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);
358 size_t unmapped, len;
359 phys_addr_t phys, next;
361 phys = iommu_iova_to_phys(domain->domain, iova);
362 if (WARN_ON(!phys)) {
368 * To optimize for fewer iommu_unmap() calls, each of which
369 * may require hardware cache flushing, try to find the
370 * largest contiguous physical memory chunk to unmap.
372 for (len = PAGE_SIZE;
373 !domain->fgsp && iova + len < end; len += PAGE_SIZE) {
374 next = iommu_iova_to_phys(domain->domain, iova + len);
375 if (next != phys + len)
379 unmapped = iommu_unmap(domain->domain, iova, len);
380 if (WARN_ON(!unmapped))
383 unlocked += vfio_unpin_pages(phys >> PAGE_SHIFT,
384 unmapped >> PAGE_SHIFT,
389 vfio_lock_acct(-unlocked);
392 static void vfio_remove_dma(struct vfio_iommu *iommu, struct vfio_dma *dma)
394 vfio_unmap_unpin(iommu, dma);
395 vfio_unlink_dma(iommu, dma);
399 static unsigned long vfio_pgsize_bitmap(struct vfio_iommu *iommu)
401 struct vfio_domain *domain;
402 unsigned long bitmap = PAGE_MASK;
404 mutex_lock(&iommu->lock);
405 list_for_each_entry(domain, &iommu->domain_list, next)
406 bitmap &= domain->domain->ops->pgsize_bitmap;
407 mutex_unlock(&iommu->lock);
412 static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
413 struct vfio_iommu_type1_dma_unmap *unmap)
416 struct vfio_dma *dma;
420 mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1;
422 if (unmap->iova & mask)
424 if (!unmap->size || unmap->size & mask)
427 WARN_ON(mask & PAGE_MASK);
429 mutex_lock(&iommu->lock);
432 * vfio-iommu-type1 (v1) - User mappings were coalesced together to
433 * avoid tracking individual mappings. This means that the granularity
434 * of the original mapping was lost and the user was allowed to attempt
435 * to unmap any range. Depending on the contiguousness of physical
436 * memory and page sizes supported by the IOMMU, arbitrary unmaps may
437 * or may not have worked. We only guaranteed unmap granularity
438 * matching the original mapping; even though it was untracked here,
439 * the original mappings are reflected in IOMMU mappings. This
440 * resulted in a couple unusual behaviors. First, if a range is not
441 * able to be unmapped, ex. a set of 4k pages that was mapped as a
442 * 2M hugepage into the IOMMU, the unmap ioctl returns success but with
443 * a zero sized unmap. Also, if an unmap request overlaps the first
444 * address of a hugepage, the IOMMU will unmap the entire hugepage.
445 * This also returns success and the returned unmap size reflects the
446 * actual size unmapped.
448 * We attempt to maintain compatibility with this "v1" interface, but
449 * we take control out of the hands of the IOMMU. Therefore, an unmap
450 * request offset from the beginning of the original mapping will
451 * return success with zero sized unmap. And an unmap request covering
452 * the first iova of mapping will unmap the entire range.
454 * The v2 version of this interface intends to be more deterministic.
455 * Unmap requests must fully cover previous mappings. Multiple
456 * mappings may still be unmaped by specifying large ranges, but there
457 * must not be any previous mappings bisected by the range. An error
458 * will be returned if these conditions are not met. The v2 interface
459 * will only return success and a size of zero if there were no
460 * mappings within the range.
463 dma = vfio_find_dma(iommu, unmap->iova, 0);
464 if (dma && dma->iova != unmap->iova) {
468 dma = vfio_find_dma(iommu, unmap->iova + unmap->size - 1, 0);
469 if (dma && dma->iova + dma->size != unmap->iova + unmap->size) {
475 while ((dma = vfio_find_dma(iommu, unmap->iova, unmap->size))) {
476 if (!iommu->v2 && unmap->iova > dma->iova)
478 unmapped += dma->size;
479 vfio_remove_dma(iommu, dma);
483 mutex_unlock(&iommu->lock);
485 /* Report how much was unmapped */
486 unmap->size = unmapped;
492 * Turns out AMD IOMMU has a page table bug where it won't map large pages
493 * to a region that previously mapped smaller pages. This should be fixed
494 * soon, so this is just a temporary workaround to break mappings down into
495 * PAGE_SIZE. Better to map smaller pages than nothing.
