2 * VFIO: IOMMU DMA mapping support for Type1 IOMMU
4 * Copyright (C) 2012 Red Hat, Inc. All rights reserved.
5 * Author: Alex Williamson <alex.williamson@redhat.com>
7 * This program is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License version 2 as
9 * published by the Free Software Foundation.
11 * Derived from original vfio:
12 * Copyright 2010 Cisco Systems, Inc. All rights reserved.
13 * Author: Tom Lyon, pugs@cisco.com
15 * We arbitrarily define a Type1 IOMMU as one matching the below code.
16 * It could be called the x86 IOMMU as it's designed for AMD-Vi & Intel
17 * VT-d, but that makes it harder to re-use as theoretically anyone
18 * implementing a similar IOMMU could make use of this. We expect the
19 * IOMMU to support the IOMMU API and have few to no restrictions around
20 * the IOVA range that can be mapped. The Type1 IOMMU is currently
21 * optimized for relatively static mappings of a userspace process with
22 * userpsace pages pinned into memory. We also assume devices and IOMMU
23 * domains are PCI based as the IOMMU API is still centered around a
24 * device/bus interface rather than a group interface.
27 #include <linux/compat.h>
28 #include <linux/device.h>
30 #include <linux/iommu.h>
31 #include <linux/module.h>
33 #include <linux/rbtree.h>
34 #include <linux/sched/signal.h>
35 #include <linux/sched/mm.h>
36 #include <linux/slab.h>
37 #include <linux/uaccess.h>
38 #include <linux/vfio.h>
39 #include <linux/workqueue.h>
40 #include <linux/mdev.h>
41 #include <linux/notifier.h>
42 #include <linux/dma-iommu.h>
43 #include <linux/irqdomain.h>
45 #define DRIVER_VERSION "0.2"
46 #define DRIVER_AUTHOR "Alex Williamson <alex.williamson@redhat.com>"
47 #define DRIVER_DESC "Type1 IOMMU driver for VFIO"
49 static bool allow_unsafe_interrupts
;
50 module_param_named(allow_unsafe_interrupts
,
51 allow_unsafe_interrupts
, bool, S_IRUGO
| S_IWUSR
);
52 MODULE_PARM_DESC(allow_unsafe_interrupts
,
53 "Enable VFIO IOMMU support for on platforms without interrupt remapping support.");
55 static bool disable_hugepages
;
56 module_param_named(disable_hugepages
,
57 disable_hugepages
, bool, S_IRUGO
| S_IWUSR
);
58 MODULE_PARM_DESC(disable_hugepages
,
59 "Disable VFIO IOMMU support for IOMMU hugepages.");
61 static unsigned int dma_entry_limit __read_mostly
= U16_MAX
;
62 module_param_named(dma_entry_limit
, dma_entry_limit
, uint
, 0644);
63 MODULE_PARM_DESC(dma_entry_limit
,
64 "Maximum number of user DMA mappings per container (65535).");
67 struct list_head domain_list
;
68 struct vfio_domain
*external_domain
; /* domain for external user */
70 struct rb_root dma_list
;
71 struct blocking_notifier_head notifier
;
72 unsigned int dma_avail
;
78 struct iommu_domain
*domain
;
79 struct list_head next
;
80 struct list_head group_list
;
81 int prot
; /* IOMMU_CACHE */
82 bool fgsp
; /* Fine-grained super pages */
87 dma_addr_t iova
; /* Device address */
88 unsigned long vaddr
; /* Process virtual addr */
89 size_t size
; /* Map size (bytes) */
90 int prot
; /* IOMMU_READ/WRITE */
92 bool lock_cap
; /* capable(CAP_IPC_LOCK) */
93 struct task_struct
*task
;
94 struct rb_root pfn_list
; /* Ex-user pinned pfn list */
98 struct iommu_group
*iommu_group
;
99 struct list_head next
;
103 * Guest RAM pinning working set or DMA target
107 dma_addr_t iova
; /* Device address */
108 unsigned long pfn
; /* Host pfn */
112 #define IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu) \
113 (!list_empty(&iommu->domain_list))
115 static int put_pfn(unsigned long pfn
, int prot
);
118 * This code handles mapping and unmapping of user data buffers
119 * into DMA'ble space using the IOMMU
122 static struct vfio_dma
*vfio_find_dma(struct vfio_iommu
*iommu
,
123 dma_addr_t start
, size_t size
)
125 struct rb_node
*node
= iommu
->dma_list
.rb_node
;
128 struct vfio_dma
*dma
= rb_entry(node
, struct vfio_dma
, node
);
130 if (start
+ size
<= dma
->iova
)
131 node
= node
->rb_left
;
132 else if (start
>= dma
->iova
+ dma
->size
)
133 node
= node
->rb_right
;
141 static void vfio_link_dma(struct vfio_iommu
*iommu
, struct vfio_dma
*new)
143 struct rb_node
**link
= &iommu
->dma_list
.rb_node
, *parent
= NULL
;
144 struct vfio_dma
*dma
;
148 dma
= rb_entry(parent
, struct vfio_dma
, node
);
150 if (new->iova
+ new->size
<= dma
->iova
)
151 link
= &(*link
)->rb_left
;
153 link
= &(*link
)->rb_right
;
156 rb_link_node(&new->node
, parent
, link
);
157 rb_insert_color(&new->node
, &iommu
->dma_list
);
160 static void vfio_unlink_dma(struct vfio_iommu
*iommu
, struct vfio_dma
*old
)
162 rb_erase(&old
->node
, &iommu
->dma_list
);
166 * Helper Functions for host iova-pfn list
168 static struct vfio_pfn
*vfio_find_vpfn(struct vfio_dma
*dma
, dma_addr_t iova
)
170 struct vfio_pfn
*vpfn
;
