1 // SPDX-License-Identifier: GPL-2.0-only
5 * Copyright (C) 2012 Red Hat, Inc. All rights reserved.
6 * Author: Alex Williamson <alex.williamson@redhat.com>
8 * Derived from original vfio:
9 * Copyright 2010 Cisco Systems, Inc. All rights reserved.
10 * Author: Tom Lyon, pugs@cisco.com
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(group->iommufd);
147 group->iommufd = iommufd;
151 /* The FD passed is not recognized. */
155 mutex_unlock(&group->group_lock);
160 static int vfio_device_group_open(struct vfio_device *device)
164 mutex_lock(&device->group->group_lock);
165 if (!vfio_group_has_iommu(device->group)) {
171 * Here we pass the KVM pointer with the group under the lock. If the
172 * device driver will use it, it must obtain a reference and release it
173 * during close_device.
175 ret = vfio_device_open(device, device->group->iommufd,
179 mutex_unlock(&device->group->group_lock);
183 void vfio_device_group_close(struct vfio_device *device)
185 mutex_lock(&device->group->group_lock);
186 vfio_device_close(device, device->group->iommufd);
187 mutex_unlock(&device->group->group_lock);
190 static struct file *vfio_device_open_file(struct vfio_device *device)
195 ret = vfio_device_group_open(device);
200 * We can't use anon_inode_getfd() because we need to modify
201 * the f_mode flags directly to allow more than just ioctls
203 filep = anon_inode_getfile("[vfio-device]", &vfio_device_fops,
206 ret = PTR_ERR(filep);
207 goto err_close_device;
211 * TODO: add an anon_inode interface to do this.
212 * Appears to be missing by lack of need rather than
213 * explicitly prevented. Now there's need.
215 filep->f_mode |= (FMODE_PREAD | FMODE_PWRITE);
217 if (device->group->type == VFIO_NO_IOMMU)
218 dev_warn(device->dev, "vfio-noiommu device opened by user "
219 "(%s:%d)\n", current->comm, task_pid_nr(current));
221 * On success the ref of device is moved to the file and
222 * put in vfio_device_fops_release()
227 vfio_device_group_close(device);
232 static int vfio_group_ioctl_get_device_fd(struct vfio_group *group,
235 struct vfio_device *device;
241 buf = strndup_user(arg, PAGE_SIZE);
245 device = vfio_device_get_from_name(group, buf);
248 return PTR_ERR(device);
250 fdno = get_unused_fd_flags(O_CLOEXEC);
256 filep = vfio_device_open_file(device);
258 ret = PTR_ERR(filep);
262 fd_install(fdno, filep);
268 vfio_device_put_registration(device);
272 static int vfio_group_ioctl_get_status(struct vfio_group *group,
273 struct vfio_group_status __user *arg)
275 unsigned long minsz = offsetofend(struct vfio_group_status, flags);
276 struct vfio_group_status status;
278 if (copy_from_user(&status, arg, minsz))
281 if (status.argsz < minsz)
286 mutex_lock(&group->group_lock);
287 if (!group->iommu_group) {
288 mutex_unlock(&group->group_lock);
293 * With the container FD the iommu_group_claim_dma_owner() is done
294 * during SET_CONTAINER but for IOMMFD this is done during
295 * VFIO_GROUP_GET_DEVICE_FD. Meaning that with iommufd
296 * VFIO_GROUP_FLAGS_VIABLE could be set but GET_DEVICE_FD will fail due
299 if (vfio_group_has_iommu(group))
300 status.flags |= VFIO_GROUP_FLAGS_CONTAINER_SET |
