1 // SPDX-License-Identifier: GPL-2.0-only
3 * VFIO: IOMMU DMA mapping support for Type1 IOMMU
5 * Copyright (C) 2012 Red Hat, Inc. All rights reserved.
8 * Derived from original vfio:
9 * Copyright 2010 Cisco Systems, Inc. All rights reserved.
12 * We arbitrarily define a Type1 IOMMU as one matching the below code.
13 * It could be called the x86 IOMMU as it's designed for AMD-Vi & Intel
14 * VT-d, but that makes it harder to re-use as theoretically anyone
15 * implementing a similar IOMMU could make use of this. We expect the
16 * IOMMU to support the IOMMU API and have few to no restrictions around
17 * the IOVA range that can be mapped. The Type1 IOMMU is currently
18 * optimized for relatively static mappings of a userspace process with
19 * userpsace pages pinned into memory. We also assume devices and IOMMU
20 * domains are PCI based as the IOMMU API is still centered around a
21 * device/bus interface rather than a group interface.
24 #include <linux/compat.h>
25 #include <linux/device.h>
27 #include <linux/iommu.h>
28 #include <linux/module.h>
30 #include <linux/rbtree.h>
31 #include <linux/sched/signal.h>
32 #include <linux/sched/mm.h>
33 #include <linux/slab.h>
34 #include <linux/uaccess.h>
35 #include <linux/vfio.h>
36 #include <linux/workqueue.h>
37 #include <linux/mdev.h>
38 #include <linux/notifier.h>
39 #include <linux/dma-iommu.h>
40 #include <linux/irqdomain.h>
42 #define DRIVER_VERSION "0.2"
44 #define DRIVER_DESC "Type1 IOMMU driver for VFIO"
46 static bool allow_unsafe_interrupts;
47 module_param_named(allow_unsafe_interrupts,
48 allow_unsafe_interrupts, bool, S_IRUGO | S_IWUSR);
49 MODULE_PARM_DESC(allow_unsafe_interrupts,
50 "Enable VFIO IOMMU support for on platforms without interrupt remapping support.");
52 static bool disable_hugepages;
53 module_param_named(disable_hugepages,
54 disable_hugepages, bool, S_IRUGO | S_IWUSR);
55 MODULE_PARM_DESC(disable_hugepages,
56 "Disable VFIO IOMMU support for IOMMU hugepages.");
58 static unsigned int dma_entry_limit __read_mostly = U16_MAX;
59 module_param_named(dma_entry_limit, dma_entry_limit, uint, 0644);
60 MODULE_PARM_DESC(dma_entry_limit,
61 "Maximum number of user DMA mappings per container (65535).");
64 struct list_head domain_list;
65 struct vfio_domain *external_domain; /* domain for external user */
67 struct rb_root dma_list;
68 struct blocking_notifier_head notifier;
69 unsigned int dma_avail;
75 struct iommu_domain *domain;
76 struct list_head next;
77 struct list_head group_list;
78 int prot; /* IOMMU_CACHE */
79 bool fgsp; /* Fine-grained super pages */
84 dma_addr_t iova; /* Device address */
85 unsigned long vaddr; /* Process virtual addr */
86 size_t size; /* Map size (bytes) */
87 int prot; /* IOMMU_READ/WRITE */
89 bool lock_cap; /* capable(CAP_IPC_LOCK) */
90 struct task_struct *task;
91 struct rb_root pfn_list; /* Ex-user pinned pfn list */
95 struct iommu_group *iommu_group;
96 struct list_head next;
97 bool mdev_group; /* An mdev group */
101 * Guest RAM pinning working set or DMA target
105 dma_addr_t iova; /* Device address */
106 unsigned long pfn; /* Host pfn */
110 struct vfio_regions {
111 struct list_head list;
117 #define IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu) \
118 (!list_empty(&iommu->domain_list))
120 static int put_pfn(unsigned long pfn, int prot);
123 * This code handles mapping and unmapping of user data buffers
124 * into DMA'ble space using the IOMMU
127 static struct vfio_dma *vfio_find_dma(struct vfio_iommu *iommu,
128 dma_addr_t start, size_t size)
130 struct rb_node *node = iommu->dma_list.rb_node;
133 struct vfio_dma *dma = rb_entry(node, struct vfio_dma, node);
135 if (start + size <= dma->iova)
136 node = node->rb_left;
137 else if (start >= dma->iova + dma->size)
138 node = node->rb_right;
146 static void vfio_link_dma(struct vfio_iommu *iommu, struct vfio_dma *new)
148 struct rb_node **link = &iommu->dma_list.rb_node, *parent = NULL;
149 struct vfio_dma *dma;
153 dma = rb_entry(parent, struct vfio_dma, node);
155 if (new->iova + new->size <= dma->iova)
156 link = &(*link)->rb_left;
158 link = &(*link)->rb_right;
161 rb_link_node(&new->node, parent, link);
162 rb_insert_color(&new->node, &iommu->dma_list);
165 static void vfio_unlink_dma(struct vfio_iommu *iommu, struct vfio_dma *old)
167 rb_erase(&old->node, &iommu->dma_list);
171 * Helper Functions for host iova-pfn list
173 static struct vfio_pfn *vfio_find_vpfn(struct vfio_dma *dma, dma_addr_t iova)
175 struct vfio_pfn *vpfn;
176 struct rb_node *node = dma->pfn_list.rb_node;
179 vpfn = rb_entry(node, struct vfio_pfn, node);
181 if (iova < vpfn->iova)
182 node = node->rb_left;
183 else if (iova > vpfn->iova)
184 node = node->rb_right;
191 static void vfio_link_pfn(struct vfio_dma *dma,
192 struct vfio_pfn *new)
194 struct rb_node **link, *parent = NULL;
195 struct vfio_pfn *vpfn;
197 link = &dma->pfn_list.rb_node;
200 vpfn = rb_entry(parent, struct vfio_pfn, node);
202 if (new->iova < vpfn->iova)
203 link = &(*link)->rb_left;
205 link = &(*link)->rb_right;
208 rb_link_node(&new->node, parent, link);
209 rb_insert_color(&new->node, &dma->pfn_list);
212 static void vfio_unlink_pfn(struct vfio_dma *dma, struct vfio_pfn *old)
214 rb_erase(&old->node, &dma->pfn_list);
217 static int vfio_add_to_pfn_list(struct vfio_dma *dma, dma_addr_t iova,
220 struct vfio_pfn *vpfn;
222 vpfn = kzalloc(sizeof(*vpfn), GFP_KERNEL);
228 atomic_set(&vpfn->ref_count, 1);
229 vfio_link_pfn(dma, vpfn);
233 static void vfio_remove_from_pfn_list(struct vfio_dma *dma,
234 struct vfio_pfn *vpfn)
236 vfio_unlink_pfn(dma, vpfn);
240 static struct vfio_pfn *vfio_iova_get_vfio_pfn(struct vfio_dma *dma,
243 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
246 atomic_inc(&vpfn->ref_count);
250 static int vfio_iova_put_vfio_pfn(struct vfio_dma *dma, struct vfio_pfn *vpfn)
254 if (atomic_dec_and_test(&vpfn->ref_count)) {
255 ret = put_pfn(vpfn->pfn, dma->prot);
256 vfio_remove_from_pfn_list(dma, vpfn);
261 static int vfio_lock_acct(struct vfio_dma *dma, long npage, bool async)
263 struct mm_struct *mm;
269 mm = async ? get_task_mm(dma->task) : dma->task->mm;
271 return -ESRCH; /* process exited */
273 ret = down_write_killable(&mm->mmap_sem);
275 ret = __account_locked_vm(mm, abs(npage), npage > 0, dma->task,
277 up_write(&mm->mmap_sem);
287 * Some mappings aren't backed by a struct page, for example an mmap'd
288 * MMIO range for our own or another device. These use a different
289 * pfn conversion and shouldn't be tracked as locked pages.