171 struct rb_node
*node
= dma
->pfn_list
.rb_node
;
174 vpfn
= rb_entry(node
, struct vfio_pfn
, node
);
176 if (iova
< vpfn
->iova
)
177 node
= node
->rb_left
;
178 else if (iova
> vpfn
->iova
)
179 node
= node
->rb_right
;
186 static void vfio_link_pfn(struct vfio_dma
*dma
,
187 struct vfio_pfn
*new)
189 struct rb_node
**link
, *parent
= NULL
;
190 struct vfio_pfn
*vpfn
;
192 link
= &dma
->pfn_list
.rb_node
;
195 vpfn
= rb_entry(parent
, struct vfio_pfn
, node
);
197 if (new->iova
< vpfn
->iova
)
198 link
= &(*link
)->rb_left
;
200 link
= &(*link
)->rb_right
;
203 rb_link_node(&new->node
, parent
, link
);
204 rb_insert_color(&new->node
, &dma
->pfn_list
);
207 static void vfio_unlink_pfn(struct vfio_dma
*dma
, struct vfio_pfn
*old
)
209 rb_erase(&old
->node
, &dma
->pfn_list
);
212 static int vfio_add_to_pfn_list(struct vfio_dma
*dma
, dma_addr_t iova
,
215 struct vfio_pfn
*vpfn
;
217 vpfn
= kzalloc(sizeof(*vpfn
), GFP_KERNEL
);
223 atomic_set(&vpfn
->ref_count
, 1);
224 vfio_link_pfn(dma
, vpfn
);
228 static void vfio_remove_from_pfn_list(struct vfio_dma
*dma
,
229 struct vfio_pfn
*vpfn
)
231 vfio_unlink_pfn(dma
, vpfn
);
235 static struct vfio_pfn
*vfio_iova_get_vfio_pfn(struct vfio_dma
*dma
,
238 struct vfio_pfn
*vpfn
= vfio_find_vpfn(dma
, iova
);
241 atomic_inc(&vpfn
->ref_count
);
245 static int vfio_iova_put_vfio_pfn(struct vfio_dma
*dma
, struct vfio_pfn
*vpfn
)
249 if (atomic_dec_and_test(&vpfn
->ref_count
)) {
250 ret
= put_pfn(vpfn
->pfn
, dma
->prot
);
251 vfio_remove_from_pfn_list(dma
, vpfn
);
256 static int vfio_lock_acct(struct vfio_dma
*dma
, long npage
, bool async
)
258 struct mm_struct
*mm
;
264 mm
= async
? get_task_mm(dma
->task
) : dma
->task
->mm
;
266 return -ESRCH
; /* process exited */
268 ret
= down_write_killable(&mm
->mmap_sem
);
271 if (!dma
->lock_cap
) {
274 limit
= task_rlimit(dma
->task
,
275 RLIMIT_MEMLOCK
) >> PAGE_SHIFT
;
277 if (mm
->locked_vm
+ npage
> limit
)
283 mm
->locked_vm
+= npage
;
285 up_write(&mm
->mmap_sem
);
295 * Some mappings aren't backed by a struct page, for example an mmap'd
296 * MMIO range for our own or another device. These use a different
297 * pfn conversion and shouldn't be tracked as locked pages.
299 static bool is_invalid_reserved_pfn(unsigned long pfn
)
301 if (pfn_valid(pfn
)) {
303 struct page
*tail
= pfn_to_page(pfn
);
304 struct page
*head
= compound_head(tail
);
305 reserved
= !!(PageReserved(head
));
308 * "head" is not a dangling pointer
309 * (compound_head takes care of that)
310 * but the hugepage may have been split
311 * from under us (and we may not hold a
312 * reference count on the head page so it can
313 * be reused before we run PageReferenced), so
314 * we've to check PageTail before returning
321 return PageReserved(tail
);
327 static int put_pfn(unsigned long pfn
, int prot
)
329 if (!is_invalid_reserved_pfn(pfn
)) {
330 struct page
*page
= pfn_to_page(pfn
);
331 if (prot
& IOMMU_WRITE
)
339 static int vaddr_get_pfn(struct mm_struct
*mm
, unsigned long vaddr
,
340 int prot
, unsigned long *pfn
)
342 struct page
*page
[1];
343 struct vm_area_struct
*vma
;
344 struct vm_area_struct
*vmas
[1];
345 unsigned int flags
= 0;
348 if (prot
& IOMMU_WRITE
)
351 down_read(&mm
->mmap_sem
);
352 if (mm
== current
->mm
) {
353 ret
= get_user_pages_longterm(vaddr
, 1, flags
, page
, vmas
);
355 ret
= get_user_pages_remote(NULL
, mm
, vaddr
, 1, flags
, page
,
358 * The lifetime of a vaddr_get_pfn() page pin is
359 * userspace-controlled. In the fs-dax case this could
360 * lead to indefinite stalls in filesystem operations.
361 * Disallow attempts to pin fs-dax pages via this
364 if (ret
> 0 && vma_is_fsdax(vmas
[0])) {
369 up_read(&mm
->mmap_sem
);
372 *pfn
= page_to_pfn(page
[0]);
376 down_read(&mm
->mmap_sem
);
378 vma
= find_vma_intersection(mm
, vaddr
, vaddr
+ 1);
380 if (vma
&& vma
->vm_flags
& VM_PFNMAP
) {
381 *pfn
= ((vaddr
- vma
->vm_start
) >> PAGE_SHIFT
) + vma
->vm_pgoff
;
382 if (is_invalid_reserved_pfn(*pfn
))
386 up_read(&mm
->mmap_sem
);
391 * Attempt to pin pages. We really don't want to track all the pfns and
392 * the iommu can only map chunks of consecutive pfns anyway, so get the
393 * first page and all consecutive pages with the same locking.
395 static long vfio_pin_pages_remote(struct vfio_dma
*dma
, unsigned long vaddr
,
396 long npage
, unsigned long *pfn_base
,
399 unsigned long pfn
= 0;
400 long ret
, pinned
= 0, lock_acct
= 0;
402 dma_addr_t iova
= vaddr
- dma
->vaddr
+ dma
->iova
;
404 /* This code path is only user initiated */
408 ret
= vaddr_get_pfn(current
->mm
, vaddr
, dma
->prot
, pfn_base
);
413 rsvd
= is_invalid_reserved_pfn(*pfn_base
);
416 * Reserved pages aren't counted against the user, externally pinned
417 * pages are already counted against the user.
