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.");
62 struct list_head domain_list
;
63 struct vfio_domain
*external_domain
; /* domain for external user */
65 struct rb_root dma_list
;
66 struct blocking_notifier_head notifier
;
72 struct iommu_domain
*domain
;
73 struct list_head next
;
74 struct list_head group_list
;
75 int prot
; /* IOMMU_CACHE */
76 bool fgsp
; /* Fine-grained super pages */
81 dma_addr_t iova
; /* Device address */
82 unsigned long vaddr
; /* Process virtual addr */
83 size_t size
; /* Map size (bytes) */
84 int prot
; /* IOMMU_READ/WRITE */
86 struct task_struct
*task
;
87 struct rb_root pfn_list
; /* Ex-user pinned pfn list */
91 struct iommu_group
*iommu_group
;
92 struct list_head next
;
96 * Guest RAM pinning working set or DMA target
100 dma_addr_t iova
; /* Device address */
101 unsigned long pfn
; /* Host pfn */
105 #define IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu) \
106 (!list_empty(&iommu->domain_list))
108 static int put_pfn(unsigned long pfn
, int prot
);
111 * This code handles mapping and unmapping of user data buffers
112 * into DMA'ble space using the IOMMU
115 static struct vfio_dma
*vfio_find_dma(struct vfio_iommu
*iommu
,
116 dma_addr_t start
, size_t size
)
118 struct rb_node
*node
= iommu
->dma_list
.rb_node
;
121 struct vfio_dma
*dma
= rb_entry(node
, struct vfio_dma
, node
);
123 if (start
+ size
<= dma
->iova
)
124 node
= node
->rb_left
;
125 else if (start
>= dma
->iova
+ dma
->size
)
126 node
= node
->rb_right
;
134 static void vfio_link_dma(struct vfio_iommu
*iommu
, struct vfio_dma
*new)
136 struct rb_node
**link
= &iommu
->dma_list
.rb_node
, *parent
= NULL
;
137 struct vfio_dma
*dma
;
141 dma
= rb_entry(parent
, struct vfio_dma
, node
);
143 if (new->iova
+ new->size
<= dma
->iova
)
144 link
= &(*link
)->rb_left
;
146 link
= &(*link
)->rb_right
;
149 rb_link_node(&new->node
, parent
, link
);
150 rb_insert_color(&new->node
, &iommu
->dma_list
);
153 static void vfio_unlink_dma(struct vfio_iommu
*iommu
, struct vfio_dma
*old
)
155 rb_erase(&old
->node
, &iommu
->dma_list
);
159 * Helper Functions for host iova-pfn list
161 static struct vfio_pfn
*vfio_find_vpfn(struct vfio_dma
*dma
, dma_addr_t iova
)
163 struct vfio_pfn
*vpfn
;
164 struct rb_node
*node
= dma
->pfn_list
.rb_node
;
167 vpfn
= rb_entry(node
, struct vfio_pfn
, node
);
169 if (iova
< vpfn
->iova
)
170 node
= node
->rb_left
;
171 else if (iova
> vpfn
->iova
)
172 node
= node
->rb_right
;
179 static void vfio_link_pfn(struct vfio_dma
*dma
,
180 struct vfio_pfn
*new)
182 struct rb_node
**link
, *parent
= NULL
;
183 struct vfio_pfn
*vpfn
;
185 link
= &dma
->pfn_list
.rb_node
;
188 vpfn
= rb_entry(parent
, struct vfio_pfn
, node
);
190 if (new->iova
< vpfn
->iova
)
191 link
= &(*link
)->rb_left
;
193 link
= &(*link
)->rb_right
;
196 rb_link_node(&new->node
, parent
, link
);
197 rb_insert_color(&new->node
, &dma
->pfn_list
);
200 static void vfio_unlink_pfn(struct vfio_dma
*dma
, struct vfio_pfn
*old
)
202 rb_erase(&old
->node
, &dma
->pfn_list
);
205 static int vfio_add_to_pfn_list(struct vfio_dma
*dma
, dma_addr_t iova
,
208 struct vfio_pfn
*vpfn
;
210 vpfn
= kzalloc(sizeof(*vpfn
), GFP_KERNEL
);
216 atomic_set(&vpfn
->ref_count
, 1);
217 vfio_link_pfn(dma
, vpfn
);
221 static void vfio_remove_from_pfn_list(struct vfio_dma
*dma
,
222 struct vfio_pfn
*vpfn
)
224 vfio_unlink_pfn(dma
, vpfn
);
228 static struct vfio_pfn
*vfio_iova_get_vfio_pfn(struct vfio_dma
*dma
,
231 struct vfio_pfn
*vpfn
= vfio_find_vpfn(dma
, iova
);
234 atomic_inc(&vpfn
->ref_count
);
238 static int vfio_iova_put_vfio_pfn(struct vfio_dma
*dma
, struct vfio_pfn
*vpfn
)
242 if (atomic_dec_and_test(&vpfn
->ref_count
)) {
243 ret
= put_pfn(vpfn
->pfn
, dma
->prot
);
244 vfio_remove_from_pfn_list(dma
, vpfn
);
249 static int vfio_lock_acct(struct task_struct
*task
, long npage
, bool *lock_cap
)
251 struct mm_struct
*mm
;
258 is_current
= (task
->mm
== current
->mm
);
260 mm
= is_current
? task
->mm
: get_task_mm(task
);
262 return -ESRCH
; /* process exited */
264 ret
= down_write_killable(&mm
->mmap_sem
);
267 if (lock_cap
? !*lock_cap
:
268 !has_capability(task
, CAP_IPC_LOCK
)) {
271 limit
= task_rlimit(task
,
272 RLIMIT_MEMLOCK
) >> PAGE_SHIFT
;
274 if (mm
->locked_vm
+ npage
> limit
)
280 mm
->locked_vm
+= npage
;
282 up_write(&mm
->mmap_sem
);
292 * Some mappings aren't backed by a struct page, for example an mmap'd
293 * MMIO range for our own or another device. These use a different
294 * pfn conversion and shouldn't be tracked as locked pages.
