2 * VFIO: IOMMU DMA mapping support for Type1 IOMMU
4 * Copyright (C) 2012 Red Hat, Inc. All rights reserved.
5 * Author: Alex Williamson <alex.williamson@redhat.com>
7 * This program is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License version 2 as
9 * published by the Free Software Foundation.
11 * Derived from original vfio:
12 * Copyright 2010 Cisco Systems, Inc. All rights reserved.
13 * Author: Tom Lyon, pugs@cisco.com
15 * We arbitrarily define a Type1 IOMMU as one matching the below code.
16 * It could be called the x86 IOMMU as it's designed for AMD-Vi & Intel
17 * VT-d, but that makes it harder to re-use as theoretically anyone
18 * implementing a similar IOMMU could make use of this. We expect the
19 * IOMMU to support the IOMMU API and have few to no restrictions around
20 * the IOVA range that can be mapped. The Type1 IOMMU is currently
21 * optimized for relatively static mappings of a userspace process with
22 * userpsace pages pinned into memory. We also assume devices and IOMMU
23 * domains are PCI based as the IOMMU API is still centered around a
24 * device/bus interface rather than a group interface.
27 #include <linux/compat.h>
28 #include <linux/device.h>
30 #include <linux/iommu.h>
31 #include <linux/module.h>
33 #include <linux/rbtree.h>
34 #include <linux/sched/signal.h>
35 #include <linux/sched/mm.h>
36 #include <linux/slab.h>
37 #include <linux/uaccess.h>
38 #include <linux/vfio.h>
39 #include <linux/workqueue.h>
40 #include <linux/mdev.h>
41 #include <linux/notifier.h>
42 #include <linux/dma-iommu.h>
43 #include <linux/irqdomain.h>
45 #define DRIVER_VERSION "0.2"
46 #define DRIVER_AUTHOR "Alex Williamson <alex.williamson@redhat.com>"
47 #define DRIVER_DESC "Type1 IOMMU driver for VFIO"
49 static bool allow_unsafe_interrupts;
50 module_param_named(allow_unsafe_interrupts,
51 allow_unsafe_interrupts, bool, S_IRUGO | S_IWUSR);
52 MODULE_PARM_DESC(allow_unsafe_interrupts,
53 "Enable VFIO IOMMU support for on platforms without interrupt remapping support.");
55 static bool disable_hugepages;
56 module_param_named(disable_hugepages,
57 disable_hugepages, bool, S_IRUGO | S_IWUSR);
58 MODULE_PARM_DESC(disable_hugepages,
59 "Disable VFIO IOMMU support for IOMMU hugepages.");
61 static unsigned int dma_entry_limit __read_mostly = U16_MAX;
62 module_param_named(dma_entry_limit, dma_entry_limit, uint, 0644);
63 MODULE_PARM_DESC(dma_entry_limit,
64 "Maximum number of user DMA mappings per container (65535).");
67 struct list_head domain_list;
68 struct vfio_domain *external_domain; /* domain for external user */
70 struct rb_root dma_list;
71 struct blocking_notifier_head notifier;
72 unsigned int dma_avail;
78 struct iommu_domain *domain;
79 struct list_head next;
80 struct list_head group_list;
81 int prot; /* IOMMU_CACHE */
82 bool fgsp; /* Fine-grained super pages */
87 dma_addr_t iova; /* Device address */
88 unsigned long vaddr; /* Process virtual addr */
89 size_t size; /* Map size (bytes) */
90 int prot; /* IOMMU_READ/WRITE */
92 bool lock_cap; /* capable(CAP_IPC_LOCK) */
93 struct task_struct *task;
94 struct rb_root pfn_list; /* Ex-user pinned pfn list */
98 struct iommu_group *iommu_group;
99 struct list_head next;
103 * Guest RAM pinning working set or DMA target
107 dma_addr_t iova; /* Device address */
108 unsigned long pfn; /* Host pfn */
112 struct vfio_regions {
113 struct list_head list;
119 #define IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu) \
120 (!list_empty(&iommu->domain_list))
122 static int put_pfn(unsigned long pfn, int prot);
125 * This code handles mapping and unmapping of user data buffers
126 * into DMA'ble space using the IOMMU
129 static struct vfio_dma *vfio_find_dma(struct vfio_iommu *iommu,
130 dma_addr_t start, size_t size)
132 struct rb_node *node = iommu->dma_list.rb_node;
135 struct vfio_dma *dma = rb_entry(node, struct vfio_dma, node);
137 if (start + size <= dma->iova)
138 node = node->rb_left;
139 else if (start >= dma->iova + dma->size)
140 node = node->rb_right;
148 static void vfio_link_dma(struct vfio_iommu *iommu, struct vfio_dma *new)
150 struct rb_node **link = &iommu->dma_list.rb_node, *parent = NULL;
151 struct vfio_dma *dma;
155 dma = rb_entry(parent, struct vfio_dma, node);
157 if (new->iova + new->size <= dma->iova)
158 link = &(*link)->rb_left;
160 link = &(*link)->rb_right;
163 rb_link_node(&new->node, parent, link);
164 rb_insert_color(&new->node, &iommu->dma_list);
167 static void vfio_unlink_dma(struct vfio_iommu *iommu, struct vfio_dma *old)
169 rb_erase(&old->node, &iommu->dma_list);
173 * Helper Functions for host iova-pfn list
175 static struct vfio_pfn *vfio_find_vpfn(struct vfio_dma *dma, dma_addr_t iova)
177 struct vfio_pfn *vpfn;
178 struct rb_node *node = dma->pfn_list.rb_node;
181 vpfn = rb_entry(node, struct vfio_pfn, node);