301 VFIO_GROUP_FLAGS_VIABLE;
302 else if (!iommu_group_dma_owner_claimed(group->iommu_group))
303 status.flags |= VFIO_GROUP_FLAGS_VIABLE;
304 mutex_unlock(&group->group_lock);
306 if (copy_to_user(arg, &status, minsz))
311 static long vfio_group_fops_unl_ioctl(struct file *filep,
312 unsigned int cmd, unsigned long arg)
314 struct vfio_group *group = filep->private_data;
315 void __user *uarg = (void __user *)arg;
318 case VFIO_GROUP_GET_DEVICE_FD:
319 return vfio_group_ioctl_get_device_fd(group, uarg);
320 case VFIO_GROUP_GET_STATUS:
321 return vfio_group_ioctl_get_status(group, uarg);
322 case VFIO_GROUP_SET_CONTAINER:
323 return vfio_group_ioctl_set_container(group, uarg);
324 case VFIO_GROUP_UNSET_CONTAINER:
325 return vfio_group_ioctl_unset_container(group);
331 static int vfio_group_fops_open(struct inode *inode, struct file *filep)
333 struct vfio_group *group =
334 container_of(inode->i_cdev, struct vfio_group, cdev);
337 mutex_lock(&group->group_lock);
340 * drivers can be zero if this races with vfio_device_remove_group(), it
341 * will be stable at 0 under the group rwsem
343 if (refcount_read(&group->drivers) == 0) {
348 if (group->type == VFIO_NO_IOMMU && !capable(CAP_SYS_RAWIO)) {
354 * Do we need multiple instances of the group open? Seems not.
356 if (group->opened_file) {
360 group->opened_file = filep;
361 filep->private_data = group;
364 mutex_unlock(&group->group_lock);
368 static int vfio_group_fops_release(struct inode *inode, struct file *filep)
370 struct vfio_group *group = filep->private_data;
372 filep->private_data = NULL;
374 mutex_lock(&group->group_lock);
376 * Device FDs hold a group file reference, therefore the group release
377 * is only called when there are no open devices.
379 WARN_ON(group->notifier.head);
380 if (group->container)
381 vfio_group_detach_container(group);
382 if (group->iommufd) {
383 iommufd_ctx_put(group->iommufd);
384 group->iommufd = NULL;
386 group->opened_file = NULL;
387 mutex_unlock(&group->group_lock);
391 static const struct file_operations vfio_group_fops = {
392 .owner = THIS_MODULE,
393 .unlocked_ioctl = vfio_group_fops_unl_ioctl,
394 .compat_ioctl = compat_ptr_ioctl,
395 .open = vfio_group_fops_open,
396 .release = vfio_group_fops_release,
400 * Group objects - create, release, get, put, search
402 static struct vfio_group *
403 vfio_group_find_from_iommu(struct iommu_group *iommu_group)
405 struct vfio_group *group;
407 lockdep_assert_held(&vfio.group_lock);
410 * group->iommu_group from the vfio.group_list cannot be NULL
411 * under the vfio.group_lock.
413 list_for_each_entry(group, &vfio.group_list, vfio_next) {
414 if (group->iommu_group == iommu_group)
420 static void vfio_group_release(struct device *dev)
422 struct vfio_group *group = container_of(dev, struct vfio_group, dev);
424 mutex_destroy(&group->device_lock);
425 mutex_destroy(&group->group_lock);
426 WARN_ON(group->iommu_group);
427 ida_free(&vfio.group_ida, MINOR(group->dev.devt));
431 static struct vfio_group *vfio_group_alloc(struct iommu_group *iommu_group,
432 enum vfio_group_type type)
434 struct vfio_group *group;
437 group = kzalloc(sizeof(*group), GFP_KERNEL);
