio_group_fops_release,
1294 static int vfio_device_fops_release(struct inode *inode, struct file *filep)
1296 struct vfio_device *device = filep->private_data;
1298 device->ops->release(device->device_data);
1300 vfio_group_try_dissolve_container(device->group);
1302 vfio_device_put(device);
1307 static long vfio_device_fops_unl_ioctl(struct file *filep,
1308 unsigned int cmd, unsigned long arg)
1310 struct vfio_device *device = filep->private_data;
1312 if (unlikely(!device->ops->ioctl))
1315 return device->ops->ioctl(device->device_data, cmd, arg);
1318 static ssize_t vfio_device_fops_read(struct file *filep, char __user *buf,
1319 size_t count, loff_t *ppos)
1321 struct vfio_device *device = filep->private_data;
1323 if (unlikely(!device->ops->read))
1326 return device->ops->read(device->device_data, buf, count, ppos);
1329 static ssize_t vfio_device_fops_write(struct file *filep,
1330 const char __user *buf,
1331 size_t count, loff_t *ppos)
1333 struct vfio_device *device = filep->private_data;
1335 if (unlikely(!device->ops->write))
1338 return device->ops->write(device->device_data, buf, count, ppos);
1341 static int vfio_device_fops_mmap(struct file *filep, struct vm_area_struct *vma)
1343 struct vfio_device *device = filep->private_data;
1345 if (unlikely(!device->ops->mmap))
1348 return device->ops->mmap(device->device_data, vma);
1351 #ifdef CONFIG_COMPAT
1352 static long vfio_device_fops_compat_ioctl(struct file *filep,
1353 unsigned int cmd, unsigned long arg)
1355 arg = (unsigned long)compat_ptr(arg);
1356 return vfio_device_fops_unl_ioctl(filep, cmd, arg);
1358 #endif /* CONFIG_COMPAT */
1360 static const struct file_operations vfio_device_fops = {
1361 .owner = THIS_MODULE,
1362 .release = vfio_device_fops_release,
1363 .read = vfio_device_fops_read,
1364 .write = vfio_device_fops_write,
1365 .unlocked_ioctl = vfio_device_fops_unl_ioctl,
1366 #ifdef CONFIG_COMPAT
1367 .compat_ioctl = vfio_device_fops_compat_ioctl,
1369 .mmap = vfio_device_fops_mmap,
1373 * Module/class support
1375 static char *vfio_devnode(struct device *dev, umode_t *mode)
1377 if (mode && (MINOR(dev->devt) == 0))
1378 *mode = S_IRUGO | S_IWUGO;
1380 return kasprintf(GFP_KERNEL, "vfio/%s", dev_name(dev));
1383 static int __init vfio_init(void)
1387 idr_init(&vfio.group_idr);
1388 mutex_init(&vfio.group_lock);
1389 mutex_init(&vfio.iommu_drivers_lock);
1390 INIT_LIST_HEAD(&vfio.group_list);
1391 INIT_LIST_HEAD(&vfio.iommu_drivers_list);
1392 init_waitqueue_head(&vfio.release_q);
1394 vfio.class = class_create(THIS_MODULE, "vfio");
1395 if (IS_ERR(vfio.class)) {
1396 ret = PTR_ERR(vfio.class);
1400 vfio.class->devnode = vfio_devnode;
1402 ret = alloc_chrdev_region(&vfio.devt, 0, MINORMASK, "vfio");
1404 goto err_base_chrdev;
1406 cdev_init(&vfio.cdev, &vfio_fops);
1407 ret = cdev_add(&vfio.cdev, vfio.devt, 1);
1411 vfio.dev = device_create(vfio.class, NULL, vfio.devt, NULL, "vfio");
1412 if (IS_ERR(vfio.dev)) {
1413 ret = PTR_ERR(vfio.dev);
1417 /* /dev/vfio/$GROUP */
1418 cdev_init(&vfio.group_cdev, &vfio_group_fops);
1419 ret = cdev_add(&vfio.group_cdev,
1420 MKDEV(MAJOR(vfio.devt), 1), MINORMASK - 1);
1422 goto err_groups_cdev;
1424 pr_info(DRIVER_DESC " version: " DRIVER_VERSION "\n");
1427 * Attempt to load known iommu-drivers. This gives us a working
1428 * environment without the user needing to explicitly load iommu
1431 request_module_nowait("vfio_iommu_type1");
1432 request_module_nowait("vfio_iommu_spapr_tce");
1437 device_destroy(vfio.class, vfio.devt);
1439 cdev_del(&vfio.cdev);
1441 unregister_chrdev_region(vfio.devt, MINORMASK);
1443 class_destroy(vfio.class);
1449 static void __exit vfio_cleanup(void)
1451 WARN_ON(!list_empty(&vfio.group_list));
1453 idr_destroy(&vfio.group_idr);
1454 cdev_del(&vfio.group_cdev);
1455 device_destroy(vfio.class, vfio.devt);
1456 cdev_del(&vfio.cdev);
1457 unregister_chrdev_region(vfio.devt, MINORMASK);
1458 class_destroy(vfio.class);
1462 module_init(vfio_init);
1463 module_exit(vfio_cleanup);
1465 MODULE_VERSION(DRIVER_VERSION);
1466 MODULE_LICENSE("GPL v2");
1467 MODULE_AUTHOR(DRIVER_AUTHOR);
1468 MODULE_DESCRIPTION(DRIVER_DESC);