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(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 if (!device->group->kvm)
166 _vfio_device_get_kvm_safe(device, device->group->kvm);
169 spin_unlock(&device->group->kvm_ref_lock);
172 static int vfio_device_group_open(struct vfio_device *device)
176 mutex_lock(&device->group->group_lock);
177 if (!vfio_group_has_iommu(device->group)) {
182 mutex_lock(&device->dev_set->lock);
185 * Before the first device open, get the KVM pointer currently
186 * associated with the group (if there is one) and obtain a reference
187 * now that will be held until the open_count reaches 0 again. Save
188 * the pointer in the device for use by drivers.
190 if (device->open_count == 0)
191 vfio_device_group_get_kvm_safe(device);
193 ret = vfio_device_open(device, device->group->iommufd);
195 if (device->open_count == 0)
196 vfio_device_put_kvm(device);
198 mutex_unlock(&device->dev_set->lock);
201 mutex_unlock(&device->group->group_lock);
205 void vfio_device_group_close(struct vfio_device *device)
207 mutex_lock(&device->group->group_lock);
208 mutex_lock(&device->dev_set->lock);
210 vfio_device_close(device, device->group->iommufd);
212 if (device->open_count == 0)
213 vfio_device_put_kvm(device);
215 mutex_unlock(&device->dev_set->lock);
216 mutex_unlock(&device->group->group_lock);
219 static struct file *vfio_device_open_file(struct vfio_device *device)
224 ret = vfio_device_group_open(device);
229 * We can't use anon_inode_getfd() because we need to modify
230 * the f_mode flags directly to allow more than just ioctls
232 filep = anon_inode_getfile("[vfio-device]", &vfio_device_fops,
235 ret = PTR_ERR(filep);
236 goto err_close_device;
240 * TODO: add an anon_inode interface to do this.
241 * Appears to be missing by lack of need rather than
242 * explicitly prevented. Now there's need.
244 filep->f_mode |= (FMODE_PREAD | FMODE_PWRITE);
246 if (device->group->type == VFIO_NO_IOMMU)
247 dev_warn(device->dev, "vfio-noiommu device opened by user "
248 "(%s:%d)\n", current->comm, task_pid_nr(current));
250 * On success the ref of device is moved to the file and
251 * put in vfio_device_fops_release()
256 vfio_device_group_close(device);
261 static int vfio_group_ioctl_get_device_fd(struct vfio_group *group,
264 struct vfio_device *device;
270 buf = strndup_user(arg, PAGE_SIZE);
274 device = vfio_device_get_from_name(group, buf);
277 return PTR_ERR(device);
279 fdno = get_unused_fd_flags(O_CLOEXEC);
285 filep = vfio_device_open_file(device);
287 ret = PTR_ERR(filep);
291 fd_install(fdno, filep);
297 vfio_device_put_registration(device);
301 static int vfio_group_ioctl_get_status(struct vfio_group *group,
302 struct vfio_group_status __user *arg)
304 unsigned long minsz = offsetofend(struct vfio_group_status, flags);
305 struct vfio_group_status status;
307 if (copy_from_user(&status, arg, minsz))
310 if (status.argsz < minsz)
315 mutex_lock(&group->group_lock);
316 if (!group->iommu_group) {