291 static bool is_invalid_reserved_pfn(unsigned long pfn)
293 if (pfn_valid(pfn)) {
295 struct page *tail = pfn_to_page(pfn);
296 struct page *head = compound_head(tail);
297 reserved = !!(PageReserved(head));
300 * "head" is not a dangling pointer
301 * (compound_head takes care of that)
302 * but the hugepage may have been split
303 * from under us (and we may not hold a
304 * reference count on the head page so it can
305 * be reused before we run PageReferenced), so
306 * we've to check PageTail before returning
313 return PageReserved(tail);
319 static int put_pfn(unsigned long pfn, int prot)
321 if (!is_invalid_reserved_pfn(pfn)) {
322 struct page *page = pfn_to_page(pfn);
323 if (prot & IOMMU_WRITE)
331 static int vaddr_get_pfn(struct mm_struct *mm, unsigned long vaddr,
332 int prot, unsigned long *pfn)
334 struct page *page[1];
335 struct vm_area_struct *vma;
336 struct vm_area_struct *vmas[1];
337 unsigned int flags = 0;
340 if (prot & IOMMU_WRITE)
343 down_read(&mm->mmap_sem);
344 if (mm == current->mm) {
345 ret = get_user_pages(vaddr, 1, flags | FOLL_LONGTERM, page,
348 ret = get_user_pages_remote(NULL, mm, vaddr, 1, flags, page,
351 * The lifetime of a vaddr_get_pfn() page pin is
352 * userspace-controlled. In the fs-dax case this could
353 * lead to indefinite stalls in filesystem operations.
354 * Disallow attempts to pin fs-dax pages via this
357 if (ret > 0 && vma_is_fsdax(vmas[0])) {
362 up_read(&mm->mmap_sem);
365 *pfn = page_to_pfn(page[0]);
369 down_read(&mm->mmap_sem);
371 vma = find_vma_intersection(mm, vaddr, vaddr + 1);
373 if (vma && vma->vm_flags & VM_PFNMAP) {
374 *pfn = ((vaddr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
375 if (is_invalid_reserved_pfn(*pfn))
379 up_read(&mm->mmap_sem);
384 * Attempt to pin pages. We really don't want to track all the pfns and
385 * the iommu can only map chunks of consecutive pfns anyway, so get the
386 * first page and all consecutive pages with the same locking.
388 static long vfio_pin_pages_remote(struct vfio_dma *dma, unsigned long vaddr,
389 long npage, unsigned long *pfn_base,
392 unsigned long pfn = 0;
393 long ret, pinned = 0, lock_acct = 0;
395 dma_addr_t iova = vaddr - dma->vaddr + dma->iova;
397 /* This code path is only user initiated */
401 ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, pfn_base);
406 rsvd = is_invalid_reserved_pfn(*pfn_base);
409 * Reserved pages aren't counted against the user, externally pinned
410 * pages are already counted against the user.
