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.h>
35 #include <linux/slab.h>
36 #include <linux/uaccess.h>
37 #include <linux/vfio.h>
38 #include <linux/workqueue.h>
40 #define DRIVER_VERSION "0.2"
41 #define DRIVER_AUTHOR "Alex Williamson <alex.williamson@redhat.com>"
42 #define DRIVER_DESC "Type1 IOMMU driver for VFIO"
44 static bool allow_unsafe_interrupts
;
45 module_param_named(allow_unsafe_interrupts
,
46 allow_unsafe_interrupts
, bool, S_IRUGO
| S_IWUSR
);
47 MODULE_PARM_DESC(allow_unsafe_interrupts
,
48 "Enable VFIO IOMMU support for on platforms without interrupt remapping support.");
50 static bool disable_hugepages
;
51 module_param_named(disable_hugepages
,
52 disable_hugepages
, bool, S_IRUGO
| S_IWUSR
);
53 MODULE_PARM_DESC(disable_hugepages
,
54 "Disable VFIO IOMMU support for IOMMU hugepages.");
57 struct list_head domain_list
;
59 struct rb_root dma_list
;
65 struct iommu_domain
*domain
;
66 struct list_head next
;
67 struct list_head group_list
;
68 int prot
; /* IOMMU_CACHE */
69 bool fgsp
; /* Fine-grained super pages */
74 dma_addr_t iova
; /* Device address */
75 unsigned long vaddr
; /* Process virtual addr */
76 size_t size
; /* Map size (bytes) */
77 int prot
; /* IOMMU_READ/WRITE */
81 struct iommu_group
*iommu_group
;
82 struct list_head next
;
86 * This code handles mapping and unmapping of user data buffers
87 * into DMA'ble space using the IOMMU
90 static struct vfio_dma
*vfio_find_dma(struct vfio_iommu
*iommu
,
91 dma_addr_t start
, size_t size
)
93 struct rb_node
*node
= iommu
->dma_list
.rb_node
;
96 struct vfio_dma
*dma
= rb_entry(node
, struct vfio_dma
, node
);
98 if (start
+ size
<= dma
->iova
)
100 else if (start
>= dma
->iova
+ dma
->size
)
101 node
= node
->rb_right
;
109 static void vfio_link_dma(struct vfio_iommu
*iommu
, struct vfio_dma
*new)
111 struct rb_node
**link
= &iommu
->dma_list
.rb_node
, *parent
= NULL
;
112 struct vfio_dma
*dma
;
116 dma
= rb_entry(parent
, struct vfio_dma
, node
);
118 if (new->iova
+ new->size
<= dma
->iova
)
119 link
= &(*link
)->rb_left
;
121 link
= &(*link
)->rb_right
;
124 rb_link_node(&new->node
, parent
, link
);
125 rb_insert_color(&new->node
, &iommu
->dma_list
);
128 static void vfio_unlink_dma(struct vfio_iommu
*iommu
, struct vfio_dma
*old
)
130 rb_erase(&old
->node
, &iommu
->dma_list
);
134 struct mm_struct
*mm
;
136 struct work_struct work
;
139 /* delayed decrement/increment for locked_vm */
140 static void vfio_lock_acct_bg(struct work_struct
*work
)
142 struct vwork
*vwork
= container_of(work
, struct vwork
, work
);
143 struct mm_struct
*mm
;
146 down_write(&mm
->mmap_sem
);
147 mm
->locked_vm
+= vwork
->npage
;
148 up_write(&mm
->mmap_sem
);
153 static void vfio_lock_acct(long npage
)
156 struct mm_struct
*mm
;
158 if (!current
->mm
|| !npage
)
159 return; /* process exited or nothing to do */
161 if (down_write_trylock(¤t
->mm
->mmap_sem
)) {
162 current
->mm
->locked_vm
+= npage
;
163 up_write(¤t
->mm
->mmap_sem
);
168 * Couldn't get mmap_sem lock, so must setup to update
169 * mm->locked_vm later. If locked_vm were atomic, we
170 * wouldn't need this silliness
172 vwork
= kmalloc(sizeof(struct vwork
), GFP_KERNEL
);
175 mm
= get_task_mm(current
);
180 INIT_WORK(&vwork
->work
, vfio_lock_acct_bg
);
182 vwork
->npage
= npage
;
183 schedule_work(&vwork
->work
);
187 * Some mappings aren't backed by a struct page, for example an mmap'd
188 * MMIO range for our own or another device. These use a different
189 * pfn conversion and shouldn't be tracked as locked pages.
