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.");
56 static unsigned int dma_entry_limit __read_mostly = U16_MAX;
57 module_param_named(dma_entry_limit, dma_entry_limit, uint, 0644);
58 MODULE_PARM_DESC(dma_entry_limit,
59 "Maximum number of user DMA mappings per container (65535).");
62 struct list_head domain_list;
64 struct rb_root dma_list;
65 unsigned int dma_avail;
71 struct iommu_domain *domain;
72 struct list_head next;
73 struct list_head group_list;
74 int prot; /* IOMMU_CACHE */
75 bool fgsp; /* Fine-grained super pages */
80 dma_addr_t iova; /* Device address */
81 unsigned long vaddr; /* Process virtual addr */
82 size_t size; /* Map size (bytes) */
83 int prot; /* IOMMU_READ/WRITE */
87 struct iommu_group *iommu_group;
88 struct list_head next;
92 * This code handles mapping and unmapping of user data buffers
93 * into DMA'ble space using the IOMMU
96 static struct vfio_dma *vfio_find_dma(struct vfio_iommu *iommu,
97 dma_addr_t start, size_t size)
99 struct rb_node *node = iommu->dma_list.rb_node;
102 struct vfio_dma *dma = rb_entry(node, struct vfio_dma, node);
104 if (start + size <= dma->iova)
105 node = node->rb_left;
106 else if (start >= dma->iova + dma->size)
107 node = node->rb_right;
115 static void vfio_link_dma(struct vfio_iommu *iommu, struct vfio_dma *new)
117 struct rb_node **link = &iommu->dma_list.rb_node, *parent = NULL;
118 struct vfio_dma *dma;
122 dma = rb_entry(parent, struct vfio_dma, node);
124 if (new->iova + new->size <= dma->iova)
125 link = &(*link)->rb_left;
127 link = &(*link)->rb_right;
130 rb_link_node(&new->node, parent, link);
131 rb_insert_color(&new->node, &iommu->dma_list);
134 static void vfio_unlink_dma(struct vfio_iommu *iommu, struct vfio_dma *old)
136 rb_erase(&old->node, &iommu->dma_list);
139 static int vfio_lock_acct(long npage, bool *lock_cap)
147 return -ESRCH; /* process exited */
149 down_write(¤t->mm->mmap_sem);
151 if (lock_cap ? !*lock_cap : !capable(CAP_IPC_LOCK)) {
154 limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
156 if (current->mm->locked_vm + npage > limit)
162 current->mm->locked_vm += npage;
164 up_write(¤t->mm->mmap_sem);
170 * Some mappings aren't backed by a struct page, for example an mmap'd
171 * MMIO range for our own or another device. These use a different
172 * pfn conversion and shouldn't be tracked as locked pages.
174 static bool is_invalid_reserved_pfn(unsigned long pfn)
176 if (pfn_valid(pfn)) {
178 struct page *tail = pfn_to_page(pfn);
179 struct page *head = compound_head(tail);
180 reserved = !!(PageReserved(head));
183 * "head" is not a dangling pointer
184 * (compound_head takes care of that)
185 * but the hugepage may have been split
186 * from under us (and we may not hold a
187 * reference count on the head page so it can
188 * be reused before we run PageReferenced), so
189 * we've to check PageTail before returning
196 return PageReserved(tail);
202 static int put_pfn(unsigned long pfn, int prot)
204 if (!is_invalid_reserved_pfn(pfn)) {
205 struct page *page = pfn_to_page(pfn);
206 if (prot & IOMMU_WRITE)
214 static int vaddr_get_pfn(unsigned long vaddr, int prot, unsigned long *pfn)
216 struct page *page[1];
217 struct vm_area_struct *vma;
220 if (get_user_pages_fast(vaddr, 1, !!(prot & IOMMU_WRITE), page) == 1) {
221 *pfn = page_to_pfn(page[0]);
225 down_read(¤t->mm->mmap_sem);
227 vma = find_vma_intersection(current->mm, vaddr, vaddr + 1);
229 if (vma && vma->vm_flags & VM_PFNMAP) {
230 if (!follow_pfn(vma, vaddr, pfn) &&
231 is_invalid_reserved_pfn(*pfn))
235 up_read(¤t->mm->mmap_sem);
241 * Attempt to pin pages. We really don't want to track all the pfns and
242 * the iommu can only map chunks of consecutive pfns anyway, so get the
243 * first page and all consecutive pages with the same locking.