317 mutex_unlock(&group->group_lock);
322 * With the container FD the iommu_group_claim_dma_owner() is done
323 * during SET_CONTAINER but for IOMMFD this is done during
324 * VFIO_GROUP_GET_DEVICE_FD. Meaning that with iommufd
325 * VFIO_GROUP_FLAGS_VIABLE could be set but GET_DEVICE_FD will fail due
328 if (vfio_group_has_iommu(group))
329 status.flags |= VFIO_GROUP_FLAGS_CONTAINER_SET |
330 VFIO_GROUP_FLAGS_VIABLE;
331 else if (!iommu_group_dma_owner_claimed(group->iommu_group))
332 status.flags |= VFIO_GROUP_FLAGS_VIABLE;
333 mutex_unlock(&group->group_lock);
335 if (copy_to_user(arg, &status, minsz))
340 static long vfio_group_fops_unl_ioctl(struct file *filep,
341 unsigned int cmd, unsigned long arg)
343 struct vfio_group *group = filep->private_data;
344 void __user *uarg = (void __user *)arg;
347 case VFIO_GROUP_GET_DEVICE_FD:
348 return vfio_group_ioctl_get_device_fd(group, uarg);
349 case VFIO_GROUP_GET_STATUS:
350 return vfio_group_ioctl_get_status(group, uarg);
351 case VFIO_GROUP_SET_CONTAINER:
352 return vfio_group_ioctl_set_container(group, uarg);
353 case VFIO_GROUP_UNSET_CONTAINER:
354 return vfio_group_ioctl_unset_container(group);
360 static int vfio_group_fops_open(struct inode *inode, struct file *filep)
362 struct vfio_group *group =
363 container_of(inode->i_cdev, struct vfio_group, cdev);
366 mutex_lock(&group->group_lock);
369 * drivers can be zero if this races with vfio_device_remove_group(), it
370 * will be stable at 0 under the group rwsem
372 if (refcount_read(&group->drivers) == 0) {
377 if (group->type == VFIO_NO_IOMMU && !capable(CAP_SYS_RAWIO)) {
383 * Do we need multiple instances of the group open? Seems not.
385 if (group->opened_file) {
389 group->opened_file = filep;
390 filep->private_data = group;
393 mutex_unlock(&group->group_lock);
397 static int vfio_group_fops_release(struct inode *inode, struct file *filep)
399 struct vfio_group *group = filep->private_data;
401 filep->private_data = NULL;
403 mutex_lock(&group->group_lock);
405 * Device FDs hold a group file reference, therefore the group release
406 * is only called when there are no open devices.
408 WARN_ON(group->notifier.head);
409 if (group->container)
410 vfio_group_detach_container(group);
411 if (group->iommufd) {
412 iommufd_ctx_put(group->iommufd);
413 group->iommufd = NULL;
415 group->opened_file = NULL;
416 mutex_unlock(&group->group_lock);
420 static const struct file_operations vfio_group_fops = {
421 .owner = THIS_MODULE,
422 .unlocked_ioctl = vfio_group_fops_unl_ioctl,
423 .compat_ioctl = compat_ptr_ioctl,
424 .open = vfio_group_fops_open,
425 .release = vfio_group_fops_release,
429 * Group objects - create, release, get, put, search
431 static struct vfio_group *
432 vfio_group_find_from_iommu(struct iommu_group *iommu_group)
434 struct vfio_group *group;
436 lockdep_assert_held(&vfio.group_lock);
439 * group->iommu_group from the vfio.group_list cannot be NULL
440 * under the vfio.group_lock.
