1 // SPDX-License-Identifier: GPL-2.0-only
5 * Copyright (C) 2012 Red Hat, Inc. All rights reserved.
8 * Derived from original vfio:
9 * Copyright 2010 Cisco Systems, Inc. All rights reserved.
13 #include <linux/vfio.h>
14 #include <linux/iommufd.h>
15 #include <linux/anon_inodes.h>
20 struct list_head group_list;
21 struct mutex group_lock; /* locks group_list */
26 static struct vfio_device *vfio_device_get_from_name(struct vfio_group *group,
29 struct vfio_device *it, *device = ERR_PTR(-ENODEV);
31 mutex_lock(&group->device_lock);
32 list_for_each_entry(it, &group->device_list, group_next) {
36 ret = it->ops->match(it, buf);
38 device = ERR_PTR(ret);
42 ret = !strcmp(dev_name(it->dev), buf);
45 if (ret && vfio_device_try_get_registration(it)) {
50 mutex_unlock(&group->device_lock);
56 * VFIO Group fd, /dev/vfio/$GROUP
58 static bool vfio_group_has_iommu(struct vfio_group *group)
60 lockdep_assert_held(&group->group_lock);
62 * There can only be users if there is a container, and if there is a
63 * container there must be users.
65 WARN_ON(!group->container != !group->container_users);
67 return group->container || group->iommufd;
71 * VFIO_GROUP_UNSET_CONTAINER should fail if there are other users or
72 * if there was no container to unset. Since the ioctl is called on
73 * the group, we know that still exists, therefore the only valid
74 * transition here is 1->0.
76 static int vfio_group_ioctl_unset_container(struct vfio_group *group)
80 mutex_lock(&group->group_lock);
81 if (!vfio_group_has_iommu(group)) {
85 if (group->container) {
86 if (group->container_users != 1) {
90 vfio_group_detach_container(group);
93 iommufd_ctx_put(group->iommufd);
94 group->iommufd = NULL;
98 mutex_unlock(&group->group_lock);
102 static int vfio_group_ioctl_set_container(struct vfio_group *group,
105 struct vfio_container *container;
106 struct iommufd_ctx *iommufd;
111 if (get_user(fd, arg))
118 mutex_lock(&group->group_lock);
119 if (vfio_group_has_iommu(group)) {
123 if (!group->iommu_group) {
128 container = vfio_container_from_file(f.file);
130 ret = vfio_container_attach_group(container, group);
134 iommufd = iommufd_ctx_from_file(f.file);
135 if (!IS_ERR(iommufd)) {
136 if (IS_ENABLED(CONFIG_VFIO_NOIOMMU) &&
137 group->type == VFIO_NO_IOMMU)
138 ret = iommufd_vfio_compat_set_no_iommu(iommufd);
140 ret = iommufd_vfio_compat_ioas_create(iommufd);
143 iommufd_ctx_put(iommufd);
147 group->iommufd = iommufd;
151 /* The FD passed is not recognized. */
155 mutex_unlock(&group->group_lock);
160 static void vfio_device_group_get_kvm_safe(struct vfio_device *device)
162 spin_lock(&device->group->kvm_ref_lock);
163 vfio_device_get_kvm_safe(device, device->group->kvm);
164 spin_unlock(&device->group->kvm_ref_lock);
167 static int vfio_df_group_open(struct vfio_device_file *df)
169 struct vfio_device *device = df->device;
172 mutex_lock(&device->group->group_lock);
173 if (!vfio_group_has_iommu(device->group)) {
178 mutex_lock(&device->dev_set->lock);
181 * Before the first device open, get the KVM pointer currently
182 * associated with the group (if there is one) and obtain a reference
183 * now that will be held until the open_count reaches 0 again. Save
184 * the pointer in the device for use by drivers.