296 static bool is_invalid_reserved_pfn(unsigned long pfn
)
298 if (pfn_valid(pfn
)) {
300 struct page
*tail
= pfn_to_page(pfn
);
301 struct page
*head
= compound_head(tail
);
302 reserved
= !!(PageReserved(head
));
305 * "head" is not a dangling pointer
306 * (compound_head takes care of that)
307 * but the hugepage may have been split
308 * from under us (and we may not hold a
309 * reference count on the head page so it can
310 * be reused before we run PageReferenced), so
311 * we've to check PageTail before returning
318 return PageReserved(tail
);
324 static int put_pfn(unsigned long pfn
, int prot
)
326 if (!is_invalid_reserved_pfn(pfn
)) {
327 struct page
*page
= pfn_to_page(pfn
);
328 if (prot
& IOMMU_WRITE
)
336 static int vaddr_get_pfn(struct mm_struct
*mm
, unsigned long vaddr
,
337 int prot
, unsigned long *pfn
)
339 struct page
*page
[1];
340 struct vm_area_struct
*vma
;
343 if (mm
== current
->mm
) {
344 ret
= get_user_pages_fast(vaddr
, 1, !!(prot
& IOMMU_WRITE
),
347 unsigned int flags
= 0;
349 if (prot
& IOMMU_WRITE
)
352 down_read(&mm
->mmap_sem
);
353 ret
= get_user_pages_remote(NULL
, mm
, vaddr
, 1, flags
, page
,
355 up_read(&mm
->mmap_sem
);
359 *pfn
= page_to_pfn(page
[0]);
363 down_read(&mm
->mmap_sem
);
365 vma
= find_vma_intersection(mm
, vaddr
, vaddr
+ 1);
367 if (vma
&& vma
->vm_flags
& VM_PFNMAP
) {
368 *pfn
= ((vaddr
- vma
->vm_start
) >> PAGE_SHIFT
) + vma
->vm_pgoff
;
369 if (is_invalid_reserved_pfn(*pfn
))
373 up_read(&mm
->mmap_sem
);
378 * Attempt to pin pages. We really don't want to track all the pfns and
379 * the iommu can only map chunks of consecutive pfns anyway, so get the
380 * first page and all consecutive pages with the same locking.
382 static long vfio_pin_pages_remote(struct vfio_dma
*dma
, unsigned long vaddr
,
383 long npage
, unsigned long *pfn_base
,
384 bool lock_cap
, unsigned long limit
)
386 unsigned long pfn
= 0;
387 long ret
, pinned
= 0, lock_acct
= 0;
389 dma_addr_t iova
= vaddr
- dma
->vaddr
+ dma
->iova
;
391 /* This code path is only user initiated */
395 ret
= vaddr_get_pfn(current
->mm
, vaddr
, dma
->prot
, pfn_base
);
400 rsvd
= is_invalid_reserved_pfn(*pfn_base
);
403 * Reserved pages aren't counted against the user, externally pinned
404 * pages are already counted against the user.
406 if (!rsvd
&& !vfio_find_vpfn(dma
, iova
)) {
407 if (!lock_cap
&& current
->mm
->locked_vm
+ 1 > limit
) {
408 put_pfn(*pfn_base
, dma
->prot
);
409 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n", __func__
,
410 limit
<< PAGE_SHIFT
);
416 if (unlikely(disable_hugepages
))
419 /* Lock all the consecutive pages from pfn_base */
420 for (vaddr
+= PAGE_SIZE
, iova
+= PAGE_SIZE
; pinned
< npage
;
421 pinned
++, vaddr
+= PAGE_SIZE
, iova
+= PAGE_SIZE
) {
422 ret
= vaddr_get_pfn(current
->mm
, vaddr
, dma
->prot
, &pfn
);
426 if (pfn
!= *pfn_base
+ pinned
||
427 rsvd
!= is_invalid_reserved_pfn(pfn
)) {
428 put_pfn(pfn
, dma
->prot
);
432 if (!rsvd
&& !vfio_find_vpfn(dma
, iova
)) {
434 current
->mm
->locked_vm
+ lock_acct
+ 1 > limit
) {
435 put_pfn(pfn
, dma
->prot
);
436 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n",
437 __func__
, limit
<< PAGE_SHIFT
);
446 ret
= vfio_lock_acct(current
, lock_acct
, &lock_cap
);
451 for (pfn
= *pfn_base
; pinned
; pfn
++, pinned
--)
452 put_pfn(pfn
, dma
->prot
);
461 static long vfio_unpin_pages_remote(struct vfio_dma
*dma
, dma_addr_t iova
,
462 unsigned long pfn
, long npage
,
465 long unlocked
= 0, locked
= 0;
468 for (i
= 0; i
< npage
; i
++, iova
+= PAGE_SIZE
) {
469 if (put_pfn(pfn
++, dma
->prot
)) {
471 if (vfio_find_vpfn(dma
, iova
))
477 vfio_lock_acct(dma
->task
, locked
- unlocked
, NULL
);
482 static int vfio_pin_page_external(struct vfio_dma
*dma
, unsigned long vaddr
,
483 unsigned long *pfn_base
, bool do_accounting
)
485 struct mm_struct
*mm
;
488 mm
= get_task_mm(dma
->task
);
492 ret
= vaddr_get_pfn(mm
, vaddr
, dma
->prot
, pfn_base
);
493 if (!ret
&& do_accounting
&& !is_invalid_reserved_pfn(*pfn_base
)) {
494 ret
= vfio_lock_acct(dma
->task
, 1, NULL
);
496 put_pfn(*pfn_base
, dma
->prot
);
498 pr_warn("%s: Task %s (%d) RLIMIT_MEMLOCK "
499 "(%ld) exceeded\n", __func__
,
500 dma
->task
->comm
, task_pid_nr(dma
->task
),
501 task_rlimit(dma
->task
, RLIMIT_MEMLOCK
));
509 static int vfio_unpin_page_external(struct vfio_dma
*dma
, dma_addr_t iova
,
513 struct vfio_pfn
*vpfn
= vfio_find_vpfn(dma
, iova
);
518 unlocked
= vfio_iova_put_vfio_pfn(dma
, vpfn
);
521 vfio_lock_acct(dma
->task
, -unlocked
, NULL
);
526 static int vfio_iommu_type1_pin_pages(void *iommu_data
,
527 unsigned long *user_pfn
,
529 unsigned long *phys_pfn
)
531 struct vfio_iommu
*iommu
= iommu_data
;
533 unsigned long remote_vaddr
;
534 struct vfio_dma
*dma
;
537 if (!iommu
|| !user_pfn
|| !phys_pfn
)
540 /* Supported for v2 version only */
544 mutex_lock(&iommu
->lock
);
546 /* Fail if notifier list is empty */
547 if ((!iommu
->external_domain
) || (!iommu
->notifier
.head
)) {
553 * If iommu capable domain exist in the container then all pages are
554 * already pinned and accounted. Accouting should be done if there is no
555 * iommu capable domain in the container.