245 static long vfio_pin_pages(unsigned long vaddr, long npage,
246 int prot, unsigned long *pfn_base)
248 unsigned long pfn = 0, limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
249 bool lock_cap = capable(CAP_IPC_LOCK);
256 ret = vaddr_get_pfn(vaddr, prot, pfn_base);
260 rsvd = is_invalid_reserved_pfn(*pfn_base);
262 if (!rsvd && !lock_cap && current->mm->locked_vm + 1 > limit) {
263 put_pfn(*pfn_base, prot);
264 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n", __func__,
265 limit << PAGE_SHIFT);
269 if (unlikely(disable_hugepages))
272 /* Lock all the consecutive pages from pfn_base */
273 for (vaddr += PAGE_SIZE; i < npage; i++, vaddr += PAGE_SIZE) {
274 ret = vaddr_get_pfn(vaddr, prot, &pfn);
278 if (pfn != *pfn_base + i ||
279 rsvd != is_invalid_reserved_pfn(pfn)) {
284 if (!rsvd && !lock_cap &&
285 current->mm->locked_vm + i + 1 > limit) {
287 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n",
288 __func__, limit << PAGE_SHIFT);
296 ret = vfio_lock_acct(i, &lock_cap);
301 for (pfn = *pfn_base ; i ; pfn++, i--)
311 static long vfio_unpin_pages(unsigned long pfn, long npage,
312 int prot, bool do_accounting)
314 unsigned long unlocked = 0;
317 for (i = 0; i < npage; i++)
318 unlocked += put_pfn(pfn++, prot);
321 vfio_lock_acct(-unlocked, NULL);
326 static void vfio_unmap_unpin(struct vfio_iommu *iommu, struct vfio_dma *dma)
328 dma_addr_t iova = dma->iova, end = dma->iova + dma->size;
329 struct vfio_domain *domain, *d;
335 * We use the IOMMU to track the physical addresses, otherwise we'd
336 * need a much more complicated tracking system. Unfortunately that
337 * means we need to use one of the iommu domains to figure out the
338 * pfns to unpin. The rest need to be unmapped in advance so we have
339 * no iommu translations remaining when the pages are unpinned.
341 domain = d = list_first_entry(&iommu->domain_list,
342 struct vfio_domain, next);
344 list_for_each_entry_continue(d, &iommu->domain_list, next) {
345 iommu_unmap(d->domain, dma->iova, dma->size);
350 size_t unmapped, len;
351 phys_addr_t phys, next;
353 phys = iommu_iova_to_phys(domain->domain, iova);
354 if (WARN_ON(!phys)) {
360 * To optimize for fewer iommu_unmap() calls, each of which
361 * may require hardware cache flushing, try to find the
362 * largest contiguous physical memory chunk to unmap.
364 for (len = PAGE_SIZE;
365 !domain->fgsp && iova + len < end; len += PAGE_SIZE) {
366 next = iommu_iova_to_phys(domain->domain, iova + len);
367 if (next != phys + len)
371 unmapped = iommu_unmap(domain->domain, iova, len);
372 if (WARN_ON(!unmapped))
375 unlocked += vfio_unpin_pages(phys >> PAGE_SHIFT,
376 unmapped >> PAGE_SHIFT,
383 vfio_lock_acct(-unlocked, NULL);
386 static void vfio_remove_dma(struct vfio_iommu *iommu, struct vfio_dma *dma)
388 vfio_unmap_unpin(iommu, dma);
389 vfio_unlink_dma(iommu, dma);
394 static unsigned long vfio_pgsize_bitmap(struct vfio_iommu *iommu)
396 struct vfio_domain *domain;
397 unsigned long bitmap = ULONG_MAX;
399 mutex_lock(&iommu->lock);
400 list_for_each_entry(domain, &iommu->domain_list, next)
401 bitmap &= domain->domain->ops->pgsize_bitmap;
402 mutex_unlock(&iommu->lock);
405 * In case the IOMMU supports page sizes smaller than PAGE_SIZE
406 * we pretend PAGE_SIZE is supported and hide sub-PAGE_SIZE sizes.