186 if (device->open_count == 0)
187 vfio_device_group_get_kvm_safe(device);
189 df->iommufd = device->group->iommufd;
190 if (df->iommufd && vfio_device_is_noiommu(device) && device->open_count == 0) {
192 * Require no compat ioas to be assigned to proceed. The basic
193 * statement is that the user cannot have done something that
194 * implies they expected translation to exist
196 if (!capable(CAP_SYS_RAWIO) ||
197 vfio_iommufd_device_has_compat_ioas(device, df->iommufd))
204 ret = vfio_df_open(df);
208 if (df->iommufd && device->open_count == 1) {
209 ret = vfio_iommufd_compat_attach_ioas(device, df->iommufd);
211 goto out_close_device;
215 * Paired with smp_load_acquire() in vfio_device_fops::ioctl/
216 * read/write/mmap and vfio_file_has_device_access()
218 smp_store_release(&df->access_granted, true);
220 mutex_unlock(&device->dev_set->lock);
221 mutex_unlock(&device->group->group_lock);
228 if (device->open_count == 0)
229 vfio_device_put_kvm(device);
230 mutex_unlock(&device->dev_set->lock);
232 mutex_unlock(&device->group->group_lock);
236 void vfio_df_group_close(struct vfio_device_file *df)
238 struct vfio_device *device = df->device;
240 mutex_lock(&device->group->group_lock);
241 mutex_lock(&device->dev_set->lock);
246 if (device->open_count == 0)
247 vfio_device_put_kvm(device);
249 mutex_unlock(&device->dev_set->lock);
250 mutex_unlock(&device->group->group_lock);
253 static struct file *vfio_device_open_file(struct vfio_device *device)
255 struct vfio_device_file *df;
259 df = vfio_allocate_device_file(device);
265 df->group = device->group;
267 ret = vfio_df_group_open(df);
272 * We can't use anon_inode_getfd() because we need to modify
273 * the f_mode flags directly to allow more than just ioctls
275 filep = anon_inode_getfile("[vfio-device]", &vfio_device_fops,
278 ret = PTR_ERR(filep);
279 goto err_close_device;
283 * TODO: add an anon_inode interface to do this.
284 * Appears to be missing by lack of need rather than
285 * explicitly prevented. Now there's need.
287 filep->f_mode |= (FMODE_PREAD | FMODE_PWRITE);
290 * Use the pseudo fs inode on the device to link all mmaps
291 * to the same address space, allowing us to unmap all vmas
292 * associated to this device using unmap_mapping_range().
294 filep->f_mapping = device->inode->i_mapping;
296 if (device->group->type == VFIO_NO_IOMMU)
297 dev_warn(device->dev, "vfio-noiommu device opened by user "
298 "(%s:%d)\n", current->comm, task_pid_nr(current));
300 * On success the ref of device is moved to the file and
301 * put in vfio_device_fops_release()
306 vfio_df_group_close(df);
313 static int vfio_group_ioctl_get_device_fd(struct vfio_group *group,
316 struct vfio_device *device;
322 buf = strndup_user(arg, PAGE_SIZE);
326 device = vfio_device_get_from_name(group, buf);
329 return PTR_ERR(device);
331 fdno = get_unused_fd_flags(O_CLOEXEC);
337 filep = vfio_device_open_file(device);
339 ret = PTR_ERR(filep);
343 fd_install(fdno, filep);
349 vfio_device_put_registration(device);
353 static int vfio_group_ioctl_get_status(struct vfio_group *group,
354 struct vfio_group_status __user *arg)
356 unsigned long minsz = offsetofend(struct vfio_group_status, flags);
357 struct vfio_group_status status;
359 if (copy_from_user(&status, arg, minsz))
362 if (status.argsz < minsz)
367 mutex_lock(&group->group_lock);
368 if (!group->iommu_group) {
369 mutex_unlock(&group->group_lock);