439 return ERR_PTR(-ENOMEM);
441 minor = ida_alloc_max(&vfio.group_ida, MINORMASK, GFP_KERNEL);
444 return ERR_PTR(minor);
447 device_initialize(&group->dev);
448 group->dev.devt = MKDEV(MAJOR(vfio.group_devt), minor);
449 group->dev.class = vfio.class;
450 group->dev.release = vfio_group_release;
451 cdev_init(&group->cdev, &vfio_group_fops);
452 group->cdev.owner = THIS_MODULE;
454 refcount_set(&group->drivers, 1);
455 mutex_init(&group->group_lock);
456 INIT_LIST_HEAD(&group->device_list);
457 mutex_init(&group->device_lock);
458 group->iommu_group = iommu_group;
459 /* put in vfio_group_release() */
460 iommu_group_ref_get(iommu_group);
462 BLOCKING_INIT_NOTIFIER_HEAD(&group->notifier);
467 static struct vfio_group *vfio_create_group(struct iommu_group *iommu_group,
468 enum vfio_group_type type)
470 struct vfio_group *group;
471 struct vfio_group *ret;
474 lockdep_assert_held(&vfio.group_lock);
476 group = vfio_group_alloc(iommu_group, type);
480 err = dev_set_name(&group->dev, "%s%d",
481 group->type == VFIO_NO_IOMMU ? "noiommu-" : "",
482 iommu_group_id(iommu_group));
488 err = cdev_device_add(&group->cdev, &group->dev);
494 list_add(&group->vfio_next, &vfio.group_list);
499 put_device(&group->dev);
503 static struct vfio_group *vfio_noiommu_group_alloc(struct device *dev,
504 enum vfio_group_type type)
506 struct iommu_group *iommu_group;
507 struct vfio_group *group;
510 iommu_group = iommu_group_alloc();
511 if (IS_ERR(iommu_group))
512 return ERR_CAST(iommu_group);
514 ret = iommu_group_set_name(iommu_group, "vfio-noiommu");
517 ret = iommu_group_add_device(iommu_group, dev);
521 mutex_lock(&vfio.group_lock);
522 group = vfio_create_group(iommu_group, type);
523 mutex_unlock(&vfio.group_lock);
525 ret = PTR_ERR(group);
526 goto out_remove_device;
528 iommu_group_put(iommu_group);
532 iommu_group_remove_device(dev);
534 iommu_group_put(iommu_group);
538 static bool vfio_group_has_device(struct vfio_group *group, struct device *dev)
540 struct vfio_device *device;
542 mutex_lock(&group->device_lock);
543 list_for_each_entry(device, &group->device_list, group_next) {
544 if (device->dev == dev) {
545 mutex_unlock(&group->device_lock);
549 mutex_unlock(&group->device_lock);
553 static struct vfio_group *vfio_group_find_or_alloc(struct device *dev)
555 struct iommu_group *iommu_group;
556 struct vfio_group *group;
558 iommu_group = iommu_group_get(dev);
559 if (!iommu_group && vfio_noiommu) {
561 * With noiommu enabled, create an IOMMU group for devices that
562 * don't already have one, implying no IOMMU hardware/driver
563 * exists. Taint the kernel because we're about to give a DMA
564 * capable device to a user without IOMMU protection.
566 group = vfio_noiommu_group_alloc(dev, VFIO_NO_IOMMU);
567 if (!IS_ERR(group)) {
568 add_taint(TAINT_USER, LOCKDEP_STILL_OK);
569 dev_warn(dev, "Adding kernel taint for vfio-noiommu group on device\n");
575 return ERR_PTR(-EINVAL);
578 * VFIO always sets IOMMU_CACHE because we offer no way for userspace to
579 * restore cache coherency. It has to be checked here because it is only
580 * valid for cases where we are using iommu groups.