407 * That way the user will be able to map/unmap buffers whose size/
408 * start address is aligned with PAGE_SIZE. Pinning code uses that
409 * granularity while iommu driver can use the sub-PAGE_SIZE size
412 if (bitmap & ~PAGE_MASK) {
420 static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
421 struct vfio_iommu_type1_dma_unmap *unmap)
424 struct vfio_dma *dma;
428 mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1;
430 if (unmap->iova & mask)
432 if (!unmap->size || unmap->size & mask)
435 WARN_ON(mask & PAGE_MASK);
437 mutex_lock(&iommu->lock);
440 * vfio-iommu-type1 (v1) - User mappings were coalesced together to
441 * avoid tracking individual mappings. This means that the granularity
442 * of the original mapping was lost and the user was allowed to attempt
443 * to unmap any range. Depending on the contiguousness of physical
444 * memory and page sizes supported by the IOMMU, arbitrary unmaps may
445 * or may not have worked. We only guaranteed unmap granularity
446 * matching the original mapping; even though it was untracked here,
447 * the original mappings are reflected in IOMMU mappings. This
448 * resulted in a couple unusual behaviors. First, if a range is not
449 * able to be unmapped, ex. a set of 4k pages that was mapped as a
450 * 2M hugepage into the IOMMU, the unmap ioctl returns success but with
451 * a zero sized unmap. Also, if an unmap request overlaps the first
452 * address of a hugepage, the IOMMU will unmap the entire hugepage.
453 * This also returns success and the returned unmap size reflects the
454 * actual size unmapped.
456 * We attempt to maintain compatibility with this "v1" interface, but
457 * we take control out of the hands of the IOMMU. Therefore, an unmap
458 * request offset from the beginning of the original mapping will
459 * return success with zero sized unmap. And an unmap request covering
460 * the first iova of mapping will unmap the entire range.
462 * The v2 version of this interface intends to be more deterministic.
463 * Unmap requests must fully cover previous mappings. Multiple
464 * mappings may still be unmaped by specifying large ranges, but there
465 * must not be any previous mappings bisected by the range. An error
466 * will be returned if these conditions are not met. The v2 interface
467 * will only return success and a size of zero if there were no
468 * mappings within the range.
471 dma = vfio_find_dma(iommu, unmap->iova, 0);
472 if (dma && dma->iova != unmap->iova) {
476 dma = vfio_find_dma(iommu, unmap->iova + unmap->size - 1, 0);
477 if (dma && dma->iova + dma->size != unmap->iova + unmap->size) {
483 while ((dma = vfio_find_dma(iommu, unmap->iova, unmap->size))) {
484 if (!iommu->v2 && unmap->iova > dma->iova)
486 unmapped += dma->size;
487 vfio_remove_dma(iommu, dma);
491 mutex_unlock(&iommu->lock);
493 /* Report how much was unmapped */
494 unmap->size = unmapped;
500 * Turns out AMD IOMMU has a page table bug where it won't map large pages
501 * to a region that previously mapped smaller pages. This should be fixed
502 * soon, so this is just a temporary workaround to break mappings down into
503 * PAGE_SIZE. Better to map smaller pages than nothing.