374 * With the container FD the iommu_group_claim_dma_owner() is done
375 * during SET_CONTAINER but for IOMMFD this is done during
376 * VFIO_GROUP_GET_DEVICE_FD. Meaning that with iommufd
377 * VFIO_GROUP_FLAGS_VIABLE could be set but GET_DEVICE_FD will fail due
380 if (vfio_group_has_iommu(group))
381 status.flags |= VFIO_GROUP_FLAGS_CONTAINER_SET |
382 VFIO_GROUP_FLAGS_VIABLE;
383 else if (!iommu_group_dma_owner_claimed(group->iommu_group))
384 status.flags |= VFIO_GROUP_FLAGS_VIABLE;
385 mutex_unlock(&group->group_lock);
387 if (copy_to_user(arg, &status, minsz))
392 static long vfio_group_fops_unl_ioctl(struct file *filep,
393 unsigned int cmd, unsigned long arg)
395 struct vfio_group *group = filep->private_data;
396 void __user *uarg = (void __user *)arg;
399 case VFIO_GROUP_GET_DEVICE_FD:
400 return vfio_group_ioctl_get_device_fd(group, uarg);
401 case VFIO_GROUP_GET_STATUS:
402 return vfio_group_ioctl_get_status(group, uarg);
403 case VFIO_GROUP_SET_CONTAINER:
404 return vfio_group_ioctl_set_container(group, uarg);
405 case VFIO_GROUP_UNSET_CONTAINER:
406 return vfio_group_ioctl_unset_container(group);
412 int vfio_device_block_group(struct vfio_device *device)
414 struct vfio_group *group = device->group;
417 mutex_lock(&group->group_lock);
418 if (group->opened_file) {
423 group->cdev_device_open_cnt++;
426 mutex_unlock(&group->group_lock);
430 void vfio_device_unblock_group(struct vfio_device *device)
432 struct vfio_group *group = device->group;
434 mutex_lock(&group->group_lock);
435 group->cdev_device_open_cnt--;
436 mutex_unlock(&group->group_lock);
439 static int vfio_group_fops_open(struct inode *inode, struct file *filep)
441 struct vfio_group *group =
442 container_of(inode->i_cdev, struct vfio_group, cdev);
445 mutex_lock(&group->group_lock);
448 * drivers can be zero if this races with vfio_device_remove_group(), it
449 * will be stable at 0 under the group rwsem
451 if (refcount_read(&group->drivers) == 0) {
456 if (group->type == VFIO_NO_IOMMU && !capable(CAP_SYS_RAWIO)) {
461 if (group->cdev_device_open_cnt) {
467 * Do we need multiple instances of the group open? Seems not.
469 if (group->opened_file) {
473 group->opened_file = filep;
474 filep->private_data = group;
477 mutex_unlock(&group->group_lock);
481 static int vfio_group_fops_release(struct inode *inode, struct file *filep)
483 struct vfio_group *group = filep->private_data;
485 filep->private_data = NULL;
487 mutex_lock(&group->group_lock);
489 * Device FDs hold a group file reference, therefore the group release
490 * is only called when there are no open devices.
492 WARN_ON(group->notifier.head);
493 if (group->container)
494 vfio_group_detach_container(group);
495 if (group->iommufd) {
496 iommufd_ctx_put(group->iommufd);
497 group->iommufd = NULL;
499 group->opened_file = NULL;
500 mutex_unlock(&group->group_lock);
504 static const struct file_operations vfio_group_fops = {
505 .owner = THIS_MODULE,
506 .unlocked_ioctl = vfio_group_fops_unl_ioctl,
507 .compat_ioctl = compat_ptr_ioctl,
508 .open = vfio_group_fops_open,
509 .release = vfio_group_fops_release,
513 * Group objects - create, release, get, put, search
515 static struct vfio_group *
516 vfio_group_find_from_iommu(struct iommu_group *iommu_group)
518 struct vfio_group *group;
520 lockdep_assert_held(&vfio.group_lock);
523 * group->iommu_group from the vfio.group_list cannot be NULL
524 * under the vfio.group_lock.