497 static int map_try_harder(struct vfio_domain *domain, dma_addr_t iova,
498 unsigned long pfn, long npage, int prot)
503 for (i = 0; i < npage; i++, pfn++, iova += PAGE_SIZE) {
504 ret = iommu_map(domain->domain, iova,
505 (phys_addr_t)pfn << PAGE_SHIFT,
506 PAGE_SIZE, prot | domain->prot);
511 for (; i < npage && i > 0; i--, iova -= PAGE_SIZE)
512 iommu_unmap(domain->domain, iova, PAGE_SIZE);
517 static int vfio_iommu_map(struct vfio_iommu *iommu, dma_addr_t iova,
518 unsigned long pfn, long npage, int prot)
520 struct vfio_domain *d;
523 list_for_each_entry(d, &iommu->domain_list, next) {
524 ret = iommu_map(d->domain, iova, (phys_addr_t)pfn << PAGE_SHIFT,
525 npage << PAGE_SHIFT, prot | d->prot);
528 map_try_harder(d, iova, pfn, npage, prot))
536 list_for_each_entry_continue_reverse(d, &iommu->domain_list, next)
537 iommu_unmap(d->domain, iova, npage << PAGE_SHIFT);
542 static int vfio_dma_do_map(struct vfio_iommu *iommu,
543 struct vfio_iommu_type1_dma_map *map)
545 dma_addr_t iova = map->iova;
546 unsigned long vaddr = map->vaddr;
547 size_t size = map->size;
549 int ret = 0, prot = 0;
551 struct vfio_dma *dma;
554 /* Verify that none of our __u64 fields overflow */
555 if (map->size != size || map->vaddr != vaddr || map->iova != iova)
558 mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1;
560 WARN_ON(mask & PAGE_MASK);
562 /* READ/WRITE from device perspective */
563 if (map->flags & VFIO_DMA_MAP_FLAG_WRITE)
565 if (map->flags & VFIO_DMA_MAP_FLAG_READ)
568 if (!prot || !size || (size | iova | vaddr) & mask)
571 /* Don't allow IOVA or virtual address wrap */
572 if (iova + size - 1 < iova || vaddr + size - 1 < vaddr)
575 mutex_lock(&iommu->lock);
577 if (vfio_find_dma(iommu, iova, size)) {
578 mutex_unlock(&iommu->lock);
582 dma = kzalloc(sizeof(*dma), GFP_KERNEL);
584 mutex_unlock(&iommu->lock);
592 /* Insert zero-sized and grow as we map chunks of it */
593 vfio_link_dma(iommu, dma);
596 /* Pin a contiguous chunk of memory */
597 npage = vfio_pin_pages(vaddr + dma->size,
598 size >> PAGE_SHIFT, prot, &pfn);
606 ret = vfio_iommu_map(iommu, iova + dma->size, pfn, npage, prot);
608 vfio_unpin_pages(pfn, npage, prot, true);
612 size -= npage << PAGE_SHIFT;
613 dma->size += npage << PAGE_SHIFT;
617 vfio_remove_dma(iommu, dma);
619 mutex_unlock(&iommu->lock);
623 static int vfio_bus_type(struct device *dev, void *data)
625 struct bus_type **bus = data;
627 if (*bus && *bus != dev->bus)
635 static int vfio_iommu_replay(struct vfio_iommu *iommu,
636 struct vfio_domain *domain)
638 struct vfio_domain *d;
642 /* Arbitrarily pick the first domain in the list for lookups */
643 d = list_first_entry(&iommu->domain_list, struct vfio_domain, next);
644 n = rb_first(&iommu->dma_list);
646 /* If there's not a domain, there better not be any mappings */
647 if (WARN_ON(n && !d))
650 for (; n; n = rb_next(n)) {
651 struct vfio_dma *dma;
654 dma = rb_entry(n, struct vfio_dma, node);
657 while (iova < dma->iova + dma->size) {
658 phys_addr_t phys = iommu_iova_to_phys(d->domain, iova);
661 if (WARN_ON(!phys)) {
668 while (iova + size < dma->iova + dma->size &&
669 phys + size == iommu_iova_to_phys(d->domain,
673 ret = iommu_map(domain->domain, iova, phys,
674 size, dma->prot | domain->prot);
686 * We change our unmap behavior slightly depending on whether the IOMMU
687 * supports fine-grained superpages. IOMMUs like AMD-Vi will use a superpage
688 * for practically any contiguous power-of-two mapping we give it. This means
689 * we don't need to look for contiguous chunks ourselves to make unmapping
690 * more efficient. On IOMMUs with coarse-grained super pages, like Intel VT-d
691 * with discrete 2M/1G/512G/1T superpages, identifying contiguous chunks
692 * significantly boosts non-hugetlbfs mappings and doesn't seem to hurt when
693 * hugetlbfs is in use.