419 if (!rsvd
&& !vfio_find_vpfn(dma
, iova
)) {
420 if (!dma
->lock_cap
&& current
->mm
->locked_vm
+ 1 > limit
) {
421 put_pfn(*pfn_base
, dma
->prot
);
422 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n", __func__
,
423 limit
<< PAGE_SHIFT
);
429 if (unlikely(disable_hugepages
))
432 /* Lock all the consecutive pages from pfn_base */
433 for (vaddr
+= PAGE_SIZE
, iova
+= PAGE_SIZE
; pinned
< npage
;
434 pinned
++, vaddr
+= PAGE_SIZE
, iova
+= PAGE_SIZE
) {
435 ret
= vaddr_get_pfn(current
->mm
, vaddr
, dma
->prot
, &pfn
);
439 if (pfn
!= *pfn_base
+ pinned
||
440 rsvd
!= is_invalid_reserved_pfn(pfn
)) {
441 put_pfn(pfn
, dma
->prot
);
445 if (!rsvd
&& !vfio_find_vpfn(dma
, iova
)) {
446 if (!dma
->lock_cap
&&
447 current
->mm
->locked_vm
+ lock_acct
+ 1 > limit
) {
448 put_pfn(pfn
, dma
->prot
);
449 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n",
450 __func__
, limit
<< PAGE_SHIFT
);
459 ret
= vfio_lock_acct(dma
, lock_acct
, false);
464 for (pfn
= *pfn_base
; pinned
; pfn
++, pinned
--)
465 put_pfn(pfn
, dma
->prot
);
474 static long vfio_unpin_pages_remote(struct vfio_dma
*dma
, dma_addr_t iova
,
475 unsigned long pfn
, long npage
,
478 long unlocked
= 0, locked
= 0;
481 for (i
= 0; i
< npage
; i
++, iova
+= PAGE_SIZE
) {
482 if (put_pfn(pfn
++, dma
->prot
)) {
484 if (vfio_find_vpfn(dma
, iova
))
490 vfio_lock_acct(dma
, locked
- unlocked
, true);
495 static int vfio_pin_page_external(struct vfio_dma
*dma
, unsigned long vaddr
,
496 unsigned long *pfn_base
, bool do_accounting
)
498 struct mm_struct
*mm
;
501 mm
= get_task_mm(dma
->task
);
505 ret
= vaddr_get_pfn(mm
, vaddr
, dma
->prot
, pfn_base
);
506 if (!ret
&& do_accounting
&& !is_invalid_reserved_pfn(*pfn_base
)) {
507 ret
= vfio_lock_acct(dma
, 1, true);
509 put_pfn(*pfn_base
, dma
->prot
);
511 pr_warn("%s: Task %s (%d) RLIMIT_MEMLOCK "
512 "(%ld) exceeded\n", __func__
,
513 dma
->task
->comm
, task_pid_nr(dma
->task
),
514 task_rlimit(dma
->task
, RLIMIT_MEMLOCK
));
522 static int vfio_unpin_page_external(struct vfio_dma
*dma
, dma_addr_t iova
,
526 struct vfio_pfn
*vpfn
= vfio_find_vpfn(dma
, iova
);
531 unlocked
= vfio_iova_put_vfio_pfn(dma
, vpfn
);
534 vfio_lock_acct(dma
, -unlocked
, true);
539 static int vfio_iommu_type1_pin_pages(void *iommu_data
,
540 unsigned long *user_pfn
,
542 unsigned long *phys_pfn
)
544 struct vfio_iommu
*iommu
= iommu_data
;
546 unsigned long remote_vaddr
;
547 struct vfio_dma
*dma
;
550 if (!iommu
|| !user_pfn
|| !phys_pfn
)
553 /* Supported for v2 version only */
557 mutex_lock(&iommu
->lock
);
559 /* Fail if notifier list is empty */
560 if ((!iommu
->external_domain
) || (!iommu
->notifier
.head
)) {
566 * If iommu capable domain exist in the container then all pages are
567 * already pinned and accounted. Accouting should be done if there is no
568 * iommu capable domain in the container.
570 do_accounting
= !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu
);
572 for (i
= 0; i
< npage
; i
++) {
574 struct vfio_pfn
*vpfn
;
576 iova
= user_pfn
[i
] << PAGE_SHIFT
;
577 dma
= vfio_find_dma(iommu
, iova
, PAGE_SIZE
);
583 if ((dma
->prot
& prot
) != prot
) {
588 vpfn
= vfio_iova_get_vfio_pfn(dma
, iova
);
590 phys_pfn
[i
] = vpfn
->pfn
;
594 remote_vaddr
= dma
->vaddr
+ iova
- dma
->iova
;
595 ret
= vfio_pin_page_external(dma
, remote_vaddr
, &phys_pfn
[i
],
600 ret
= vfio_add_to_pfn_list(dma
, iova
, phys_pfn
[i
]);
602 vfio_unpin_page_external(dma
, iova
, do_accounting
);
612 for (j
= 0; j
< i
; j
++) {
615 iova
= user_pfn
[j
] << PAGE_SHIFT
;
616 dma
= vfio_find_dma(iommu
, iova
, PAGE_SIZE
);
617 vfio_unpin_page_external(dma
, iova
, do_accounting
);
621 mutex_unlock(&iommu
->lock
);
625 static int vfio_iommu_type1_unpin_pages(void *iommu_data
,
626 unsigned long *user_pfn
,
629 struct vfio_iommu
*iommu
= iommu_data
;
633 if (!iommu
|| !user_pfn
)
636 /* Supported for v2 version only */
640 mutex_lock(&iommu
->lock
);
642 if (!iommu
->external_domain
) {
643 mutex_unlock(&iommu
->lock
);
647 do_accounting
= !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu
);
648 for (i
= 0; i
< npage
; i
++) {
649 struct vfio_dma
*dma
;
652 iova
= user_pfn
[i
] << PAGE_SHIFT
;
653 dma
= vfio_find_dma(iommu
, iova
, PAGE_SIZE
);
656 vfio_unpin_page_external(dma
, iova
, do_accounting
);
660 mutex_unlock(&iommu
->lock
);
661 return i
> npage
? npage
: (i
> 0 ? i
: -EINVAL
);
664 static long vfio_unmap_unpin(struct vfio_iommu
*iommu
, struct vfio_dma
*dma
,
667 dma_addr_t iova
= dma
->iova
, end
= dma
->iova
+ dma
->size
;
668 struct vfio_domain
*domain
, *d
;
674 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu
))
678 * We use the IOMMU to track the physical addresses, otherwise we'd
679 * need a much more complicated tracking system. Unfortunately that
680 * means we need to use one of the iommu domains to figure out the
681 * pfns to unpin. The rest need to be unmapped in advance so we have
682 * no iommu translations remaining when the pages are unpinned.