526 list_for_each_entry(group, &vfio.group_list, vfio_next) {
527 if (group->iommu_group == iommu_group)
533 static void vfio_group_release(struct device *dev)
535 struct vfio_group *group = container_of(dev, struct vfio_group, dev);
537 mutex_destroy(&group->device_lock);
538 mutex_destroy(&group->group_lock);
539 WARN_ON(group->iommu_group);
540 WARN_ON(group->cdev_device_open_cnt);
541 ida_free(&vfio.group_ida, MINOR(group->dev.devt));
545 static struct vfio_group *vfio_group_alloc(struct iommu_group *iommu_group,
546 enum vfio_group_type type)
548 struct vfio_group *group;
551 group = kzalloc(sizeof(*group), GFP_KERNEL);
553 return ERR_PTR(-ENOMEM);
555 minor = ida_alloc_max(&vfio.group_ida, MINORMASK, GFP_KERNEL);
558 return ERR_PTR(minor);
561 device_initialize(&group->dev);
562 group->dev.devt = MKDEV(MAJOR(vfio.group_devt), minor);
563 group->dev.class = vfio.class;
564 group->dev.release = vfio_group_release;
565 cdev_init(&group->cdev, &vfio_group_fops);
566 group->cdev.owner = THIS_MODULE;
568 refcount_set(&group->drivers, 1);
569 mutex_init(&group->group_lock);
570 spin_lock_init(&group->kvm_ref_lock);
571 INIT_LIST_HEAD(&group->device_list);
572 mutex_init(&group->device_lock);
573 group->iommu_group = iommu_group;
574 /* put in vfio_group_release() */
575 iommu_group_ref_get(iommu_group);
577 BLOCKING_INIT_NOTIFIER_HEAD(&group->notifier);
582 static struct vfio_group *vfio_create_group(struct iommu_group *iommu_group,
583 enum vfio_group_type type)
585 struct vfio_group *group;
586 struct vfio_group *ret;
589 lockdep_assert_held(&vfio.group_lock);
591 group = vfio_group_alloc(iommu_group, type);
595 err = dev_set_name(&group->dev, "%s%d",
596 group->type == VFIO_NO_IOMMU ? "noiommu-" : "",
597 iommu_group_id(iommu_group));
603 err = cdev_device_add(&group->cdev, &group->dev);
609 list_add(&group->vfio_next, &vfio.group_list);
614 put_device(&group->dev);
618 static struct vfio_group *vfio_noiommu_group_alloc(struct device *dev,
619 enum vfio_group_type type)
621 struct iommu_group *iommu_group;
622 struct vfio_group *group;
625 iommu_group = iommu_group_alloc();
626 if (IS_ERR(iommu_group))
627 return ERR_CAST(iommu_group);
629 ret = iommu_group_set_name(iommu_group, "vfio-noiommu");
632 ret = iommu_group_add_device(iommu_group, dev);
636 mutex_lock(&vfio.group_lock);
637 group = vfio_create_group(iommu_group, type);
638 mutex_unlock(&vfio.group_lock);
640 ret = PTR_ERR(group);
641 goto out_remove_device;
643 iommu_group_put(iommu_group);
647 iommu_group_remove_device(dev);
649 iommu_group_put(iommu_group);
653 static bool vfio_group_has_device(struct vfio_group *group, struct device *dev)
655 struct vfio_device *device;
657 mutex_lock(&group->device_lock);
658 list_for_each_entry(device, &group->device_list, group_next) {
659 if (device->dev == dev) {
660 mutex_unlock(&group->device_lock);
664 mutex_unlock(&group->device_lock);
668 static struct vfio_group *vfio_group_find_or_alloc(struct device *dev)
670 struct iommu_group *iommu_group;
671 struct vfio_group *group;
673 iommu_group = iommu_group_get(dev);
674 if (!iommu_group && vfio_noiommu) {
676 * With noiommu enabled, create an IOMMU group for devices that
677 * don't already have one, implying no IOMMU hardware/driver
678 * exists. Taint the kernel because we're about to give a DMA
679 * capable device to a user without IOMMU protection.
