1 // SPDX-License-Identifier: GPL-2.0-only
3 * VFIO: IOMMU DMA mapping support for Type1 IOMMU
5 * Copyright (C) 2012 Red Hat, Inc. All rights reserved.
6 * Author: Alex Williamson <alex.williamson@redhat.com>
8 * Derived from original vfio:
9 * Copyright 2010 Cisco Systems, Inc. All rights reserved.
10 * Author: Tom Lyon, pugs@cisco.com
12 * We arbitrarily define a Type1 IOMMU as one matching the below code.
13 * It could be called the x86 IOMMU as it's designed for AMD-Vi & Intel
14 * VT-d, but that makes it harder to re-use as theoretically anyone
15 * implementing a similar IOMMU could make use of this. We expect the
16 * IOMMU to support the IOMMU API and have few to no restrictions around
17 * the IOVA range that can be mapped. The Type1 IOMMU is currently
18 * optimized for relatively static mappings of a userspace process with
19 * userpsace pages pinned into memory. We also assume devices and IOMMU
20 * domains are PCI based as the IOMMU API is still centered around a
21 * device/bus interface rather than a group interface.
24 #include <linux/compat.h>
25 #include <linux/device.h>
27 #include <linux/highmem.h>
28 #include <linux/iommu.h>
29 #include <linux/module.h>
31 #include <linux/kthread.h>
32 #include <linux/rbtree.h>
33 #include <linux/sched/signal.h>
34 #include <linux/sched/mm.h>
35 #include <linux/slab.h>
36 #include <linux/uaccess.h>
37 #include <linux/vfio.h>
38 #include <linux/workqueue.h>
39 #include <linux/mdev.h>
40 #include <linux/notifier.h>
41 #include <linux/dma-iommu.h>
42 #include <linux/irqdomain.h>
44 #define DRIVER_VERSION "0.2"
45 #define DRIVER_AUTHOR "Alex Williamson <alex.williamson@redhat.com>"
46 #define DRIVER_DESC "Type1 IOMMU driver for VFIO"
48 static bool allow_unsafe_interrupts;
49 module_param_named(allow_unsafe_interrupts,
50 allow_unsafe_interrupts, bool, S_IRUGO | S_IWUSR);
51 MODULE_PARM_DESC(allow_unsafe_interrupts,
52 "Enable VFIO IOMMU support for on platforms without interrupt remapping support.");
54 static bool disable_hugepages;
55 module_param_named(disable_hugepages,
56 disable_hugepages, bool, S_IRUGO | S_IWUSR);
57 MODULE_PARM_DESC(disable_hugepages,
58 "Disable VFIO IOMMU support for IOMMU hugepages.");
60 static unsigned int dma_entry_limit __read_mostly = U16_MAX;
61 module_param_named(dma_entry_limit, dma_entry_limit, uint, 0644);
62 MODULE_PARM_DESC(dma_entry_limit,
63 "Maximum number of user DMA mappings per container (65535).");
66 struct list_head domain_list;
67 struct list_head iova_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;
73 uint64_t pgsize_bitmap;
76 bool dirty_page_tracking;
77 bool pinned_page_dirty_scope;
81 struct iommu_domain *domain;
82 struct list_head next;
83 struct list_head group_list;
84 int prot; /* IOMMU_CACHE */
85 bool fgsp; /* Fine-grained super pages */
90 dma_addr_t iova; /* Device address */
91 unsigned long vaddr; /* Process virtual addr */
92 size_t size; /* Map size (bytes) */
93 int prot; /* IOMMU_READ/WRITE */
95 bool lock_cap; /* capable(CAP_IPC_LOCK) */
96 struct task_struct *task;
97 struct rb_root pfn_list; /* Ex-user pinned pfn list */
98 unsigned long *bitmap;
102 struct page **pages; /* for pin_user_pages_remote */
103 struct page *fallback_page; /* if pages alloc fails */
104 int capacity; /* length of pages array */
108 struct iommu_group *iommu_group;
109 struct list_head next;
110 bool mdev_group; /* An mdev group */
111 bool pinned_page_dirty_scope;
115 struct list_head list;
121 * Guest RAM pinning working set or DMA target
125 dma_addr_t iova; /* Device address */
126 unsigned long pfn; /* Host pfn */
127 unsigned int ref_count;
130 struct vfio_regions {
131 struct list_head list;
137 #define IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu) \
138 (!list_empty(&iommu->domain_list))
140 #define DIRTY_BITMAP_BYTES(n) (ALIGN(n, BITS_PER_TYPE(u64)) / BITS_PER_BYTE)
143 * Input argument of number of bits to bitmap_set() is unsigned integer, which
144 * further casts to signed integer for unaligned multi-bit operation,
146 * Then maximum bitmap size supported is 2^31 bits divided by 2^3 bits/byte,
147 * that is 2^28 (256 MB) which maps to 2^31 * 2^12 = 2^43 (8TB) on 4K page
150 #define DIRTY_BITMAP_PAGES_MAX ((u64)INT_MAX)
151 #define DIRTY_BITMAP_SIZE_MAX DIRTY_BITMAP_BYTES(DIRTY_BITMAP_PAGES_MAX)
153 static int put_pfn(unsigned long pfn, int prot);
155 static struct vfio_group *vfio_iommu_find_iommu_group(struct vfio_iommu *iommu,
156 struct iommu_group *iommu_group);
158 static void update_pinned_page_dirty_scope(struct vfio_iommu *iommu);
160 * This code handles mapping and unmapping of user data buffers
161 * into DMA'ble space using the IOMMU
164 static struct vfio_dma *vfio_find_dma(struct vfio_iommu *iommu,
165 dma_addr_t start, size_t size)
167 struct rb_node *node = iommu->dma_list.rb_node;
170 struct vfio_dma *dma = rb_entry(node, struct vfio_dma, node);
172 if (start + size <= dma->iova)
173 node = node->rb_left;
174 else if (start >= dma->iova + dma->size)
175 node = node->rb_right;
183 static void vfio_link_dma(struct vfio_iommu *iommu, struct vfio_dma *new)
185 struct rb_node **link = &iommu->dma_list.rb_node, *parent = NULL;
186 struct vfio_dma *dma;
190 dma = rb_entry(parent, struct vfio_dma, node);
192 if (new->iova + new->size <= dma->iova)
193 link = &(*link)->rb_left;
195 link = &(*link)->rb_right;
198 rb_link_node(&new->node, parent, link);
199 rb_insert_color(&new->node, &iommu->dma_list);
202 static void vfio_unlink_dma(struct vfio_iommu *iommu, struct vfio_dma *old)
204 rb_erase(&old->node, &iommu->dma_list);
208 static int vfio_dma_bitmap_alloc(struct vfio_dma *dma, size_t pgsize)
210 uint64_t npages = dma->size / pgsize;
212 if (npages > DIRTY_BITMAP_PAGES_MAX)
216 * Allocate extra 64 bits that are used to calculate shift required for
217 * bitmap_shift_left() to manipulate and club unaligned number of pages
218 * in adjacent vfio_dma ranges.