191 static bool is_invalid_reserved_pfn(unsigned long pfn
)
193 if (pfn_valid(pfn
)) {
195 struct page
*tail
= pfn_to_page(pfn
);
196 struct page
*head
= compound_head(tail
);
197 reserved
= !!(PageReserved(head
));
200 * "head" is not a dangling pointer
201 * (compound_head takes care of that)
202 * but the hugepage may have been split
203 * from under us (and we may not hold a
204 * reference count on the head page so it can
205 * be reused before we run PageReferenced), so
206 * we've to check PageTail before returning
213 return PageReserved(tail
);
219 static int put_pfn(unsigned long pfn
, int prot
)
221 if (!is_invalid_reserved_pfn(pfn
)) {
222 struct page
*page
= pfn_to_page(pfn
);
223 if (prot
& IOMMU_WRITE
)
231 static int vaddr_get_pfn(unsigned long vaddr
, int prot
, unsigned long *pfn
)
233 struct page
*page
[1];
234 struct vm_area_struct
*vma
;
237 if (get_user_pages_fast(vaddr
, 1, !!(prot
& IOMMU_WRITE
), page
) == 1) {
238 *pfn
= page_to_pfn(page
[0]);
242 down_read(¤t
->mm
->mmap_sem
);
244 vma
= find_vma_intersection(current
->mm
, vaddr
, vaddr
+ 1);
246 if (vma
&& vma
->vm_flags
& VM_PFNMAP
) {
247 *pfn
= ((vaddr
- vma
->vm_start
) >> PAGE_SHIFT
) + vma
->vm_pgoff
;
248 if (is_invalid_reserved_pfn(*pfn
))
252 up_read(¤t
->mm
->mmap_sem
);
258 * Attempt to pin pages. We really don't want to track all the pfns and
259 * the iommu can only map chunks of consecutive pfns anyway, so get the
260 * first page and all consecutive pages with the same locking.
262 static long vfio_pin_pages(unsigned long vaddr
, long npage
,
263 int prot
, unsigned long *pfn_base
)
265 unsigned long limit
= rlimit(RLIMIT_MEMLOCK
) >> PAGE_SHIFT
;
266 bool lock_cap
= capable(CAP_IPC_LOCK
);
273 ret
= vaddr_get_pfn(vaddr
, prot
, pfn_base
);
277 rsvd
= is_invalid_reserved_pfn(*pfn_base
);
279 if (!rsvd
&& !lock_cap
&& current
->mm
->locked_vm
+ 1 > limit
) {
280 put_pfn(*pfn_base
, prot
);
281 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n", __func__
,
282 limit
<< PAGE_SHIFT
);
286 if (unlikely(disable_hugepages
)) {
292 /* Lock all the consecutive pages from pfn_base */
293 for (i
= 1, vaddr
+= PAGE_SIZE
; i
< npage
; i
++, vaddr
+= PAGE_SIZE
) {
294 unsigned long pfn
= 0;
296 ret
= vaddr_get_pfn(vaddr
, prot
, &pfn
);
300 if (pfn
!= *pfn_base
+ i
||
301 rsvd
!= is_invalid_reserved_pfn(pfn
)) {
306 if (!rsvd
&& !lock_cap
&&
307 current
->mm
->locked_vm
+ i
+ 1 > limit
) {
309 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n",
310 __func__
, limit
<< PAGE_SHIFT
);
321 static long vfio_unpin_pages(unsigned long pfn
, long npage
,
322 int prot
, bool do_accounting
)
324 unsigned long unlocked
= 0;
327 for (i
= 0; i
< npage
; i
++)
328 unlocked
+= put_pfn(pfn
++, prot
);
331 vfio_lock_acct(-unlocked
);
336 static void vfio_unmap_unpin(struct vfio_iommu
*iommu
, struct vfio_dma
*dma
)
338 dma_addr_t iova
= dma
->iova
, end
= dma
->iova
+ dma
->size
;
339 struct vfio_domain
*domain
, *d
;
345 * We use the IOMMU to track the physical addresses, otherwise we'd
346 * need a much more complicated tracking system. Unfortunately that
347 * means we need to use one of the iommu domains to figure out the
348 * pfns to unpin. The rest need to be unmapped in advance so we have
349 * no iommu translations remaining when the pages are unpinned.