684 domain
= d
= list_first_entry(&iommu
->domain_list
,
685 struct vfio_domain
, next
);
687 list_for_each_entry_continue(d
, &iommu
->domain_list
, next
) {
688 iommu_unmap(d
->domain
, dma
->iova
, dma
->size
);
693 size_t unmapped
, len
;
694 phys_addr_t phys
, next
;
696 phys
= iommu_iova_to_phys(domain
->domain
, iova
);
697 if (WARN_ON(!phys
)) {
703 * To optimize for fewer iommu_unmap() calls, each of which
704 * may require hardware cache flushing, try to find the
705 * largest contiguous physical memory chunk to unmap.
707 for (len
= PAGE_SIZE
;
708 !domain
->fgsp
&& iova
+ len
< end
; len
+= PAGE_SIZE
) {
709 next
= iommu_iova_to_phys(domain
->domain
, iova
+ len
);
710 if (next
!= phys
+ len
)
714 unmapped
= iommu_unmap(domain
->domain
, iova
, len
);
715 if (WARN_ON(!unmapped
))
718 unlocked
+= vfio_unpin_pages_remote(dma
, iova
,
720 unmapped
>> PAGE_SHIFT
,
727 dma
->iommu_mapped
= false;
729 vfio_lock_acct(dma
, -unlocked
, true);
735 static void vfio_remove_dma(struct vfio_iommu
*iommu
, struct vfio_dma
*dma
)
737 vfio_unmap_unpin(iommu
, dma
, true);
738 vfio_unlink_dma(iommu
, dma
);
739 put_task_struct(dma
->task
);
744 static unsigned long vfio_pgsize_bitmap(struct vfio_iommu
*iommu
)
746 struct vfio_domain
*domain
;
747 unsigned long bitmap
= ULONG_MAX
;
749 mutex_lock(&iommu
->lock
);
750 list_for_each_entry(domain
, &iommu
->domain_list
, next
)
751 bitmap
&= domain
->domain
->pgsize_bitmap
;
752 mutex_unlock(&iommu
->lock
);
755 * In case the IOMMU supports page sizes smaller than PAGE_SIZE
756 * we pretend PAGE_SIZE is supported and hide sub-PAGE_SIZE sizes.
757 * That way the user will be able to map/unmap buffers whose size/
758 * start address is aligned with PAGE_SIZE. Pinning code uses that
759 * granularity while iommu driver can use the sub-PAGE_SIZE size
762 if (bitmap
& ~PAGE_MASK
) {
770 static int vfio_dma_do_unmap(struct vfio_iommu
*iommu
,
771 struct vfio_iommu_type1_dma_unmap
*unmap
)
774 struct vfio_dma
*dma
, *dma_last
= NULL
;
776 int ret
= 0, retries
= 0;
778 mask
= ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu
))) - 1;
780 if (unmap
->iova
& mask
)
782 if (!unmap
->size
|| unmap
->size
& mask
)
784 if (unmap
->iova
+ unmap
->size
< unmap
->iova
||
785 unmap
->size
> SIZE_MAX
)
788 WARN_ON(mask
& PAGE_MASK
);
790 mutex_lock(&iommu
->lock
);
793 * vfio-iommu-type1 (v1) - User mappings were coalesced together to
794 * avoid tracking individual mappings. This means that the granularity
795 * of the original mapping was lost and the user was allowed to attempt
796 * to unmap any range. Depending on the contiguousness of physical
797 * memory and page sizes supported by the IOMMU, arbitrary unmaps may
798 * or may not have worked. We only guaranteed unmap granularity
799 * matching the original mapping; even though it was untracked here,
800 * the original mappings are reflected in IOMMU mappings. This
801 * resulted in a couple unusual behaviors. First, if a range is not
802 * able to be unmapped, ex. a set of 4k pages that was mapped as a
803 * 2M hugepage into the IOMMU, the unmap ioctl returns success but with
804 * a zero sized unmap. Also, if an unmap request overlaps the first
805 * address of a hugepage, the IOMMU will unmap the entire hugepage.
806 * This also returns success and the returned unmap size reflects the
807 * actual size unmapped.
809 * We attempt to maintain compatibility with this "v1" interface, but
810 * we take control out of the hands of the IOMMU. Therefore, an unmap
811 * request offset from the beginning of the original mapping will
812 * return success with zero sized unmap. And an unmap request covering
813 * the first iova of mapping will unmap the entire range.
815 * The v2 version of this interface intends to be more deterministic.
816 * Unmap requests must fully cover previous mappings. Multiple
817 * mappings may still be unmaped by specifying large ranges, but there
818 * must not be any previous mappings bisected by the range. An error
819 * will be returned if these conditions are not met. The v2 interface
820 * will only return success and a size of zero if there were no
821 * mappings within the range.
824 dma
= vfio_find_dma(iommu
, unmap
->iova
, 1);
825 if (dma
&& dma
->iova
!= unmap
->iova
) {
829 dma
= vfio_find_dma(iommu
, unmap
->iova
+ unmap
->size
- 1, 0);
830 if (dma
&& dma
->iova
+ dma
->size
!= unmap
->iova
+ unmap
->size
) {
836 while ((dma
= vfio_find_dma(iommu
, unmap
->iova
, unmap
->size
))) {
837 if (!iommu
->v2
&& unmap
->iova
> dma
->iova
)
840 * Task with same address space who mapped this iova range is
841 * allowed to unmap the iova range.
843 if (dma
->task
->mm
!= current
->mm
)
846 if (!RB_EMPTY_ROOT(&dma
->pfn_list
)) {
847 struct vfio_iommu_type1_dma_unmap nb_unmap
;
849 if (dma_last
== dma
) {
850 BUG_ON(++retries
> 10);
856 nb_unmap
.iova
= dma
->iova
;
857 nb_unmap
.size
= dma
->size
;
860 * Notify anyone (mdev vendor drivers) to invalidate and
861 * unmap iovas within the range we're about to unmap.
862 * Vendor drivers MUST unpin pages in response to an
865 mutex_unlock(&iommu
->lock
);
866 blocking_notifier_call_chain(&iommu
->notifier
,
867 VFIO_IOMMU_NOTIFY_DMA_UNMAP
,
871 unmapped
+= dma
->size
;
872 vfio_remove_dma(iommu
, dma
);
876 mutex_unlock(&iommu
->lock
);
878 /* Report how much was unmapped */
879 unmap
->size
= unmapped
;
885 * Turns out AMD IOMMU has a page table bug where it won't map large pages
886 * to a region that previously mapped smaller pages. This should be fixed
887 * soon, so this is just a temporary workaround to break mappings down into
888 * PAGE_SIZE. Better to map smaller pages than nothing.