505 static int map_try_harder(struct vfio_domain *domain, dma_addr_t iova,
506 unsigned long pfn, long npage, int prot)
511 for (i = 0; i < npage; i++, pfn++, iova += PAGE_SIZE) {
512 ret = iommu_map(domain->domain, iova,
513 (phys_addr_t)pfn << PAGE_SHIFT,
514 PAGE_SIZE, prot | domain->prot);
519 for (; i < npage && i > 0; i--, iova -= PAGE_SIZE)
520 iommu_unmap(domain->domain, iova, PAGE_SIZE);
525 static int vfio_iommu_map(struct vfio_iommu *iommu, dma_addr_t iova,
526 unsigned long pfn, long npage, int prot)
528 struct vfio_domain *d;
531 list_for_each_entry(d, &iommu->domain_list, next) {
532 ret = iommu_map(d->domain, iova, (phys_addr_t)pfn << PAGE_SHIFT,
533 npage << PAGE_SHIFT, prot | d->prot);
536 map_try_harder(d, iova, pfn, npage, prot))
546 list_for_each_entry_continue_reverse(d, &iommu->domain_list, next)
547 iommu_unmap(d->domain, iova, npage << PAGE_SHIFT);
552 static int vfio_dma_do_map(struct vfio_iommu *iommu,
553 struct vfio_iommu_type1_dma_map *map)
555 dma_addr_t iova = map->iova;
556 unsigned long vaddr = map->vaddr;
557 size_t size = map->size;
559 int ret = 0, prot = 0;
561 struct vfio_dma *dma;
564 /* Verify that none of our __u64 fields overflow */
565 if (map->size != size || map->vaddr != vaddr || map->iova != iova)
568 mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1;
570 WARN_ON(mask & PAGE_MASK);
572 /* READ/WRITE from device perspective */
573 if (map->flags & VFIO_DMA_MAP_FLAG_WRITE)
575 if (map->flags & VFIO_DMA_MAP_FLAG_READ)
578 if (!prot || !size || (size | iova | vaddr) & mask)
581 /* Don't allow IOVA or virtual address wrap */
582 if (iova + size - 1 < iova || vaddr + size - 1 < vaddr)
585 mutex_lock(&iommu->lock);
587 if (vfio_find_dma(iommu, iova, size)) {
588 mutex_unlock(&iommu->lock);
592 if (!iommu->dma_avail) {
593 mutex_unlock(&iommu->lock);
597 dma = kzalloc(sizeof(*dma), GFP_KERNEL);
599 mutex_unlock(&iommu->lock);
608 /* Insert zero-sized and grow as we map chunks of it */
609 vfio_link_dma(iommu, dma);
612 /* Pin a contiguous chunk of memory */
613 npage = vfio_pin_pages(vaddr + dma->size,
614 size >> PAGE_SHIFT, prot, &pfn);
622 ret = vfio_iommu_map(iommu, iova + dma->size, pfn, npage, prot);
624 vfio_unpin_pages(pfn, npage, prot, true);
628 size -= npage << PAGE_SHIFT;
629 dma->size += npage << PAGE_SHIFT;
633 vfio_remove_dma(iommu, dma);
635 mutex_unlock(&iommu->lock);
639 static int vfio_bus_type(struct device *dev, void *data)
641 struct bus_type **bus = data;
643 if (*bus && *bus != dev->bus)
651 static int vfio_iommu_replay(struct vfio_iommu *iommu,
652 struct vfio_domain *domain)
654 struct vfio_domain *d;
658 /* Arbitrarily pick the first domain in the list for lookups */
659 d = list_first_entry(&iommu->domain_list, struct vfio_domain, next);
660 n = rb_first(&iommu->dma_list);
662 /* If there's not a domain, there better not be any mappings */
663 if (WARN_ON(n && !d))
666 for (; n; n = rb_next(n)) {
667 struct vfio_dma *dma;
670 dma = rb_entry(n, struct vfio_dma, node);
673 while (iova < dma->iova + dma->size) {
674 phys_addr_t phys = iommu_iova_to_phys(d->domain, iova);
677 if (WARN_ON(!phys)) {
684 while (iova + size < dma->iova + dma->size &&
685 phys + size == iommu_iova_to_phys(d->domain,
689 ret = iommu_map(domain->domain, iova, phys,
690 size, dma->prot | domain->prot);
702 * We change our unmap behavior slightly depending on whether the IOMMU
703 * supports fine-grained superpages. IOMMUs like AMD-Vi will use a superpage
704 * for practically any contiguous power-of-two mapping we give it. This means
705 * we don't need to look for contiguous chunks ourselves to make unmapping
706 * more efficient. On IOMMUs with coarse-grained super pages, like Intel VT-d
707 * with discrete 2M/1G/512G/1T superpages, identifying contiguous chunks
708 * significantly boosts non-hugetlbfs mappings and doesn't seem to hurt when
709 * hugetlbfs is in use.