183 if (iova < vpfn->iova)
184 node = node->rb_left;
185 else if (iova > vpfn->iova)
186 node = node->rb_right;
193 static void vfio_link_pfn(struct vfio_dma *dma,
194 struct vfio_pfn *new)
196 struct rb_node **link, *parent = NULL;
197 struct vfio_pfn *vpfn;
199 link = &dma->pfn_list.rb_node;
202 vpfn = rb_entry(parent, struct vfio_pfn, node);
204 if (new->iova < vpfn->iova)
205 link = &(*link)->rb_left;
207 link = &(*link)->rb_right;
210 rb_link_node(&new->node, parent, link);
211 rb_insert_color(&new->node, &dma->pfn_list);
214 static void vfio_unlink_pfn(struct vfio_dma *dma, struct vfio_pfn *old)
216 rb_erase(&old->node, &dma->pfn_list);
219 static int vfio_add_to_pfn_list(struct vfio_dma *dma, dma_addr_t iova,
222 struct vfio_pfn *vpfn;
224 vpfn = kzalloc(sizeof(*vpfn), GFP_KERNEL);
230 atomic_set(&vpfn->ref_count, 1);
231 vfio_link_pfn(dma, vpfn);
235 static void vfio_remove_from_pfn_list(struct vfio_dma *dma,
236 struct vfio_pfn *vpfn)
238 vfio_unlink_pfn(dma, vpfn);
242 static struct vfio_pfn *vfio_iova_get_vfio_pfn(struct vfio_dma *dma,
245 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
248 atomic_inc(&vpfn->ref_count);
252 static int vfio_iova_put_vfio_pfn(struct vfio_dma *dma, struct vfio_pfn *vpfn)
256 if (atomic_dec_and_test(&vpfn->ref_count)) {
257 ret = put_pfn(vpfn->pfn, dma->prot);
258 vfio_remove_from_pfn_list(dma, vpfn);
263 static int vfio_lock_acct(struct vfio_dma *dma, long npage, bool async)
265 struct mm_struct *mm;
271 mm = async ? get_task_mm(dma->task) : dma->task->mm;
273 return -ESRCH; /* process exited */
275 ret = down_write_killable(&mm->mmap_sem);
278 if (!dma->lock_cap) {
281 limit = task_rlimit(dma->task,
282 RLIMIT_MEMLOCK) >> PAGE_SHIFT;
284 if (mm->locked_vm + npage > limit)
290 mm->locked_vm += npage;
292 up_write(&mm->mmap_sem);
302 * Some mappings aren't backed by a struct page, for example an mmap'd
303 * MMIO range for our own or another device. These use a different
304 * pfn conversion and shouldn't be tracked as locked pages.
306 static bool is_invalid_reserved_pfn(unsigned long pfn)
308 if (pfn_valid(pfn)) {
310 struct page *tail = pfn_to_page(pfn);
311 struct page *head = compound_head(tail);
312 reserved = !!(PageReserved(head));
315 * "head" is not a dangling pointer
316 * (compound_head takes care of that)
317 * but the hugepage may have been split
318 * from under us (and we may not hold a
319 * reference count on the head page so it can
320 * be reused before we run PageReferenced), so
321 * we've to check PageTail before returning
328 return PageReserved(tail);
334 static int put_pfn(unsigned long pfn, int prot)
336 if (!is_invalid_reserved_pfn(pfn)) {
337 struct page *page = pfn_to_page(pfn);
338 if (prot & IOMMU_WRITE)
346 static int follow_fault_pfn(struct vm_area_struct *vma, struct mm_struct *mm,
347 unsigned long vaddr, unsigned long *pfn,
352 ret = follow_pfn(vma, vaddr, pfn);
354 bool unlocked = false;
356 ret = fixup_user_fault(NULL, mm, vaddr,
358 (write_fault ? FAULT_FLAG_WRITE : 0),
366 ret = follow_pfn(vma, vaddr, pfn);
372 static int vaddr_get_pfn(struct mm_struct *mm, unsigned long vaddr,
373 int prot, unsigned long *pfn)
375 struct page *page[1];
376 struct vm_area_struct *vma;
377 struct vm_area_struct *vmas[1];
378 unsigned int flags = 0;
381 if (prot & IOMMU_WRITE)
384 down_read(&mm->mmap_sem);
385 if (mm == current->mm) {
386 ret = get_user_pages_longterm(vaddr, 1, flags, page, vmas);
388 ret = get_user_pages_remote(NULL, mm, vaddr, 1, flags, page,
391 * The lifetime of a vaddr_get_pfn() page pin is
392 * userspace-controlled. In the fs-dax case this could
393 * lead to indefinite stalls in filesystem operations.
394 * Disallow attempts to pin fs-dax pages via this
397 if (ret > 0 && vma_is_fsdax(vmas[0])) {
402 up_read(&mm->mmap_sem);
405 *pfn = page_to_pfn(page[0]);
409 down_read(&mm->mmap_sem);
412 vma = find_vma_intersection(mm, vaddr, vaddr + 1);
414 if (vma && vma->vm_flags & VM_PFNMAP) {
415 ret = follow_fault_pfn(vma, mm, vaddr, pfn, prot & IOMMU_WRITE);
419 if (!ret && !is_invalid_reserved_pfn(*pfn))
423 up_read(&mm->mmap_sem);
428 * Attempt to pin pages. We really don't want to track all the pfns and
429 * the iommu can only map chunks of consecutive pfns anyway, so get the
430 * first page and all consecutive pages with the same locking.
432 static long vfio_pin_pages_remote(struct vfio_dma *dma, unsigned long vaddr,
433 long npage, unsigned long *pfn_base,
436 unsigned long pfn = 0;
437 long ret, pinned = 0, lock_acct = 0;
439 dma_addr_t iova = vaddr - dma->vaddr + dma->iova;
441 /* This code path is only user initiated */
445 ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, pfn_base);
450 rsvd = is_invalid_reserved_pfn(*pfn_base);
453 * Reserved pages aren't counted against the user, externally pinned
454 * pages are already counted against the user.