681 group = vfio_noiommu_group_alloc(dev, VFIO_NO_IOMMU);
682 if (!IS_ERR(group)) {
683 add_taint(TAINT_USER, LOCKDEP_STILL_OK);
684 dev_warn(dev, "Adding kernel taint for vfio-noiommu group on device\n");
690 return ERR_PTR(-EINVAL);
692 mutex_lock(&vfio.group_lock);
693 group = vfio_group_find_from_iommu(iommu_group);
695 if (WARN_ON(vfio_group_has_device(group, dev)))
696 group = ERR_PTR(-EINVAL);
698 refcount_inc(&group->drivers);
700 group = vfio_create_group(iommu_group, VFIO_IOMMU);
702 mutex_unlock(&vfio.group_lock);
704 /* The vfio_group holds a reference to the iommu_group */
705 iommu_group_put(iommu_group);
709 int vfio_device_set_group(struct vfio_device *device,
710 enum vfio_group_type type)
712 struct vfio_group *group;
714 if (type == VFIO_IOMMU)
715 group = vfio_group_find_or_alloc(device->dev);
717 group = vfio_noiommu_group_alloc(device->dev, type);
720 return PTR_ERR(group);
722 /* Our reference on group is moved to the device */
723 device->group = group;
727 void vfio_device_remove_group(struct vfio_device *device)
729 struct vfio_group *group = device->group;
730 struct iommu_group *iommu_group;
732 if (group->type == VFIO_NO_IOMMU || group->type == VFIO_EMULATED_IOMMU)
733 iommu_group_remove_device(device->dev);
735 /* Pairs with vfio_create_group() / vfio_group_get_from_iommu() */
736 if (!refcount_dec_and_mutex_lock(&group->drivers, &vfio.group_lock))
738 list_del(&group->vfio_next);
741 * We could concurrently probe another driver in the group that might
742 * race vfio_device_remove_group() with vfio_get_group(), so we have to
743 * ensure that the sysfs is all cleaned up under lock otherwise the
744 * cdev_device_add() will fail due to the name aready existing.
746 cdev_device_del(&group->cdev, &group->dev);
748 mutex_lock(&group->group_lock);
750 * These data structures all have paired operations that can only be
751 * undone when the caller holds a live reference on the device. Since
752 * all pairs must be undone these WARN_ON's indicate some caller did not
753 * properly hold the group reference.
755 WARN_ON(!list_empty(&group->device_list));
756 WARN_ON(group->notifier.head);
759 * Revoke all users of group->iommu_group. At this point we know there
760 * are no devices active because we are unplugging the last one. Setting
761 * iommu_group to NULL blocks all new users.
763 if (group->container)
764 vfio_group_detach_container(group);
765 iommu_group = group->iommu_group;
766 group->iommu_group = NULL;
767 mutex_unlock(&group->group_lock);
768 mutex_unlock(&vfio.group_lock);
770 iommu_group_put(iommu_group);
771 put_device(&group->dev);
774 void vfio_device_group_register(struct vfio_device *device)
776 mutex_lock(&device->group->device_lock);
777 list_add(&device->group_next, &device->group->device_list);
778 mutex_unlock(&device->group->device_lock);
781 void vfio_device_group_unregister(struct vfio_device *device)
783 mutex_lock(&device->group->device_lock);
784 list_del(&device->group_next);
785 mutex_unlock(&device->group->device_lock);
788 int vfio_device_group_use_iommu(struct vfio_device *device)
790 struct vfio_group *group = device->group;
793 lockdep_assert_held(&group->group_lock);
795 if (WARN_ON(!group->container))
798 ret = vfio_group_use_container(group);
801 vfio_device_container_register(device);
805 void vfio_device_group_unuse_iommu(struct vfio_device *device)
807 struct vfio_group *group = device->group;
809 lockdep_assert_held(&group->group_lock);
811 if (WARN_ON(!group->container))
814 vfio_device_container_unregister(device);
815 vfio_group_unuse_container(group);
818 bool vfio_device_has_container(struct vfio_device *device)
820 return device->group->container;
823 struct vfio_group *vfio_group_from_file(struct file *file)
825 struct vfio_group *group = file->private_data;
827 if (file->f_op != &vfio_group_fops)
833 * vfio_file_iommu_group - Return the struct iommu_group for the vfio group file
834 * @file: VFIO group file
836 * The returned iommu_group is valid as long as a ref is held on the file. This
837 * returns a reference on the group. This function is deprecated, only the SPAPR
838 * path in kvm should call it.