582 if (!device_iommu_capable(dev, IOMMU_CAP_CACHE_COHERENCY)) {
583 iommu_group_put(iommu_group);
584 return ERR_PTR(-EINVAL);
587 mutex_lock(&vfio.group_lock);
588 group = vfio_group_find_from_iommu(iommu_group);
590 if (WARN_ON(vfio_group_has_device(group, dev)))
591 group = ERR_PTR(-EINVAL);
593 refcount_inc(&group->drivers);
595 group = vfio_create_group(iommu_group, VFIO_IOMMU);
597 mutex_unlock(&vfio.group_lock);
599 /* The vfio_group holds a reference to the iommu_group */
600 iommu_group_put(iommu_group);
604 int vfio_device_set_group(struct vfio_device *device,
605 enum vfio_group_type type)
607 struct vfio_group *group;
609 if (type == VFIO_IOMMU)
610 group = vfio_group_find_or_alloc(device->dev);
612 group = vfio_noiommu_group_alloc(device->dev, type);
615 return PTR_ERR(group);
617 /* Our reference on group is moved to the device */
618 device->group = group;
622 void vfio_device_remove_group(struct vfio_device *device)
624 struct vfio_group *group = device->group;
625 struct iommu_group *iommu_group;
627 if (group->type == VFIO_NO_IOMMU || group->type == VFIO_EMULATED_IOMMU)
628 iommu_group_remove_device(device->dev);
630 /* Pairs with vfio_create_group() / vfio_group_get_from_iommu() */
631 if (!refcount_dec_and_mutex_lock(&group->drivers, &vfio.group_lock))
633 list_del(&group->vfio_next);
636 * We could concurrently probe another driver in the group that might
637 * race vfio_device_remove_group() with vfio_get_group(), so we have to
638 * ensure that the sysfs is all cleaned up under lock otherwise the
639 * cdev_device_add() will fail due to the name aready existing.
641 cdev_device_del(&group->cdev, &group->dev);
643 mutex_lock(&group->group_lock);
645 * These data structures all have paired operations that can only be
646 * undone when the caller holds a live reference on the device. Since
647 * all pairs must be undone these WARN_ON's indicate some caller did not
648 * properly hold the group reference.
650 WARN_ON(!list_empty(&group->device_list));
651 WARN_ON(group->notifier.head);
654 * Revoke all users of group->iommu_group. At this point we know there
655 * are no devices active because we are unplugging the last one. Setting
656 * iommu_group to NULL blocks all new users.
658 if (group->container)
659 vfio_group_detach_container(group);
660 iommu_group = group->iommu_group;
661 group->iommu_group = NULL;
662 mutex_unlock(&group->group_lock);
663 mutex_unlock(&vfio.group_lock);
665 iommu_group_put(iommu_group);
666 put_device(&group->dev);
669 void vfio_device_group_register(struct vfio_device *device)
671 mutex_lock(&device->group->device_lock);
672 list_add(&device->group_next, &device->group->device_list);
673 mutex_unlock(&device->group->device_lock);
676 void vfio_device_group_unregister(struct vfio_device *device)
678 mutex_lock(&device->group->device_lock);
679 list_del(&device->group_next);
680 mutex_unlock(&device->group->device_lock);
683 int vfio_device_group_use_iommu(struct vfio_device *device)
685 struct vfio_group *group = device->group;
688 lockdep_assert_held(&group->group_lock);
690 if (WARN_ON(!group->container))
693 ret = vfio_group_use_container(group);
696 vfio_device_container_register(device);
700 void vfio_device_group_unuse_iommu(struct vfio_device *device)
702 struct vfio_group *group = device->group;
704 lockdep_assert_held(&group->group_lock);
706 if (WARN_ON(!group->container))
709 vfio_device_container_unregister(device);
710 vfio_group_unuse_container(group);
713 bool vfio_device_has_container(struct vfio_device *device)
715 return device->group->container;
719 * vfio_file_iommu_group - Return the struct iommu_group for the vfio group file
720 * @file: VFIO group file
722 * The returned iommu_group is valid as long as a ref is held on the file. This
723 * returns a reference on the group. This function is deprecated, only the SPAPR
724 * path in kvm should call it.