456 if (!rsvd && !vfio_find_vpfn(dma, iova)) {
457 if (!dma->lock_cap && current->mm->locked_vm + 1 > limit) {
458 put_pfn(*pfn_base, dma->prot);
459 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n", __func__,
460 limit << PAGE_SHIFT);
466 if (unlikely(disable_hugepages))
469 /* Lock all the consecutive pages from pfn_base */
470 for (vaddr += PAGE_SIZE, iova += PAGE_SIZE; pinned < npage;
471 pinned++, vaddr += PAGE_SIZE, iova += PAGE_SIZE) {
472 ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, &pfn);
476 if (pfn != *pfn_base + pinned ||
477 rsvd != is_invalid_reserved_pfn(pfn)) {
478 put_pfn(pfn, dma->prot);
482 if (!rsvd && !vfio_find_vpfn(dma, iova)) {
483 if (!dma->lock_cap &&
484 current->mm->locked_vm + lock_acct + 1 > limit) {
485 put_pfn(pfn, dma->prot);
486 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n",
487 __func__, limit << PAGE_SHIFT);
496 ret = vfio_lock_acct(dma, lock_acct, false);
501 for (pfn = *pfn_base ; pinned ; pfn++, pinned--)
502 put_pfn(pfn, dma->prot);
511 static long vfio_unpin_pages_remote(struct vfio_dma *dma, dma_addr_t iova,
512 unsigned long pfn, long npage,
515 long unlocked = 0, locked = 0;
518 for (i = 0; i < npage; i++, iova += PAGE_SIZE) {
519 if (put_pfn(pfn++, dma->prot)) {
521 if (vfio_find_vpfn(dma, iova))
527 vfio_lock_acct(dma, locked - unlocked, true);
532 static int vfio_pin_page_external(struct vfio_dma *dma, unsigned long vaddr,
533 unsigned long *pfn_base, bool do_accounting)
535 struct mm_struct *mm;
538 mm = get_task_mm(dma->task);
542 ret = vaddr_get_pfn(mm, vaddr, dma->prot, pfn_base);
543 if (!ret && do_accounting && !is_invalid_reserved_pfn(*pfn_base)) {
544 ret = vfio_lock_acct(dma, 1, true);
546 put_pfn(*pfn_base, dma->prot);
548 pr_warn("%s: Task %s (%d) RLIMIT_MEMLOCK "
549 "(%ld) exceeded\n", __func__,
550 dma->task->comm, task_pid_nr(dma->task),
551 task_rlimit(dma->task, RLIMIT_MEMLOCK));
559 static int vfio_unpin_page_external(struct vfio_dma *dma, dma_addr_t iova,
563 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
568 unlocked = vfio_iova_put_vfio_pfn(dma, vpfn);
571 vfio_lock_acct(dma, -unlocked, true);
576 static int vfio_iommu_type1_pin_pages(void *iommu_data,
577 unsigned long *user_pfn,
579 unsigned long *phys_pfn)
581 struct vfio_iommu *iommu = iommu_data;
583 unsigned long remote_vaddr;
584 struct vfio_dma *dma;
587 if (!iommu || !user_pfn || !phys_pfn)
590 /* Supported for v2 version only */
594 mutex_lock(&iommu->lock);
596 /* Fail if notifier list is empty */
597 if ((!iommu->external_domain) || (!iommu->notifier.head)) {
603 * If iommu capable domain exist in the container then all pages are
604 * already pinned and accounted. Accouting should be done if there is no
605 * iommu capable domain in the container.
607 do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu);
609 for (i = 0; i < npage; i++) {
611 struct vfio_pfn *vpfn;
613 iova = user_pfn[i] << PAGE_SHIFT;
614 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
620 if ((dma->prot & prot) != prot) {
625 vpfn = vfio_iova_get_vfio_pfn(dma, iova);
627 phys_pfn[i] = vpfn->pfn;
631 remote_vaddr = dma->vaddr + (iova - dma->iova);
632 ret = vfio_pin_page_external(dma, remote_vaddr, &phys_pfn[i],
637 ret = vfio_add_to_pfn_list(dma, iova, phys_pfn[i]);
639 if (put_pfn(phys_pfn[i], dma->prot) && do_accounting)
640 vfio_lock_acct(dma, -1, true);
650 for (j = 0; j < i; j++) {
653 iova = user_pfn[j] << PAGE_SHIFT;
654 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
655 vfio_unpin_page_external(dma, iova, do_accounting);
659 mutex_unlock(&iommu->lock);
663 static int vfio_iommu_type1_unpin_pages(void *iommu_data,
664 unsigned long *user_pfn,
667 struct vfio_iommu *iommu = iommu_data;
671 if (!iommu || !user_pfn)
674 /* Supported for v2 version only */
678 mutex_lock(&iommu->lock);
680 if (!iommu->external_domain) {
681 mutex_unlock(&iommu->lock);
685 do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu);
686 for (i = 0; i < npage; i++) {
687 struct vfio_dma *dma;
690 iova = user_pfn[i] << PAGE_SHIFT;
691 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
694 vfio_unpin_page_external(dma, iova, do_accounting);
698 mutex_unlock(&iommu->lock);
699 return i > npage ? npage : (i > 0 ? i : -EINVAL);
702 static long vfio_sync_unpin(struct vfio_dma *dma, struct vfio_domain *domain,
703 struct list_head *regions)
706 struct vfio_regions *entry, *next;
708 iommu_tlb_sync(domain->domain);
710 list_for_each_entry_safe(entry, next, regions, list) {
711 unlocked += vfio_unpin_pages_remote(dma,
713 entry->phys >> PAGE_SHIFT,
714 entry->len >> PAGE_SHIFT,
716 list_del(&entry->list);
726 * Generally, VFIO needs to unpin remote pages after each IOTLB flush.
727 * Therefore, when using IOTLB flush sync interface, VFIO need to keep track
728 * of these regions (currently using a list).
730 * This value specifies maximum number of regions for each IOTLB flush sync.
732 #define VFIO_IOMMU_TLB_SYNC_MAX 512
734 static size_t unmap_unpin_fast(struct vfio_domain *domain,
735 struct vfio_dma *dma, dma_addr_t *iova,
736 size_t len, phys_addr_t phys, long *unlocked,
737 struct list_head *unmapped_list,
741 struct vfio_regions *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
744 unmapped = iommu_unmap_fast(domain->domain, *iova, len);
749 iommu_tlb_range_add(domain->domain, *iova, unmapped);
752 entry->len = unmapped;
753 list_add_tail(&entry->list, unmapped_list);
761 * Sync if the number of fast-unmap regions hits the limit
762 * or in case of errors.