412 if (!rsvd && !vfio_find_vpfn(dma, iova)) {
413 if (!dma->lock_cap && current->mm->locked_vm + 1 > limit) {
414 put_pfn(*pfn_base, dma->prot);
415 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n", __func__,
416 limit << PAGE_SHIFT);
422 if (unlikely(disable_hugepages))
425 /* Lock all the consecutive pages from pfn_base */
426 for (vaddr += PAGE_SIZE, iova += PAGE_SIZE; pinned < npage;
427 pinned++, vaddr += PAGE_SIZE, iova += PAGE_SIZE) {
428 ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, &pfn);
432 if (pfn != *pfn_base + pinned ||
433 rsvd != is_invalid_reserved_pfn(pfn)) {
434 put_pfn(pfn, dma->prot);
438 if (!rsvd && !vfio_find_vpfn(dma, iova)) {
439 if (!dma->lock_cap &&
440 current->mm->locked_vm + lock_acct + 1 > limit) {
441 put_pfn(pfn, dma->prot);
442 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n",
443 __func__, limit << PAGE_SHIFT);
452 ret = vfio_lock_acct(dma, lock_acct, false);
457 for (pfn = *pfn_base ; pinned ; pfn++, pinned--)
458 put_pfn(pfn, dma->prot);
467 static long vfio_unpin_pages_remote(struct vfio_dma *dma, dma_addr_t iova,
468 unsigned long pfn, long npage,
471 long unlocked = 0, locked = 0;
474 for (i = 0; i < npage; i++, iova += PAGE_SIZE) {
475 if (put_pfn(pfn++, dma->prot)) {
477 if (vfio_find_vpfn(dma, iova))
483 vfio_lock_acct(dma, locked - unlocked, true);
488 static int vfio_pin_page_external(struct vfio_dma *dma, unsigned long vaddr,
489 unsigned long *pfn_base, bool do_accounting)
491 struct mm_struct *mm;
494 mm = get_task_mm(dma->task);
498 ret = vaddr_get_pfn(mm, vaddr, dma->prot, pfn_base);
499 if (!ret && do_accounting && !is_invalid_reserved_pfn(*pfn_base)) {
500 ret = vfio_lock_acct(dma, 1, true);
502 put_pfn(*pfn_base, dma->prot);
504 pr_warn("%s: Task %s (%d) RLIMIT_MEMLOCK "
505 "(%ld) exceeded\n", __func__,
506 dma->task->comm, task_pid_nr(dma->task),
507 task_rlimit(dma->task, RLIMIT_MEMLOCK));
515 static int vfio_unpin_page_external(struct vfio_dma *dma, dma_addr_t iova,
519 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
524 unlocked = vfio_iova_put_vfio_pfn(dma, vpfn);
527 vfio_lock_acct(dma, -unlocked, true);
532 static int vfio_iommu_type1_pin_pages(void *iommu_data,
533 unsigned long *user_pfn,
535 unsigned long *phys_pfn)
537 struct vfio_iommu *iommu = iommu_data;
539 unsigned long remote_vaddr;
540 struct vfio_dma *dma;
543 if (!iommu || !user_pfn || !phys_pfn)
546 /* Supported for v2 version only */
550 mutex_lock(&iommu->lock);
552 /* Fail if notifier list is empty */
553 if (!iommu->notifier.head) {
559 * If iommu capable domain exist in the container then all pages are
560 * already pinned and accounted. Accouting should be done if there is no
561 * iommu capable domain in the container.
563 do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu);
565 for (i = 0; i < npage; i++) {
567 struct vfio_pfn *vpfn;
569 iova = user_pfn[i] << PAGE_SHIFT;
570 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
576 if ((dma->prot & prot) != prot) {
581 vpfn = vfio_iova_get_vfio_pfn(dma, iova);
583 phys_pfn[i] = vpfn->pfn;
587 remote_vaddr = dma->vaddr + iova - dma->iova;
588 ret = vfio_pin_page_external(dma, remote_vaddr, &phys_pfn[i],
593 ret = vfio_add_to_pfn_list(dma, iova, phys_pfn[i]);
595 vfio_unpin_page_external(dma, iova, do_accounting);
605 for (j = 0; j < i; j++) {
608 iova = user_pfn[j] << PAGE_SHIFT;
609 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
610 vfio_unpin_page_external(dma, iova, do_accounting);
614 mutex_unlock(&iommu->lock);
618 static int vfio_iommu_type1_unpin_pages(void *iommu_data,
619 unsigned long *user_pfn,
622 struct vfio_iommu *iommu = iommu_data;
626 if (!iommu || !user_pfn)
629 /* Supported for v2 version only */
633 mutex_lock(&iommu->lock);
635 do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu);
636 for (i = 0; i < npage; i++) {
637 struct vfio_dma *dma;
640 iova = user_pfn[i] << PAGE_SHIFT;
641 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
644 vfio_unpin_page_external(dma, iova, do_accounting);
648 mutex_unlock(&iommu->lock);
649 return i > npage ? npage : (i > 0 ? i : -EINVAL);
652 static long vfio_sync_unpin(struct vfio_dma *dma, struct vfio_domain *domain,
653 struct list_head *regions)
656 struct vfio_regions *entry, *next;
658 iommu_tlb_sync(domain->domain);
660 list_for_each_entry_safe(entry, next, regions, list) {
661 unlocked += vfio_unpin_pages_remote(dma,
663 entry->phys >> PAGE_SHIFT,
664 entry->len >> PAGE_SHIFT,
666 list_del(&entry->list);
676 * Generally, VFIO needs to unpin remote pages after each IOTLB flush.
677 * Therefore, when using IOTLB flush sync interface, VFIO need to keep track
678 * of these regions (currently using a list).
680 * This value specifies maximum number of regions for each IOTLB flush sync.
682 #define VFIO_IOMMU_TLB_SYNC_MAX 512
684 static size_t unmap_unpin_fast(struct vfio_domain *domain,
685 struct vfio_dma *dma, dma_addr_t *iova,
686 size_t len, phys_addr_t phys, long *unlocked,
687 struct list_head *unmapped_list,
691 struct vfio_regions *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
694 unmapped = iommu_unmap_fast(domain->domain, *iova, len);
699 iommu_tlb_range_add(domain->domain, *iova, unmapped);
702 entry->len = unmapped;
703 list_add_tail(&entry->list, unmapped_list);
711 * Sync if the number of fast-unmap regions hits the limit
712 * or in case of errors.
714 if (*unmapped_cnt >= VFIO_IOMMU_TLB_SYNC_MAX || !unmapped) {
715 *unlocked += vfio_sync_unpin(dma, domain,
723 static size_t unmap_unpin_slow(struct vfio_domain *domain,
724 struct vfio_dma *dma, dma_addr_t *iova,
725 size_t len, phys_addr_t phys,
728 size_t unmapped = iommu_unmap(domain->domain, *iova, len);
731 *unlocked += vfio_unpin_pages_remote(dma, *iova,
733 unmapped >> PAGE_SHIFT,
741 static long vfio_unmap_unpin(struct vfio_iommu *iommu, struct vfio_dma *dma,
744 dma_addr_t iova = dma->iova, end = dma->iova + dma->size;
745 struct vfio_domain *domain, *d;
746 LIST_HEAD(unmapped_region_list);
747 int unmapped_region_cnt = 0;
753 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
757 * We use the IOMMU to track the physical addresses, otherwise we'd
758 * need a much more complicated tracking system. Unfortunately that
759 * means we need to use one of the iommu domains to figure out the
760 * pfns to unpin. The rest need to be unmapped in advance so we have
761 * no iommu translations remaining when the pages are unpinned.