351 domain
= d
= list_first_entry(&iommu
->domain_list
,
352 struct vfio_domain
, next
);
354 list_for_each_entry_continue(d
, &iommu
->domain_list
, next
) {
355 iommu_unmap(d
->domain
, dma
->iova
, dma
->size
);
360 size_t unmapped
, len
;
361 phys_addr_t phys
, next
;
363 phys
= iommu_iova_to_phys(domain
->domain
, iova
);
364 if (WARN_ON(!phys
)) {
370 * To optimize for fewer iommu_unmap() calls, each of which
371 * may require hardware cache flushing, try to find the
372 * largest contiguous physical memory chunk to unmap.
374 for (len
= PAGE_SIZE
;
375 !domain
->fgsp
&& iova
+ len
< end
; len
+= PAGE_SIZE
) {
376 next
= iommu_iova_to_phys(domain
->domain
, iova
+ len
);
377 if (next
!= phys
+ len
)
381 unmapped
= iommu_unmap(domain
->domain
, iova
, len
);
382 if (WARN_ON(!unmapped
))
385 unlocked
+= vfio_unpin_pages(phys
>> PAGE_SHIFT
,
386 unmapped
>> PAGE_SHIFT
,
393 vfio_lock_acct(-unlocked
);
396 static void vfio_remove_dma(struct vfio_iommu
*iommu
, struct vfio_dma
*dma
)
398 vfio_unmap_unpin(iommu
, dma
);
399 vfio_unlink_dma(iommu
, dma
);
403 static unsigned long vfio_pgsize_bitmap(struct vfio_iommu
*iommu
)
405 struct vfio_domain
*domain
;
406 unsigned long bitmap
= ULONG_MAX
;
408 mutex_lock(&iommu
->lock
);
409 list_for_each_entry(domain
, &iommu
->domain_list
, next
)
410 bitmap
&= domain
->domain
->ops
->pgsize_bitmap
;
411 mutex_unlock(&iommu
->lock
);
414 * In case the IOMMU supports page sizes smaller than PAGE_SIZE
415 * we pretend PAGE_SIZE is supported and hide sub-PAGE_SIZE sizes.
416 * That way the user will be able to map/unmap buffers whose size/
417 * start address is aligned with PAGE_SIZE. Pinning code uses that
418 * granularity while iommu driver can use the sub-PAGE_SIZE size
421 if (bitmap
& ~PAGE_MASK
) {
429 static int vfio_dma_do_unmap(struct vfio_iommu
*iommu
,
430 struct vfio_iommu_type1_dma_unmap
*unmap
)
433 struct vfio_dma
*dma
;
437 mask
= ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu
))) - 1;
439 if (unmap
->iova
& mask
)
441 if (!unmap
->size
|| unmap
->size
& mask
)
444 WARN_ON(mask
& PAGE_MASK
);
446 mutex_lock(&iommu
->lock
);
449 * vfio-iommu-type1 (v1) - User mappings were coalesced together to
450 * avoid tracking individual mappings. This means that the granularity
451 * of the original mapping was lost and the user was allowed to attempt
452 * to unmap any range. Depending on the contiguousness of physical
453 * memory and page sizes supported by the IOMMU, arbitrary unmaps may
454 * or may not have worked. We only guaranteed unmap granularity
455 * matching the original mapping; even though it was untracked here,
456 * the original mappings are reflected in IOMMU mappings. This
457 * resulted in a couple unusual behaviors. First, if a range is not
458 * able to be unmapped, ex. a set of 4k pages that was mapped as a
459 * 2M hugepage into the IOMMU, the unmap ioctl returns success but with
460 * a zero sized unmap. Also, if an unmap request overlaps the first
461 * address of a hugepage, the IOMMU will unmap the entire hugepage.
462 * This also returns success and the returned unmap size reflects the
463 * actual size unmapped.
465 * We attempt to maintain compatibility with this "v1" interface, but
466 * we take control out of the hands of the IOMMU. Therefore, an unmap
467 * request offset from the beginning of the original mapping will
468 * return success with zero sized unmap. And an unmap request covering
469 * the first iova of mapping will unmap the entire range.
471 * The v2 version of this interface intends to be more deterministic.