764 if (*unmapped_cnt >= VFIO_IOMMU_TLB_SYNC_MAX || !unmapped) {
765 *unlocked += vfio_sync_unpin(dma, domain,
773 static size_t unmap_unpin_slow(struct vfio_domain *domain,
774 struct vfio_dma *dma, dma_addr_t *iova,
775 size_t len, phys_addr_t phys,
778 size_t unmapped = iommu_unmap(domain->domain, *iova, len);
781 *unlocked += vfio_unpin_pages_remote(dma, *iova,
783 unmapped >> PAGE_SHIFT,
791 static long vfio_unmap_unpin(struct vfio_iommu *iommu, struct vfio_dma *dma,
794 dma_addr_t iova = dma->iova, end = dma->iova + dma->size;
795 struct vfio_domain *domain, *d;
796 LIST_HEAD(unmapped_region_list);
797 int unmapped_region_cnt = 0;
803 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
807 * We use the IOMMU to track the physical addresses, otherwise we'd
808 * need a much more complicated tracking system. Unfortunately that
809 * means we need to use one of the iommu domains to figure out the
810 * pfns to unpin. The rest need to be unmapped in advance so we have
811 * no iommu translations remaining when the pages are unpinned.
813 domain = d = list_first_entry(&iommu->domain_list,
814 struct vfio_domain, next);
816 list_for_each_entry_continue(d, &iommu->domain_list, next) {
817 iommu_unmap(d->domain, dma->iova, dma->size);
822 size_t unmapped, len;
823 phys_addr_t phys, next;
825 phys = iommu_iova_to_phys(domain->domain, iova);
826 if (WARN_ON(!phys)) {
832 * To optimize for fewer iommu_unmap() calls, each of which
833 * may require hardware cache flushing, try to find the
834 * largest contiguous physical memory chunk to unmap.
836 for (len = PAGE_SIZE;
837 !domain->fgsp && iova + len < end; len += PAGE_SIZE) {
838 next = iommu_iova_to_phys(domain->domain, iova + len);
839 if (next != phys + len)
844 * First, try to use fast unmap/unpin. In case of failure,
845 * switch to slow unmap/unpin path.
847 unmapped = unmap_unpin_fast(domain, dma, &iova, len, phys,
848 &unlocked, &unmapped_region_list,
849 &unmapped_region_cnt);
851 unmapped = unmap_unpin_slow(domain, dma, &iova, len,
853 if (WARN_ON(!unmapped))
858 dma->iommu_mapped = false;
860 if (unmapped_region_cnt)
861 unlocked += vfio_sync_unpin(dma, domain, &unmapped_region_list);
864 vfio_lock_acct(dma, -unlocked, true);
870 static void vfio_remove_dma(struct vfio_iommu *iommu, struct vfio_dma *dma)
872 vfio_unmap_unpin(iommu, dma, true);
873 vfio_unlink_dma(iommu, dma);
874 put_task_struct(dma->task);
879 static unsigned long vfio_pgsize_bitmap(struct vfio_iommu *iommu)
881 struct vfio_domain *domain;
882 unsigned long bitmap = ULONG_MAX;
884 mutex_lock(&iommu->lock);
885 list_for_each_entry(domain, &iommu->domain_list, next)
886 bitmap &= domain->domain->pgsize_bitmap;
887 mutex_unlock(&iommu->lock);
890 * In case the IOMMU supports page sizes smaller than PAGE_SIZE
891 * we pretend PAGE_SIZE is supported and hide sub-PAGE_SIZE sizes.
892 * That way the user will be able to map/unmap buffers whose size/
893 * start address is aligned with PAGE_SIZE. Pinning code uses that
894 * granularity while iommu driver can use the sub-PAGE_SIZE size
897 if (bitmap & ~PAGE_MASK) {
905 static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
906 struct vfio_iommu_type1_dma_unmap *unmap)
909 struct vfio_dma *dma, *dma_last = NULL;
911 int ret = 0, retries = 0;
913 mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1;
915 if (unmap->iova & mask)
917 if (!unmap->size || unmap->size & mask)
919 if (unmap->iova + unmap->size - 1 < unmap->iova ||
920 unmap->size > SIZE_MAX)
923 WARN_ON(mask & PAGE_MASK);
925 mutex_lock(&iommu->lock);
928 * vfio-iommu-type1 (v1) - User mappings were coalesced together to
929 * avoid tracking individual mappings. This means that the granularity
930 * of the original mapping was lost and the user was allowed to attempt
931 * to unmap any range. Depending on the contiguousness of physical
932 * memory and page sizes supported by the IOMMU, arbitrary unmaps may
933 * or may not have worked. We only guaranteed unmap granularity
934 * matching the original mapping; even though it was untracked here,
935 * the original mappings are reflected in IOMMU mappings. This
936 * resulted in a couple unusual behaviors. First, if a range is not
937 * able to be unmapped, ex. a set of 4k pages that was mapped as a
938 * 2M hugepage into the IOMMU, the unmap ioctl returns success but with
939 * a zero sized unmap. Also, if an unmap request overlaps the first
940 * address of a hugepage, the IOMMU will unmap the entire hugepage.
941 * This also returns success and the returned unmap size reflects the
942 * actual size unmapped.
944 * We attempt to maintain compatibility with this "v1" interface, but
945 * we take control out of the hands of the IOMMU. Therefore, an unmap
946 * request offset from the beginning of the original mapping will
947 * return success with zero sized unmap. And an unmap request covering
948 * the first iova of mapping will unmap the entire range.
950 * The v2 version of this interface intends to be more deterministic.
951 * Unmap requests must fully cover previous mappings. Multiple
952 * mappings may still be unmaped by specifying large ranges, but there
953 * must not be any previous mappings bisected by the range. An error
954 * will be returned if these conditions are not met. The v2 interface
955 * will only return success and a size of zero if there were no
956 * mappings within the range.
959 dma = vfio_find_dma(iommu, unmap->iova, 1);
960 if (dma && dma->iova != unmap->iova) {
964 dma = vfio_find_dma(iommu, unmap->iova + unmap->size - 1, 0);
965 if (dma && dma->iova + dma->size != unmap->iova + unmap->size) {
971 while ((dma = vfio_find_dma(iommu, unmap->iova, unmap->size))) {
972 if (!iommu->v2 && unmap->iova > dma->iova)
975 * Task with same address space who mapped this iova range is
976 * allowed to unmap the iova range.