557 do_accounting
= !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu
);
559 for (i
= 0; i
< npage
; i
++) {
561 struct vfio_pfn
*vpfn
;
563 iova
= user_pfn
[i
] << PAGE_SHIFT
;
564 dma
= vfio_find_dma(iommu
, iova
, PAGE_SIZE
);
570 if ((dma
->prot
& prot
) != prot
) {
575 vpfn
= vfio_iova_get_vfio_pfn(dma
, iova
);
577 phys_pfn
[i
] = vpfn
->pfn
;
581 remote_vaddr
= dma
->vaddr
+ iova
- dma
->iova
;
582 ret
= vfio_pin_page_external(dma
, remote_vaddr
, &phys_pfn
[i
],
587 ret
= vfio_add_to_pfn_list(dma
, iova
, phys_pfn
[i
]);
589 vfio_unpin_page_external(dma
, iova
, do_accounting
);
599 for (j
= 0; j
< i
; j
++) {
602 iova
= user_pfn
[j
] << PAGE_SHIFT
;
603 dma
= vfio_find_dma(iommu
, iova
, PAGE_SIZE
);
604 vfio_unpin_page_external(dma
, iova
, do_accounting
);
608 mutex_unlock(&iommu
->lock
);
612 static int vfio_iommu_type1_unpin_pages(void *iommu_data
,
613 unsigned long *user_pfn
,
616 struct vfio_iommu
*iommu
= iommu_data
;
620 if (!iommu
|| !user_pfn
)
623 /* Supported for v2 version only */
627 mutex_lock(&iommu
->lock
);
629 if (!iommu
->external_domain
) {
630 mutex_unlock(&iommu
->lock
);
634 do_accounting
= !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu
);
635 for (i
= 0; i
< npage
; i
++) {
636 struct vfio_dma
*dma
;
639 iova
= user_pfn
[i
] << PAGE_SHIFT
;
640 dma
= vfio_find_dma(iommu
, iova
, PAGE_SIZE
);
643 vfio_unpin_page_external(dma
, iova
, do_accounting
);
647 mutex_unlock(&iommu
->lock
);
648 return i
> npage
? npage
: (i
> 0 ? i
: -EINVAL
);
651 static long vfio_unmap_unpin(struct vfio_iommu
*iommu
, struct vfio_dma
*dma
,
654 dma_addr_t iova
= dma
->iova
, end
= dma
->iova
+ dma
->size
;
655 struct vfio_domain
*domain
, *d
;
661 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu
))
665 * We use the IOMMU to track the physical addresses, otherwise we'd
666 * need a much more complicated tracking system. Unfortunately that
667 * means we need to use one of the iommu domains to figure out the
668 * pfns to unpin. The rest need to be unmapped in advance so we have
669 * no iommu translations remaining when the pages are unpinned.
671 domain
= d
= list_first_entry(&iommu
->domain_list
,
672 struct vfio_domain
, next
);
674 list_for_each_entry_continue(d
, &iommu
->domain_list
, next
) {
675 iommu_unmap(d
->domain
, dma
->iova
, dma
->size
);
680 size_t unmapped
, len
;
681 phys_addr_t phys
, next
;
683 phys
= iommu_iova_to_phys(domain
->domain
, iova
);
684 if (WARN_ON(!phys
)) {
690 * To optimize for fewer iommu_unmap() calls, each of which
691 * may require hardware cache flushing, try to find the
692 * largest contiguous physical memory chunk to unmap.
694 for (len
= PAGE_SIZE
;
695 !domain
->fgsp
&& iova
+ len
< end
; len
+= PAGE_SIZE
) {
696 next
= iommu_iova_to_phys(domain
->domain
, iova
+ len
);
697 if (next
!= phys
+ len
)
701 unmapped
= iommu_unmap(domain
->domain
, iova
, len
);
702 if (WARN_ON(!unmapped
))
705 unlocked
+= vfio_unpin_pages_remote(dma
, iova
,
707 unmapped
>> PAGE_SHIFT
,
714 dma
->iommu_mapped
= false;
716 vfio_lock_acct(dma
->task
, -unlocked
, NULL
);
722 static void vfio_remove_dma(struct vfio_iommu
*iommu
, struct vfio_dma
*dma
)
724 vfio_unmap_unpin(iommu
, dma
, true);
725 vfio_unlink_dma(iommu
, dma
);
726 put_task_struct(dma
->task
);
730 static unsigned long vfio_pgsize_bitmap(struct vfio_iommu
*iommu
)
732 struct vfio_domain
*domain
;
733 unsigned long bitmap
= ULONG_MAX
;
735 mutex_lock(&iommu
->lock
);
736 list_for_each_entry(domain
, &iommu
->domain_list
, next
)
737 bitmap
&= domain
->domain
->pgsize_bitmap
;
738 mutex_unlock(&iommu
->lock
);
741 * In case the IOMMU supports page sizes smaller than PAGE_SIZE
742 * we pretend PAGE_SIZE is supported and hide sub-PAGE_SIZE sizes.