695 static void vfio_test_domain_fgsp(struct vfio_domain *domain)
698 int ret, order = get_order(PAGE_SIZE * 2);
700 pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order);
704 ret = iommu_map(domain->domain, 0, page_to_phys(pages), PAGE_SIZE * 2,
705 IOMMU_READ | IOMMU_WRITE | domain->prot);
707 size_t unmapped = iommu_unmap(domain->domain, 0, PAGE_SIZE);
709 if (unmapped == PAGE_SIZE)
710 iommu_unmap(domain->domain, PAGE_SIZE, PAGE_SIZE);
715 __free_pages(pages, order);
718 static int vfio_iommu_type1_attach_group(void *iommu_data,
719 struct iommu_group *iommu_group)
721 struct vfio_iommu *iommu = iommu_data;
722 struct vfio_group *group, *g;
723 struct vfio_domain *domain, *d;
724 struct bus_type *bus = NULL;
727 mutex_lock(&iommu->lock);
729 list_for_each_entry(d, &iommu->domain_list, next) {
730 list_for_each_entry(g, &d->group_list, next) {
731 if (g->iommu_group != iommu_group)
734 mutex_unlock(&iommu->lock);
739 group = kzalloc(sizeof(*group), GFP_KERNEL);
740 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
741 if (!group || !domain) {
746 group->iommu_group = iommu_group;
748 /* Determine bus_type in order to allocate a domain */
749 ret = iommu_group_for_each_dev(iommu_group, &bus, vfio_bus_type);
753 domain->domain = iommu_domain_alloc(bus);
754 if (!domain->domain) {
759 if (iommu->nesting) {
762 ret = iommu_domain_set_attr(domain->domain, DOMAIN_ATTR_NESTING,
768 ret = iommu_attach_group(domain->domain, iommu_group);
772 INIT_LIST_HEAD(&domain->group_list);
773 list_add(&group->next, &domain->group_list);
775 if (!allow_unsafe_interrupts &&
776 !iommu_capable(bus, IOMMU_CAP_INTR_REMAP)) {
777 pr_warn("%s: No interrupt remapping support. Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n",
783 if (iommu_capable(bus, IOMMU_CAP_CACHE_COHERENCY))
784 domain->prot |= IOMMU_CACHE;
787 * Try to match an existing compatible domain. We don't want to
788 * preclude an IOMMU driver supporting multiple bus_types and being
789 * able to include different bus_types in the same IOMMU domain, so
790 * we test whether the domains use the same iommu_ops rather than
791 * testing if they're on the same bus_type.
793 list_for_each_entry(d, &iommu->domain_list, next) {
794 if (d->domain->ops == domain->domain->ops &&
795 d->prot == domain->prot) {
796 iommu_detach_group(domain->domain, iommu_group);
797 if (!iommu_attach_group(d->domain, iommu_group)) {
798 list_add(&group->next, &d->group_list);
799 iommu_domain_free(domain->domain);
801 mutex_unlock(&iommu->lock);
805 ret = iommu_attach_group(domain->domain, iommu_group);
811 vfio_test_domain_fgsp(domain);
813 /* replay mappings on new domains */
814 ret = vfio_iommu_replay(iommu, domain);
818 list_add(&domain->next, &iommu->domain_list);
820 mutex_unlock(&iommu->lock);
825 iommu_detach_group(domain->domain, iommu_group);
827 iommu_domain_free(domain->domain);
831 mutex_unlock(&iommu->lock);
835 static void vfio_iommu_unmap_unpin_all(struct vfio_iommu *iommu)
837 struct rb_node *node;
839 while ((node = rb_first(&iommu->dma_list)))
840 vfio_remove_dma(iommu, rb_entry(node, struct vfio_dma, node));
843 static void vfio_iommu_type1_detach_group(void *iommu_data,
844 struct iommu_group *iommu_group)
846 struct vfio_iommu *iommu = iommu_data;
847 struct vfio_domain *domain;
848 struct vfio_group *group;
850 mutex_lock(&iommu->lock);
852 list_for_each_entry(domain, &iommu->domain_list, next) {
853 list_for_each_entry(group, &domain->group_list, next) {
854 if (group->iommu_group != iommu_group)
857 iommu_detach_group(domain->domain, iommu_group);
858 list_del(&group->next);
861 * Group ownership provides privilege, if the group
862 * list is empty, the domain goes away. If it's the
863 * last domain, then all the mappings go away too.