220 dma->bitmap = kvzalloc(DIRTY_BITMAP_BYTES(npages) + sizeof(u64),
228 static void vfio_dma_bitmap_free(struct vfio_dma *dma)
234 static void vfio_dma_populate_bitmap(struct vfio_dma *dma, size_t pgsize)
237 unsigned long pgshift = __ffs(pgsize);
239 for (p = rb_first(&dma->pfn_list); p; p = rb_next(p)) {
240 struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn, node);
242 bitmap_set(dma->bitmap, (vpfn->iova - dma->iova) >> pgshift, 1);
246 static void vfio_iommu_populate_bitmap_full(struct vfio_iommu *iommu)
249 unsigned long pgshift = __ffs(iommu->pgsize_bitmap);
251 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
252 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
254 bitmap_set(dma->bitmap, 0, dma->size >> pgshift);
258 static int vfio_dma_bitmap_alloc_all(struct vfio_iommu *iommu, size_t pgsize)
262 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
263 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
266 ret = vfio_dma_bitmap_alloc(dma, pgsize);
270 for (p = rb_prev(n); p; p = rb_prev(p)) {
271 struct vfio_dma *dma = rb_entry(n,
272 struct vfio_dma, node);
274 vfio_dma_bitmap_free(dma);
278 vfio_dma_populate_bitmap(dma, pgsize);
283 static void vfio_dma_bitmap_free_all(struct vfio_iommu *iommu)
287 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
288 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
290 vfio_dma_bitmap_free(dma);
295 * Helper Functions for host iova-pfn list
297 static struct vfio_pfn *vfio_find_vpfn(struct vfio_dma *dma, dma_addr_t iova)
299 struct vfio_pfn *vpfn;
300 struct rb_node *node = dma->pfn_list.rb_node;
303 vpfn = rb_entry(node, struct vfio_pfn, node);
305 if (iova < vpfn->iova)
306 node = node->rb_left;
307 else if (iova > vpfn->iova)
308 node = node->rb_right;
315 static void vfio_link_pfn(struct vfio_dma *dma,
316 struct vfio_pfn *new)
318 struct rb_node **link, *parent = NULL;
319 struct vfio_pfn *vpfn;
321 link = &dma->pfn_list.rb_node;
324 vpfn = rb_entry(parent, struct vfio_pfn, node);
326 if (new->iova < vpfn->iova)
327 link = &(*link)->rb_left;
329 link = &(*link)->rb_right;
332 rb_link_node(&new->node, parent, link);
333 rb_insert_color(&new->node, &dma->pfn_list);
336 static void vfio_unlink_pfn(struct vfio_dma *dma, struct vfio_pfn *old)
338 rb_erase(&old->node, &dma->pfn_list);
341 static int vfio_add_to_pfn_list(struct vfio_dma *dma, dma_addr_t iova,
344 struct vfio_pfn *vpfn;
346 vpfn = kzalloc(sizeof(*vpfn), GFP_KERNEL);
353 vfio_link_pfn(dma, vpfn);
357 static void vfio_remove_from_pfn_list(struct vfio_dma *dma,
358 struct vfio_pfn *vpfn)
360 vfio_unlink_pfn(dma, vpfn);
364 static struct vfio_pfn *vfio_iova_get_vfio_pfn(struct vfio_dma *dma,
367 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
374 static int vfio_iova_put_vfio_pfn(struct vfio_dma *dma, struct vfio_pfn *vpfn)
379 if (!vpfn->ref_count) {
380 ret = put_pfn(vpfn->pfn, dma->prot);
381 vfio_remove_from_pfn_list(dma, vpfn);
386 static int vfio_lock_acct(struct vfio_dma *dma, long npage, bool async)
388 struct mm_struct *mm;
394 mm = async ? get_task_mm(dma->task) : dma->task->mm;
396 return -ESRCH; /* process exited */
398 ret = mmap_write_lock_killable(mm);
400 ret = __account_locked_vm(mm, abs(npage), npage > 0, dma->task,
402 mmap_write_unlock(mm);
412 * Some mappings aren't backed by a struct page, for example an mmap'd
413 * MMIO range for our own or another device. These use a different
414 * pfn conversion and shouldn't be tracked as locked pages.
415 * For compound pages, any driver that sets the reserved bit in head
416 * page needs to set the reserved bit in all subpages to be safe.
418 static bool is_invalid_reserved_pfn(unsigned long pfn)
421 return PageReserved(pfn_to_page(pfn));
426 static int put_pfn(unsigned long pfn, int prot)
428 if (!is_invalid_reserved_pfn(pfn)) {
429 struct page *page = pfn_to_page(pfn);
431 unpin_user_pages_dirty_lock(&page, 1, prot & IOMMU_WRITE);
437 #define VFIO_BATCH_MAX_CAPACITY (PAGE_SIZE / sizeof(struct page *))
439 static void vfio_batch_init(struct vfio_batch *batch)
441 if (unlikely(disable_hugepages))
444 batch->pages = (struct page **) __get_free_page(GFP_KERNEL);
448 batch->capacity = VFIO_BATCH_MAX_CAPACITY;
452 batch->pages = &batch->fallback_page;
456 static void vfio_batch_fini(struct vfio_batch *batch)
458 if (batch->capacity == VFIO_BATCH_MAX_CAPACITY)
459 free_page((unsigned long)batch->pages);
462 static int follow_fault_pfn(struct vm_area_struct *vma, struct mm_struct *mm,
463 unsigned long vaddr, unsigned long *pfn,
470 ret = follow_pte(vma->vm_mm, vaddr, &ptep, &ptl);
472 bool unlocked = false;
474 ret = fixup_user_fault(mm, vaddr,
476 (write_fault ? FAULT_FLAG_WRITE : 0),
484 ret = follow_pte(vma->vm_mm, vaddr, &ptep, &ptl);
489 if (write_fault && !pte_write(*ptep))
492 *pfn = pte_pfn(*ptep);
494 pte_unmap_unlock(ptep, ptl);
499 * Returns the positive number of pfns successfully obtained or a negative
502 static int vaddr_get_pfns(struct mm_struct *mm, unsigned long vaddr,
503 long npages, int prot, unsigned long *pfn,
506 struct vm_area_struct *vma;
507 unsigned int flags = 0;
510 if (prot & IOMMU_WRITE)
514 ret = pin_user_pages_remote(mm, vaddr, npages, flags | FOLL_LONGTERM,
520 * The zero page is always resident, we don't need to pin it
521 * and it falls into our invalid/reserved test so we don't
522 * unpin in put_pfn(). Unpin all zero pages in the batch here.
524 for (i = 0 ; i < ret; i++) {
525 if (unlikely(is_zero_pfn(page_to_pfn(pages[i]))))
526 unpin_user_page(pages[i]);
529 *pfn = page_to_pfn(pages[0]);
533 vaddr = untagged_addr(vaddr);
536 vma = find_vma_intersection(mm, vaddr, vaddr + 1);
538 if (vma && vma->vm_flags & VM_PFNMAP) {
539 ret = follow_fault_pfn(vma, mm, vaddr, pfn, prot & IOMMU_WRITE);
544 if (is_invalid_reserved_pfn(*pfn))
551 mmap_read_unlock(mm);
556 * Attempt to pin pages. We really don't want to track all the pfns and
557 * the iommu can only map chunks of consecutive pfns anyway, so get the
558 * first page and all consecutive pages with the same locking.
560 static long vfio_pin_pages_remote(struct vfio_dma *dma, unsigned long vaddr,
561 long npage, unsigned long *pfn_base,
562 unsigned long limit, struct vfio_batch *batch)
564 unsigned long pfn = 0;
565 long ret, pinned = 0, lock_acct = 0;
567 dma_addr_t iova = vaddr - dma->vaddr + dma->iova;
569 /* This code path is only user initiated */
573 ret = vaddr_get_pfns(current->mm, vaddr, 1, dma->prot, pfn_base,
579 rsvd = is_invalid_reserved_pfn(*pfn_base);
582 * Reserved pages aren't counted against the user, externally pinned
583 * pages are already counted against the user.
585 if (!rsvd && !vfio_find_vpfn(dma, iova)) {
586 if (!dma->lock_cap && current->mm->locked_vm + 1 > limit) {
587 put_pfn(*pfn_base, dma->prot);
588 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n", __func__,
589 limit << PAGE_SHIFT);
595 if (unlikely(disable_hugepages))
598 /* Lock all the consecutive pages from pfn_base */
599 for (vaddr += PAGE_SIZE, iova += PAGE_SIZE; pinned < npage;
600 pinned++, vaddr += PAGE_SIZE, iova += PAGE_SIZE) {
601 ret = vaddr_get_pfns(current->mm, vaddr, 1, dma->prot, &pfn,
606 if (pfn != *pfn_base + pinned ||
607 rsvd != is_invalid_reserved_pfn(pfn)) {
608 put_pfn(pfn, dma->prot);
612 if (!rsvd && !vfio_find_vpfn(dma, iova)) {
613 if (!dma->lock_cap &&
614 current->mm->locked_vm + lock_acct + 1 > limit) {
615 put_pfn(pfn, dma->prot);
616 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n",
617 __func__, limit << PAGE_SHIFT);
626 ret = vfio_lock_acct(dma, lock_acct, false);
631 for (pfn = *pfn_base ; pinned ; pfn++, pinned--)
632 put_pfn(pfn, dma->prot);
641 static long vfio_unpin_pages_remote(struct vfio_dma *dma, dma_addr_t iova,
642 unsigned long pfn, long npage,
645 long unlocked = 0, locked = 0;
648 for (i = 0; i < npage; i++, iova += PAGE_SIZE) {
649 if (put_pfn(pfn++, dma->prot)) {
651 if (vfio_find_vpfn(dma, iova))
657 vfio_lock_acct(dma, locked - unlocked, true);
662 static int vfio_pin_page_external(struct vfio_dma *dma, unsigned long vaddr,
663 unsigned long *pfn_base, bool do_accounting)
665 struct page *pages[1];
666 struct mm_struct *mm;
669 mm = get_task_mm(dma->task);
673 ret = vaddr_get_pfns(mm, vaddr, 1, dma->prot, pfn_base, pages);
679 if (do_accounting && !is_invalid_reserved_pfn(*pfn_base)) {
680 ret = vfio_lock_acct(dma, 1, true);
682 put_pfn(*pfn_base, dma->prot);
684 pr_warn("%s: Task %s (%d) RLIMIT_MEMLOCK "
685 "(%ld) exceeded\n", __func__,
686 dma->task->comm, task_pid_nr(dma->task),
687 task_rlimit(dma->task, RLIMIT_MEMLOCK));
696 static int vfio_unpin_page_external(struct vfio_dma *dma, dma_addr_t iova,
700 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
705 unlocked = vfio_iova_put_vfio_pfn(dma, vpfn);
708 vfio_lock_acct(dma, -unlocked, true);
713 static int vfio_iommu_type1_pin_pages(void *iommu_data,
714 struct iommu_group *iommu_group,
715 unsigned long *user_pfn,
717 unsigned long *phys_pfn)
719 struct vfio_iommu *iommu = iommu_data;
720 struct vfio_group *group;
722 unsigned long remote_vaddr;
723 struct vfio_dma *dma;
726 if (!iommu || !user_pfn || !phys_pfn)
729 /* Supported for v2 version only */
733 mutex_lock(&iommu->lock);
735 /* Fail if notifier list is empty */
736 if (!iommu->notifier.head) {
742 * If iommu capable domain exist in the container then all pages are
743 * already pinned and accounted. Accouting should be done if there is no
744 * iommu capable domain in the container.