743 * That way the user will be able to map/unmap buffers whose size/
744 * start address is aligned with PAGE_SIZE. Pinning code uses that
745 * granularity while iommu driver can use the sub-PAGE_SIZE size
748 if (bitmap
& ~PAGE_MASK
) {
756 static int vfio_dma_do_unmap(struct vfio_iommu
*iommu
,
757 struct vfio_iommu_type1_dma_unmap
*unmap
)
760 struct vfio_dma
*dma
, *dma_last
= NULL
;
762 int ret
= 0, retries
= 0;
764 mask
= ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu
))) - 1;
766 if (unmap
->iova
& mask
)
768 if (!unmap
->size
|| unmap
->size
& mask
)
770 if (unmap
->iova
+ unmap
->size
< unmap
->iova
||
771 unmap
->size
> SIZE_MAX
)
774 WARN_ON(mask
& PAGE_MASK
);
776 mutex_lock(&iommu
->lock
);
779 * vfio-iommu-type1 (v1) - User mappings were coalesced together to
780 * avoid tracking individual mappings. This means that the granularity
781 * of the original mapping was lost and the user was allowed to attempt
782 * to unmap any range. Depending on the contiguousness of physical
783 * memory and page sizes supported by the IOMMU, arbitrary unmaps may
784 * or may not have worked. We only guaranteed unmap granularity
785 * matching the original mapping; even though it was untracked here,
786 * the original mappings are reflected in IOMMU mappings. This
787 * resulted in a couple unusual behaviors. First, if a range is not
788 * able to be unmapped, ex. a set of 4k pages that was mapped as a
789 * 2M hugepage into the IOMMU, the unmap ioctl returns success but with
790 * a zero sized unmap. Also, if an unmap request overlaps the first
791 * address of a hugepage, the IOMMU will unmap the entire hugepage.
792 * This also returns success and the returned unmap size reflects the
793 * actual size unmapped.
795 * We attempt to maintain compatibility with this "v1" interface, but
796 * we take control out of the hands of the IOMMU. Therefore, an unmap
797 * request offset from the beginning of the original mapping will
798 * return success with zero sized unmap. And an unmap request covering
799 * the first iova of mapping will unmap the entire range.
801 * The v2 version of this interface intends to be more deterministic.
802 * Unmap requests must fully cover previous mappings. Multiple
803 * mappings may still be unmaped by specifying large ranges, but there
804 * must not be any previous mappings bisected by the range. An error
805 * will be returned if these conditions are not met. The v2 interface
806 * will only return success and a size of zero if there were no
807 * mappings within the range.
810 dma
= vfio_find_dma(iommu
, unmap
->iova
, 1);
811 if (dma
&& dma
->iova
!= unmap
->iova
) {
815 dma
= vfio_find_dma(iommu
, unmap
->iova
+ unmap
->size
- 1, 0);
816 if (dma
&& dma
->iova
+ dma
->size
!= unmap
->iova
+ unmap
->size
) {
822 while ((dma
= vfio_find_dma(iommu
, unmap
->iova
, unmap
->size
))) {
823 if (!iommu
->v2
&& unmap
->iova
> dma
->iova
)
826 * Task with same address space who mapped this iova range is
827 * allowed to unmap the iova range.
829 if (dma
->task
->mm
!= current
->mm
)
832 if (!RB_EMPTY_ROOT(&dma
->pfn_list
)) {
833 struct vfio_iommu_type1_dma_unmap nb_unmap
;
835 if (dma_last
== dma
) {
836 BUG_ON(++retries
> 10);
842 nb_unmap
.iova
= dma
->iova
;
843 nb_unmap
.size
= dma
->size
;
846 * Notify anyone (mdev vendor drivers) to invalidate and
847 * unmap iovas within the range we're about to unmap.
848 * Vendor drivers MUST unpin pages in response to an
851 mutex_unlock(&iommu
->lock
);
852 blocking_notifier_call_chain(&iommu
->notifier
,
853 VFIO_IOMMU_NOTIFY_DMA_UNMAP
,
857 unmapped
+= dma
->size
;
858 vfio_remove_dma(iommu
, dma
);
862 mutex_unlock(&iommu
->lock
);
864 /* Report how much was unmapped */
865 unmap
->size
= unmapped
;
871 * Turns out AMD IOMMU has a page table bug where it won't map large pages
872 * to a region that previously mapped smaller pages. This should be fixed
873 * soon, so this is just a temporary workaround to break mappings down into
874 * PAGE_SIZE. Better to map smaller pages than nothing.