711 static void vfio_test_domain_fgsp(struct vfio_domain *domain)
714 int ret, order = get_order(PAGE_SIZE * 2);
716 pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order);
720 ret = iommu_map(domain->domain, 0, page_to_phys(pages), PAGE_SIZE * 2,
721 IOMMU_READ | IOMMU_WRITE | domain->prot);
723 size_t unmapped = iommu_unmap(domain->domain, 0, PAGE_SIZE);
725 if (unmapped == PAGE_SIZE)
726 iommu_unmap(domain->domain, PAGE_SIZE, PAGE_SIZE);
731 __free_pages(pages, order);
734 static int vfio_iommu_type1_attach_group(void *iommu_data,
735 struct iommu_group *iommu_group)
737 struct vfio_iommu *iommu = iommu_data;
738 struct vfio_group *group, *g;
739 struct vfio_domain *domain, *d;
740 struct bus_type *bus = NULL;
743 mutex_lock(&iommu->lock);
745 list_for_each_entry(d, &iommu->domain_list, next) {
746 list_for_each_entry(g, &d->group_list, next) {
747 if (g->iommu_group != iommu_group)
750 mutex_unlock(&iommu->lock);
755 group = kzalloc(sizeof(*group), GFP_KERNEL);
756 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
757 if (!group || !domain) {
762 group->iommu_group = iommu_group;
764 /* Determine bus_type in order to allocate a domain */
765 ret = iommu_group_for_each_dev(iommu_group, &bus, vfio_bus_type);
769 domain->domain = iommu_domain_alloc(bus);
770 if (!domain->domain) {
775 if (iommu->nesting) {
778 ret = iommu_domain_set_attr(domain->domain, DOMAIN_ATTR_NESTING,
784 ret = iommu_attach_group(domain->domain, iommu_group);
788 INIT_LIST_HEAD(&domain->group_list);
789 list_add(&group->next, &domain->group_list);
791 if (!allow_unsafe_interrupts &&
792 !iommu_capable(bus, IOMMU_CAP_INTR_REMAP)) {
793 pr_warn("%s: No interrupt remapping support. Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n",
799 if (iommu_capable(bus, IOMMU_CAP_CACHE_COHERENCY))
800 domain->prot |= IOMMU_CACHE;
803 * Try to match an existing compatible domain. We don't want to
804 * preclude an IOMMU driver supporting multiple bus_types and being
805 * able to include different bus_types in the same IOMMU domain, so
806 * we test whether the domains use the same iommu_ops rather than
807 * testing if they're on the same bus_type.
809 list_for_each_entry(d, &iommu->domain_list, next) {
810 if (d->domain->ops == domain->domain->ops &&
811 d->prot == domain->prot) {
812 iommu_detach_group(domain->domain, iommu_group);
813 if (!iommu_attach_group(d->domain, iommu_group)) {
814 list_add(&group->next, &d->group_list);
815 iommu_domain_free(domain->domain);
817 mutex_unlock(&iommu->lock);
821 ret = iommu_attach_group(domain->domain, iommu_group);
827 vfio_test_domain_fgsp(domain);
829 /* replay mappings on new domains */
830 ret = vfio_iommu_replay(iommu, domain);
834 list_add(&domain->next, &iommu->domain_list);
836 mutex_unlock(&iommu->lock);
841 iommu_detach_group(domain->domain, iommu_group);
843 iommu_domain_free(domain->domain);
847 mutex_unlock(&iommu->lock);
851 static void vfio_iommu_unmap_unpin_all(struct vfio_iommu *iommu)
853 struct rb_node *node;
855 while ((node = rb_first(&iommu->dma_list)))
856 vfio_remove_dma(iommu, rb_entry(node, struct vfio_dma, node));
859 static void vfio_iommu_type1_detach_group(void *iommu_data,
860 struct iommu_group *iommu_group)
862 struct vfio_iommu *iommu = iommu_data;
863 struct vfio_domain *domain;
864 struct vfio_group *group;
866 mutex_lock(&iommu->lock);
868 list_for_each_entry(domain, &iommu->domain_list, next) {
869 list_for_each_entry(group, &domain->group_list, next) {
870 if (group->iommu_group != iommu_group)
873 iommu_detach_group(domain->domain, iommu_group);
874 list_del(&group->next);
877 * Group ownership provides privilege, if the group
878 * list is empty, the domain goes away. If it's the
879 * last domain, then all the mappings go away too.