746 do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu);
748 for (i = 0; i < npage; i++) {
750 struct vfio_pfn *vpfn;
752 iova = user_pfn[i] << PAGE_SHIFT;
753 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
759 if ((dma->prot & prot) != prot) {
764 vpfn = vfio_iova_get_vfio_pfn(dma, iova);
766 phys_pfn[i] = vpfn->pfn;
770 remote_vaddr = dma->vaddr + (iova - dma->iova);
771 ret = vfio_pin_page_external(dma, remote_vaddr, &phys_pfn[i],
776 ret = vfio_add_to_pfn_list(dma, iova, phys_pfn[i]);
778 if (put_pfn(phys_pfn[i], dma->prot) && do_accounting)
779 vfio_lock_acct(dma, -1, true);
783 if (iommu->dirty_page_tracking) {
784 unsigned long pgshift = __ffs(iommu->pgsize_bitmap);
787 * Bitmap populated with the smallest supported page
790 bitmap_set(dma->bitmap,
791 (iova - dma->iova) >> pgshift, 1);
796 group = vfio_iommu_find_iommu_group(iommu, iommu_group);
797 if (!group->pinned_page_dirty_scope) {
798 group->pinned_page_dirty_scope = true;
799 update_pinned_page_dirty_scope(iommu);
806 for (j = 0; j < i; j++) {
809 iova = user_pfn[j] << PAGE_SHIFT;
810 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
811 vfio_unpin_page_external(dma, iova, do_accounting);
815 mutex_unlock(&iommu->lock);
819 static int vfio_iommu_type1_unpin_pages(void *iommu_data,
820 unsigned long *user_pfn,
823 struct vfio_iommu *iommu = iommu_data;
827 if (!iommu || !user_pfn)
830 /* Supported for v2 version only */
834 mutex_lock(&iommu->lock);
836 do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu);
837 for (i = 0; i < npage; i++) {
838 struct vfio_dma *dma;
841 iova = user_pfn[i] << PAGE_SHIFT;
842 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
845 vfio_unpin_page_external(dma, iova, do_accounting);
849 mutex_unlock(&iommu->lock);
850 return i > npage ? npage : (i > 0 ? i : -EINVAL);
853 static long vfio_sync_unpin(struct vfio_dma *dma, struct vfio_domain *domain,
854 struct list_head *regions,
855 struct iommu_iotlb_gather *iotlb_gather)
858 struct vfio_regions *entry, *next;
860 iommu_iotlb_sync(domain->domain, iotlb_gather);
862 list_for_each_entry_safe(entry, next, regions, list) {
863 unlocked += vfio_unpin_pages_remote(dma,
865 entry->phys >> PAGE_SHIFT,
866 entry->len >> PAGE_SHIFT,
868 list_del(&entry->list);
878 * Generally, VFIO needs to unpin remote pages after each IOTLB flush.
879 * Therefore, when using IOTLB flush sync interface, VFIO need to keep track
880 * of these regions (currently using a list).
882 * This value specifies maximum number of regions for each IOTLB flush sync.
884 #define VFIO_IOMMU_TLB_SYNC_MAX 512
886 static size_t unmap_unpin_fast(struct vfio_domain *domain,
887 struct vfio_dma *dma, dma_addr_t *iova,
888 size_t len, phys_addr_t phys, long *unlocked,
889 struct list_head *unmapped_list,
891 struct iommu_iotlb_gather *iotlb_gather)
894 struct vfio_regions *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
897 unmapped = iommu_unmap_fast(domain->domain, *iova, len,
905 entry->len = unmapped;
906 list_add_tail(&entry->list, unmapped_list);
914 * Sync if the number of fast-unmap regions hits the limit
915 * or in case of errors.
917 if (*unmapped_cnt >= VFIO_IOMMU_TLB_SYNC_MAX || !unmapped) {
918 *unlocked += vfio_sync_unpin(dma, domain, unmapped_list,
926 static size_t unmap_unpin_slow(struct vfio_domain *domain,
927 struct vfio_dma *dma, dma_addr_t *iova,
928 size_t len, phys_addr_t phys,
931 size_t unmapped = iommu_unmap(domain->domain, *iova, len);
934 *unlocked += vfio_unpin_pages_remote(dma, *iova,
936 unmapped >> PAGE_SHIFT,
944 static long vfio_unmap_unpin(struct vfio_iommu *iommu, struct vfio_dma *dma,
947 dma_addr_t iova = dma->iova, end = dma->iova + dma->size;
948 struct vfio_domain *domain, *d;
949 LIST_HEAD(unmapped_region_list);
950 struct iommu_iotlb_gather iotlb_gather;
951 int unmapped_region_cnt = 0;
957 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
961 * We use the IOMMU to track the physical addresses, otherwise we'd
962 * need a much more complicated tracking system. Unfortunately that
963 * means we need to use one of the iommu domains to figure out the
964 * pfns to unpin. The rest need to be unmapped in advance so we have
965 * no iommu translations remaining when the pages are unpinned.
967 domain = d = list_first_entry(&iommu->domain_list,
968 struct vfio_domain, next);
970 list_for_each_entry_continue(d, &iommu->domain_list, next) {
971 iommu_unmap(d->domain, dma->iova, dma->size);
975 iommu_iotlb_gather_init(&iotlb_gather);
977 size_t unmapped, len;
978 phys_addr_t phys, next;
980 phys = iommu_iova_to_phys(domain->domain, iova);
981 if (WARN_ON(!phys)) {
987 * To optimize for fewer iommu_unmap() calls, each of which
988 * may require hardware cache flushing, try to find the
989 * largest contiguous physical memory chunk to unmap.
991 for (len = PAGE_SIZE;
992 !domain->fgsp && iova + len < end; len += PAGE_SIZE) {
993 next = iommu_iova_to_phys(domain->domain, iova + len);
994 if (next != phys + len)
999 * First, try to use fast unmap/unpin. In case of failure,
1000 * switch to slow unmap/unpin path.
1002 unmapped = unmap_unpin_fast(domain, dma, &iova, len, phys,
1003 &unlocked, &unmapped_region_list,
1004 &unmapped_region_cnt,
1007 unmapped = unmap_unpin_slow(domain, dma, &iova, len,
1009 if (WARN_ON(!unmapped))
1014 dma->iommu_mapped = false;
1016 if (unmapped_region_cnt) {
1017 unlocked += vfio_sync_unpin(dma, domain, &unmapped_region_list,
1021 if (do_accounting) {
1022 vfio_lock_acct(dma, -unlocked, true);
1028 static void vfio_remove_dma(struct vfio_iommu *iommu, struct vfio_dma *dma)
1030 WARN_ON(!RB_EMPTY_ROOT(&dma->pfn_list));
1031 vfio_unmap_unpin(iommu, dma, true);
1032 vfio_unlink_dma(iommu, dma);
1033 put_task_struct(dma->task);
1034 vfio_dma_bitmap_free(dma);
1039 static void vfio_update_pgsize_bitmap(struct vfio_iommu *iommu)
1041 struct vfio_domain *domain;
1043 iommu->pgsize_bitmap = ULONG_MAX;
1045 list_for_each_entry(domain, &iommu->domain_list, next)
1046 iommu->pgsize_bitmap &= domain->domain->pgsize_bitmap;
1049 * In case the IOMMU supports page sizes smaller than PAGE_SIZE
1050 * we pretend PAGE_SIZE is supported and hide sub-PAGE_SIZE sizes.