876 static int map_try_harder(struct vfio_domain
*domain
, dma_addr_t iova
,
877 unsigned long pfn
, long npage
, int prot
)
882 for (i
= 0; i
< npage
; i
++, pfn
++, iova
+= PAGE_SIZE
) {
883 ret
= iommu_map(domain
->domain
, iova
,
884 (phys_addr_t
)pfn
<< PAGE_SHIFT
,
885 PAGE_SIZE
, prot
| domain
->prot
);
890 for (; i
< npage
&& i
> 0; i
--, iova
-= PAGE_SIZE
)
891 iommu_unmap(domain
->domain
, iova
, PAGE_SIZE
);
896 static int vfio_iommu_map(struct vfio_iommu
*iommu
, dma_addr_t iova
,
897 unsigned long pfn
, long npage
, int prot
)
899 struct vfio_domain
*d
;
902 list_for_each_entry(d
, &iommu
->domain_list
, next
) {
903 ret
= iommu_map(d
->domain
, iova
, (phys_addr_t
)pfn
<< PAGE_SHIFT
,
904 npage
<< PAGE_SHIFT
, prot
| d
->prot
);
907 map_try_harder(d
, iova
, pfn
, npage
, prot
))
917 list_for_each_entry_continue_reverse(d
, &iommu
->domain_list
, next
)
918 iommu_unmap(d
->domain
, iova
, npage
<< PAGE_SHIFT
);
923 static int vfio_pin_map_dma(struct vfio_iommu
*iommu
, struct vfio_dma
*dma
,
926 dma_addr_t iova
= dma
->iova
;
927 unsigned long vaddr
= dma
->vaddr
;
928 size_t size
= map_size
;
930 unsigned long pfn
, limit
= rlimit(RLIMIT_MEMLOCK
) >> PAGE_SHIFT
;
931 bool lock_cap
= capable(CAP_IPC_LOCK
);
935 /* Pin a contiguous chunk of memory */
936 npage
= vfio_pin_pages_remote(dma
, vaddr
+ dma
->size
,
937 size
>> PAGE_SHIFT
, &pfn
,
946 ret
= vfio_iommu_map(iommu
, iova
+ dma
->size
, pfn
, npage
,
949 vfio_unpin_pages_remote(dma
, iova
+ dma
->size
, pfn
,
954 size
-= npage
<< PAGE_SHIFT
;
955 dma
->size
+= npage
<< PAGE_SHIFT
;
958 dma
->iommu_mapped
= true;
961 vfio_remove_dma(iommu
, dma
);
966 static int vfio_dma_do_map(struct vfio_iommu
*iommu
,
967 struct vfio_iommu_type1_dma_map
*map
)
969 dma_addr_t iova
= map
->iova
;
970 unsigned long vaddr
= map
->vaddr
;
971 size_t size
= map
->size
;
972 int ret
= 0, prot
= 0;
974 struct vfio_dma
*dma
;
976 /* Verify that none of our __u64 fields overflow */
977 if (map
->size
!= size
|| map
->vaddr
!= vaddr
|| map
->iova
!= iova
)
980 mask
= ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu
))) - 1;
982 WARN_ON(mask
& PAGE_MASK
);
984 /* READ/WRITE from device perspective */
985 if (map
->flags
& VFIO_DMA_MAP_FLAG_WRITE
)
987 if (map
->flags
& VFIO_DMA_MAP_FLAG_READ
)
990 if (!prot
|| !size
|| (size
| iova
| vaddr
) & mask
)
993 /* Don't allow IOVA or virtual address wrap */
994 if (iova
+ size
- 1 < iova
|| vaddr
+ size
- 1 < vaddr
)
997 mutex_lock(&iommu
->lock
);
999 if (vfio_find_dma(iommu
, iova
, size
)) {
1004 dma
= kzalloc(sizeof(*dma
), GFP_KERNEL
);
1013 get_task_struct(current
);
1014 dma
->task
= current
;
1015 dma
->pfn_list
= RB_ROOT
;
1017 /* Insert zero-sized and grow as we map chunks of it */
1018 vfio_link_dma(iommu
, dma
);
1020 /* Don't pin and map if container doesn't contain IOMMU capable domain*/
1021 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu
))
1024 ret
= vfio_pin_map_dma(iommu
, dma
, size
);
1027 mutex_unlock(&iommu
->lock
);
1031 static int vfio_bus_type(struct device
*dev
, void *data
)
1033 struct bus_type
**bus
= data
;
1035 if (*bus
&& *bus
!= dev
->bus
)
1043 static int vfio_iommu_replay(struct vfio_iommu
*iommu
,
1044 struct vfio_domain
*domain
)
1046 struct vfio_domain
*d
;
1048 unsigned long limit
= rlimit(RLIMIT_MEMLOCK
) >> PAGE_SHIFT
;
1049 bool lock_cap
= capable(CAP_IPC_LOCK
);
1052 /* Arbitrarily pick the first domain in the list for lookups */
1053 d
= list_first_entry(&iommu
->domain_list
, struct vfio_domain
, next
);
1054 n
= rb_first(&iommu
->dma_list
);
1056 for (; n
; n
= rb_next(n
)) {
1057 struct vfio_dma
*dma
;
1060 dma
= rb_entry(n
, struct vfio_dma
, node
);
1063 while (iova
< dma
->iova
+ dma
->size
) {
1067 if (dma
->iommu_mapped
) {
1071 phys
= iommu_iova_to_phys(d
->domain
, iova
);
1073 if (WARN_ON(!phys
)) {
1081 while (i
< dma
->iova
+ dma
->size
&&
1082 p
== iommu_iova_to_phys(d
->domain
, i
)) {
1089 unsigned long vaddr
= dma
->vaddr
+
1091 size_t n
= dma
->iova
+ dma
->size
- iova
;
1094 npage
= vfio_pin_pages_remote(dma
, vaddr
,
1104 phys
= pfn
<< PAGE_SHIFT
;
1105 size
= npage
<< PAGE_SHIFT
;
1108 ret
= iommu_map(domain
->domain
, iova
, phys
,
1109 size
, dma
->prot
| domain
->prot
);
1115 dma
->iommu_mapped
= true;
1121 * We change our unmap behavior slightly depending on whether the IOMMU
1122 * supports fine-grained superpages. IOMMUs like AMD-Vi will use a superpage
1123 * for practically any contiguous power-of-two mapping we give it. This means
1124 * we don't need to look for contiguous chunks ourselves to make unmapping
1125 * more efficient. On IOMMUs with coarse-grained super pages, like Intel VT-d
1126 * with discrete 2M/1G/512G/1T superpages, identifying contiguous chunks
1127 * significantly boosts non-hugetlbfs mappings and doesn't seem to hurt when
1128 * hugetlbfs is in use.