978 if (dma->task->mm != current->mm)
981 if (!RB_EMPTY_ROOT(&dma->pfn_list)) {
982 struct vfio_iommu_type1_dma_unmap nb_unmap;
984 if (dma_last == dma) {
985 BUG_ON(++retries > 10);
991 nb_unmap.iova = dma->iova;
992 nb_unmap.size = dma->size;
995 * Notify anyone (mdev vendor drivers) to invalidate and
996 * unmap iovas within the range we're about to unmap.
997 * Vendor drivers MUST unpin pages in response to an
1000 mutex_unlock(&iommu->lock);
1001 blocking_notifier_call_chain(&iommu->notifier,
1002 VFIO_IOMMU_NOTIFY_DMA_UNMAP,
1006 unmapped += dma->size;
1007 vfio_remove_dma(iommu, dma);
1011 mutex_unlock(&iommu->lock);
1013 /* Report how much was unmapped */
1014 unmap->size = unmapped;
1020 * Turns out AMD IOMMU has a page table bug where it won't map large pages
1021 * to a region that previously mapped smaller pages. This should be fixed
1022 * soon, so this is just a temporary workaround to break mappings down into
1023 * PAGE_SIZE. Better to map smaller pages than nothing.
1025 static int map_try_harder(struct vfio_domain *domain, dma_addr_t iova,
1026 unsigned long pfn, long npage, int prot)
1031 for (i = 0; i < npage; i++, pfn++, iova += PAGE_SIZE) {
1032 ret = iommu_map(domain->domain, iova,
1033 (phys_addr_t)pfn << PAGE_SHIFT,
1034 PAGE_SIZE, prot | domain->prot);
1039 for (; i < npage && i > 0; i--, iova -= PAGE_SIZE)
1040 iommu_unmap(domain->domain, iova, PAGE_SIZE);
1045 static int vfio_iommu_map(struct vfio_iommu *iommu, dma_addr_t iova,
1046 unsigned long pfn, long npage, int prot)
1048 struct vfio_domain *d;
1051 list_for_each_entry(d, &iommu->domain_list, next) {
1052 ret = iommu_map(d->domain, iova, (phys_addr_t)pfn << PAGE_SHIFT,
1053 npage << PAGE_SHIFT, prot | d->prot);
1055 if (ret != -EBUSY ||
1056 map_try_harder(d, iova, pfn, npage, prot))
1066 list_for_each_entry_continue_reverse(d, &iommu->domain_list, next)
1067 iommu_unmap(d->domain, iova, npage << PAGE_SHIFT);
1072 static int vfio_pin_map_dma(struct vfio_iommu *iommu, struct vfio_dma *dma,
1075 dma_addr_t iova = dma->iova;
1076 unsigned long vaddr = dma->vaddr;
1077 size_t size = map_size;
1079 unsigned long pfn, limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1083 /* Pin a contiguous chunk of memory */
1084 npage = vfio_pin_pages_remote(dma, vaddr + dma->size,
1085 size >> PAGE_SHIFT, &pfn, limit);
1093 ret = vfio_iommu_map(iommu, iova + dma->size, pfn, npage,
1096 vfio_unpin_pages_remote(dma, iova + dma->size, pfn,
1101 size -= npage << PAGE_SHIFT;
1102 dma->size += npage << PAGE_SHIFT;
1105 dma->iommu_mapped = true;
1108 vfio_remove_dma(iommu, dma);
1113 static int vfio_dma_do_map(struct vfio_iommu *iommu,
1114 struct vfio_iommu_type1_dma_map *map)
1116 dma_addr_t iova = map->iova;
1117 unsigned long vaddr = map->vaddr;
1118 size_t size = map->size;
1119 int ret = 0, prot = 0;
1121 struct vfio_dma *dma;
1123 /* Verify that none of our __u64 fields overflow */
1124 if (map->size != size || map->vaddr != vaddr || map->iova != iova)
1127 mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1;
1129 WARN_ON(mask & PAGE_MASK);
1131 /* READ/WRITE from device perspective */
1132 if (map->flags & VFIO_DMA_MAP_FLAG_WRITE)
1133 prot |= IOMMU_WRITE;
1134 if (map->flags & VFIO_DMA_MAP_FLAG_READ)
1137 if (!prot || !size || (size | iova | vaddr) & mask)
1140 /* Don't allow IOVA or virtual address wrap */
1141 if (iova + size - 1 < iova || vaddr + size - 1 < vaddr)
1144 mutex_lock(&iommu->lock);
1146 if (vfio_find_dma(iommu, iova, size)) {
1151 if (!iommu->dma_avail) {
1156 dma = kzalloc(sizeof(*dma), GFP_KERNEL);
1168 * We need to be able to both add to a task's locked memory and test
1169 * against the locked memory limit and we need to be able to do both
1170 * outside of this call path as pinning can be asynchronous via the
1171 * external interfaces for mdev devices. RLIMIT_MEMLOCK requires a
1172 * task_struct and VM locked pages requires an mm_struct, however
1173 * holding an indefinite mm reference is not recommended, therefore we
1174 * only hold a reference to a task. We could hold a reference to
1175 * current, however QEMU uses this call path through vCPU threads,
1176 * which can be killed resulting in a NULL mm and failure in the unmap
1177 * path when called via a different thread. Avoid this problem by
1178 * using the group_leader as threads within the same group require
1179 * both CLONE_THREAD and CLONE_VM and will therefore use the same
1182 * Previously we also used the task for testing CAP_IPC_LOCK at the
1183 * time of pinning and accounting, however has_capability() makes use
1184 * of real_cred, a copy-on-write field, so we can't guarantee that it
1185 * matches group_leader, or in fact that it might not change by the
1186 * time it's evaluated. If a process were to call MAP_DMA with
1187 * CAP_IPC_LOCK but later drop it, it doesn't make sense that they
1188 * possibly see different results for an iommu_mapped vfio_dma vs
1189 * externally mapped. Therefore track CAP_IPC_LOCK in vfio_dma at the
1190 * time of calling MAP_DMA.