1051 * That way the user will be able to map/unmap buffers whose size/
1052 * start address is aligned with PAGE_SIZE. Pinning code uses that
1053 * granularity while iommu driver can use the sub-PAGE_SIZE size
1054 * to map the buffer.
1056 if (iommu->pgsize_bitmap & ~PAGE_MASK) {
1057 iommu->pgsize_bitmap &= PAGE_MASK;
1058 iommu->pgsize_bitmap |= PAGE_SIZE;
1062 static int update_user_bitmap(u64 __user *bitmap, struct vfio_iommu *iommu,
1063 struct vfio_dma *dma, dma_addr_t base_iova,
1066 unsigned long pgshift = __ffs(pgsize);
1067 unsigned long nbits = dma->size >> pgshift;
1068 unsigned long bit_offset = (dma->iova - base_iova) >> pgshift;
1069 unsigned long copy_offset = bit_offset / BITS_PER_LONG;
1070 unsigned long shift = bit_offset % BITS_PER_LONG;
1071 unsigned long leftover;
1074 * mark all pages dirty if any IOMMU capable device is not able
1075 * to report dirty pages and all pages are pinned and mapped.
1077 if (!iommu->pinned_page_dirty_scope && dma->iommu_mapped)
1078 bitmap_set(dma->bitmap, 0, nbits);
1081 bitmap_shift_left(dma->bitmap, dma->bitmap, shift,
1084 if (copy_from_user(&leftover,
1085 (void __user *)(bitmap + copy_offset),
1089 bitmap_or(dma->bitmap, dma->bitmap, &leftover, shift);
1092 if (copy_to_user((void __user *)(bitmap + copy_offset), dma->bitmap,
1093 DIRTY_BITMAP_BYTES(nbits + shift)))
1099 static int vfio_iova_dirty_bitmap(u64 __user *bitmap, struct vfio_iommu *iommu,
1100 dma_addr_t iova, size_t size, size_t pgsize)
1102 struct vfio_dma *dma;
1104 unsigned long pgshift = __ffs(pgsize);
1108 * GET_BITMAP request must fully cover vfio_dma mappings. Multiple
1109 * vfio_dma mappings may be clubbed by specifying large ranges, but
1110 * there must not be any previous mappings bisected by the range.
1111 * An error will be returned if these conditions are not met.
1113 dma = vfio_find_dma(iommu, iova, 1);
1114 if (dma && dma->iova != iova)
1117 dma = vfio_find_dma(iommu, iova + size - 1, 0);
1118 if (dma && dma->iova + dma->size != iova + size)
1121 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
1122 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1124 if (dma->iova < iova)
1127 if (dma->iova > iova + size - 1)
1130 ret = update_user_bitmap(bitmap, iommu, dma, iova, pgsize);
1135 * Re-populate bitmap to include all pinned pages which are
1136 * considered as dirty but exclude pages which are unpinned and
1137 * pages which are marked dirty by vfio_dma_rw()
1139 bitmap_clear(dma->bitmap, 0, dma->size >> pgshift);
1140 vfio_dma_populate_bitmap(dma, pgsize);
1145 static int verify_bitmap_size(uint64_t npages, uint64_t bitmap_size)
1147 if (!npages || !bitmap_size || (bitmap_size > DIRTY_BITMAP_SIZE_MAX) ||
1148 (bitmap_size < DIRTY_BITMAP_BYTES(npages)))
1154 static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
1155 struct vfio_iommu_type1_dma_unmap *unmap,
1156 struct vfio_bitmap *bitmap)
1158 struct vfio_dma *dma, *dma_last = NULL;
1159 size_t unmapped = 0, pgsize;
1160 int ret = 0, retries = 0;
1161 unsigned long pgshift;
1163 mutex_lock(&iommu->lock);
1165 pgshift = __ffs(iommu->pgsize_bitmap);
1166 pgsize = (size_t)1 << pgshift;
1168 if (unmap->iova & (pgsize - 1)) {
1173 if (!unmap->size || unmap->size & (pgsize - 1)) {
1178 if (unmap->iova + unmap->size - 1 < unmap->iova ||
1179 unmap->size > SIZE_MAX) {
1184 /* When dirty tracking is enabled, allow only min supported pgsize */
1185 if ((unmap->flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) &&
1186 (!iommu->dirty_page_tracking || (bitmap->pgsize != pgsize))) {
1191 WARN_ON((pgsize - 1) & PAGE_MASK);
1194 * vfio-iommu-type1 (v1) - User mappings were coalesced together to
1195 * avoid tracking individual mappings. This means that the granularity
1196 * of the original mapping was lost and the user was allowed to attempt
1197 * to unmap any range. Depending on the contiguousness of physical
1198 * memory and page sizes supported by the IOMMU, arbitrary unmaps may
1199 * or may not have worked. We only guaranteed unmap granularity
1200 * matching the original mapping; even though it was untracked here,
1201 * the original mappings are reflected in IOMMU mappings. This
1202 * resulted in a couple unusual behaviors. First, if a range is not
1203 * able to be unmapped, ex. a set of 4k pages that was mapped as a
1204 * 2M hugepage into the IOMMU, the unmap ioctl returns success but with
1205 * a zero sized unmap. Also, if an unmap request overlaps the first
1206 * address of a hugepage, the IOMMU will unmap the entire hugepage.
1207 * This also returns success and the returned unmap size reflects the
1208 * actual size unmapped.
1210 * We attempt to maintain compatibility with this "v1" interface, but
1211 * we take control out of the hands of the IOMMU. Therefore, an unmap
1212 * request offset from the beginning of the original mapping will
1213 * return success with zero sized unmap. And an unmap request covering
1214 * the first iova of mapping will unmap the entire range.
1216 * The v2 version of this interface intends to be more deterministic.
1217 * Unmap requests must fully cover previous mappings. Multiple
1218 * mappings may still be unmaped by specifying large ranges, but there
1219 * must not be any previous mappings bisected by the range. An error
1220 * will be returned if these conditions are not met. The v2 interface
1221 * will only return success and a size of zero if there were no
1222 * mappings within the range.
1225 dma = vfio_find_dma(iommu, unmap->iova, 1);
1226 if (dma && dma->iova != unmap->iova) {
1230 dma = vfio_find_dma(iommu, unmap->iova + unmap->size - 1, 0);
1231 if (dma && dma->iova + dma->size != unmap->iova + unmap->size) {
1237 while ((dma = vfio_find_dma(iommu, unmap->iova, unmap->size))) {
1238 if (!iommu->v2 && unmap->iova > dma->iova)
1241 * Task with same address space who mapped this iova range is
1242 * allowed to unmap the iova range.
1244 if (dma->task->mm != current->mm)
1247 if (!RB_EMPTY_ROOT(&dma->pfn_list)) {
1248 struct vfio_iommu_type1_dma_unmap nb_unmap;
1250 if (dma_last == dma) {
1251 BUG_ON(++retries > 10);
1257 nb_unmap.iova = dma->iova;
1258 nb_unmap.size = dma->size;
1261 * Notify anyone (mdev vendor drivers) to invalidate and
1262 * unmap iovas within the range we're about to unmap.