763 domain = d = list_first_entry(&iommu->domain_list,
764 struct vfio_domain, next);
766 list_for_each_entry_continue(d, &iommu->domain_list, next) {
767 iommu_unmap(d->domain, dma->iova, dma->size);
772 size_t unmapped, len;
773 phys_addr_t phys, next;
775 phys = iommu_iova_to_phys(domain->domain, iova);
776 if (WARN_ON(!phys)) {
782 * To optimize for fewer iommu_unmap() calls, each of which
783 * may require hardware cache flushing, try to find the
784 * largest contiguous physical memory chunk to unmap.
786 for (len = PAGE_SIZE;
787 !domain->fgsp && iova + len < end; len += PAGE_SIZE) {
788 next = iommu_iova_to_phys(domain->domain, iova + len);
789 if (next != phys + len)
794 * First, try to use fast unmap/unpin. In case of failure,
795 * switch to slow unmap/unpin path.
797 unmapped = unmap_unpin_fast(domain, dma, &iova, len, phys,
798 &unlocked, &unmapped_region_list,
799 &unmapped_region_cnt);
801 unmapped = unmap_unpin_slow(domain, dma, &iova, len,
803 if (WARN_ON(!unmapped))
808 dma->iommu_mapped = false;
810 if (unmapped_region_cnt)
811 unlocked += vfio_sync_unpin(dma, domain, &unmapped_region_list);
814 vfio_lock_acct(dma, -unlocked, true);
820 static void vfio_remove_dma(struct vfio_iommu *iommu, struct vfio_dma *dma)
822 vfio_unmap_unpin(iommu, dma, true);
823 vfio_unlink_dma(iommu, dma);
824 put_task_struct(dma->task);
829 static unsigned long vfio_pgsize_bitmap(struct vfio_iommu *iommu)
831 struct vfio_domain *domain;
832 unsigned long bitmap = ULONG_MAX;
834 mutex_lock(&iommu->lock);
835 list_for_each_entry(domain, &iommu->domain_list, next)
836 bitmap &= domain->domain->pgsize_bitmap;
837 mutex_unlock(&iommu->lock);
840 * In case the IOMMU supports page sizes smaller than PAGE_SIZE
841 * we pretend PAGE_SIZE is supported and hide sub-PAGE_SIZE sizes.
842 * That way the user will be able to map/unmap buffers whose size/
843 * start address is aligned with PAGE_SIZE. Pinning code uses that
844 * granularity while iommu driver can use the sub-PAGE_SIZE size
847 if (bitmap & ~PAGE_MASK) {
855 static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
856 struct vfio_iommu_type1_dma_unmap *unmap)
859 struct vfio_dma *dma, *dma_last = NULL;
861 int ret = 0, retries = 0;
863 mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1;
865 if (unmap->iova & mask)
867 if (!unmap->size || unmap->size & mask)
869 if (unmap->iova + unmap->size - 1 < unmap->iova ||
870 unmap->size > SIZE_MAX)
873 WARN_ON(mask & PAGE_MASK);
875 mutex_lock(&iommu->lock);
878 * vfio-iommu-type1 (v1) - User mappings were coalesced together to
879 * avoid tracking individual mappings. This means that the granularity
880 * of the original mapping was lost and the user was allowed to attempt
881 * to unmap any range. Depending on the contiguousness of physical
882 * memory and page sizes supported by the IOMMU, arbitrary unmaps may
883 * or may not have worked. We only guaranteed unmap granularity
884 * matching the original mapping; even though it was untracked here,
885 * the original mappings are reflected in IOMMU mappings. This
886 * resulted in a couple unusual behaviors. First, if a range is not
887 * able to be unmapped, ex. a set of 4k pages that was mapped as a
888 * 2M hugepage into the IOMMU, the unmap ioctl returns success but with
889 * a zero sized unmap. Also, if an unmap request overlaps the first
890 * address of a hugepage, the IOMMU will unmap the entire hugepage.
891 * This also returns success and the returned unmap size reflects the
892 * actual size unmapped.
894 * We attempt to maintain compatibility with this "v1" interface, but
895 * we take control out of the hands of the IOMMU. Therefore, an unmap
896 * request offset from the beginning of the original mapping will
897 * return success with zero sized unmap. And an unmap request covering
898 * the first iova of mapping will unmap the entire range.
900 * The v2 version of this interface intends to be more deterministic.
901 * Unmap requests must fully cover previous mappings. Multiple
902 * mappings may still be unmaped by specifying large ranges, but there
903 * must not be any previous mappings bisected by the range. An error
904 * will be returned if these conditions are not met. The v2 interface
905 * will only return success and a size of zero if there were no
906 * mappings within the range.
909 dma = vfio_find_dma(iommu, unmap->iova, 1);
910 if (dma && dma->iova != unmap->iova) {
914 dma = vfio_find_dma(iommu, unmap->iova + unmap->size - 1, 0);
915 if (dma && dma->iova + dma->size != unmap->iova + unmap->size) {
921 while ((dma = vfio_find_dma(iommu, unmap->iova, unmap->size))) {
922 if (!iommu->v2 && unmap->iova > dma->iova)
925 * Task with same address space who mapped this iova range is
926 * allowed to unmap the iova range.
928 if (dma->task->mm != current->mm)
931 if (!RB_EMPTY_ROOT(&dma->pfn_list)) {
932 struct vfio_iommu_type1_dma_unmap nb_unmap;
934 if (dma_last == dma) {
935 BUG_ON(++retries > 10);
941 nb_unmap.iova = dma->iova;
942 nb_unmap.size = dma->size;
945 * Notify anyone (mdev vendor drivers) to invalidate and
946 * unmap iovas within the range we're about to unmap.