472 * Unmap requests must fully cover previous mappings. Multiple
473 * mappings may still be unmaped by specifying large ranges, but there
474 * must not be any previous mappings bisected by the range. An error
475 * will be returned if these conditions are not met. The v2 interface
476 * will only return success and a size of zero if there were no
477 * mappings within the range.
480 dma
= vfio_find_dma(iommu
, unmap
->iova
, 0);
481 if (dma
&& dma
->iova
!= unmap
->iova
) {
485 dma
= vfio_find_dma(iommu
, unmap
->iova
+ unmap
->size
- 1, 0);
486 if (dma
&& dma
->iova
+ dma
->size
!= unmap
->iova
+ unmap
->size
) {
492 while ((dma
= vfio_find_dma(iommu
, unmap
->iova
, unmap
->size
))) {
493 if (!iommu
->v2
&& unmap
->iova
> dma
->iova
)
495 unmapped
+= dma
->size
;
496 vfio_remove_dma(iommu
, dma
);
500 mutex_unlock(&iommu
->lock
);
502 /* Report how much was unmapped */
503 unmap
->size
= unmapped
;
509 * Turns out AMD IOMMU has a page table bug where it won't map large pages
510 * to a region that previously mapped smaller pages. This should be fixed
511 * soon, so this is just a temporary workaround to break mappings down into
512 * PAGE_SIZE. Better to map smaller pages than nothing.
514 static int map_try_harder(struct vfio_domain
*domain
, dma_addr_t iova
,
515 unsigned long pfn
, long npage
, int prot
)
520 for (i
= 0; i
< npage
; i
++, pfn
++, iova
+= PAGE_SIZE
) {
521 ret
= iommu_map(domain
->domain
, iova
,
522 (phys_addr_t
)pfn
<< PAGE_SHIFT
,
523 PAGE_SIZE
, prot
| domain
->prot
);
528 for (; i
< npage
&& i
> 0; i
--, iova
-= PAGE_SIZE
)
529 iommu_unmap(domain
->domain
, iova
, PAGE_SIZE
);
534 static int vfio_iommu_map(struct vfio_iommu
*iommu
, dma_addr_t iova
,
535 unsigned long pfn
, long npage
, int prot
)
537 struct vfio_domain
*d
;
540 list_for_each_entry(d
, &iommu
->domain_list
, next
) {
541 ret
= iommu_map(d
->domain
, iova
, (phys_addr_t
)pfn
<< PAGE_SHIFT
,
542 npage
<< PAGE_SHIFT
, prot
| d
->prot
);
545 map_try_harder(d
, iova
, pfn
, npage
, prot
))
555 list_for_each_entry_continue_reverse(d
, &iommu
->domain_list
, next
)
556 iommu_unmap(d
->domain
, iova
, npage
<< PAGE_SHIFT
);
561 static int vfio_dma_do_map(struct vfio_iommu
*iommu
,
562 struct vfio_iommu_type1_dma_map
*map
)
564 dma_addr_t iova
= map
->iova
;
565 unsigned long vaddr
= map
->vaddr
;
566 size_t size
= map
->size
;
568 int ret
= 0, prot
= 0;
570 struct vfio_dma
*dma
;
573 /* Verify that none of our __u64 fields overflow */
574 if (map
->size
!= size
|| map
->vaddr
!= vaddr
|| map
->iova
!= iova
)
577 mask
= ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu
))) - 1;
579 WARN_ON(mask
& PAGE_MASK
);
581 /* READ/WRITE from device perspective */
582 if (map
->flags
& VFIO_DMA_MAP_FLAG_WRITE
)
584 if (map
->flags
& VFIO_DMA_MAP_FLAG_READ
)
587 if (!prot
|| !size
|| (size
| iova
| vaddr
) & mask
)
590 /* Don't allow IOVA or virtual address wrap */
591 if (iova
+ size
- 1 < iova
|| vaddr
+ size
- 1 < vaddr
)
594 mutex_lock(&iommu
->lock
);
596 if (vfio_find_dma(iommu
, iova
, size
)) {
597 mutex_unlock(&iommu
->lock
);
601 dma
= kzalloc(sizeof(*dma
), GFP_KERNEL
);
603 mutex_unlock(&iommu
->lock
);
611 /* Insert zero-sized and grow as we map chunks of it */
612 vfio_link_dma(iommu
, dma
);
615 /* Pin a contiguous chunk of memory */
616 npage
= vfio_pin_pages(vaddr
+ dma
->size
,
617 size
>> PAGE_SHIFT
, prot
, &pfn
);
625 ret
= vfio_iommu_map(iommu
, iova
+ dma
->size
, pfn
, npage
, prot
);
627 vfio_unpin_pages(pfn
, npage
, prot
, true);
631 size
-= npage
<< PAGE_SHIFT
;
632 dma
->size
+= npage
<< PAGE_SHIFT
;
636 vfio_remove_dma(iommu
, dma
);
638 mutex_unlock(&iommu
->lock
);
642 static int vfio_bus_type(struct device
*dev
, void *data
)