1192 get_task_struct(current->group_leader);
1193 dma->task = current->group_leader;
1194 dma->lock_cap = capable(CAP_IPC_LOCK);
1196 dma->pfn_list = RB_ROOT;
1198 /* Insert zero-sized and grow as we map chunks of it */
1199 vfio_link_dma(iommu, dma);
1201 /* Don't pin and map if container doesn't contain IOMMU capable domain*/
1202 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
1205 ret = vfio_pin_map_dma(iommu, dma, size);
1208 mutex_unlock(&iommu->lock);
1212 static int vfio_bus_type(struct device *dev, void *data)
1214 struct bus_type **bus = data;
1216 if (*bus && *bus != dev->bus)
1224 static int vfio_iommu_replay(struct vfio_iommu *iommu,
1225 struct vfio_domain *domain)
1227 struct vfio_domain *d = NULL;
1229 unsigned long limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1232 /* Arbitrarily pick the first domain in the list for lookups */
1233 if (!list_empty(&iommu->domain_list))
1234 d = list_first_entry(&iommu->domain_list,
1235 struct vfio_domain, next);
1237 n = rb_first(&iommu->dma_list);
1239 for (; n; n = rb_next(n)) {
1240 struct vfio_dma *dma;
1243 dma = rb_entry(n, struct vfio_dma, node);
1246 while (iova < dma->iova + dma->size) {
1250 if (dma->iommu_mapped) {
1254 if (WARN_ON(!d)) { /* mapped w/o a domain?! */
1259 phys = iommu_iova_to_phys(d->domain, iova);
1261 if (WARN_ON(!phys)) {
1269 while (i < dma->iova + dma->size &&
1270 p == iommu_iova_to_phys(d->domain, i)) {
1277 unsigned long vaddr = dma->vaddr +
1279 size_t n = dma->iova + dma->size - iova;
1282 npage = vfio_pin_pages_remote(dma, vaddr,
1291 phys = pfn << PAGE_SHIFT;
1292 size = npage << PAGE_SHIFT;
1295 ret = iommu_map(domain->domain, iova, phys,
1296 size, dma->prot | domain->prot);
1298 if (!dma->iommu_mapped)
1299 vfio_unpin_pages_remote(dma, iova,
1310 /* All dmas are now mapped, defer to second tree walk for unwind */
1311 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
1312 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1314 dma->iommu_mapped = true;
1320 for (; n; n = rb_prev(n)) {
1321 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1324 if (dma->iommu_mapped) {
1325 iommu_unmap(domain->domain, dma->iova, dma->size);
1330 while (iova < dma->iova + dma->size) {
1331 phys_addr_t phys, p;
1335 phys = iommu_iova_to_phys(domain->domain, iova);
1344 while (i < dma->iova + dma->size &&
1345 p == iommu_iova_to_phys(domain->domain, i)) {
1351 iommu_unmap(domain->domain, iova, size);
1352 vfio_unpin_pages_remote(dma, iova, phys >> PAGE_SHIFT,
1353 size >> PAGE_SHIFT, true);
1361 * We change our unmap behavior slightly depending on whether the IOMMU
1362 * supports fine-grained superpages. IOMMUs like AMD-Vi will use a superpage
1363 * for practically any contiguous power-of-two mapping we give it. This means
1364 * we don't need to look for contiguous chunks ourselves to make unmapping
1365 * more efficient. On IOMMUs with coarse-grained super pages, like Intel VT-d
1366 * with discrete 2M/1G/512G/1T superpages, identifying contiguous chunks
1367 * significantly boosts non-hugetlbfs mappings and doesn't seem to hurt when
1368 * hugetlbfs is in use.
1370 static void vfio_test_domain_fgsp(struct vfio_domain *domain)
1373 int ret, order = get_order(PAGE_SIZE * 2);
1375 pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order);
1379 ret = iommu_map(domain->domain, 0, page_to_phys(pages), PAGE_SIZE * 2,
1380 IOMMU_READ | IOMMU_WRITE | domain->prot);
1382 size_t unmapped = iommu_unmap(domain->domain, 0, PAGE_SIZE);
1384 if (unmapped == PAGE_SIZE)
1385 iommu_unmap(domain->domain, PAGE_SIZE, PAGE_SIZE);
1387 domain->fgsp = true;
1390 __free_pages(pages, order);
1393 static struct vfio_group *find_iommu_group(struct vfio_domain *domain,
1394 struct iommu_group *iommu_group)
1396 struct vfio_group *g;
1398 list_for_each_entry(g, &domain->group_list, next) {
1399 if (g->iommu_group == iommu_group)
1406 static bool vfio_iommu_has_sw_msi(struct iommu_group *group, phys_addr_t *base)
1408 struct list_head group_resv_regions;
1409 struct iommu_resv_region *region, *next;
1412 INIT_LIST_HEAD(&group_resv_regions);
1413 iommu_get_group_resv_regions(group, &group_resv_regions);
1414 list_for_each_entry(region, &group_resv_regions, list) {
1416 * The presence of any 'real' MSI regions should take
1417 * precedence over the software-managed one if the
1418 * IOMMU driver happens to advertise both types.