890 static int map_try_harder(struct vfio_domain
*domain
, dma_addr_t iova
,
891 unsigned long pfn
, long npage
, int prot
)
896 for (i
= 0; i
< npage
; i
++, pfn
++, iova
+= PAGE_SIZE
) {
897 ret
= iommu_map(domain
->domain
, iova
,
898 (phys_addr_t
)pfn
<< PAGE_SHIFT
,
899 PAGE_SIZE
, prot
| domain
->prot
);
904 for (; i
< npage
&& i
> 0; i
--, iova
-= PAGE_SIZE
)
905 iommu_unmap(domain
->domain
, iova
, PAGE_SIZE
);
910 static int vfio_iommu_map(struct vfio_iommu
*iommu
, dma_addr_t iova
,
911 unsigned long pfn
, long npage
, int prot
)
913 struct vfio_domain
*d
;
916 list_for_each_entry(d
, &iommu
->domain_list
, next
) {
917 ret
= iommu_map(d
->domain
, iova
, (phys_addr_t
)pfn
<< PAGE_SHIFT
,
918 npage
<< PAGE_SHIFT
, prot
| d
->prot
);
921 map_try_harder(d
, iova
, pfn
, npage
, prot
))
931 list_for_each_entry_continue_reverse(d
, &iommu
->domain_list
, next
)
932 iommu_unmap(d
->domain
, iova
, npage
<< PAGE_SHIFT
);
937 static int vfio_pin_map_dma(struct vfio_iommu
*iommu
, struct vfio_dma
*dma
,
940 dma_addr_t iova
= dma
->iova
;
941 unsigned long vaddr
= dma
->vaddr
;
942 size_t size
= map_size
;
944 unsigned long pfn
, limit
= rlimit(RLIMIT_MEMLOCK
) >> PAGE_SHIFT
;
948 /* Pin a contiguous chunk of memory */
949 npage
= vfio_pin_pages_remote(dma
, vaddr
+ dma
->size
,
950 size
>> PAGE_SHIFT
, &pfn
, limit
);
958 ret
= vfio_iommu_map(iommu
, iova
+ dma
->size
, pfn
, npage
,
961 vfio_unpin_pages_remote(dma
, iova
+ dma
->size
, pfn
,
966 size
-= npage
<< PAGE_SHIFT
;
967 dma
->size
+= npage
<< PAGE_SHIFT
;
970 dma
->iommu_mapped
= true;
973 vfio_remove_dma(iommu
, dma
);
978 static int vfio_dma_do_map(struct vfio_iommu
*iommu
,
979 struct vfio_iommu_type1_dma_map
*map
)
981 dma_addr_t iova
= map
->iova
;
982 unsigned long vaddr
= map
->vaddr
;
983 size_t size
= map
->size
;
984 int ret
= 0, prot
= 0;
986 struct vfio_dma
*dma
;
988 /* Verify that none of our __u64 fields overflow */
989 if (map
->size
!= size
|| map
->vaddr
!= vaddr
|| map
->iova
!= iova
)
992 mask
= ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu
))) - 1;
994 WARN_ON(mask
& PAGE_MASK
);
996 /* READ/WRITE from device perspective */
997 if (map
->flags
& VFIO_DMA_MAP_FLAG_WRITE
)
999 if (map
->flags
& VFIO_DMA_MAP_FLAG_READ
)
1002 if (!prot
|| !size
|| (size
| iova
| vaddr
) & mask
)
1005 /* Don't allow IOVA or virtual address wrap */
1006 if (iova
+ size
- 1 < iova
|| vaddr
+ size
- 1 < vaddr
)
1009 mutex_lock(&iommu
->lock
);
1011 if (vfio_find_dma(iommu
, iova
, size
)) {
1016 if (!iommu
->dma_avail
) {
1021 dma
= kzalloc(sizeof(*dma
), GFP_KERNEL
);
1033 * We need to be able to both add to a task's locked memory and test
1034 * against the locked memory limit and we need to be able to do both
1035 * outside of this call path as pinning can be asynchronous via the
1036 * external interfaces for mdev devices. RLIMIT_MEMLOCK requires a
1037 * task_struct and VM locked pages requires an mm_struct, however
1038 * holding an indefinite mm reference is not recommended, therefore we
1039 * only hold a reference to a task. We could hold a reference to
1040 * current, however QEMU uses this call path through vCPU threads,
1041 * which can be killed resulting in a NULL mm and failure in the unmap
1042 * path when called via a different thread. Avoid this problem by
1043 * using the group_leader as threads within the same group require
1044 * both CLONE_THREAD and CLONE_VM and will therefore use the same
1047 * Previously we also used the task for testing CAP_IPC_LOCK at the
1048 * time of pinning and accounting, however has_capability() makes use
1049 * of real_cred, a copy-on-write field, so we can't guarantee that it
1050 * matches group_leader, or in fact that it might not change by the
1051 * time it's evaluated. If a process were to call MAP_DMA with
1052 * CAP_IPC_LOCK but later drop it, it doesn't make sense that they
1053 * possibly see different results for an iommu_mapped vfio_dma vs
1054 * externally mapped. Therefore track CAP_IPC_LOCK in vfio_dma at the
1055 * time of calling MAP_DMA.