644 struct bus_type
**bus
= data
;
646 if (*bus
&& *bus
!= dev
->bus
)
654 static int vfio_iommu_replay(struct vfio_iommu
*iommu
,
655 struct vfio_domain
*domain
)
657 struct vfio_domain
*d
;
661 /* Arbitrarily pick the first domain in the list for lookups */
662 d
= list_first_entry(&iommu
->domain_list
, struct vfio_domain
, next
);
663 n
= rb_first(&iommu
->dma_list
);
665 /* If there's not a domain, there better not be any mappings */
666 if (WARN_ON(n
&& !d
))
669 for (; n
; n
= rb_next(n
)) {
670 struct vfio_dma
*dma
;
673 dma
= rb_entry(n
, struct vfio_dma
, node
);
676 while (iova
< dma
->iova
+ dma
->size
) {
677 phys_addr_t phys
= iommu_iova_to_phys(d
->domain
, iova
);
680 if (WARN_ON(!phys
)) {
687 while (iova
+ size
< dma
->iova
+ dma
->size
&&
688 phys
+ size
== iommu_iova_to_phys(d
->domain
,
692 ret
= iommu_map(domain
->domain
, iova
, phys
,
693 size
, dma
->prot
| domain
->prot
);
705 * We change our unmap behavior slightly depending on whether the IOMMU
706 * supports fine-grained superpages. IOMMUs like AMD-Vi will use a superpage
707 * for practically any contiguous power-of-two mapping we give it. This means
708 * we don't need to look for contiguous chunks ourselves to make unmapping
709 * more efficient. On IOMMUs with coarse-grained super pages, like Intel VT-d
710 * with discrete 2M/1G/512G/1T superpages, identifying contiguous chunks
711 * significantly boosts non-hugetlbfs mappings and doesn't seem to hurt when
712 * hugetlbfs is in use.
714 static void vfio_test_domain_fgsp(struct vfio_domain
*domain
)
717 int ret
, order
= get_order(PAGE_SIZE
* 2);
719 pages
= alloc_pages(GFP_KERNEL
| __GFP_ZERO
, order
);
723 ret
= iommu_map(domain
->domain
, 0, page_to_phys(pages
), PAGE_SIZE
* 2,
724 IOMMU_READ
| IOMMU_WRITE
| domain
->prot
);
726 size_t unmapped
= iommu_unmap(domain
->domain
, 0, PAGE_SIZE
);
728 if (unmapped
== PAGE_SIZE
)
729 iommu_unmap(domain
->domain
, PAGE_SIZE
, PAGE_SIZE
);
734 __free_pages(pages
, order
);
737 static int vfio_iommu_type1_attach_group(void *iommu_data
,
738 struct iommu_group
*iommu_group
)
740 struct vfio_iommu
*iommu
= iommu_data
;
741 struct vfio_group
*group
, *g
;
742 struct vfio_domain
*domain
, *d
;
743 struct bus_type
*bus
= NULL
;
746 mutex_lock(&iommu
->lock
);
748 list_for_each_entry(d
, &iommu
->domain_list
, next
) {
749 list_for_each_entry(g
, &d
->group_list
, next
) {
750 if (g
->iommu_group
!= iommu_group
)
753 mutex_unlock(&iommu
->lock
);
758 group
= kzalloc(sizeof(*group
), GFP_KERNEL
);
759 domain
= kzalloc(sizeof(*domain
), GFP_KERNEL
);
760 if (!group
|| !domain
) {
765 group
->iommu_group
= iommu_group
;
767 /* Determine bus_type in order to allocate a domain */
768 ret
= iommu_group_for_each_dev(iommu_group
, &bus
, vfio_bus_type
);
772 domain
->domain
= iommu_domain_alloc(bus
);
773 if (!domain
->domain
) {
778 if (iommu
->nesting
) {
781 ret
= iommu_domain_set_attr(domain
->domain
, DOMAIN_ATTR_NESTING
,
787 ret
= iommu_attach_group(domain
->domain
, iommu_group
);
791 INIT_LIST_HEAD(&domain
->group_list
);
792 list_add(&group
->next
, &domain
->group_list
);
794 if (!allow_unsafe_interrupts
&&
795 !iommu_capable(bus
, IOMMU_CAP_INTR_REMAP
)) {
796 pr_warn("%s: No interrupt remapping support. Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n",
802 if (iommu_capable(bus
, IOMMU_CAP_CACHE_COHERENCY
))
803 domain
->prot
|= IOMMU_CACHE
;
806 * Try to match an existing compatible domain. We don't want to
807 * preclude an IOMMU driver supporting multiple bus_types and being
808 * able to include different bus_types in the same IOMMU domain, so
809 * we test whether the domains use the same iommu_ops rather than
810 * testing if they're on the same bus_type.