1263 * Vendor drivers MUST unpin pages in response to an
1266 mutex_unlock(&iommu->lock);
1267 blocking_notifier_call_chain(&iommu->notifier,
1268 VFIO_IOMMU_NOTIFY_DMA_UNMAP,
1270 mutex_lock(&iommu->lock);
1274 if (unmap->flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) {
1275 ret = update_user_bitmap(bitmap->data, iommu, dma,
1276 unmap->iova, pgsize);
1281 unmapped += dma->size;
1282 vfio_remove_dma(iommu, dma);
1286 mutex_unlock(&iommu->lock);
1288 /* Report how much was unmapped */
1289 unmap->size = unmapped;
1294 static int vfio_iommu_map(struct vfio_iommu *iommu, dma_addr_t iova,
1295 unsigned long pfn, long npage, int prot)
1297 struct vfio_domain *d;
1300 list_for_each_entry(d, &iommu->domain_list, next) {
1301 ret = iommu_map(d->domain, iova, (phys_addr_t)pfn << PAGE_SHIFT,
1302 npage << PAGE_SHIFT, prot | d->prot);
1312 list_for_each_entry_continue_reverse(d, &iommu->domain_list, next) {
1313 iommu_unmap(d->domain, iova, npage << PAGE_SHIFT);
1320 static int vfio_pin_map_dma(struct vfio_iommu *iommu, struct vfio_dma *dma,
1323 dma_addr_t iova = dma->iova;
1324 unsigned long vaddr = dma->vaddr;
1325 struct vfio_batch batch;
1326 size_t size = map_size;
1328 unsigned long pfn, limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1331 vfio_batch_init(&batch);
1334 /* Pin a contiguous chunk of memory */
1335 npage = vfio_pin_pages_remote(dma, vaddr + dma->size,
1336 size >> PAGE_SHIFT, &pfn, limit,
1345 ret = vfio_iommu_map(iommu, iova + dma->size, pfn, npage,
1348 vfio_unpin_pages_remote(dma, iova + dma->size, pfn,
1353 size -= npage << PAGE_SHIFT;
1354 dma->size += npage << PAGE_SHIFT;
1357 vfio_batch_fini(&batch);
1358 dma->iommu_mapped = true;
1361 vfio_remove_dma(iommu, dma);
1367 * Check dma map request is within a valid iova range
1369 static bool vfio_iommu_iova_dma_valid(struct vfio_iommu *iommu,
1370 dma_addr_t start, dma_addr_t end)
1372 struct list_head *iova = &iommu->iova_list;
1373 struct vfio_iova *node;
1375 list_for_each_entry(node, iova, list) {
1376 if (start >= node->start && end <= node->end)
1381 * Check for list_empty() as well since a container with
1382 * a single mdev device will have an empty list.
1384 return list_empty(iova);
1387 static int vfio_dma_do_map(struct vfio_iommu *iommu,
1388 struct vfio_iommu_type1_dma_map *map)
1390 dma_addr_t iova = map->iova;
1391 unsigned long vaddr = map->vaddr;
1392 size_t size = map->size;
1393 int ret = 0, prot = 0;
1395 struct vfio_dma *dma;
1397 /* Verify that none of our __u64 fields overflow */
1398 if (map->size != size || map->vaddr != vaddr || map->iova != iova)
1401 /* READ/WRITE from device perspective */
1402 if (map->flags & VFIO_DMA_MAP_FLAG_WRITE)
1403 prot |= IOMMU_WRITE;
1404 if (map->flags & VFIO_DMA_MAP_FLAG_READ)
1407 mutex_lock(&iommu->lock);
1409 pgsize = (size_t)1 << __ffs(iommu->pgsize_bitmap);
1411 WARN_ON((pgsize - 1) & PAGE_MASK);
1413 if (!prot || !size || (size | iova | vaddr) & (pgsize - 1)) {
1418 /* Don't allow IOVA or virtual address wrap */
1419 if (iova + size - 1 < iova || vaddr + size - 1 < vaddr) {
1424 if (vfio_find_dma(iommu, iova, size)) {
1429 if (!iommu->dma_avail) {
1434 if (!vfio_iommu_iova_dma_valid(iommu, iova, iova + size - 1)) {
1439 dma = kzalloc(sizeof(*dma), GFP_KERNEL);
1451 * We need to be able to both add to a task's locked memory and test
1452 * against the locked memory limit and we need to be able to do both
1453 * outside of this call path as pinning can be asynchronous via the
1454 * external interfaces for mdev devices. RLIMIT_MEMLOCK requires a
1455 * task_struct and VM locked pages requires an mm_struct, however
1456 * holding an indefinite mm reference is not recommended, therefore we
1457 * only hold a reference to a task. We could hold a reference to
1458 * current, however QEMU uses this call path through vCPU threads,
1459 * which can be killed resulting in a NULL mm and failure in the unmap
1460 * path when called via a different thread. Avoid this problem by
1461 * using the group_leader as threads within the same group require
1462 * both CLONE_THREAD and CLONE_VM and will therefore use the same
1465 * Previously we also used the task for testing CAP_IPC_LOCK at the
1466 * time of pinning and accounting, however has_capability() makes use
1467 * of real_cred, a copy-on-write field, so we can't guarantee that it
1468 * matches group_leader, or in fact that it might not change by the
1469 * time it's evaluated. If a process were to call MAP_DMA with
1470 * CAP_IPC_LOCK but later drop it, it doesn't make sense that they
1471 * possibly see different results for an iommu_mapped vfio_dma vs
1472 * externally mapped. Therefore track CAP_IPC_LOCK in vfio_dma at the
1473 * time of calling MAP_DMA.
1475 get_task_struct(current->group_leader);
1476 dma->task = current->group_leader;
1477 dma->lock_cap = capable(CAP_IPC_LOCK);
1479 dma->pfn_list = RB_ROOT;
1481 /* Insert zero-sized and grow as we map chunks of it */
1482 vfio_link_dma(iommu, dma);
1484 /* Don't pin and map if container doesn't contain IOMMU capable domain*/
1485 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
1488 ret = vfio_pin_map_dma(iommu, dma, size);
1490 if (!ret && iommu->dirty_page_tracking) {
1491 ret = vfio_dma_bitmap_alloc(dma, pgsize);
1493 vfio_remove_dma(iommu, dma);
1497 mutex_unlock(&iommu->lock);
1501 static int vfio_bus_type(struct device *dev, void *data)
1503 struct bus_type **bus = data;
1505 if (*bus && *bus != dev->bus)
1513 static int vfio_iommu_replay(struct vfio_iommu *iommu,
1514 struct vfio_domain *domain)
1516 struct vfio_batch batch;
1517 struct vfio_domain *d = NULL;
1519 unsigned long limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1522 /* Arbitrarily pick the first domain in the list for lookups */
1523 if (!list_empty(&iommu->domain_list))
1524 d = list_first_entry(&iommu->domain_list,
1525 struct vfio_domain, next);
1527 vfio_batch_init(&batch);
1529 n = rb_first(&iommu->dma_list);
1531 for (; n; n = rb_next(n)) {
1532 struct vfio_dma *dma;
1535 dma = rb_entry(n, struct vfio_dma, node);
1538 while (iova < dma->iova + dma->size) {
1542 if (dma->iommu_mapped) {
1546 if (WARN_ON(!d)) { /* mapped w/o a domain?! */
1551 phys = iommu_iova_to_phys(d->domain, iova);
1553 if (WARN_ON(!phys)) {
1561 while (i < dma->iova + dma->size &&
1562 p == iommu_iova_to_phys(d->domain, i)) {
1569 unsigned long vaddr = dma->vaddr +
1571 size_t n = dma->iova + dma->size - iova;
1574 npage = vfio_pin_pages_remote(dma, vaddr,
1584 phys = pfn << PAGE_SHIFT;
1585 size = npage << PAGE_SHIFT;
1588 ret = iommu_map(domain->domain, iova, phys,
1589 size, dma->prot | domain->prot);
1591 if (!dma->iommu_mapped)
1592 vfio_unpin_pages_remote(dma, iova,
1603 /* All dmas are now mapped, defer to second tree walk for unwind */
1604 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
1605 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1607 dma->iommu_mapped = true;
1610 vfio_batch_fini(&batch);
1614 for (; n; n = rb_prev(n)) {
1615 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1618 if (dma->iommu_mapped) {
1619 iommu_unmap(domain->domain, dma->iova, dma->size);
1624 while (iova < dma->iova + dma->size) {
1625 phys_addr_t phys, p;
1629 phys = iommu_iova_to_phys(domain->domain, iova);
1638 while (i < dma->iova + dma->size &&
1639 p == iommu_iova_to_phys(domain->domain, i)) {
1645 iommu_unmap(domain->domain, iova, size);
1646 vfio_unpin_pages_remote(dma, iova, phys >> PAGE_SHIFT,
1647 size >> PAGE_SHIFT, true);
1651 vfio_batch_fini(&batch);
1656 * We change our unmap behavior slightly depending on whether the IOMMU
1657 * supports fine-grained superpages. IOMMUs like AMD-Vi will use a superpage
1658 * for practically any contiguous power-of-two mapping we give it. This means
1659 * we don't need to look for contiguous chunks ourselves to make unmapping
1660 * more efficient. On IOMMUs with coarse-grained super pages, like Intel VT-d
1661 * with discrete 2M/1G/512G/1T superpages, identifying contiguous chunks
1662 * significantly boosts non-hugetlbfs mappings and doesn't seem to hurt when
1663 * hugetlbfs is in use.