1057 get_task_struct(current
->group_leader
);
1058 dma
->task
= current
->group_leader
;
1059 dma
->lock_cap
= capable(CAP_IPC_LOCK
);
1061 dma
->pfn_list
= RB_ROOT
;
1063 /* Insert zero-sized and grow as we map chunks of it */
1064 vfio_link_dma(iommu
, dma
);
1066 /* Don't pin and map if container doesn't contain IOMMU capable domain*/
1067 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu
))
1070 ret
= vfio_pin_map_dma(iommu
, dma
, size
);
1073 mutex_unlock(&iommu
->lock
);
1077 static int vfio_bus_type(struct device
*dev
, void *data
)
1079 struct bus_type
**bus
= data
;
1081 if (*bus
&& *bus
!= dev
->bus
)
1089 static int vfio_iommu_replay(struct vfio_iommu
*iommu
,
1090 struct vfio_domain
*domain
)
1092 struct vfio_domain
*d
;
1094 unsigned long limit
= rlimit(RLIMIT_MEMLOCK
) >> PAGE_SHIFT
;
1097 /* Arbitrarily pick the first domain in the list for lookups */
1098 d
= list_first_entry(&iommu
->domain_list
, struct vfio_domain
, next
);
1099 n
= rb_first(&iommu
->dma_list
);
1101 for (; n
; n
= rb_next(n
)) {
1102 struct vfio_dma
*dma
;
1105 dma
= rb_entry(n
, struct vfio_dma
, node
);
1108 while (iova
< dma
->iova
+ dma
->size
) {
1112 if (dma
->iommu_mapped
) {
1116 phys
= iommu_iova_to_phys(d
->domain
, iova
);
1118 if (WARN_ON(!phys
)) {
1126 while (i
< dma
->iova
+ dma
->size
&&
1127 p
== iommu_iova_to_phys(d
->domain
, i
)) {
1134 unsigned long vaddr
= dma
->vaddr
+
1136 size_t n
= dma
->iova
+ dma
->size
- iova
;
1139 npage
= vfio_pin_pages_remote(dma
, vaddr
,
1148 phys
= pfn
<< PAGE_SHIFT
;
1149 size
= npage
<< PAGE_SHIFT
;
1152 ret
= iommu_map(domain
->domain
, iova
, phys
,
1153 size
, dma
->prot
| domain
->prot
);
1159 dma
->iommu_mapped
= true;
1165 * We change our unmap behavior slightly depending on whether the IOMMU
1166 * supports fine-grained superpages. IOMMUs like AMD-Vi will use a superpage
1167 * for practically any contiguous power-of-two mapping we give it. This means
1168 * we don't need to look for contiguous chunks ourselves to make unmapping
1169 * more efficient. On IOMMUs with coarse-grained super pages, like Intel VT-d
1170 * with discrete 2M/1G/512G/1T superpages, identifying contiguous chunks
1171 * significantly boosts non-hugetlbfs mappings and doesn't seem to hurt when
1172 * hugetlbfs is in use.
1174 static void vfio_test_domain_fgsp(struct vfio_domain
*domain
)
1177 int ret
, order
= get_order(PAGE_SIZE
* 2);
1179 pages
= alloc_pages(GFP_KERNEL
| __GFP_ZERO
, order
);
1183 ret
= iommu_map(domain
->domain
, 0, page_to_phys(pages
), PAGE_SIZE
* 2,
1184 IOMMU_READ
| IOMMU_WRITE
| domain
->prot
);
1186 size_t unmapped
= iommu_unmap(domain
->domain
, 0, PAGE_SIZE
);
1188 if (unmapped
== PAGE_SIZE
)
1189 iommu_unmap(domain
->domain
, PAGE_SIZE
, PAGE_SIZE
);
1191 domain
->fgsp
= true;
1194 __free_pages(pages
, order
);
1197 static struct vfio_group
*find_iommu_group(struct vfio_domain
*domain
,
1198 struct iommu_group
*iommu_group
)
1200 struct vfio_group
*g
;
1202 list_for_each_entry(g
, &domain
->group_list
, next
) {
1203 if (g
->iommu_group
== iommu_group
)
1210 static bool vfio_iommu_has_sw_msi(struct iommu_group
*group
, phys_addr_t
*base
)
1212 struct list_head group_resv_regions
;
1213 struct iommu_resv_region
*region
, *next
;
1216 INIT_LIST_HEAD(&group_resv_regions
);
1217 iommu_get_group_resv_regions(group
, &group_resv_regions
);
1218 list_for_each_entry(region
, &group_resv_regions
, list
) {
1220 * The presence of any 'real' MSI regions should take
1221 * precedence over the software-managed one if the
1222 * IOMMU driver happens to advertise both types.
1224 if (region
->type
== IOMMU_RESV_MSI
) {
1229 if (region
->type
== IOMMU_RESV_SW_MSI
) {
1230 *base
= region
->start
;
1234 list_for_each_entry_safe(region
, next
, &group_resv_regions
, list
)
1239 static int vfio_iommu_type1_attach_group(void *iommu_data
,
1240 struct iommu_group
*iommu_group
)
1242 struct vfio_iommu
*iommu
= iommu_data
;
1243 struct vfio_group
*group
;
1244 struct vfio_domain
*domain
, *d
;
1245 struct bus_type
*bus
= NULL
, *mdev_bus
;
1247 bool resv_msi
, msi_remap
;
1248 phys_addr_t resv_msi_base
;
1250 mutex_lock(&iommu
->lock
);
1252 list_for_each_entry(d
, &iommu
->domain_list
, next
) {
1253 if (find_iommu_group(d
, iommu_group
)) {
1254 mutex_unlock(&iommu
->lock
);
1259 if (iommu
->external_domain
) {
1260 if (find_iommu_group(iommu
->external_domain
, iommu_group
)) {
1261 mutex_unlock(&iommu
->lock
);
1266 group
= kzalloc(sizeof(*group
), GFP_KERNEL
);
1267 domain
= kzalloc(sizeof(*domain
), GFP_KERNEL
);
1268 if (!group
|| !domain
) {
1273 group
->iommu_group
= iommu_group
;
1275 /* Determine bus_type in order to allocate a domain */
1276 ret
= iommu_group_for_each_dev(iommu_group
, &bus
, vfio_bus_type
);
1280 mdev_bus
= symbol_get(mdev_bus_type
);
1283 if ((bus
== mdev_bus
) && !iommu_present(bus
)) {
1284 symbol_put(mdev_bus_type
);
1285 if (!iommu
->external_domain
) {
1286 INIT_LIST_HEAD(&domain
->group_list
);
1287 iommu
->external_domain
= domain
;
1291 list_add(&group
->next
,
1292 &iommu
->external_domain
->group_list
);
1293 mutex_unlock(&iommu
->lock
);
1296 symbol_put(mdev_bus_type
);
1299 domain
->domain
= iommu_domain_alloc(bus
);
1300 if (!domain
->domain
) {
1305 if (iommu
->nesting
) {
1308 ret
= iommu_domain_set_attr(domain
->domain
, DOMAIN_ATTR_NESTING
,
1314 ret
= iommu_attach_group(domain
->domain
, iommu_group
);
1318 resv_msi
= vfio_iommu_has_sw_msi(iommu_group
, &resv_msi_base
);
1320 INIT_LIST_HEAD(&domain
->group_list
);
1321 list_add(&group
->next
, &domain
->group_list
);
1323 msi_remap
= irq_domain_check_msi_remap() ||
1324 iommu_capable(bus
, IOMMU_CAP_INTR_REMAP
);
1326 if (!allow_unsafe_interrupts
&& !msi_remap
) {
1327 pr_warn("%s: No interrupt remapping support. Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n",
1333 if (iommu_capable(bus
, IOMMU_CAP_CACHE_COHERENCY
))
1334 domain
->prot
|= IOMMU_CACHE
;
1337 * Try to match an existing compatible domain. We don't want to
1338 * preclude an IOMMU driver supporting multiple bus_types and being
1339 * able to include different bus_types in the same IOMMU domain, so
1340 * we test whether the domains use the same iommu_ops rather than
1341 * testing if they're on the same bus_type.