1130 static void vfio_test_domain_fgsp(struct vfio_domain
*domain
)
1133 int ret
, order
= get_order(PAGE_SIZE
* 2);
1135 pages
= alloc_pages(GFP_KERNEL
| __GFP_ZERO
, order
);
1139 ret
= iommu_map(domain
->domain
, 0, page_to_phys(pages
), PAGE_SIZE
* 2,
1140 IOMMU_READ
| IOMMU_WRITE
| domain
->prot
);
1142 size_t unmapped
= iommu_unmap(domain
->domain
, 0, PAGE_SIZE
);
1144 if (unmapped
== PAGE_SIZE
)
1145 iommu_unmap(domain
->domain
, PAGE_SIZE
, PAGE_SIZE
);
1147 domain
->fgsp
= true;
1150 __free_pages(pages
, order
);
1153 static struct vfio_group
*find_iommu_group(struct vfio_domain
*domain
,
1154 struct iommu_group
*iommu_group
)
1156 struct vfio_group
*g
;
1158 list_for_each_entry(g
, &domain
->group_list
, next
) {
1159 if (g
->iommu_group
== iommu_group
)
1166 static bool vfio_iommu_has_sw_msi(struct iommu_group
*group
, phys_addr_t
*base
)
1168 struct list_head group_resv_regions
;
1169 struct iommu_resv_region
*region
, *next
;
1172 INIT_LIST_HEAD(&group_resv_regions
);
1173 iommu_get_group_resv_regions(group
, &group_resv_regions
);
1174 list_for_each_entry(region
, &group_resv_regions
, list
) {
1176 * The presence of any 'real' MSI regions should take
1177 * precedence over the software-managed one if the
1178 * IOMMU driver happens to advertise both types.
1180 if (region
->type
== IOMMU_RESV_MSI
) {
1185 if (region
->type
== IOMMU_RESV_SW_MSI
) {
1186 *base
= region
->start
;
1190 list_for_each_entry_safe(region
, next
, &group_resv_regions
, list
)
1195 static int vfio_iommu_type1_attach_group(void *iommu_data
,
1196 struct iommu_group
*iommu_group
)
1198 struct vfio_iommu
*iommu
= iommu_data
;
1199 struct vfio_group
*group
;
1200 struct vfio_domain
*domain
, *d
;
1201 struct bus_type
*bus
= NULL
, *mdev_bus
;
1203 bool resv_msi
, msi_remap
;
1204 phys_addr_t resv_msi_base
;
1206 mutex_lock(&iommu
->lock
);
1208 list_for_each_entry(d
, &iommu
->domain_list
, next
) {
1209 if (find_iommu_group(d
, iommu_group
)) {
1210 mutex_unlock(&iommu
->lock
);
1215 if (iommu
->external_domain
) {
1216 if (find_iommu_group(iommu
->external_domain
, iommu_group
)) {
1217 mutex_unlock(&iommu
->lock
);
1222 group
= kzalloc(sizeof(*group
), GFP_KERNEL
);
1223 domain
= kzalloc(sizeof(*domain
), GFP_KERNEL
);
1224 if (!group
|| !domain
) {
1229 group
->iommu_group
= iommu_group
;
1231 /* Determine bus_type in order to allocate a domain */
1232 ret
= iommu_group_for_each_dev(iommu_group
, &bus
, vfio_bus_type
);
1236 mdev_bus
= symbol_get(mdev_bus_type
);
1239 if ((bus
== mdev_bus
) && !iommu_present(bus
)) {
1240 symbol_put(mdev_bus_type
);
1241 if (!iommu
->external_domain
) {
1242 INIT_LIST_HEAD(&domain
->group_list
);
1243 iommu
->external_domain
= domain
;
1247 list_add(&group
->next
,
1248 &iommu
->external_domain
->group_list
);
1249 mutex_unlock(&iommu
->lock
);
1252 symbol_put(mdev_bus_type
);
1255 domain
->domain
= iommu_domain_alloc(bus
);
1256 if (!domain
->domain
) {
1261 if (iommu
->nesting
) {
1264 ret
= iommu_domain_set_attr(domain
->domain
, DOMAIN_ATTR_NESTING
,
1270 ret
= iommu_attach_group(domain
->domain
, iommu_group
);
1274 resv_msi
= vfio_iommu_has_sw_msi(iommu_group
, &resv_msi_base
);
1276 INIT_LIST_HEAD(&domain
->group_list
);
1277 list_add(&group
->next
, &domain
->group_list
);
1279 msi_remap
= irq_domain_check_msi_remap() ||
1280 iommu_capable(bus
, IOMMU_CAP_INTR_REMAP
);
1282 if (!allow_unsafe_interrupts
&& !msi_remap
) {
1283 pr_warn("%s: No interrupt remapping support. Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n",
1289 if (iommu_capable(bus
, IOMMU_CAP_CACHE_COHERENCY
))
1290 domain
->prot
|= IOMMU_CACHE
;
1293 * Try to match an existing compatible domain. We don't want to
1294 * preclude an IOMMU driver supporting multiple bus_types and being
1295 * able to include different bus_types in the same IOMMU domain, so
1296 * we test whether the domains use the same iommu_ops rather than
1297 * testing if they're on the same bus_type.