840 struct iommu_group *vfio_file_iommu_group(struct file *file)
842 struct vfio_group *group = vfio_group_from_file(file);
843 struct iommu_group *iommu_group = NULL;
845 if (!IS_ENABLED(CONFIG_SPAPR_TCE_IOMMU))
851 mutex_lock(&group->group_lock);
852 if (group->iommu_group) {
853 iommu_group = group->iommu_group;
854 iommu_group_ref_get(iommu_group);
856 mutex_unlock(&group->group_lock);
859 EXPORT_SYMBOL_GPL(vfio_file_iommu_group);
862 * vfio_file_is_group - True if the file is a vfio group file
863 * @file: VFIO group file
865 bool vfio_file_is_group(struct file *file)
867 return vfio_group_from_file(file);
869 EXPORT_SYMBOL_GPL(vfio_file_is_group);
871 bool vfio_group_enforced_coherent(struct vfio_group *group)
873 struct vfio_device *device;
877 * If the device does not have IOMMU_CAP_ENFORCE_CACHE_COHERENCY then
878 * any domain later attached to it will also not support it. If the cap
879 * is set then the iommu_domain eventually attached to the device/group
880 * must use a domain with enforce_cache_coherency().
882 mutex_lock(&group->device_lock);
883 list_for_each_entry(device, &group->device_list, group_next) {
884 if (!device_iommu_capable(device->dev,
885 IOMMU_CAP_ENFORCE_CACHE_COHERENCY)) {
890 mutex_unlock(&group->device_lock);
894 void vfio_group_set_kvm(struct vfio_group *group, struct kvm *kvm)
896 spin_lock(&group->kvm_ref_lock);
898 spin_unlock(&group->kvm_ref_lock);
902 * vfio_file_has_dev - True if the VFIO file is a handle for device
903 * @file: VFIO file to check
904 * @device: Device that must be part of the file
906 * Returns true if given file has permission to manipulate the given device.
908 bool vfio_file_has_dev(struct file *file, struct vfio_device *device)
910 struct vfio_group *group = vfio_group_from_file(file);
915 return group == device->group;
917 EXPORT_SYMBOL_GPL(vfio_file_has_dev);
919 static char *vfio_devnode(const struct device *dev, umode_t *mode)
921 return kasprintf(GFP_KERNEL, "vfio/%s", dev_name(dev));
924 int __init vfio_group_init(void)
928 ida_init(&vfio.group_ida);
929 mutex_init(&vfio.group_lock);
930 INIT_LIST_HEAD(&vfio.group_list);
932 ret = vfio_container_init();
936 /* /dev/vfio/$GROUP */
937 vfio.class = class_create("vfio");
938 if (IS_ERR(vfio.class)) {
939 ret = PTR_ERR(vfio.class);
940 goto err_group_class;
943 vfio.class->devnode = vfio_devnode;
945 ret = alloc_chrdev_region(&vfio.group_devt, 0, MINORMASK + 1, "vfio");
947 goto err_alloc_chrdev;
951 class_destroy(vfio.class);
954 vfio_container_cleanup();
958 void vfio_group_cleanup(void)
960 WARN_ON(!list_empty(&vfio.group_list));
961 ida_destroy(&vfio.group_ida);
962 unregister_chrdev_region(vfio.group_devt, MINORMASK + 1);
963 class_destroy(vfio.class);
965 vfio_container_cleanup();