1343 list_for_each_entry(d
, &iommu
->domain_list
, next
) {
1344 if (d
->domain
->ops
== domain
->domain
->ops
&&
1345 d
->prot
== domain
->prot
) {
1346 iommu_detach_group(domain
->domain
, iommu_group
);
1347 if (!iommu_attach_group(d
->domain
, iommu_group
)) {
1348 list_add(&group
->next
, &d
->group_list
);
1349 iommu_domain_free(domain
->domain
);
1351 mutex_unlock(&iommu
->lock
);
1355 ret
= iommu_attach_group(domain
->domain
, iommu_group
);
1361 vfio_test_domain_fgsp(domain
);
1363 /* replay mappings on new domains */
1364 ret
= vfio_iommu_replay(iommu
, domain
);
1369 ret
= iommu_get_msi_cookie(domain
->domain
, resv_msi_base
);
1374 list_add(&domain
->next
, &iommu
->domain_list
);
1376 mutex_unlock(&iommu
->lock
);
1381 iommu_detach_group(domain
->domain
, iommu_group
);
1383 iommu_domain_free(domain
->domain
);
1387 mutex_unlock(&iommu
->lock
);
1391 static void vfio_iommu_unmap_unpin_all(struct vfio_iommu
*iommu
)
1393 struct rb_node
*node
;
1395 while ((node
= rb_first(&iommu
->dma_list
)))
1396 vfio_remove_dma(iommu
, rb_entry(node
, struct vfio_dma
, node
));
1399 static void vfio_iommu_unmap_unpin_reaccount(struct vfio_iommu
*iommu
)
1401 struct rb_node
*n
, *p
;
1403 n
= rb_first(&iommu
->dma_list
);
1404 for (; n
; n
= rb_next(n
)) {
1405 struct vfio_dma
*dma
;
1406 long locked
= 0, unlocked
= 0;
1408 dma
= rb_entry(n
, struct vfio_dma
, node
);
1409 unlocked
+= vfio_unmap_unpin(iommu
, dma
, false);
1410 p
= rb_first(&dma
->pfn_list
);
1411 for (; p
; p
= rb_next(p
)) {
1412 struct vfio_pfn
*vpfn
= rb_entry(p
, struct vfio_pfn
,
1415 if (!is_invalid_reserved_pfn(vpfn
->pfn
))
1418 vfio_lock_acct(dma
, locked
- unlocked
, true);
1422 static void vfio_sanity_check_pfn_list(struct vfio_iommu
*iommu
)
1426 n
= rb_first(&iommu
->dma_list
);
1427 for (; n
; n
= rb_next(n
)) {
1428 struct vfio_dma
*dma
;
1430 dma
= rb_entry(n
, struct vfio_dma
, node
);
1432 if (WARN_ON(!RB_EMPTY_ROOT(&dma
->pfn_list
)))
1435 /* mdev vendor driver must unregister notifier */
1436 WARN_ON(iommu
->notifier
.head
);
1439 static void vfio_iommu_type1_detach_group(void *iommu_data
,
1440 struct iommu_group
*iommu_group
)
1442 struct vfio_iommu
*iommu
= iommu_data
;
1443 struct vfio_domain
*domain
;
1444 struct vfio_group
*group
;
1446 mutex_lock(&iommu
->lock
);
1448 if (iommu
->external_domain
) {
1449 group
= find_iommu_group(iommu
->external_domain
, iommu_group
);
1451 list_del(&group
->next
);
1454 if (list_empty(&iommu
->external_domain
->group_list
)) {
1455 vfio_sanity_check_pfn_list(iommu
);
1457 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu
))
1458 vfio_iommu_unmap_unpin_all(iommu
);
1460 kfree(iommu
->external_domain
);
1461 iommu
->external_domain
= NULL
;
1463 goto detach_group_done
;
1467 list_for_each_entry(domain
, &iommu
->domain_list
, next
) {
1468 group
= find_iommu_group(domain
, iommu_group
);
1472 iommu_detach_group(domain
->domain
, iommu_group
);
1473 list_del(&group
->next
);
1476 * Group ownership provides privilege, if the group list is
1477 * empty, the domain goes away. If it's the last domain with
1478 * iommu and external domain doesn't exist, then all the
1479 * mappings go away too. If it's the last domain with iommu and
1480 * external domain exist, update accounting
1482 if (list_empty(&domain
->group_list
)) {
1483 if (list_is_singular(&iommu
->domain_list
)) {
1484 if (!iommu
->external_domain
)
1485 vfio_iommu_unmap_unpin_all(iommu
);
1487 vfio_iommu_unmap_unpin_reaccount(iommu
);
1489 iommu_domain_free(domain
->domain
);
1490 list_del(&domain
->next
);
1497 mutex_unlock(&iommu
->lock
);
1500 static void *vfio_iommu_type1_open(unsigned long arg
)
1502 struct vfio_iommu
*iommu
;
1504 iommu
= kzalloc(sizeof(*iommu
), GFP_KERNEL
);
1506 return ERR_PTR(-ENOMEM
);
1509 case VFIO_TYPE1_IOMMU
:
1511 case VFIO_TYPE1_NESTING_IOMMU
:
1512 iommu
->nesting
= true;
1513 case VFIO_TYPE1v2_IOMMU
:
1518 return ERR_PTR(-EINVAL
);
1521 INIT_LIST_HEAD(&iommu
->domain_list
);
1522 iommu
->dma_list
= RB_ROOT
;
1523 iommu
->dma_avail
= dma_entry_limit
;
1524 mutex_init(&iommu
->lock
);
1525 BLOCKING_INIT_NOTIFIER_HEAD(&iommu
->notifier
);
1530 static void vfio_release_domain(struct vfio_domain
*domain
, bool external
)
1532 struct vfio_group
*group
, *group_tmp
;