812 list_for_each_entry(d
, &iommu
->domain_list
, next
) {
813 if (d
->domain
->ops
== domain
->domain
->ops
&&
814 d
->prot
== domain
->prot
) {
815 iommu_detach_group(domain
->domain
, iommu_group
);
816 if (!iommu_attach_group(d
->domain
, iommu_group
)) {
817 list_add(&group
->next
, &d
->group_list
);
818 iommu_domain_free(domain
->domain
);
820 mutex_unlock(&iommu
->lock
);
824 ret
= iommu_attach_group(domain
->domain
, iommu_group
);
830 vfio_test_domain_fgsp(domain
);
832 /* replay mappings on new domains */
833 ret
= vfio_iommu_replay(iommu
, domain
);
837 list_add(&domain
->next
, &iommu
->domain_list
);
839 mutex_unlock(&iommu
->lock
);
844 iommu_detach_group(domain
->domain
, iommu_group
);
846 iommu_domain_free(domain
->domain
);
850 mutex_unlock(&iommu
->lock
);
854 static void vfio_iommu_unmap_unpin_all(struct vfio_iommu
*iommu
)
856 struct rb_node
*node
;
858 while ((node
= rb_first(&iommu
->dma_list
)))
859 vfio_remove_dma(iommu
, rb_entry(node
, struct vfio_dma
, node
));
862 static void vfio_iommu_type1_detach_group(void *iommu_data
,
863 struct iommu_group
*iommu_group
)
865 struct vfio_iommu
*iommu
= iommu_data
;
866 struct vfio_domain
*domain
;
867 struct vfio_group
*group
;
869 mutex_lock(&iommu
->lock
);
871 list_for_each_entry(domain
, &iommu
->domain_list
, next
) {
872 list_for_each_entry(group
, &domain
->group_list
, next
) {
873 if (group
->iommu_group
!= iommu_group
)
876 iommu_detach_group(domain
->domain
, iommu_group
);
877 list_del(&group
->next
);
880 * Group ownership provides privilege, if the group
881 * list is empty, the domain goes away. If it's the
882 * last domain, then all the mappings go away too.
884 if (list_empty(&domain
->group_list
)) {
885 if (list_is_singular(&iommu
->domain_list
))
886 vfio_iommu_unmap_unpin_all(iommu
);
887 iommu_domain_free(domain
->domain
);
888 list_del(&domain
->next
);
896 mutex_unlock(&iommu
->lock
);
899 static void *vfio_iommu_type1_open(unsigned long arg
)
901 struct vfio_iommu
*iommu
;
903 iommu
= kzalloc(sizeof(*iommu
), GFP_KERNEL
);
905 return ERR_PTR(-ENOMEM
);
908 case VFIO_TYPE1_IOMMU
:
910 case VFIO_TYPE1_NESTING_IOMMU
:
911 iommu
->nesting
= true;
912 case VFIO_TYPE1v2_IOMMU
:
917 return ERR_PTR(-EINVAL
);
920 INIT_LIST_HEAD(&iommu
->domain_list
);
921 iommu
->dma_list
= RB_ROOT
;
922 mutex_init(&iommu
->lock
);
927 static void vfio_iommu_type1_release(void *iommu_data
)
929 struct vfio_iommu
*iommu
= iommu_data
;
930 struct vfio_domain
*domain
, *domain_tmp
;
931 struct vfio_group
*group
, *group_tmp
;
933 vfio_iommu_unmap_unpin_all(iommu
);
935 list_for_each_entry_safe(domain
, domain_tmp
,
936 &iommu
->domain_list
, next
) {
937 list_for_each_entry_safe(group
, group_tmp
,
938 &domain
->group_list
, next
) {
939 iommu_detach_group(domain