442 list_for_each_entry(group, &vfio.group_list, vfio_next) {
443 if (group->iommu_group == iommu_group)
449 static void vfio_group_release(struct device *dev)
451 struct vfio_group *group = container_of(dev, struct vfio_group, dev);
453 mutex_destroy(&group->device_lock);
454 mutex_destroy(&group->group_lock);
455 WARN_ON(group->iommu_group);
456 ida_free(&vfio.group_ida, MINOR(group->dev.devt));
460 static struct vfio_group *vfio_group_alloc(struct iommu_group *iommu_group,
461 enum vfio_group_type type)
463 struct vfio_group *group;
466 group = kzalloc(sizeof(*group), GFP_KERNEL);
468 return ERR_PTR(-ENOMEM);
470 minor = ida_alloc_max(&vfio.group_ida, MINORMASK, GFP_KERNEL);
473 return ERR_PTR(minor);
476 device_initialize(&group->dev);
477 group->dev.devt = MKDEV(MAJOR(vfio.group_devt), minor);
478 group->dev.class = vfio.class;
479 group->dev.release = vfio_group_release;
480 cdev_init(&group->cdev, &vfio_group_fops);
481 group->cdev.owner = THIS_MODULE;
483 refcount_set(&group->drivers, 1);
484 mutex_init(&group->group_lock);
485 spin_lock_init(&group->kvm_ref_lock);
486 INIT_LIST_HEAD(&group->device_list);
487 mutex_init(&group->device_lock);
488 group->iommu_group = iommu_group;
489 /* put in vfio_group_release() */
490 iommu_group_ref_get(iommu_group);
492 BLOCKING_INIT_NOTIFIER_HEAD(&group->notifier);
497 static struct vfio_group *vfio_create_group(struct iommu_group *iommu_group,
498 enum vfio_group_type type)
500 struct vfio_group *group;
501 struct vfio_group *ret;
504 lockdep_assert_held(&vfio.group_lock);
506 group = vfio_group_alloc(iommu_group, type);
510 err = dev_set_name(&group->dev, "%s%d",
511 group->type == VFIO_NO_IOMMU ? "noiommu-" : "",
512 iommu_group_id(iommu_group));
518 err = cdev_device_add(&group->cdev, &group->dev);
524 list_add(&group->vfio_next, &vfio.group_list);
529 put_device(&group->dev);
533 static struct vfio_group *vfio_noiommu_group_alloc(struct device *dev,
534 enum vfio_group_type type)
536 struct iommu_group *iommu_group;
537 struct vfio_group *group;
540 iommu_group = iommu_group_alloc();
541 if (IS_ERR(iommu_group))
542 return ERR_CAST(iommu_group);
544 ret = iommu_group_set_name(iommu_group, "vfio-noiommu");
547 ret = iommu_group_add_device(iommu_group, dev);
551 mutex_lock(&vfio.group_lock);
552 group = vfio_create_group(iommu_group, type);
553 mutex_unlock(&vfio.group_lock);
555 ret = PTR_ERR(group);
556 goto out_remove_device;
558 iommu_group_put(iommu_group);
562 iommu_group_remove_device(dev);
564 iommu_group_put(iommu_group);
568 static bool vfio_group_has_device(struct vfio_group *group, struct device *dev)
570 struct vfio_device *device;
572 mutex_lock(&group->device_lock);
573 list_for_each_entry(device, &group->device_list, group_next) {
574 if (device->dev == dev) {
575 mutex_unlock(&group->device_lock);
579 mutex_unlock(&group->device_lock);
583 static struct vfio_group *vfio_group_find_or_alloc(struct device *dev)
585 struct iommu_group *iommu_group;
586 struct vfio_group *group;
588 iommu_group = iommu_group_get(dev);
589 if (!iommu_group && vfio_noiommu) {
591 * With noiommu enabled, create an IOMMU group for devices that
592 * don't already have one, implying no IOMMU hardware/driver
593 * exists. Taint the kernel because we're about to give a DMA
594 * capable device to a user without IOMMU protection.
596 group = vfio_noiommu_group_alloc(dev, VFIO_NO_IOMMU);
597 if (!IS_ERR(group)) {
598 add_taint(TAINT_USER, LOCKDEP_STILL_OK);
599 dev_warn(dev, "Adding kernel taint for vfio-noiommu group on device\n");
605 return ERR_PTR(-EINVAL);
608 * VFIO always sets IOMMU_CACHE because we offer no way for userspace to
609 * restore cache coherency. It has to be checked here because it is only
610 * valid for cases where we are using iommu groups.