1534 list_for_each_entry_safe(group
, group_tmp
,
1535 &domain
->group_list
, next
) {
1537 iommu_detach_group(domain
->domain
, group
->iommu_group
);
1538 list_del(&group
->next
);
1543 iommu_domain_free(domain
->domain
);
1546 static void vfio_iommu_type1_release(void *iommu_data
)
1548 struct vfio_iommu
*iommu
= iommu_data
;
1549 struct vfio_domain
*domain
, *domain_tmp
;
1551 if (iommu
->external_domain
) {
1552 vfio_release_domain(iommu
->external_domain
, true);
1553 vfio_sanity_check_pfn_list(iommu
);
1554 kfree(iommu
->external_domain
);
1557 vfio_iommu_unmap_unpin_all(iommu
);
1559 list_for_each_entry_safe(domain
, domain_tmp
,
1560 &iommu
->domain_list
, next
) {
1561 vfio_release_domain(domain
, false);
1562 list_del(&domain
->next
);
1568 static int vfio_domains_have_iommu_cache(struct vfio_iommu
*iommu
)
1570 struct vfio_domain
*domain
;
1573 mutex_lock(&iommu
->lock
);
1574 list_for_each_entry(domain
, &iommu
->domain_list
, next
) {
1575 if (!(domain
->prot
& IOMMU_CACHE
)) {
1580 mutex_unlock(&iommu
->lock
);
1585 static long vfio_iommu_type1_ioctl(void *iommu_data
,
1586 unsigned int cmd
, unsigned long arg
)
1588 struct vfio_iommu
*iommu
= iommu_data
;
1589 unsigned long minsz
;
1591 if (cmd
== VFIO_CHECK_EXTENSION
) {
1593 case VFIO_TYPE1_IOMMU
:
1594 case VFIO_TYPE1v2_IOMMU
:
1595 case VFIO_TYPE1_NESTING_IOMMU
:
1597 case VFIO_DMA_CC_IOMMU
:
1600 return vfio_domains_have_iommu_cache(iommu
);
1604 } else if (cmd
== VFIO_IOMMU_GET_INFO
) {
1605 struct vfio_iommu_type1_info info
;
1607 minsz
= offsetofend(struct vfio_iommu_type1_info
, iova_pgsizes
);
1609 if (copy_from_user(&info
, (void __user
*)arg
, minsz
))
1612 if (info
.argsz
< minsz
)
1615 info
.flags
= VFIO_IOMMU_INFO_PGSIZES
;
1617 info
.iova_pgsizes
= vfio_pgsize_bitmap(iommu
);
1619 return copy_to_user((void __user
*)arg
, &info
, minsz
) ?
1622 } else if (cmd
== VFIO_IOMMU_MAP_DMA
) {
1623 struct vfio_iommu_type1_dma_map map
;
1624 uint32_t mask
= VFIO_DMA_MAP_FLAG_READ
|
1625 VFIO_DMA_MAP_FLAG_WRITE
;
1627 minsz
= offsetofend(struct vfio_iommu_type1_dma_map
, size
);
1629 if (copy_from_user(&map
, (void __user
*)arg
, minsz
))
1632 if (map
.argsz
< minsz
|| map
.flags
& ~mask
)
1635 return vfio_dma_do_map(iommu
, &map
);
1637 } else if (cmd
== VFIO_IOMMU_UNMAP_DMA
) {
1638 struct vfio_iommu_type1_dma_unmap unmap
;
1641 minsz
= offsetofend(struct vfio_iommu_type1_dma_unmap
, size
);
1643 if (copy_from_user(&unmap
, (void __user
*)arg
, minsz
))
1646 if (unmap
.argsz
< minsz
|| unmap
.flags
)
1649 ret
= vfio_dma_do_unmap(iommu
, &unmap
);
1653 return copy_to_user((void __user
*)arg
, &unmap
, minsz
) ?
1660 static int vfio_iommu_type1_register_notifier(void *iommu_data
,
1661 unsigned long *events
,
1662 struct notifier_block
*nb
)
1664 struct vfio_iommu
*iommu
= iommu_data
;
1666 /* clear known events */
1667 *events
&= ~VFIO_IOMMU_NOTIFY_DMA_UNMAP
;
1669 /* refuse to register if still events remaining */
1673 return blocking_notifier_chain_register(&iommu
->notifier
, nb
);
1676 static int vfio_iommu_type1_unregister_notifier(void *iommu_data
,
1677 struct notifier_block
*nb
)
1679 struct vfio_iommu
*iommu
= iommu_data
;
1681 return blocking_notifier_chain_unregister(&iommu
->notifier
, nb
);
1684 static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1
= {
1685 .name
= "vfio-iommu-type1",
1686 .owner
= THIS_MODULE
,
1687 .open
= vfio_iommu_type1_open
,
1688 .release
= vfio_iommu_type1_release
,
1689 .ioctl
= vfio_iommu_type1_ioctl
,
1690 .attach_group
= vfio_iommu_type1_attach_group
,
1691 .detach_group
= vfio_iommu_type1_detach_group
,
1692 .pin_pages
= vfio_iommu_type1_pin_pages
,
1693 .unpin_pages
= vfio_iommu_type1_unpin_pages
,
1694 .register_notifier
= vfio_iommu_type1_register_notifier
,
1695 .unregister_notifier
= vfio_iommu_type1_unregister_notifier
,
1698 static int __init
vfio_iommu_type1_init(void)
1700 return vfio_register_iommu_driver(&vfio_iommu_driver_ops_type1
);
1703 static void __exit
vfio_iommu_type1_cleanup(void)
1705 vfio_unregister_iommu_driver(&vfio_iommu_driver_ops_type1
);
1708 module_init(vfio_iommu_type1_init
);
1709 module_exit(vfio_iommu_type1_cleanup
);
1711 MODULE_VERSION(DRIVER_VERSION
);
1712 MODULE_LICENSE("GPL v2");
1713 MODULE_AUTHOR(DRIVER_AUTHOR
);
1714 MODULE_DESCRIPTION(DRIVER_DESC
);