1665 static void vfio_test_domain_fgsp(struct vfio_domain *domain)
1668 int ret, order = get_order(PAGE_SIZE * 2);
1670 pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order);
1674 ret = iommu_map(domain->domain, 0, page_to_phys(pages), PAGE_SIZE * 2,
1675 IOMMU_READ | IOMMU_WRITE | domain->prot);
1677 size_t unmapped = iommu_unmap(domain->domain, 0, PAGE_SIZE);
1679 if (unmapped == PAGE_SIZE)
1680 iommu_unmap(domain->domain, PAGE_SIZE, PAGE_SIZE);
1682 domain->fgsp = true;
1685 __free_pages(pages, order);
1688 static struct vfio_group *find_iommu_group(struct vfio_domain *domain,
1689 struct iommu_group *iommu_group)
1691 struct vfio_group *g;
1693 list_for_each_entry(g, &domain->group_list, next) {
1694 if (g->iommu_group == iommu_group)
1701 static struct vfio_group *vfio_iommu_find_iommu_group(struct vfio_iommu *iommu,
1702 struct iommu_group *iommu_group)
1704 struct vfio_domain *domain;
1705 struct vfio_group *group = NULL;
1707 list_for_each_entry(domain, &iommu->domain_list, next) {
1708 group = find_iommu_group(domain, iommu_group);
1713 if (iommu->external_domain)
1714 group = find_iommu_group(iommu->external_domain, iommu_group);
1719 static void update_pinned_page_dirty_scope(struct vfio_iommu *iommu)
1721 struct vfio_domain *domain;
1722 struct vfio_group *group;
1724 list_for_each_entry(domain, &iommu->domain_list, next) {
1725 list_for_each_entry(group, &domain->group_list, next) {
1726 if (!group->pinned_page_dirty_scope) {
1727 iommu->pinned_page_dirty_scope = false;
1733 if (iommu->external_domain) {
1734 domain = iommu->external_domain;
1735 list_for_each_entry(group, &domain->group_list, next) {
1736 if (!group->pinned_page_dirty_scope) {
1737 iommu->pinned_page_dirty_scope = false;
1743 iommu->pinned_page_dirty_scope = true;
1746 static bool vfio_iommu_has_sw_msi(struct list_head *group_resv_regions,
1749 struct iommu_resv_region *region;
1752 list_for_each_entry(region, group_resv_regions, list) {
1754 * The presence of any 'real' MSI regions should take
1755 * precedence over the software-managed one if the
1756 * IOMMU driver happens to advertise both types.
1758 if (region->type == IOMMU_RESV_MSI) {
1763 if (region->type == IOMMU_RESV_SW_MSI) {
1764 *base = region->start;
1772 static struct device *vfio_mdev_get_iommu_device(struct device *dev)
1774 struct device *(*fn)(struct device *dev);
1775 struct device *iommu_device;
1777 fn = symbol_get(mdev_get_iommu_device);
1779 iommu_device = fn(dev);
1780 symbol_put(mdev_get_iommu_device);
1782 return iommu_device;
1788 static int vfio_mdev_attach_domain(struct device *dev, void *data)
1790 struct iommu_domain *domain = data;
1791 struct device *iommu_device;
1793 iommu_device = vfio_mdev_get_iommu_device(dev);
1795 if (iommu_dev_feature_enabled(iommu_device, IOMMU_DEV_FEAT_AUX))
1796 return iommu_aux_attach_device(domain, iommu_device);
1798 return iommu_attach_device(domain, iommu_device);
1804 static int vfio_mdev_detach_domain(struct device *dev, void *data)
1806 struct iommu_domain *domain = data;
1807 struct device *iommu_device;
1809 iommu_device = vfio_mdev_get_iommu_device(dev);
1811 if (iommu_dev_feature_enabled(iommu_device, IOMMU_DEV_FEAT_AUX))
1812 iommu_aux_detach_device(domain, iommu_device);
1814 iommu_detach_device(domain, iommu_device);
1820 static int vfio_iommu_attach_group(struct vfio_domain *domain,
1821 struct vfio_group *group)
1823 if (group->mdev_group)
1824 return iommu_group_for_each_dev(group->iommu_group,
1826 vfio_mdev_attach_domain);
1828 return iommu_attach_group(domain->domain, group->iommu_group);
1831 static void vfio_iommu_detach_group(struct vfio_domain *domain,
1832 struct vfio_group *group)
1834 if (group->mdev_group)
1835 iommu_group_for_each_dev(group->iommu_group, domain->domain,
1836 vfio_mdev_detach_domain);
1838 iommu_detach_group(domain->domain, group->iommu_group);
1841 static bool vfio_bus_is_mdev(struct bus_type *bus)
1843 struct bus_type *mdev_bus;
1846 mdev_bus = symbol_get(mdev_bus_type);
1848 ret = (bus == mdev_bus);
1849 symbol_put(mdev_bus_type);
1855 static int vfio_mdev_iommu_device(struct device *dev, void *data)
1857 struct device **old = data, *new;
1859 new = vfio_mdev_get_iommu_device(dev);
1860 if (!new || (*old && *old != new))
1869 * This is a helper function to insert an address range to iova list.
1870 * The list is initially created with a single entry corresponding to
1871 * the IOMMU domain geometry to which the device group is attached.
1872 * The list aperture gets modified when a new domain is added to the
1873 * container if the new aperture doesn't conflict with the current one
1874 * or with any existing dma mappings. The list is also modified to
1875 * exclude any reserved regions associated with the device group.
1877 static int vfio_iommu_iova_insert(struct list_head *head,
1878 dma_addr_t start, dma_addr_t end)
1880 struct vfio_iova *region;
1882 region = kmalloc(sizeof(*region), GFP_KERNEL);
1886 INIT_LIST_HEAD(®ion->list);
1887 region->start = start;
1890 list_add_tail(®ion->list, head);
1895 * Check the new iommu aperture conflicts with existing aper or with any
1896 * existing dma mappings.
1898 static bool vfio_iommu_aper_conflict(struct vfio_iommu *iommu,
1899 dma_addr_t start, dma_addr_t end)
1901 struct vfio_iova *first, *last;
1902 struct list_head *iova = &iommu->iova_list;
1904 if (list_empty(iova))
1907 /* Disjoint sets, return conflict */
1908 first = list_first_entry(iova, struct vfio_iova, list);
1909 last = list_last_entry(iova, struct vfio_iova, list);
1910 if (start > last->end || end < first->start)
1913 /* Check for any existing dma mappings below the new start */
1914 if (start > first->start) {
1915 if (vfio_find_dma(iommu, first->start, start - first->start))
1919 /* Check for any existing dma mappings beyond the new end */
1920 if (end < last->end) {
1921 if (vfio_find_dma(iommu, end + 1, last->end - end))
1929 * Resize iommu iova aperture window. This is called only if the new
1930 * aperture has no conflict with existing aperture and dma mappings.
1932 static int vfio_iommu_aper_resize(struct list_head *iova,
1933 dma_addr_t start, dma_addr_t end)
1935 struct vfio_iova *node, *next;
1937 if (list_empty(iova))
1938 return vfio_iommu_iova_insert(iova, start, end);
1940 /* Adjust iova list start */
1941 list_for_each_entry_safe(node, next, iova, list) {
1942 if (start < node->start)
1944 if (start >= node->start && start < node->end) {
1945 node->start = start;
1948 /* Delete nodes before new start */
1949 list_del(&node->list);
1953 /* Adjust iova list end */
1954 list_for_each_entry_safe(node, next, iova, list) {
1955 if (end > node->end)
1957 if (end > node->start && end <= node->end) {
1961 /* Delete nodes after new end */
1962 list_del(&node->list);
1970 * Check reserved region conflicts with existing dma mappings
1972 static bool vfio_iommu_resv_conflict(struct vfio_iommu *iommu,
1973 struct list_head *resv_regions)
1975 struct iommu_resv_region *region;
1977 /* Check for conflict with existing dma mappings */
1978 list_for_each_entry(region, resv_regions, list) {
1979 if (region->type == IOMMU_RESV_DIRECT_RELAXABLE)
1982 if (vfio_find_dma(iommu, region->start, region->length))
1990 * Check iova region overlap with reserved regions and
1991 * exclude them from the iommu iova range
1993 static int vfio_iommu_resv_exclude(struct list_head *iova,
1994 struct list_head *resv_regions)
1996 struct iommu_resv_region *resv;
1997 struct vfio_iova *n, *next;
1999 list_for_each_entry(resv, resv_regions, list) {
2000 phys_addr_t start, end;
2002 if (resv->type == IOMMU_RESV_DIRECT_RELAXABLE)
2005 start = resv->start;
2006 end = resv->start + resv->length - 1;
2008 list_for_each_entry_safe(n, next, iova, list) {
2012 if (start > n->end || end < n->start)
2015 * Insert a new node if current node overlaps with the
2016 * reserve region to exlude that from valid iova range.