1299 list_for_each_entry(d
, &iommu
->domain_list
, next
) {
1300 if (d
->domain
->ops
== domain
->domain
->ops
&&
1301 d
->prot
== domain
->prot
) {
1302 iommu_detach_group(domain
->domain
, iommu_group
);
1303 if (!iommu_attach_group(d
->domain
, iommu_group
)) {
1304 list_add(&group
->next
, &d
->group_list
);
1305 iommu_domain_free(domain
->domain
);
1307 mutex_unlock(&iommu
->lock
);
1311 ret
= iommu_attach_group(domain
->domain
, iommu_group
);
1317 vfio_test_domain_fgsp(domain
);
1319 /* replay mappings on new domains */
1320 ret
= vfio_iommu_replay(iommu
, domain
);
1325 ret
= iommu_get_msi_cookie(domain
->domain
, resv_msi_base
);
1330 list_add(&domain
->next
, &iommu
->domain_list
);
1332 mutex_unlock(&iommu
->lock
);
1337 iommu_detach_group(domain
->domain
, iommu_group
);
1339 iommu_domain_free(domain
->domain
);
1343 mutex_unlock(&iommu
->lock
);
1347 static void vfio_iommu_unmap_unpin_all(struct vfio_iommu
*iommu
)
1349 struct rb_node
*node
;
1351 while ((node
= rb_first(&iommu
->dma_list
)))
1352 vfio_remove_dma(iommu
, rb_entry(node
, struct vfio_dma
, node
));
1355 static void vfio_iommu_unmap_unpin_reaccount(struct vfio_iommu
*iommu
)
1357 struct rb_node
*n
, *p
;
1359 n
= rb_first(&iommu
->dma_list
);
1360 for (; n
; n
= rb_next(n
)) {
1361 struct vfio_dma
*dma
;
1362 long locked
= 0, unlocked
= 0;
1364 dma
= rb_entry(n
, struct vfio_dma
, node
);
1365 unlocked
+= vfio_unmap_unpin(iommu
, dma
, false);
1366 p
= rb_first(&dma
->pfn_list
);
1367 for (; p
; p
= rb_next(p
)) {
1368 struct vfio_pfn
*vpfn
= rb_entry(p
, struct vfio_pfn
,
1371 if (!is_invalid_reserved_pfn(vpfn
->pfn
))
1374 vfio_lock_acct(dma
->task
, locked
- unlocked
, NULL
);
1378 static void vfio_sanity_check_pfn_list(struct vfio_iommu
*iommu
)
1382 n
= rb_first(&iommu
->dma_list
);
1383 for (; n
; n
= rb_next(n
)) {
1384 struct vfio_dma
*dma
;
1386 dma
= rb_entry(n
, struct vfio_dma
, node
);
1388 if (WARN_ON(!RB_EMPTY_ROOT(&dma
->pfn_list
)))
1391 /* mdev vendor driver must unregister notifier */
1392 WARN_ON(iommu
->notifier
.head
);
1395 static void vfio_iommu_type1_detach_group(void *iommu_data
,
1396 struct iommu_group
*iommu_group
)
1398 struct vfio_iommu
*iommu
= iommu_data
;
1399 struct vfio_domain
*domain
;
1400 struct vfio_group
*group
;
1402 mutex_lock(&iommu
->lock
);
1404 if (iommu
->external_domain
) {
1405 group
= find_iommu_group(iommu
->external_domain
, iommu_group
);
1407 list_del(&group
->next
);
1410 if (list_empty(&iommu
->external_domain
->group_list
)) {
1411 vfio_sanity_check_pfn_list(iommu
);
1413 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu
))
1414 vfio_iommu_unmap_unpin_all(iommu
);
1416 kfree(iommu
->external_domain
);
1417 iommu
->external_domain
= NULL
;
1419 goto detach_group_done
;
1423 list_for_each_entry(domain
, &iommu
->domain_list
, next
) {
1424 group
= find_iommu_group(domain
, iommu_group
);
1428 iommu_detach_group(domain
->domain
, iommu_group
);
1429 list_del(&group
->next
);
1432 * Group ownership provides privilege, if the group list is
1433 * empty, the domain goes away. If it's the last domain with
1434 * iommu and external domain doesn't exist, then all the
1435 * mappings go away too. If it's the last domain with iommu and
1436 * external domain exist, update accounting
1438 if (list_empty(&domain
->group_list
)) {
1439 if (list_is_singular(&iommu
->domain_list
)) {
1440 if (!iommu
->external_domain
)
1441 vfio_iommu_unmap_unpin_all(iommu
);
1443 vfio_iommu_unmap_unpin_reaccount(iommu
);
1445 iommu_domain_free(domain
->domain
);
1446 list_del(&domain
->next
);
1453 mutex_unlock(&iommu
->lock
);
1456 static void *vfio_iommu_type1_open(unsigned long arg
)
1458 struct vfio_iommu
*iommu
;
1460 iommu
= kzalloc(sizeof(*iommu
), GFP_KERNEL
);
1462 return ERR_PTR(-ENOMEM
);
1465 case VFIO_TYPE1_IOMMU
:
1467 case VFIO_TYPE1_NESTING_IOMMU
:
1468 iommu
->nesting
= true;
1469 case VFIO_TYPE1v2_IOMMU
:
1474 return ERR_PTR(-EINVAL
);
1477 INIT_LIST_HEAD(&iommu
->domain_list
);
1478 iommu
->dma_list
= RB_ROOT
;
1479 mutex_init(&iommu
->lock
);
1480 BLOCKING_INIT_NOTIFIER_HEAD(&iommu
->notifier
);
1485 static void vfio_release_domain(struct vfio_domain
*domain
, bool external
)
1487 struct vfio_group
*group
, *group_tmp
;
1489 list_for_each_entry_safe(group