1420 if (region->type == IOMMU_RESV_MSI) {
1425 if (region->type == IOMMU_RESV_SW_MSI) {
1426 *base = region->start;
1430 list_for_each_entry_safe(region, next, &group_resv_regions, list)
1435 static int vfio_iommu_type1_attach_group(void *iommu_data,
1436 struct iommu_group *iommu_group)
1438 struct vfio_iommu *iommu = iommu_data;
1439 struct vfio_group *group;
1440 struct vfio_domain *domain, *d;
1441 struct bus_type *bus = NULL, *mdev_bus;
1443 bool resv_msi, msi_remap;
1444 phys_addr_t resv_msi_base;
1446 mutex_lock(&iommu->lock);
1448 list_for_each_entry(d, &iommu->domain_list, next) {
1449 if (find_iommu_group(d, iommu_group)) {
1450 mutex_unlock(&iommu->lock);
1455 if (iommu->external_domain) {
1456 if (find_iommu_group(iommu->external_domain, iommu_group)) {
1457 mutex_unlock(&iommu->lock);
1462 group = kzalloc(sizeof(*group), GFP_KERNEL);
1463 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
1464 if (!group || !domain) {
1469 group->iommu_group = iommu_group;
1471 /* Determine bus_type in order to allocate a domain */
1472 ret = iommu_group_for_each_dev(iommu_group, &bus, vfio_bus_type);
1476 mdev_bus = symbol_get(mdev_bus_type);
1479 if ((bus == mdev_bus) && !iommu_present(bus)) {
1480 symbol_put(mdev_bus_type);
1481 if (!iommu->external_domain) {
1482 INIT_LIST_HEAD(&domain->group_list);
1483 iommu->external_domain = domain;
1487 list_add(&group->next,
1488 &iommu->external_domain->group_list);
1489 mutex_unlock(&iommu->lock);
1492 symbol_put(mdev_bus_type);
1495 domain->domain = iommu_domain_alloc(bus);
1496 if (!domain->domain) {
1501 if (iommu->nesting) {
1504 ret = iommu_domain_set_attr(domain->domain, DOMAIN_ATTR_NESTING,
1510 ret = iommu_attach_group(domain->domain, iommu_group);
1514 resv_msi = vfio_iommu_has_sw_msi(iommu_group, &resv_msi_base);
1516 INIT_LIST_HEAD(&domain->group_list);
1517 list_add(&group->next, &domain->group_list);
1519 msi_remap = irq_domain_check_msi_remap() ||
1520 iommu_capable(bus, IOMMU_CAP_INTR_REMAP);
1522 if (!allow_unsafe_interrupts && !msi_remap) {
1523 pr_warn("%s: No interrupt remapping support. Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n",
1529 if (iommu_capable(bus, IOMMU_CAP_CACHE_COHERENCY))
1530 domain->prot |= IOMMU_CACHE;
1533 * Try to match an existing compatible domain. We don't want to
1534 * preclude an IOMMU driver supporting multiple bus_types and being
1535 * able to include different bus_types in the same IOMMU domain, so
1536 * we test whether the domains use the same iommu_ops rather than
1537 * testing if they're on the same bus_type.
1539 list_for_each_entry(d, &iommu->domain_list, next) {
1540 if (d->domain->ops == domain->domain->ops &&
1541 d->prot == domain->prot) {
1542 iommu_detach_group(domain->domain, iommu_group);
1543 if (!iommu_attach_group(d->domain, iommu_group)) {
1544 list_add(&group->next, &d->group_list);
1545 iommu_domain_free(domain->domain);
1547 mutex_unlock(&iommu->lock);
1551 ret = iommu_attach_group(domain->domain, iommu_group);
1557 vfio_test_domain_fgsp(domain);
1559 /* replay mappings on new domains */
1560 ret = vfio_iommu_replay(iommu, domain);
1565 ret = iommu_get_msi_cookie(domain->domain, resv_msi_base);
1570 list_add(&domain->next, &iommu->domain_list);
1572 mutex_unlock(&iommu->lock);
1577 iommu_detach_group(domain->domain, iommu_group);
1579 iommu_domain_free(domain->domain);
1583 mutex_unlock(&iommu->lock);
1587 static void vfio_iommu_unmap_unpin_all(struct vfio_iommu *iommu)
1589 struct rb_node *node;
1591 while ((node = rb_first(&iommu->dma_list)))
1592 vfio_remove_dma(iommu, rb_entry(node, struct vfio_dma, node));
1595 static void vfio_iommu_unmap_unpin_reaccount(struct vfio_iommu *iommu)
1597 struct rb_node *n, *p;
1599 n = rb_first(&iommu->dma_list);
1600 for (; n; n = rb_next(n)) {
1601 struct vfio_dma *dma;
1602 long locked = 0, unlocked = 0;
1604 dma = rb_entry(n, struct vfio_dma, node);
1605 unlocked += vfio_unmap_unpin(iommu, dma, false);
1606 p = rb_first(&dma->pfn_list);
1607 for (; p; p = rb_next(p)) {
1608 struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn,
1611 if (!is_invalid_reserved_pfn(vpfn->pfn))
1614 vfio_lock_acct(dma, locked - unlocked, true);
1618 static void vfio_sanity_check_pfn_list(struct vfio_iommu *iommu)
1622 n = rb_first(&iommu->dma_list);
1623 for (; n; n = rb_next(n)) {
1624 struct vfio_dma *dma;
1626 dma = rb_entry(n, struct vfio_dma, node);
1628 if (WARN_ON(!RB_EMPTY_ROOT(&dma->pfn_list)))
1631 /* mdev vendor driver must unregister notifier */
1632 WARN_ON(iommu->notifier.head);
1635 static void vfio_iommu_type1_detach_group(void *iommu_data,
1636 struct iommu_group *iommu_group)
1638 struct vfio_iommu *iommu = iommu_data;
1639 struct vfio_domain *domain;
1640 struct vfio_group *group;
1642 mutex_lock(&iommu->lock);
1644 if (iommu->external_domain) {
1645 group = find_iommu_group(iommu->external_domain, iommu_group);
1647 list_del(&group->next);
1650 if (list_empty(&iommu->external_domain->group_list)) {
1651 vfio_sanity_check_pfn_list(iommu);
1653 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
1654 vfio_iommu_unmap_unpin_all(iommu);
1656 kfree(iommu->external_domain);
1657 iommu->external_domain = NULL;
1659 goto detach_group_done;
1663 list_for_each_entry(domain, &iommu->domain_list, next) {
1664 group = find_iommu_group(domain, iommu_group);
1668 iommu_detach_group(domain->domain, iommu_group);
1669 list_del(&group->next);