947 * Vendor drivers MUST unpin pages in response to an
950 mutex_unlock(&iommu->lock);
951 blocking_notifier_call_chain(&iommu->notifier,
952 VFIO_IOMMU_NOTIFY_DMA_UNMAP,
956 unmapped += dma->size;
957 vfio_remove_dma(iommu, dma);
961 mutex_unlock(&iommu->lock);
963 /* Report how much was unmapped */
964 unmap->size = unmapped;
969 static int vfio_iommu_map(struct vfio_iommu *iommu, dma_addr_t iova,
970 unsigned long pfn, long npage, int prot)
972 struct vfio_domain *d;
975 list_for_each_entry(d, &iommu->domain_list, next) {
976 ret = iommu_map(d->domain, iova, (phys_addr_t)pfn << PAGE_SHIFT,
977 npage << PAGE_SHIFT, prot | d->prot);
987 list_for_each_entry_continue_reverse(d, &iommu->domain_list, next)
988 iommu_unmap(d->domain, iova, npage << PAGE_SHIFT);
993 static int vfio_pin_map_dma(struct vfio_iommu *iommu, struct vfio_dma *dma,
996 dma_addr_t iova = dma->iova;
997 unsigned long vaddr = dma->vaddr;
998 size_t size = map_size;
1000 unsigned long pfn, limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1004 /* Pin a contiguous chunk of memory */
1005 npage = vfio_pin_pages_remote(dma, vaddr + dma->size,
1006 size >> PAGE_SHIFT, &pfn, limit);
1014 ret = vfio_iommu_map(iommu, iova + dma->size, pfn, npage,
1017 vfio_unpin_pages_remote(dma, iova + dma->size, pfn,
1022 size -= npage << PAGE_SHIFT;
1023 dma->size += npage << PAGE_SHIFT;
1026 dma->iommu_mapped = true;
1029 vfio_remove_dma(iommu, dma);
1034 static int vfio_dma_do_map(struct vfio_iommu *iommu,
1035 struct vfio_iommu_type1_dma_map *map)
1037 dma_addr_t iova = map->iova;
1038 unsigned long vaddr = map->vaddr;
1039 size_t size = map->size;
1040 int ret = 0, prot = 0;
1042 struct vfio_dma *dma;
1044 /* Verify that none of our __u64 fields overflow */
1045 if (map->size != size || map->vaddr != vaddr || map->iova != iova)
1048 mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1;
1050 WARN_ON(mask & PAGE_MASK);
1052 /* READ/WRITE from device perspective */
1053 if (map->flags & VFIO_DMA_MAP_FLAG_WRITE)
1054 prot |= IOMMU_WRITE;
1055 if (map->flags & VFIO_DMA_MAP_FLAG_READ)
1058 if (!prot || !size || (size | iova | vaddr) & mask)
1061 /* Don't allow IOVA or virtual address wrap */
1062 if (iova + size - 1 < iova || vaddr + size - 1 < vaddr)
1065 mutex_lock(&iommu->lock);
1067 if (vfio_find_dma(iommu, iova, size)) {
1072 if (!iommu->dma_avail) {
1077 dma = kzalloc(sizeof(*dma), GFP_KERNEL);
1089 * We need to be able to both add to a task's locked memory and test
1090 * against the locked memory limit and we need to be able to do both
1091 * outside of this call path as pinning can be asynchronous via the
1092 * external interfaces for mdev devices. RLIMIT_MEMLOCK requires a
1093 * task_struct and VM locked pages requires an mm_struct, however
1094 * holding an indefinite mm reference is not recommended, therefore we
1095 * only hold a reference to a task. We could hold a reference to
1096 * current, however QEMU uses this call path through vCPU threads,
1097 * which can be killed resulting in a NULL mm and failure in the unmap
1098 * path when called via a different thread. Avoid this problem by
1099 * using the group_leader as threads within the same group require
1100 * both CLONE_THREAD and CLONE_VM and will therefore use the same
1103 * Previously we also used the task for testing CAP_IPC_LOCK at the
1104 * time of pinning and accounting, however has_capability() makes use
1105 * of real_cred, a copy-on-write field, so we can't guarantee that it
1106 * matches group_leader, or in fact that it might not change by the
1107 * time it's evaluated. If a process were to call MAP_DMA with
1108 * CAP_IPC_LOCK but later drop it, it doesn't make sense that they
1109 * possibly see different results for an iommu_mapped vfio_dma vs
1110 * externally mapped. Therefore track CAP_IPC_LOCK in vfio_dma at the
1111 * time of calling MAP_DMA.
1113 get_task_struct(current->group_leader);
1114 dma->task = current->group_leader;
1115 dma->lock_cap = capable(CAP_IPC_LOCK);
1117 dma->pfn_list = RB_ROOT;
1119 /* Insert zero-sized and grow as we map chunks of it */
1120 vfio_link_dma(iommu, dma);
1122 /* Don't pin and map if container doesn't contain IOMMU capable domain*/
1123 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
1126 ret = vfio_pin_map_dma(iommu, dma, size);
1129 mutex_unlock(&iommu->lock);
1133 static int vfio_bus_type(struct device *dev, void *data)
1135 struct bus_type **bus = data;
1137 if (*bus && *bus != dev->bus)
1145 static int vfio_iommu_replay(struct vfio_iommu *iommu,
1146 struct vfio_domain *domain)
1148 struct vfio_domain *d;
1150 unsigned long limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1153 /* Arbitrarily pick the first domain in the list for lookups */
1154 d = list_first_entry(&iommu->domain_list, struct vfio_domain, next);
1155 n = rb_first(&iommu->dma_list);
1157 for (; n; n = rb_next(n)) {
1158 struct vfio_dma *dma;
1161 dma = rb_entry(n, struct vfio_dma, node);
1164 while (iova < dma->iova + dma->size) {
1168 if (dma->iommu_mapped) {
1172 phys = iommu_iova_to_phys(d->domain, iova);
1174 if (WARN_ON(!phys)) {
1182 while (i < dma->iova + dma->size &&
1183 p == iommu_iova_to_phys(d->domain, i)) {
1190 unsigned long vaddr = dma->vaddr +
1192 size_t n = dma->iova + dma->size - iova;
1195 npage = vfio_pin_pages_remote(dma, vaddr,
1204 phys = pfn << PAGE_SHIFT;
1205 size = npage << PAGE_SHIFT;
1208 ret = iommu_map(domain->domain, iova, phys,
1209 size, dma->prot | domain->prot);
1215 dma->iommu_mapped = true;
1221 * We change our unmap behavior slightly depending on whether the IOMMU
1222 * supports fine-grained superpages. IOMMUs like AMD-Vi will use a superpage
1223 * for practically any contiguous power-of-two mapping we give it. This means
1224 * we don't need to look for contiguous chunks ourselves to make unmapping
1225 * more efficient. On IOMMUs with coarse-grained super pages, like Intel VT-d
1226 * with discrete 2M/1G/512G/1T superpages, identifying contiguous chunks
1227 * significantly boosts non-hugetlbfs mappings and doesn't seem to hurt when
1228 * hugetlbfs is in use.