881 if (list_empty(&domain->group_list)) {
882 if (list_is_singular(&iommu->domain_list))
883 vfio_iommu_unmap_unpin_all(iommu);
884 iommu_domain_free(domain->domain);
885 list_del(&domain->next);
893 mutex_unlock(&iommu->lock);
896 static void *vfio_iommu_type1_open(unsigned long arg)
898 struct vfio_iommu *iommu;
900 iommu = kzalloc(sizeof(*iommu), GFP_KERNEL);
902 return ERR_PTR(-ENOMEM);
905 case VFIO_TYPE1_IOMMU:
907 case VFIO_TYPE1_NESTING_IOMMU:
908 iommu->nesting = true;
909 case VFIO_TYPE1v2_IOMMU:
914 return ERR_PTR(-EINVAL);
917 INIT_LIST_HEAD(&iommu->domain_list);
918 iommu->dma_list = RB_ROOT;
919 iommu->dma_avail = dma_entry_limit;
920 mutex_init(&iommu->lock);
925 static void vfio_iommu_type1_release(void *iommu_data)
927 struct vfio_iommu *iommu = iommu_data;
928 struct vfio_domain *domain, *domain_tmp;
929 struct vfio_group *group, *group_tmp;
931 vfio_iommu_unmap_unpin_all(iommu);
933 list_for_each_entry_safe(domain, domain_tmp,
934 &iommu->domain_list, next) {
935 list_for_each_entry_safe(group, group_tmp,
936 &domain->group_list, next) {
937 iommu_detach_group(domain->domain, group->iommu_group);
938 list_del(&group->next);
941 iommu_domain_free(domain->domain);
942 list_del(&domain->next);
949 static int vfio_domains_have_iommu_cache(struct vfio_iommu *iommu)
951 struct vfio_domain *domain;
954 mutex_lock(&iommu->lock);
955 list_for_each_entry(domain, &iommu->domain_list, next) {
956 if (!(domain->prot & IOMMU_CACHE)) {
961 mutex_unlock(&iommu->lock);
966 static long vfio_iommu_type1_ioctl(void *iommu_data,
967 unsigned int cmd, unsigned long arg)
969 struct vfio_iommu *iommu = iommu_data;
972 if (cmd == VFIO_CHECK_EXTENSION) {
974 case VFIO_TYPE1_IOMMU:
975 case VFIO_TYPE1v2_IOMMU:
976 case VFIO_TYPE1_NESTING_IOMMU:
978 case VFIO_DMA_CC_IOMMU:
981 return vfio_domains_have_iommu_cache(iommu);
985 } else if (cmd == VFIO_IOMMU_GET_INFO) {
986 struct vfio_iommu_type1_info info;
988 minsz = offsetofend(struct vfio_iommu_type1_info, iova_pgsizes);
990 if (copy_from_user(&info, (void __user *)arg, minsz))
993 if (info.argsz < minsz)
998 info.iova_pgsizes = vfio_pgsize_bitmap(iommu);
1000 return copy_to_user((void __user *)arg, &info, minsz) ?
1003 } else if (cmd == VFIO_IOMMU_MAP_DMA) {
1004 struct vfio_iommu_type1_dma_map map;
1005 uint32_t mask = VFIO_DMA_MAP_FLAG_READ |
1006 VFIO_DMA_MAP_FLAG_WRITE;
1008 minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
1010 if (copy_from_user(&map, (void __user *)arg, minsz))
1013 if (map.argsz < minsz || map.flags & ~mask)
1016 return vfio_dma_do_map(iommu, &map);
1018 } else if (cmd == VFIO_IOMMU_UNMAP_DMA) {
1019 struct vfio_iommu_type1_dma_unmap unmap;
1022 minsz = offsetofend(struct vfio_iommu_type1_dma_unmap, size);
1024 if (copy_from_user(&unmap, (void __user *)arg, minsz))
1027 if (unmap.argsz < minsz || unmap.flags)
1030 ret = vfio_dma_do_unmap(iommu, &unmap);
1034 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);