2017 * Note that, new node is inserted before the current
2018 * node and finally the current node is deleted keeping
2019 * the list updated and sorted.
2021 if (start > n->start)
2022 ret = vfio_iommu_iova_insert(&n->list, n->start,
2024 if (!ret && end < n->end)
2025 ret = vfio_iommu_iova_insert(&n->list, end + 1,
2035 if (list_empty(iova))
2041 static void vfio_iommu_resv_free(struct list_head *resv_regions)
2043 struct iommu_resv_region *n, *next;
2045 list_for_each_entry_safe(n, next, resv_regions, list) {
2051 static void vfio_iommu_iova_free(struct list_head *iova)
2053 struct vfio_iova *n, *next;
2055 list_for_each_entry_safe(n, next, iova, list) {
2061 static int vfio_iommu_iova_get_copy(struct vfio_iommu *iommu,
2062 struct list_head *iova_copy)
2064 struct list_head *iova = &iommu->iova_list;
2065 struct vfio_iova *n;
2068 list_for_each_entry(n, iova, list) {
2069 ret = vfio_iommu_iova_insert(iova_copy, n->start, n->end);
2077 vfio_iommu_iova_free(iova_copy);
2081 static void vfio_iommu_iova_insert_copy(struct vfio_iommu *iommu,
2082 struct list_head *iova_copy)
2084 struct list_head *iova = &iommu->iova_list;
2086 vfio_iommu_iova_free(iova);
2088 list_splice_tail(iova_copy, iova);
2091 static int vfio_iommu_type1_attach_group(void *iommu_data,
2092 struct iommu_group *iommu_group)
2094 struct vfio_iommu *iommu = iommu_data;
2095 struct vfio_group *group;
2096 struct vfio_domain *domain, *d;
2097 struct bus_type *bus = NULL;
2099 bool resv_msi, msi_remap;
2100 phys_addr_t resv_msi_base = 0;
2101 struct iommu_domain_geometry geo;
2102 LIST_HEAD(iova_copy);
2103 LIST_HEAD(group_resv_regions);
2105 mutex_lock(&iommu->lock);
2107 /* Check for duplicates */
2108 if (vfio_iommu_find_iommu_group(iommu, iommu_group)) {
2109 mutex_unlock(&iommu->lock);
2113 group = kzalloc(sizeof(*group), GFP_KERNEL);
2114 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2115 if (!group || !domain) {
2120 group->iommu_group = iommu_group;
2122 /* Determine bus_type in order to allocate a domain */
2123 ret = iommu_group_for_each_dev(iommu_group, &bus, vfio_bus_type);
2127 if (vfio_bus_is_mdev(bus)) {
2128 struct device *iommu_device = NULL;
2130 group->mdev_group = true;
2132 /* Determine the isolation type */
2133 ret = iommu_group_for_each_dev(iommu_group, &iommu_device,
2134 vfio_mdev_iommu_device);
2135 if (ret || !iommu_device) {
2136 if (!iommu->external_domain) {
2137 INIT_LIST_HEAD(&domain->group_list);
2138 iommu->external_domain = domain;
2139 vfio_update_pgsize_bitmap(iommu);
2144 list_add(&group->next,
2145 &iommu->external_domain->group_list);
2147 * Non-iommu backed group cannot dirty memory directly,
2148 * it can only use interfaces that provide dirty
2150 * The iommu scope can only be promoted with the
2151 * addition of a dirty tracking group.
2153 group->pinned_page_dirty_scope = true;
2154 if (!iommu->pinned_page_dirty_scope)
2155 update_pinned_page_dirty_scope(iommu);
2156 mutex_unlock(&iommu->lock);
2161 bus = iommu_device->bus;
2164 domain->domain = iommu_domain_alloc(bus);
2165 if (!domain->domain) {
2170 if (iommu->nesting) {
2173 ret = iommu_domain_set_attr(domain->domain, DOMAIN_ATTR_NESTING,
2179 ret = vfio_iommu_attach_group(domain, group);
2183 /* Get aperture info */
2184 iommu_domain_get_attr(domain->domain, DOMAIN_ATTR_GEOMETRY, &geo);
2186 if (vfio_iommu_aper_conflict(iommu, geo.aperture_start,
2187 geo.aperture_end)) {
2192 ret = iommu_get_group_resv_regions(iommu_group, &group_resv_regions);
2196 if (vfio_iommu_resv_conflict(iommu, &group_resv_regions)) {
2202 * We don't want to work on the original iova list as the list
2203 * gets modified and in case of failure we have to retain the
2204 * original list. Get a copy here.
2206 ret = vfio_iommu_iova_get_copy(iommu, &iova_copy);
2210 ret = vfio_iommu_aper_resize(&iova_copy, geo.aperture_start,
2215 ret = vfio_iommu_resv_exclude(&iova_copy, &group_resv_regions);
2219 resv_msi = vfio_iommu_has_sw_msi(&group_resv_regions, &resv_msi_base);
2221 INIT_LIST_HEAD(&domain->group_list);
2222 list_add(&group->next, &domain->group_list);
2224 msi_remap = irq_domain_check_msi_remap() ||
2225 iommu_capable(bus, IOMMU_CAP_INTR_REMAP);
2227 if (!allow_unsafe_interrupts && !msi_remap) {
2228 pr_warn("%s: No interrupt remapping support. Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n",
2234 if (iommu_capable(bus, IOMMU_CAP_CACHE_COHERENCY))
2235 domain->prot |= IOMMU_CACHE;
2238 * Try to match an existing compatible domain. We don't want to
2239 * preclude an IOMMU driver supporting multiple bus_types and being
2240 * able to include different bus_types in the same IOMMU domain, so
2241 * we test whether the domains use the same iommu_ops rather than
2242 * testing if they're on the same bus_type.
2244 list_for_each_entry(d, &iommu->domain_list, next) {
2245 if (d->domain->ops == domain->domain->ops &&
2246 d->prot == domain->prot) {
2247 vfio_iommu_detach_group(domain, group);
2248 if (!vfio_iommu_attach_group(d, group)) {
2249 list_add(&group->next, &d->group_list);
2250 iommu_domain_free(domain->domain);
2255 ret = vfio_iommu_attach_group(domain, group);
2261 vfio_test_domain_fgsp(domain);
2263 /* replay mappings on new domains */
2264 ret = vfio_iommu_replay(iommu, domain);
2269 ret = iommu_get_msi_cookie(domain->domain, resv_msi_base);
2270 if (ret && ret != -ENODEV)
2274 list_add(&domain->next, &iommu->domain_list);
2275 vfio_update_pgsize_bitmap(iommu);
2277 /* Delete the old one and insert new iova list */
2278 vfio_iommu_iova_insert_copy(iommu, &iova_copy);
2281 * An iommu backed group can dirty memory directly and therefore
2282 * demotes the iommu scope until it declares itself dirty tracking
2283 * capable via the page pinning interface.
2285 iommu->pinned_page_dirty_scope = false;
2286 mutex_unlock(&iommu->lock);
2287 vfio_iommu_resv_free(&group_resv_regions);
2292 vfio_iommu_detach_group(domain, group);
2294 iommu_domain_free(domain->domain);
2295 vfio_iommu_iova_free(&iova_copy);
2296 vfio_iommu_resv_free(&group_resv_regions);
2300 mutex_unlock(&iommu->lock);
2304 static void vfio_iommu_unmap_unpin_all(struct vfio_iommu *iommu)
2306 struct rb_node *node;
2308 while ((node = rb_first(&iommu->dma_list)))
2309 vfio_remove_dma(iommu, rb_entry(node, struct vfio_dma, node));
2312 static void vfio_iommu_unmap_unpin_reaccount(struct vfio_iommu *iommu)
2314 struct rb_node *n, *p;
2316 n = rb_first(&iommu->dma_list);
2317 for (; n; n = rb_next(n)) {
2318 struct vfio_dma *dma;
2319 long locked = 0, unlocked = 0;
2321 dma = rb_entry(n, struct vfio_dma, node);
2322 unlocked += vfio_unmap_unpin(iommu, dma, false);
2323 p = rb_first(&dma->pfn_list);
2324 for (; p; p = rb_next(p)) {
2325 struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn,
2328 if (!is_invalid_reserved_pfn(vpfn->pfn))
2331 vfio_lock_acct(dma, locked - unlocked, true);
2336 * Called when a domain is removed in detach. It is possible that
2337 * the removed domain decided the iova aperture window. Modify the
2338 * iova aperture with the smallest window among existing domains.