1230 static void vfio_test_domain_fgsp(struct vfio_domain *domain)
1233 int ret, order = get_order(PAGE_SIZE * 2);
1235 pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order);
1239 ret = iommu_map(domain->domain, 0, page_to_phys(pages), PAGE_SIZE * 2,
1240 IOMMU_READ | IOMMU_WRITE | domain->prot);
1242 size_t unmapped = iommu_unmap(domain->domain, 0, PAGE_SIZE);
1244 if (unmapped == PAGE_SIZE)
1245 iommu_unmap(domain->domain, PAGE_SIZE, PAGE_SIZE);
1247 domain->fgsp = true;
1250 __free_pages(pages, order);
1253 static struct vfio_group *find_iommu_group(struct vfio_domain *domain,
1254 struct iommu_group *iommu_group)
1256 struct vfio_group *g;
1258 list_for_each_entry(g, &domain->group_list, next) {
1259 if (g->iommu_group == iommu_group)
1266 static bool vfio_iommu_has_sw_msi(struct iommu_group *group, phys_addr_t *base)
1268 struct list_head group_resv_regions;
1269 struct iommu_resv_region *region, *next;
1272 INIT_LIST_HEAD(&group_resv_regions);
1273 iommu_get_group_resv_regions(group, &group_resv_regions);
1274 list_for_each_entry(region, &group_resv_regions, list) {
1276 * The presence of any 'real' MSI regions should take
1277 * precedence over the software-managed one if the
1278 * IOMMU driver happens to advertise both types.
1280 if (region->type == IOMMU_RESV_MSI) {
1285 if (region->type == IOMMU_RESV_SW_MSI) {
1286 *base = region->start;
1290 list_for_each_entry_safe(region, next, &group_resv_regions, list)
1295 static struct device *vfio_mdev_get_iommu_device(struct device *dev)
1297 struct device *(*fn)(struct device *dev);
1298 struct device *iommu_device;
1300 fn = symbol_get(mdev_get_iommu_device);
1302 iommu_device = fn(dev);
1303 symbol_put(mdev_get_iommu_device);
1305 return iommu_device;
1311 static int vfio_mdev_attach_domain(struct device *dev, void *data)
1313 struct iommu_domain *domain = data;
1314 struct device *iommu_device;
1316 iommu_device = vfio_mdev_get_iommu_device(dev);
1318 if (iommu_dev_feature_enabled(iommu_device, IOMMU_DEV_FEAT_AUX))
1319 return iommu_aux_attach_device(domain, iommu_device);
1321 return iommu_attach_device(domain, iommu_device);
1327 static int vfio_mdev_detach_domain(struct device *dev, void *data)
1329 struct iommu_domain *domain = data;
1330 struct device *iommu_device;
1332 iommu_device = vfio_mdev_get_iommu_device(dev);
1334 if (iommu_dev_feature_enabled(iommu_device, IOMMU_DEV_FEAT_AUX))
1335 iommu_aux_detach_device(domain, iommu_device);
1337 iommu_detach_device(domain, iommu_device);
1343 static int vfio_iommu_attach_group(struct vfio_domain *domain,
1344 struct vfio_group *group)
1346 if (group->mdev_group)
1347 return iommu_group_for_each_dev(group->iommu_group,
1349 vfio_mdev_attach_domain);
1351 return iommu_attach_group(domain->domain, group->iommu_group);
1354 static void vfio_iommu_detach_group(struct vfio_domain *domain,
1355 struct vfio_group *group)
1357 if (group->mdev_group)
1358 iommu_group_for_each_dev(group->iommu_group, domain->domain,
1359 vfio_mdev_detach_domain);
1361 iommu_detach_group(domain->domain, group->iommu_group);
1364 static bool vfio_bus_is_mdev(struct bus_type *bus)
1366 struct bus_type *mdev_bus;
1369 mdev_bus = symbol_get(mdev_bus_type);
1371 ret = (bus == mdev_bus);
1372 symbol_put(mdev_bus_type);
1378 static int vfio_mdev_iommu_device(struct device *dev, void *data)
1380 struct device **old = data, *new;
1382 new = vfio_mdev_get_iommu_device(dev);
1383 if (!new || (*old && *old != new))
1391 static int vfio_iommu_type1_attach_group(void *iommu_data,
1392 struct iommu_group *iommu_group)
1394 struct vfio_iommu *iommu = iommu_data;
1395 struct vfio_group *group;
1396 struct vfio_domain *domain, *d;
1397 struct bus_type *bus = NULL;
1399 bool resv_msi, msi_remap;
1400 phys_addr_t resv_msi_base;
1402 mutex_lock(&iommu->lock);
1404 list_for_each_entry(d, &iommu->domain_list, next) {
1405 if (find_iommu_group(d, iommu_group)) {
1406 mutex_unlock(&iommu->lock);
1411 if (iommu->external_domain) {
1412 if (find_iommu_group(iommu->external_domain, iommu_group)) {
1413 mutex_unlock(&iommu->lock);
1418 group = kzalloc(sizeof(*group), GFP_KERNEL);
1419 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
1420 if (!group || !domain) {
1425 group->iommu_group = iommu_group;
1427 /* Determine bus_type in order to allocate a domain */
1428 ret = iommu_group_for_each_dev(iommu_group, &bus, vfio_bus_type);
1432 if (vfio_bus_is_mdev(bus)) {
1433 struct device *iommu_device = NULL;
1435 group->mdev_group = true;
1437 /* Determine the isolation type */
1438 ret = iommu_group_for_each_dev(iommu_group, &iommu_device,
1439 vfio_mdev_iommu_device);
1440 if (ret || !iommu_device) {
1441 if (!iommu->external_domain) {
1442 INIT_LIST_HEAD(&domain->group_list);
1443 iommu->external_domain = domain;
1448 list_add(&group->next,
1449 &iommu->external_domain->group_list);
1450 mutex_unlock(&iommu->lock);
1455 bus = iommu_device->bus;
1458 domain->domain = iommu_domain_alloc(bus);
1459 if (!domain->domain) {
1464 if (iommu->nesting) {
1467 ret = iommu_domain_set_attr(domain->domain, DOMAIN_ATTR_NESTING,
1473 ret = vfio_iommu_attach_group(domain, group);
1477 resv_msi = vfio_iommu_has_sw_msi(iommu_group, &resv_msi_base);
1479 INIT_LIST_HEAD(&domain->group_list);
1480 list_add(&group->next, &domain->group_list);
1482 msi_remap = irq_domain_check_msi_remap() ||
1483 iommu_capable(bus, IOMMU_CAP_INTR_REMAP);
1485 if (!allow_unsafe_interrupts && !msi_remap) {
1486 pr_warn("%s: No interrupt remapping support. Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n",
1492 if (iommu_capable(bus, IOMMU_CAP_CACHE_COHERENCY))
1493 domain->prot |= IOMMU_CACHE;
1496 * Try to match an existing compatible domain. We don't want to
1497 * preclude an IOMMU driver supporting multiple bus_types and being
1498 * able to include different bus_types in the same IOMMU domain, so
1499 * we test whether the domains use the same iommu_ops rather than
1500 * testing if they're on the same bus_type.