, group_tmp
,
1490 &domain
->group_list
, next
) {
1492 iommu_detach_group(domain
->domain
, group
->iommu_group
);
1493 list_del(&group
->next
);
1498 iommu_domain_free(domain
->domain
);
1501 static void vfio_iommu_type1_release(void *iommu_data
)
1503 struct vfio_iommu
*iommu
= iommu_data
;
1504 struct vfio_domain
*domain
, *domain_tmp
;
1506 if (iommu
->external_domain
) {
1507 vfio_release_domain(iommu
->external_domain
, true);
1508 vfio_sanity_check_pfn_list(iommu
);
1509 kfree(iommu
->external_domain
);
1512 vfio_iommu_unmap_unpin_all(iommu
);
1514 list_for_each_entry_safe(domain
, domain_tmp
,
1515 &iommu
->domain_list
, next
) {
1516 vfio_release_domain(domain
, false);
1517 list_del(&domain
->next
);
1523 static int vfio_domains_have_iommu_cache(struct vfio_iommu
*iommu
)
1525 struct vfio_domain
*domain
;
1528 mutex_lock(&iommu
->lock
);
1529 list_for_each_entry(domain
, &iommu
->domain_list
, next
) {
1530 if (!(domain
->prot
& IOMMU_CACHE
)) {
1535 mutex_unlock(&iommu
->lock
);
1540 static long vfio_iommu_type1_ioctl(void *iommu_data
,
1541 unsigned int cmd
, unsigned long arg
)
1543 struct vfio_iommu
*iommu
= iommu_data
;
1544 unsigned long minsz
;
1546 if (cmd
== VFIO_CHECK_EXTENSION
) {
1548 case VFIO_TYPE1_IOMMU
:
1549 case VFIO_TYPE1v2_IOMMU
:
1550 case VFIO_TYPE1_NESTING_IOMMU
:
1552 case VFIO_DMA_CC_IOMMU
:
1555 return vfio_domains_have_iommu_cache(iommu
);
1559 } else if (cmd
== VFIO_IOMMU_GET_INFO
) {
1560 struct vfio_iommu_type1_info info
;
1562 minsz
= offsetofend(struct vfio_iommu_type1_info
, iova_pgsizes
);
1564 if (copy_from_user(&info
, (void __user
*)arg
, minsz
))
1567 if (info
.argsz
< minsz
)
1570 info
.flags
= VFIO_IOMMU_INFO_PGSIZES
;
1572 info
.iova_pgsizes
= vfio_pgsize_bitmap(iommu
);
1574 return copy_to_user((void __user
*)arg
, &info
, minsz
) ?
1577 } else if (cmd
== VFIO_IOMMU_MAP_DMA
) {
1578 struct vfio_iommu_type1_dma_map map
;
1579 uint32_t mask
= VFIO_DMA_MAP_FLAG_READ
|
1580 VFIO_DMA_MAP_FLAG_WRITE
;
1582 minsz
= offsetofend(struct vfio_iommu_type1_dma_map
, size
);
1584 if (copy_from_user(&map
, (void __user
*)arg
, minsz
))
1587 if (map
.argsz
< minsz
|| map
.flags
& ~mask
)
1590 return vfio_dma_do_map(iommu
, &map
);
1592 } else if (cmd
== VFIO_IOMMU_UNMAP_DMA
) {
1593 struct vfio_iommu_type1_dma_unmap unmap
;
1596 minsz
= offsetofend(struct vfio_iommu_type1_dma_unmap
, size
);
1598 if (copy_from_user(&unmap
, (void __user
*)arg
, minsz
))
1601 if (unmap
.argsz
< minsz
|| unmap
.flags
)
1604 ret
= vfio_dma_do_unmap(iommu
, &unmap
);
1608 return copy_to_user((void __user
*)arg
, &unmap
, minsz
) ?
1615 static int vfio_iommu_type1_register_notifier(void *iommu_data
,
1616 unsigned long *events
,
1617 struct notifier_block
*nb
)
1619 struct vfio_iommu
*iommu
= iommu_data
;
1621 /* clear known events */
1622 *events
&= ~VFIO_IOMMU_NOTIFY_DMA_UNMAP
;
1624 /* refuse to register if still events remaining */
1628 return blocking_notifier_chain_register(&iommu
->notifier
, nb
);
1631 static int vfio_iommu_type1_unregister_notifier(void *iommu_data
,
1632 struct notifier_block
*nb
)
1634 struct vfio_iommu
*iommu
= iommu_data
;
1636 return blocking_notifier_chain_unregister(&iommu
->notifier
, nb
);
1639 static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1
= {
1640 .name
= "vfio-iommu-type1",
1641 .owner
= THIS_MODULE
,
1642 .open
= vfio_iommu_type1_open
,
1643 .release
= vfio_iommu_type1_release
,
1644 .ioctl
= vfio_iommu_type1_ioctl
,
1645 .attach_group
= vfio_iommu_type1_attach_group
,
1646 .detach_group
= vfio_iommu_type1_detach_group
,
1647 .pin_pages
= vfio_iommu_type1_pin_pages
,
1648 .unpin_pages
= vfio_iommu_type1_unpin_pages
,
1649 .register_notifier
= vfio_iommu_type1_register_notifier
,
1650 .unregister_notifier
= vfio_iommu_type1_unregister_notifier
,
1653 static int __init
vfio_iommu_type1_init(void)
1655 return vfio_register_iommu_driver(&vfio_iommu_driver_ops_type1
);
1658 static void __exit
vfio_iommu_type1_cleanup(void)
1660 vfio_unregister_iommu_driver(&vfio_iommu_driver_ops_type1
);
1663 module_init(vfio_iommu_type1_init
);
1664 module_exit(vfio_iommu_type1_cleanup
);
1666 MODULE_VERSION(DRIVER_VERSION
);
1667 MODULE_LICENSE("GPL v2");
1668 MODULE_AUTHOR(DRIVER_AUTHOR
);
1669 MODULE_DESCRIPTION(DRIVER_DESC
);