865 if (list_empty(&domain->group_list)) {
866 if (list_is_singular(&iommu->domain_list))
867 vfio_iommu_unmap_unpin_all(iommu);
868 iommu_domain_free(domain->domain);
869 list_del(&domain->next);
877 mutex_unlock(&iommu->lock);
880 static void *vfio_iommu_type1_open(unsigned long arg)
882 struct vfio_iommu *iommu;
884 iommu = kzalloc(sizeof(*iommu), GFP_KERNEL);
886 return ERR_PTR(-ENOMEM);
889 case VFIO_TYPE1_IOMMU:
891 case VFIO_TYPE1_NESTING_IOMMU:
892 iommu->nesting = true;
893 case VFIO_TYPE1v2_IOMMU:
898 return ERR_PTR(-EINVAL);
901 INIT_LIST_HEAD(&iommu->domain_list);
902 iommu->dma_list = RB_ROOT;
903 mutex_init(&iommu->lock);
908 static void vfio_iommu_type1_release(void *iommu_data)
910 struct vfio_iommu *iommu = iommu_data;
911 struct vfio_domain *domain, *domain_tmp;
912 struct vfio_group *group, *group_tmp;
914 vfio_iommu_unmap_unpin_all(iommu);
916 list_for_each_entry_safe(domain, domain_tmp,
917 &iommu->domain_list, next) {
918 list_for_each_entry_safe(group, group_tmp,
919 &domain->group_list, next) {
920 iommu_detach_group(domain->domain, group->iommu_group);
921 list_del(&group->next);
924 iommu_domain_free(domain->domain);
925 list_del(&domain->next);
932 static int vfio_domains_have_iommu_cache(struct vfio_iommu *iommu)
934 struct vfio_domain *domain;
937 mutex_lock(&iommu->lock);
938 list_for_each_entry(domain, &iommu->domain_list, next) {
939 if (!(domain->prot & IOMMU_CACHE)) {
944 mutex_unlock(&iommu->lock);
949 static long vfio_iommu_type1_ioctl(void *iommu_data,
950 unsigned int cmd, unsigned long arg)
952 struct vfio_iommu *iommu = iommu_data;
955 if (cmd == VFIO_CHECK_EXTENSION) {
957 case VFIO_TYPE1_IOMMU:
958 case VFIO_TYPE1v2_IOMMU:
959 case VFIO_TYPE1_NESTING_IOMMU:
961 case VFIO_DMA_CC_IOMMU:
964 return vfio_domains_have_iommu_cache(iommu);
968 } else if (cmd == VFIO_IOMMU_GET_INFO) {
969 struct vfio_iommu_type1_info info;
971 minsz = offsetofend(struct vfio_iommu_type1_info, iova_pgsizes);
973 if (copy_from_user(&info, (void __user *)arg, minsz))
976 if (info.argsz < minsz)
981 info.iova_pgsizes = vfio_pgsize_bitmap(iommu);
983 return copy_to_user((void __user *)arg, &info, minsz);
985 } else if (cmd == VFIO_IOMMU_MAP_DMA) {
986 struct vfio_iommu_type1_dma_map map;
987 uint32_t mask = VFIO_DMA_MAP_FLAG_READ |
988 VFIO_DMA_MAP_FLAG_WRITE;
990 minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
992 if (copy_from_user(&map, (void __user *)arg, minsz))
995 if (map.argsz < minsz || map.flags & ~mask)
998 return vfio_dma_do_map(iommu, &map);
1000 } else if (cmd == VFIO_IOMMU_UNMAP_DMA) {
1001 struct vfio_iommu_type1_dma_unmap unmap;
1004 minsz = offsetofend(struct vfio_iommu_type1_dma_unmap, size);
1006 if (copy_from_user(&unmap, (void __user *)arg, minsz))
1009 if (unmap.argsz < minsz || unmap.flags)
1012 ret = vfio_dma_do_unmap(iommu, &unmap);
1016 return copy_to_user((void __user *)arg, &unmap, minsz);
1022 static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = {
1023 .name = "vfio-iommu-type1",
1024 .owner = THIS_MODULE,
1025 .open = vfio_iommu_type1_open,
1026 .release = vfio_iommu_type1_release,
1027 .ioctl = vfio_iommu_type1_ioctl,
1028 .attach_group = vfio_iommu_type1_attach_group,
1029 .detach_group = vfio_iommu_type1_detach_group,
1032 static int __init vfio_iommu_type1_init(void)
1034 return vfio_register_iommu_driver(&vfio_iommu_driver_ops_type1);
1037 static void __exit vfio_iommu_type1_cleanup(void)
1039 vfio_unregister_iommu_driver(&vfio_iommu_driver_ops_type1);
1042 module_init(vfio_iommu_type1_init);
1043 module_exit(vfio_iommu_type1_cleanup);
1045 MODULE_VERSION(DRIVER_VERSION);
1046 MODULE_LICENSE("GPL v2");
1047 MODULE_AUTHOR(DRIVER_AUTHOR);
1048 MODULE_DESCRIPTION(DRIVER_DESC);