1502 list_for_each_entry(d, &iommu->domain_list, next) {
1503 if (d->domain->ops == domain->domain->ops &&
1504 d->prot == domain->prot) {
1505 vfio_iommu_detach_group(domain, group);
1506 if (!vfio_iommu_attach_group(d, group)) {
1507 list_add(&group->next, &d->group_list);
1508 iommu_domain_free(domain->domain);
1510 mutex_unlock(&iommu->lock);
1514 ret = vfio_iommu_attach_group(domain, group);
1520 vfio_test_domain_fgsp(domain);
1522 /* replay mappings on new domains */
1523 ret = vfio_iommu_replay(iommu, domain);
1528 ret = iommu_get_msi_cookie(domain->domain, resv_msi_base);
1533 list_add(&domain->next, &iommu->domain_list);
1535 mutex_unlock(&iommu->lock);
1540 vfio_iommu_detach_group(domain, group);
1542 iommu_domain_free(domain->domain);
1546 mutex_unlock(&iommu->lock);
1550 static void vfio_iommu_unmap_unpin_all(struct vfio_iommu *iommu)
1552 struct rb_node *node;
1554 while ((node = rb_first(&iommu->dma_list)))
1555 vfio_remove_dma(iommu, rb_entry(node, struct vfio_dma, node));
1558 static void vfio_iommu_unmap_unpin_reaccount(struct vfio_iommu *iommu)
1560 struct rb_node *n, *p;
1562 n = rb_first(&iommu->dma_list);
1563 for (; n; n = rb_next(n)) {
1564 struct vfio_dma *dma;
1565 long locked = 0, unlocked = 0;
1567 dma = rb_entry(n, struct vfio_dma, node);
1568 unlocked += vfio_unmap_unpin(iommu, dma, false);
1569 p = rb_first(&dma->pfn_list);
1570 for (; p; p = rb_next(p)) {
1571 struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn,
1574 if (!is_invalid_reserved_pfn(vpfn->pfn))
1577 vfio_lock_acct(dma, locked - unlocked, true);
1581 static void vfio_sanity_check_pfn_list(struct vfio_iommu *iommu)
1585 n = rb_first(&iommu->dma_list);
1586 for (; n; n = rb_next(n)) {
1587 struct vfio_dma *dma;
1589 dma = rb_entry(n, struct vfio_dma, node);
1591 if (WARN_ON(!RB_EMPTY_ROOT(&dma->pfn_list)))
1594 /* mdev vendor driver must unregister notifier */
1595 WARN_ON(iommu->notifier.head);
1598 static void vfio_iommu_type1_detach_group(void *iommu_data,
1599 struct iommu_group *iommu_group)
1601 struct vfio_iommu *iommu = iommu_data;
1602 struct vfio_domain *domain;
1603 struct vfio_group *group;
1605 mutex_lock(&iommu->lock);
1607 if (iommu->external_domain) {
1608 group = find_iommu_group(iommu->external_domain, iommu_group);
1610 list_del(&group->next);
1613 if (list_empty(&iommu->external_domain->group_list)) {
1614 vfio_sanity_check_pfn_list(iommu);
1616 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
1617 vfio_iommu_unmap_unpin_all(iommu);
1619 kfree(iommu->external_domain);
1620 iommu->external_domain = NULL;
1622 goto detach_group_done;
1626 list_for_each_entry(domain, &iommu->domain_list, next) {
1627 group = find_iommu_group(domain, iommu_group);
1631 vfio_iommu_detach_group(domain, group);
1632 list_del(&group->next);
1635 * Group ownership provides privilege, if the group list is
1636 * empty, the domain goes away. If it's the last domain with
1637 * iommu and external domain doesn't exist, then all the
1638 * mappings go away too. If it's the last domain with iommu and
1639 * external domain exist, update accounting
1641 if (list_empty(&domain->group_list)) {
1642 if (list_is_singular(&iommu->domain_list)) {
1643 if (!iommu->external_domain)
1644 vfio_iommu_unmap_unpin_all(iommu);
1646 vfio_iommu_unmap_unpin_reaccount(iommu);
1648 iommu_domain_free(domain->domain);
1649 list_del(&domain->next);
1656 mutex_unlock(&iommu->lock);
1659 static void *vfio_iommu_type1_open(unsigned long arg)
1661 struct vfio_iommu *iommu;
1663 iommu = kzalloc(sizeof(*iommu), GFP_KERNEL);
1665 return ERR_PTR(-ENOMEM);
1668 case VFIO_TYPE1_IOMMU:
1670 case VFIO_TYPE1_NESTING_IOMMU:
1671 iommu->nesting = true;
1673 case VFIO_TYPE1v2_IOMMU:
1678 return ERR_PTR(-EINVAL);
1681 INIT_LIST_HEAD(&iommu->domain_list);
1682 iommu->dma_list = RB_ROOT;
1683 iommu->dma_avail = dma_entry_limit;
1684 mutex_init(&iommu->lock);
1685 BLOCKING_INIT_NOTIFIER_HEAD(&iommu->notifier);
1690 static void vfio_release_domain(struct vfio_domain *domain, bool external)
1692 struct vfio_group *group, *group_tmp;
1694 list_for_each_entry_safe(group, group_tmp,