612 if (!device_iommu_capable(dev, IOMMU_CAP_CACHE_COHERENCY)) {
613 iommu_group_put(iommu_group);
614 return ERR_PTR(-EINVAL);
617 mutex_lock(&vfio.group_lock);
618 group = vfio_group_find_from_iommu(iommu_group);
620 if (WARN_ON(vfio_group_has_device(group, dev)))
621 group = ERR_PTR(-EINVAL);
623 refcount_inc(&group->drivers);
625 group = vfio_create_group(iommu_group, VFIO_IOMMU);
627 mutex_unlock(&vfio.group_lock);
629 /* The vfio_group holds a reference to the iommu_group */
630 iommu_group_put(iommu_group);
634 int vfio_device_set_group(struct vfio_device *device,
635 enum vfio_group_type type)
637 struct vfio_group *group;
639 if (type == VFIO_IOMMU)
640 group = vfio_group_find_or_alloc(device->dev);
642 group = vfio_noiommu_group_alloc(device->dev, type);
645 return PTR_ERR(group);
647 /* Our reference on group is moved to the device */
648 device->group = group;
652 void vfio_device_remove_group(struct vfio_device *device)
654 struct vfio_group *group = device->group;
655 struct iommu_group *iommu_group;
657 if (group->type == VFIO_NO_IOMMU || group->type == VFIO_EMULATED_IOMMU)
658 iommu_group_remove_device(device->dev);
660 /* Pairs with vfio_create_group() / vfio_group_get_from_iommu() */
661 if (!refcount_dec_and_mutex_lock(&group->drivers, &vfio.group_lock))
663 list_del(&group->vfio_next);
666 * We could concurrently probe another driver in the group that might
667 * race vfio_device_remove_group() with vfio_get_group(), so we have to
668 * ensure that the sysfs is all cleaned up under lock otherwise the
669 * cdev_device_add() will fail due to the name aready existing.
671 cdev_device_del(&group->cdev, &group->dev);
673 mutex_lock(&group->group_lock);
675 * These data structures all have paired operations that can only be
676 * undone when the caller holds a live reference on the device. Since
677 * all pairs must be undone these WARN_ON's indicate some caller did not
678 * properly hold the group reference.
680 WARN_ON(!list_empty(&group->device_list));
681 WARN_ON(group->notifier.head);
684 * Revoke all users of group->iommu_group. At this point we know there
685 * are no devices active because we are unplugging the last one. Setting
686 * iommu_group to NULL blocks all new users.
688 if (group->container)
689 vfio_group_detach_container(group);
690 iommu_group = group->iommu_group;
691 group->iommu_group = NULL;
692 mutex_unlock(&group->group_lock);
693 mutex_unlock(&vfio.group_lock);
695 iommu_group_put(iommu_group);
696 put_device(&group->dev);
699 void vfio_device_group_register(struct vfio_device *device)
701 mutex_lock(&device->group->device_lock);
702 list_add(&device->group_next, &device->group->device_list);
703 mutex_unlock(&device->group->device_lock);
706 void vfio_device_group_unregister(struct vfio_device *device)
708 mutex_lock(&device->group->device_lock);
709 list_del(&device->group_next);
710 mutex_unlock(&device->group->device_lock);
713 int vfio_device_group_use_iommu(struct vfio_device *device)
715 struct vfio_group *group = device->group;
718 lockdep_assert_held(&group->group_lock);
720 if (WARN_ON(!group->container))
723 ret = vfio_group_use_container(group);
726 vfio_device_container_register(device);
730 void vfio_device_group_unuse_iommu(struct vfio_device *device)
732 struct vfio_group *group = device->group;
734 lockdep_assert_held(&group->group_lock);
736 if (WARN_ON(!group->container))
739 vfio_device_container_unregister(device);
740 vfio_group_unuse_container(group);
743 bool vfio_device_has_container(struct vfio_device *device)
745 return device->group->container;
749 * vfio_file_iommu_group - Return the struct iommu_group for the vfio group file
750 * @file: VFIO group file
752 * The returned iommu_group is valid as long as a ref is held on the file. This
753 * returns a reference on the group. This function is deprecated, only the SPAPR
754 * path in kvm should call it.