1672 * Group ownership provides privilege, if the group list is
1673 * empty, the domain goes away. If it's the last domain with
1674 * iommu and external domain doesn't exist, then all the
1675 * mappings go away too. If it's the last domain with iommu and
1676 * external domain exist, update accounting
1678 if (list_empty(&domain->group_list)) {
1679 if (list_is_singular(&iommu->domain_list)) {
1680 if (!iommu->external_domain)
1681 vfio_iommu_unmap_unpin_all(iommu);
1683 vfio_iommu_unmap_unpin_reaccount(iommu);
1685 iommu_domain_free(domain->domain);
1686 list_del(&domain->next);
1693 mutex_unlock(&iommu->lock);
1696 static void *vfio_iommu_type1_open(unsigned long arg)
1698 struct vfio_iommu *iommu;
1700 iommu = kzalloc(sizeof(*iommu), GFP_KERNEL);
1702 return ERR_PTR(-ENOMEM);
1705 case VFIO_TYPE1_IOMMU:
1707 case VFIO_TYPE1_NESTING_IOMMU:
1708 iommu->nesting = true;
1710 case VFIO_TYPE1v2_IOMMU:
1715 return ERR_PTR(-EINVAL);
1718 INIT_LIST_HEAD(&iommu->domain_list);
1719 iommu->dma_list = RB_ROOT;
1720 iommu->dma_avail = dma_entry_limit;
1721 mutex_init(&iommu->lock);
1722 BLOCKING_INIT_NOTIFIER_HEAD(&iommu->notifier);
1727 static void vfio_release_domain(struct vfio_domain *domain, bool external)
1729 struct vfio_group *group, *group_tmp;
1731 list_for_each_entry_safe(group, group_tmp,
1732 &domain->group_list, next) {
1734 iommu_detach_group(domain->domain, group->iommu_group);
1735 list_del(&group->next);
1740 iommu_domain_free(domain->domain);
1743 static void vfio_iommu_type1_release(void *iommu_data)
1745 struct vfio_iommu *iommu = iommu_data;
1746 struct vfio_domain *domain, *domain_tmp;
1748 if (iommu->external_domain) {
1749 vfio_release_domain(iommu->external_domain, true);
1750 vfio_sanity_check_pfn_list(iommu);
1751 kfree(iommu->external_domain);
1754 vfio_iommu_unmap_unpin_all(iommu);
1756 list_for_each_entry_safe(domain, domain_tmp,
1757 &iommu->domain_list, next) {
1758 vfio_release_domain(domain, false);
1759 list_del(&domain->next);
1765 static int vfio_domains_have_iommu_cache(struct vfio_iommu *iommu)
1767 struct vfio_domain *domain;
1770 mutex_lock(&iommu->lock);
1771 list_for_each_entry(domain, &iommu->domain_list, next) {
1772 if (!(domain->prot & IOMMU_CACHE)) {
1777 mutex_unlock(&iommu->lock);
1782 static long vfio_iommu_type1_ioctl(void *iommu_data,
1783 unsigned int cmd, unsigned long arg)
1785 struct vfio_iommu *iommu = iommu_data;
1786 unsigned long minsz;
1788 if (cmd == VFIO_CHECK_EXTENSION) {
1790 case VFIO_TYPE1_IOMMU:
1791 case VFIO_TYPE1v2_IOMMU:
1792 case VFIO_TYPE1_NESTING_IOMMU:
1794 case VFIO_DMA_CC_IOMMU:
1797 return vfio_domains_have_iommu_cache(iommu);
1801 } else if (cmd == VFIO_IOMMU_GET_INFO) {
1802 struct vfio_iommu_type1_info info;
1804 minsz = offsetofend(struct vfio_iommu_type1_info, iova_pgsizes);
1806 if (copy_from_user(&info, (void __user *)arg, minsz))
1809 if (info.argsz < minsz)
1812 info.flags = VFIO_IOMMU_INFO_PGSIZES;
1814 info.iova_pgsizes = vfio_pgsize_bitmap(iommu);
1816 return copy_to_user((void __user *)arg, &info, minsz) ?
1819 } else if (cmd == VFIO_IOMMU_MAP_DMA) {
1820 struct vfio_iommu_type1_dma_map map;
1821 uint32_t mask = VFIO_DMA_MAP_FLAG_READ |
1822 VFIO_DMA_MAP_FLAG_WRITE;
1824 minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
1826 if (copy_from_user(&map, (void __user *)arg, minsz))
1829 if (map.argsz < minsz || map.flags & ~mask)
1832 return vfio_dma_do_map(iommu, &map);
1834 } else if (cmd == VFIO_IOMMU_UNMAP_DMA) {
1835 struct vfio_iommu_type1_dma_unmap unmap;
1838 minsz = offsetofend(struct vfio_iommu_type1_dma_unmap, size);
1840 if (copy_from_user(&unmap, (void __user *)arg, minsz))
1843 if (unmap.argsz < minsz || unmap.flags)
1846 ret = vfio_dma_do_unmap(iommu, &unmap);
1850 return copy_to_user((void __user *)arg, &unmap, minsz) ?
1857 static int vfio_iommu_type1_register_notifier(void *iommu_data,
1858 unsigned long *events,
1859 struct notifier_block *nb)
1861 struct vfio_iommu *iommu = iommu_data;
1863 /* clear known events */
1864 *events &= ~VFIO_IOMMU_NOTIFY_DMA_UNMAP;
1866 /* refuse to register if still events remaining */
1870 return blocking_notifier_chain_register(&iommu->notifier, nb);
1873 static int vfio_iommu_type1_unregister_notifier(void *iommu_data,
1874 struct notifier_block *nb)
1876 struct vfio_iommu *iommu = iommu_data;
1878 return blocking_notifier_chain_unregister(&iommu->notifier, nb);
1881 static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = {
1882 .name = "vfio-iommu-type1",
1883 .owner = THIS_MODULE,
1884 .open = vfio_iommu_type1_open,
1885 .release = vfio_iommu_type1_release,
1886 .ioctl = vfio_iommu_type1_ioctl,
1887 .attach_group = vfio_iommu_type1_attach_group,
1888 .detach_group = vfio_iommu_type1_detach_group,
1889 .pin_pages = vfio_iommu_type1_pin_pages,
1890 .unpin_pages = vfio_iommu_type1_unpin_pages,
1891 .register_notifier = vfio_iommu_type1_register_notifier,
1892 .unregister_notifier = vfio_iommu_type1_unregister_notifier,
1895 static int __init vfio_iommu_type1_init(void)
1897 return vfio_register_iommu_driver(&vfio_iommu_driver_ops_type1);
1900 static void __exit vfio_iommu_type1_cleanup(void)
1902 vfio_unregister_iommu_driver(&vfio_iommu_driver_ops_type1);
1905 module_init(vfio_iommu_type1_init);
1906 module_exit(vfio_iommu_type1_cleanup);
1908 MODULE_VERSION(DRIVER_VERSION);
1909 MODULE_LICENSE("GPL v2");
1910 MODULE_AUTHOR(DRIVER_AUTHOR);
1911 MODULE_DESCRIPTION(DRIVER_DESC);