->domain
, group
->iommu_group
);
940 list_del(&group
->next
);
943 iommu_domain_free(domain
->domain
);
944 list_del(&domain
->next
);
951 static int vfio_domains_have_iommu_cache(struct vfio_iommu
*iommu
)
953 struct vfio_domain
*domain
;
956 mutex_lock(&iommu
->lock
);
957 list_for_each_entry(domain
, &iommu
->domain_list
, next
) {
958 if (!(domain
->prot
& IOMMU_CACHE
)) {
963 mutex_unlock(&iommu
->lock
);
968 static long vfio_iommu_type1_ioctl(void *iommu_data
,
969 unsigned int cmd
, unsigned long arg
)
971 struct vfio_iommu
*iommu
= iommu_data
;
974 if (cmd
== VFIO_CHECK_EXTENSION
) {
976 case VFIO_TYPE1_IOMMU
:
977 case VFIO_TYPE1v2_IOMMU
:
978 case VFIO_TYPE1_NESTING_IOMMU
:
980 case VFIO_DMA_CC_IOMMU
:
983 return vfio_domains_have_iommu_cache(iommu
);
987 } else if (cmd
== VFIO_IOMMU_GET_INFO
) {
988 struct vfio_iommu_type1_info info
;
990 minsz
= offsetofend(struct vfio_iommu_type1_info
, iova_pgsizes
);
992 if (copy_from_user(&info
, (void __user
*)arg
, minsz
))
995 if (info
.argsz
< minsz
)
998 info
.flags
= VFIO_IOMMU_INFO_PGSIZES
;
1000 info
.iova_pgsizes
= vfio_pgsize_bitmap(iommu
);
1002 return copy_to_user((void __user
*)arg
, &info
, minsz
);
1004 } else if (cmd
== VFIO_IOMMU_MAP_DMA
) {
1005 struct vfio_iommu_type1_dma_map map
;
1006 uint32_t mask
= VFIO_DMA_MAP_FLAG_READ
|
1007 VFIO_DMA_MAP_FLAG_WRITE
;
1009 minsz
= offsetofend(struct vfio_iommu_type1_dma_map
, size
);
1011 if (copy_from_user(&map
, (void __user
*)arg
, minsz
))
1014 if (map
.argsz
< minsz
|| map
.flags
& ~mask
)
1017 return vfio_dma_do_map(iommu
, &map
);
1019 } else if (cmd
== VFIO_IOMMU_UNMAP_DMA
) {
1020 struct vfio_iommu_type1_dma_unmap unmap
;
1023 minsz
= offsetofend(struct vfio_iommu_type1_dma_unmap
, size
);
1025 if (copy_from_user(&unmap
, (void __user
*)arg
, minsz
))
1028 if (unmap
.argsz
< minsz
|| unmap
.flags
)
1031 ret
= vfio_dma_do_unmap(iommu
, &unmap
);
1035 return copy_to_user((void __user
*)arg
, &unmap
, minsz
);
1041 static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1
= {
1042 .name
= "vfio-iommu-type1",
1043 .owner
= THIS_MODULE
,
1044 .open
= vfio_iommu_type1_open
,
1045 .release
= vfio_iommu_type1_release
,
1046 .ioctl
= vfio_iommu_type1_ioctl
,
1047 .attach_group
= vfio_iommu_type1_attach_group
,
1048 .detach_group
= vfio_iommu_type1_detach_group
,
1051 static int __init
vfio_iommu_type1_init(void)
1053 return vfio_register_iommu_driver(&vfio_iommu_driver_ops_type1
);
1056 static void __exit
vfio_iommu_type1_cleanup(void)
1058 vfio_unregister_iommu_driver(&vfio_iommu_driver_ops_type1
);
1061 module_init(vfio_iommu_type1_init
);
1062 module_exit(vfio_iommu_type1_cleanup
);
1064 MODULE_VERSION(DRIVER_VERSION
);
1065 MODULE_LICENSE("GPL v2");
1066 MODULE_AUTHOR(DRIVER_AUTHOR
);
1067 MODULE_DESCRIPTION(DRIVER_DESC
);