756 struct iommu_group *vfio_file_iommu_group(struct file *file)
758 struct vfio_group *group = file->private_data;
759 struct iommu_group *iommu_group = NULL;
761 if (!IS_ENABLED(CONFIG_SPAPR_TCE_IOMMU))
764 if (!vfio_file_is_group(file))
767 mutex_lock(&group->group_lock);
768 if (group->iommu_group) {
769 iommu_group = group->iommu_group;
770 iommu_group_ref_get(iommu_group);
772 mutex_unlock(&group->group_lock);
775 EXPORT_SYMBOL_GPL(vfio_file_iommu_group);
778 * vfio_file_is_group - True if the file is usable with VFIO aPIS
779 * @file: VFIO group file
781 bool vfio_file_is_group(struct file *file)
783 return file->f_op == &vfio_group_fops;
785 EXPORT_SYMBOL_GPL(vfio_file_is_group);
788 * vfio_file_enforced_coherent - True if the DMA associated with the VFIO file
789 * is always CPU cache coherent
790 * @file: VFIO group file
792 * Enforced coherency means that the IOMMU ignores things like the PCIe no-snoop
793 * bit in DMA transactions. A return of false indicates that the user has
794 * rights to access additional instructions such as wbinvd on x86.
796 bool vfio_file_enforced_coherent(struct file *file)
798 struct vfio_group *group = file->private_data;
799 struct vfio_device *device;
802 if (!vfio_file_is_group(file))
806 * If the device does not have IOMMU_CAP_ENFORCE_CACHE_COHERENCY then
807 * any domain later attached to it will also not support it. If the cap
808 * is set then the iommu_domain eventually attached to the device/group
809 * must use a domain with enforce_cache_coherency().
811 mutex_lock(&group->device_lock);
812 list_for_each_entry(device, &group->device_list, group_next) {
813 if (!device_iommu_capable(device->dev,
814 IOMMU_CAP_ENFORCE_CACHE_COHERENCY)) {
819 mutex_unlock(&group->device_lock);
822 EXPORT_SYMBOL_GPL(vfio_file_enforced_coherent);
825 * vfio_file_set_kvm - Link a kvm with VFIO drivers
826 * @file: VFIO group file
829 * When a VFIO device is first opened the KVM will be available in
830 * device->kvm if one was associated with the group.
832 void vfio_file_set_kvm(struct file *file, struct kvm *kvm)
834 struct vfio_group *group = file->private_data;
836 if (!vfio_file_is_group(file))
839 spin_lock(&group->kvm_ref_lock);
841 spin_unlock(&group->kvm_ref_lock);
843 EXPORT_SYMBOL_GPL(vfio_file_set_kvm);
846 * vfio_file_has_dev - True if the VFIO file is a handle for device
847 * @file: VFIO file to check
848 * @device: Device that must be part of the file
850 * Returns true if given file has permission to manipulate the given device.
852 bool vfio_file_has_dev(struct file *file, struct vfio_device *device)
854 struct vfio_group *group = file->private_data;
856 if (!vfio_file_is_group(file))
859 return group == device->group;
861 EXPORT_SYMBOL_GPL(vfio_file_has_dev);
863 static char *vfio_devnode(const struct device *dev, umode_t *mode)
865 return kasprintf(GFP_KERNEL, "vfio/%s", dev_name(dev));
868 int __init vfio_group_init(void)
872 ida_init(&vfio.group_ida);
873 mutex_init(&vfio.group_lock);
874 INIT_LIST_HEAD(&vfio.group_list);
876 ret = vfio_container_init();
880 /* /dev/vfio/$GROUP */
881 vfio.class = class_create("vfio");
882 if (IS_ERR(vfio.class)) {
883 ret = PTR_ERR(vfio.class);
884 goto err_group_class;
887 vfio.class->devnode = vfio_devnode;
889 ret = alloc_chrdev_region(&vfio.group_devt, 0, MINORMASK + 1, "vfio");
891 goto err_alloc_chrdev;
895 class_destroy(vfio.class);
898 vfio_container_cleanup();
902 void vfio_group_cleanup(void)
904 WARN_ON(!list_empty(&vfio.group_list));
905 ida_destroy(&vfio.group_ida);
906 unregister_chrdev_region(vfio.group_devt, MINORMASK + 1);
907 class_destroy(vfio.class);
909 vfio_container_cleanup();