1695 &domain->group_list, next) {
1697 vfio_iommu_detach_group(domain, group);
1698 list_del(&group->next);
1703 iommu_domain_free(domain->domain);
1706 static void vfio_iommu_type1_release(void *iommu_data)
1708 struct vfio_iommu *iommu = iommu_data;
1709 struct vfio_domain *domain, *domain_tmp;
1711 if (iommu->external_domain) {
1712 vfio_release_domain(iommu->external_domain, true);
1713 vfio_sanity_check_pfn_list(iommu);
1714 kfree(iommu->external_domain);
1717 vfio_iommu_unmap_unpin_all(iommu);
1719 list_for_each_entry_safe(domain, domain_tmp,
1720 &iommu->domain_list, next) {
1721 vfio_release_domain(domain, false);
1722 list_del(&domain->next);
1728 static int vfio_domains_have_iommu_cache(struct vfio_iommu *iommu)
1730 struct vfio_domain *domain;
1733 mutex_lock(&iommu->lock);
1734 list_for_each_entry(domain, &iommu->domain_list, next) {
1735 if (!(domain->prot & IOMMU_CACHE)) {
1740 mutex_unlock(&iommu->lock);
1745 static long vfio_iommu_type1_ioctl(void *iommu_data,
1746 unsigned int cmd, unsigned long arg)
1748 struct vfio_iommu *iommu = iommu_data;
1749 unsigned long minsz;
1751 if (cmd == VFIO_CHECK_EXTENSION) {
1753 case VFIO_TYPE1_IOMMU:
1754 case VFIO_TYPE1v2_IOMMU:
1755 case VFIO_TYPE1_NESTING_IOMMU:
1757 case VFIO_DMA_CC_IOMMU:
1760 return vfio_domains_have_iommu_cache(iommu);
1764 } else if (cmd == VFIO_IOMMU_GET_INFO) {
1765 struct vfio_iommu_type1_info info;
1767 minsz = offsetofend(struct vfio_iommu_type1_info, iova_pgsizes);
1769 if (copy_from_user(&info, (void __user *)arg, minsz))
1772 if (info.argsz < minsz)
1775 info.flags = VFIO_IOMMU_INFO_PGSIZES;
1777 info.iova_pgsizes = vfio_pgsize_bitmap(iommu);
1779 return copy_to_user((void __user *)arg, &info, minsz) ?
1782 } else if (cmd == VFIO_IOMMU_MAP_DMA) {
1783 struct vfio_iommu_type1_dma_map map;
1784 uint32_t mask = VFIO_DMA_MAP_FLAG_READ |
1785 VFIO_DMA_MAP_FLAG_WRITE;
1787 minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
1789 if (copy_from_user(&map, (void __user *)arg, minsz))
1792 if (map.argsz < minsz || map.flags & ~mask)
1795 return vfio_dma_do_map(iommu, &map);
1797 } else if (cmd == VFIO_IOMMU_UNMAP_DMA) {
1798 struct vfio_iommu_type1_dma_unmap unmap;
1801 minsz = offsetofend(struct vfio_iommu_type1_dma_unmap, size);
1803 if (copy_from_user(&unmap, (void __user *)arg, minsz))
1806 if (unmap.argsz < minsz || unmap.flags)
1809 ret = vfio_dma_do_unmap(iommu, &unmap);
1813 return copy_to_user((void __user *)arg, &unmap, minsz) ?
1820 static int vfio_iommu_type1_register_notifier(void *iommu_data,
1821 unsigned long *events,
1822 struct notifier_block *nb)
1824 struct vfio_iommu *iommu = iommu_data;
1826 /* clear known events */
1827 *events &= ~VFIO_IOMMU_NOTIFY_DMA_UNMAP;
1829 /* refuse to register if still events remaining */
1833 return blocking_notifier_chain_register(&iommu->notifier, nb);
1836 static int vfio_iommu_type1_unregister_notifier(void *iommu_data,
1837 struct notifier_block *nb)
1839 struct vfio_iommu *iommu = iommu_data;
1841 return blocking_notifier_chain_unregister(&iommu->notifier, nb);
1844 static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = {
1845 .name = "vfio-iommu-type1",
1846 .owner = THIS_MODULE,
1847 .open = vfio_iommu_type1_open,
1848 .release = vfio_iommu_type1_release,
1849 .ioctl = vfio_iommu_type1_ioctl,
1850 .attach_group = vfio_iommu_type1_attach_group,
1851 .detach_group = vfio_iommu_type1_detach_group,
1852 .pin_pages = vfio_iommu_type1_pin_pages,
1853 .unpin_pages = vfio_iommu_type1_unpin_pages,
1854 .register_notifier = vfio_iommu_type1_register_notifier,
1855 .unregister_notifier = vfio_iommu_type1_unregister_notifier,
1858 static int __init vfio_iommu_type1_init(void)
1860 return vfio_register_iommu_driver(&vfio_iommu_driver_ops_type1);
1863 static void __exit vfio_iommu_type1_cleanup(void)
1865 vfio_unregister_iommu_driver(&vfio_iommu_driver_ops_type1);
1868 module_init(vfio_iommu_type1_init);
1869 module_exit(vfio_iommu_type1_cleanup);
1871 MODULE_VERSION(DRIVER_VERSION);
1872 MODULE_LICENSE("GPL v2");
1873 MODULE_AUTHOR(DRIVER_AUTHOR);
1874 MODULE_DESCRIPTION(DRIVER_DESC);