726 struct iommu_group *vfio_file_iommu_group(struct file *file)
728 struct vfio_group *group = file->private_data;
729 struct iommu_group *iommu_group = NULL;
731 if (!IS_ENABLED(CONFIG_SPAPR_TCE_IOMMU))
734 if (!vfio_file_is_group(file))
737 mutex_lock(&group->group_lock);
738 if (group->iommu_group) {
739 iommu_group = group->iommu_group;
740 iommu_group_ref_get(iommu_group);
742 mutex_unlock(&group->group_lock);
745 EXPORT_SYMBOL_GPL(vfio_file_iommu_group);
748 * vfio_file_is_group - True if the file is usable with VFIO aPIS
749 * @file: VFIO group file
751 bool vfio_file_is_group(struct file *file)
753 return file->f_op == &vfio_group_fops;
755 EXPORT_SYMBOL_GPL(vfio_file_is_group);
758 * vfio_file_enforced_coherent - True if the DMA associated with the VFIO file
759 * is always CPU cache coherent
760 * @file: VFIO group file
762 * Enforced coherency means that the IOMMU ignores things like the PCIe no-snoop
763 * bit in DMA transactions. A return of false indicates that the user has
764 * rights to access additional instructions such as wbinvd on x86.
766 bool vfio_file_enforced_coherent(struct file *file)
768 struct vfio_group *group = file->private_data;
769 struct vfio_device *device;
772 if (!vfio_file_is_group(file))
776 * If the device does not have IOMMU_CAP_ENFORCE_CACHE_COHERENCY then
777 * any domain later attached to it will also not support it. If the cap
778 * is set then the iommu_domain eventually attached to the device/group
779 * must use a domain with enforce_cache_coherency().
781 mutex_lock(&group->device_lock);
782 list_for_each_entry(device, &group->device_list, group_next) {
783 if (!device_iommu_capable(device->dev,
784 IOMMU_CAP_ENFORCE_CACHE_COHERENCY)) {
789 mutex_unlock(&group->device_lock);
792 EXPORT_SYMBOL_GPL(vfio_file_enforced_coherent);
795 * vfio_file_set_kvm - Link a kvm with VFIO drivers
796 * @file: VFIO group file
799 * When a VFIO device is first opened the KVM will be available in
800 * device->kvm if one was associated with the group.
802 void vfio_file_set_kvm(struct file *file, struct kvm *kvm)
804 struct vfio_group *group = file->private_data;
806 if (!vfio_file_is_group(file))
809 mutex_lock(&group->group_lock);
811 mutex_unlock(&group->group_lock);
813 EXPORT_SYMBOL_GPL(vfio_file_set_kvm);
816 * vfio_file_has_dev - True if the VFIO file is a handle for device
817 * @file: VFIO file to check
818 * @device: Device that must be part of the file
820 * Returns true if given file has permission to manipulate the given device.
822 bool vfio_file_has_dev(struct file *file, struct vfio_device *device)
824 struct vfio_group *group = file->private_data;
826 if (!vfio_file_is_group(file))
829 return group == device->group;
831 EXPORT_SYMBOL_GPL(vfio_file_has_dev);
833 static char *vfio_devnode(const struct device *dev, umode_t *mode)
835 return kasprintf(GFP_KERNEL, "vfio/%s", dev_name(dev));
838 int __init vfio_group_init(void)
842 ida_init(&vfio.group_ida);
843 mutex_init(&vfio.group_lock);
844 INIT_LIST_HEAD(&vfio.group_list);
846 ret = vfio_container_init();
850 /* /dev/vfio/$GROUP */
851 vfio.class = class_create(THIS_MODULE, "vfio");
852 if (IS_ERR(vfio.class)) {
853 ret = PTR_ERR(vfio.class);
854 goto err_group_class;
857 vfio.class->devnode = vfio_devnode;
859 ret = alloc_chrdev_region(&vfio.group_devt, 0, MINORMASK + 1, "vfio");
861 goto err_alloc_chrdev;
865 class_destroy(vfio.class);
868 vfio_container_cleanup();
872 void vfio_group_cleanup(void)
874 WARN_ON(!list_empty(&vfio.group_list));
875 ida_destroy(&vfio.group_ida);
876 unregister_chrdev_region(vfio.group_devt, MINORMASK + 1);
877 class_destroy(vfio.class);
879 vfio_container_cleanup();