2340 static void vfio_iommu_aper_expand(struct vfio_iommu *iommu,
2341 struct list_head *iova_copy)
2343 struct vfio_domain *domain;
2344 struct iommu_domain_geometry geo;
2345 struct vfio_iova *node;
2346 dma_addr_t start = 0;
2347 dma_addr_t end = (dma_addr_t)~0;
2349 if (list_empty(iova_copy))
2352 list_for_each_entry(domain, &iommu->domain_list, next) {
2353 iommu_domain_get_attr(domain->domain, DOMAIN_ATTR_GEOMETRY,
2355 if (geo.aperture_start > start)
2356 start = geo.aperture_start;
2357 if (geo.aperture_end < end)
2358 end = geo.aperture_end;
2361 /* Modify aperture limits. The new aper is either same or bigger */
2362 node = list_first_entry(iova_copy, struct vfio_iova, list);
2363 node->start = start;
2364 node = list_last_entry(iova_copy, struct vfio_iova, list);
2369 * Called when a group is detached. The reserved regions for that
2370 * group can be part of valid iova now. But since reserved regions
2371 * may be duplicated among groups, populate the iova valid regions
2374 static int vfio_iommu_resv_refresh(struct vfio_iommu *iommu,
2375 struct list_head *iova_copy)
2377 struct vfio_domain *d;
2378 struct vfio_group *g;
2379 struct vfio_iova *node;
2380 dma_addr_t start, end;
2381 LIST_HEAD(resv_regions);
2384 if (list_empty(iova_copy))
2387 list_for_each_entry(d, &iommu->domain_list, next) {
2388 list_for_each_entry(g, &d->group_list, next) {
2389 ret = iommu_get_group_resv_regions(g->iommu_group,
2396 node = list_first_entry(iova_copy, struct vfio_iova, list);
2397 start = node->start;
2398 node = list_last_entry(iova_copy, struct vfio_iova, list);
2401 /* purge the iova list and create new one */
2402 vfio_iommu_iova_free(iova_copy);
2404 ret = vfio_iommu_aper_resize(iova_copy, start, end);
2408 /* Exclude current reserved regions from iova ranges */
2409 ret = vfio_iommu_resv_exclude(iova_copy, &resv_regions);
2411 vfio_iommu_resv_free(&resv_regions);
2415 static void vfio_iommu_type1_detach_group(void *iommu_data,
2416 struct iommu_group *iommu_group)
2418 struct vfio_iommu *iommu = iommu_data;
2419 struct vfio_domain *domain;
2420 struct vfio_group *group;
2421 bool update_dirty_scope = false;
2422 LIST_HEAD(iova_copy);
2424 mutex_lock(&iommu->lock);
2426 if (iommu->external_domain) {
2427 group = find_iommu_group(iommu->external_domain, iommu_group);
2429 update_dirty_scope = !group->pinned_page_dirty_scope;
2430 list_del(&group->next);
2433 if (list_empty(&iommu->external_domain->group_list)) {
2434 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu)) {
2435 WARN_ON(iommu->notifier.head);
2436 vfio_iommu_unmap_unpin_all(iommu);
2439 kfree(iommu->external_domain);
2440 iommu->external_domain = NULL;
2442 goto detach_group_done;
2447 * Get a copy of iova list. This will be used to update
2448 * and to replace the current one later. Please note that
2449 * we will leave the original list as it is if update fails.
2451 vfio_iommu_iova_get_copy(iommu, &iova_copy);
2453 list_for_each_entry(domain, &iommu->domain_list, next) {
2454 group = find_iommu_group(domain, iommu_group);
2458 vfio_iommu_detach_group(domain, group);
2459 update_dirty_scope = !group->pinned_page_dirty_scope;
2460 list_del(&group->next);
2463 * Group ownership provides privilege, if the group list is
2464 * empty, the domain goes away. If it's the last domain with
2465 * iommu and external domain doesn't exist, then all the
2466 * mappings go away too. If it's the last domain with iommu and
2467 * external domain exist, update accounting
2469 if (list_empty(&domain->group_list)) {
2470 if (list_is_singular(&iommu->domain_list)) {
2471 if (!iommu->external_domain) {
2472 WARN_ON(iommu->notifier.head);
2473 vfio_iommu_unmap_unpin_all(iommu);
2475 vfio_iommu_unmap_unpin_reaccount(iommu);
2478 iommu_domain_free(domain->domain);
2479 list_del(&domain->next);
2481 vfio_iommu_aper_expand(iommu, &iova_copy);
2482 vfio_update_pgsize_bitmap(iommu);
2487 if (!vfio_iommu_resv_refresh(iommu, &iova_copy))
2488 vfio_iommu_iova_insert_copy(iommu, &iova_copy);
2490 vfio_iommu_iova_free(&iova_copy);
2494 * Removal of a group without dirty tracking may allow the iommu scope
2497 if (update_dirty_scope) {
2498 update_pinned_page_dirty_scope(iommu);
2499 if (iommu->dirty_page_tracking)
2500 vfio_iommu_populate_bitmap_full(iommu);
2502 mutex_unlock(&iommu->lock);
2505 static void *vfio_iommu_type1_open(unsigned long arg)
2507 struct vfio_iommu *iommu;
2509 iommu = kzalloc(sizeof(*iommu), GFP_KERNEL);
2511 return ERR_PTR(-ENOMEM);
2514 case VFIO_TYPE1_IOMMU:
2516 case VFIO_TYPE1_NESTING_IOMMU:
2517 iommu->nesting = true;
2519 case VFIO_TYPE1v2_IOMMU:
2524 return ERR_PTR(-EINVAL);
2527 INIT_LIST_HEAD(&iommu->domain_list);
2528 INIT_LIST_HEAD(&iommu->iova_list);
2529 iommu->dma_list = RB_ROOT;
2530 iommu->dma_avail = dma_entry_limit;
2531 mutex_init(&iommu->lock);
2532 BLOCKING_INIT_NOTIFIER_HEAD(&iommu->notifier);
2537 static void vfio_release_domain(struct vfio_domain *domain, bool external)
2539 struct vfio_group *group, *group_tmp;
2541 list_for_each_entry_safe(group, group_tmp,
2542 &domain->group_list, next) {
2544 vfio_iommu_detach_group(domain, group);
2545 list_del(&group->next);
2550 iommu_domain_free(domain->domain);
2553 static void vfio_iommu_type1_release(void *iommu_data)
2555 struct vfio_iommu *iommu = iommu_data;
2556 struct vfio_domain *domain, *domain_tmp;
2558 if (iommu->external_domain) {
2559 vfio_release_domain(iommu->external_domain, true);
2560 kfree(iommu->external_domain);
2563 vfio_iommu_unmap_unpin_all(iommu);
2565 list_for_each_entry_safe(domain, domain_tmp,
2566 &iommu->domain_list, next) {
2567 vfio_release_domain(domain, false);
2568 list_del(&domain->next);
2572 vfio_iommu_iova_free(&iommu->iova_list);
2577 static int vfio_domains_have_iommu_cache(struct vfio_iommu *iommu)
2579 struct vfio_domain *domain;
2582 mutex_lock(&iommu->lock);
2583 list_for_each_entry(domain, &iommu->domain_list, next) {
2584 if (!(domain->prot & IOMMU_CACHE)) {
2589 mutex_unlock(&iommu->lock);
2594 static int vfio_iommu_type1_check_extension(struct vfio_iommu *iommu,
2598 case VFIO_TYPE1_IOMMU:
2599 case VFIO_TYPE1v2_IOMMU:
2600 case VFIO_TYPE1_NESTING_IOMMU:
2602 case VFIO_DMA_CC_IOMMU:
2605 return vfio_domains_have_iommu_cache(iommu);
2611 static int vfio_iommu_iova_add_cap(struct vfio_info_cap *caps,
2612 struct vfio_iommu_type1_info_cap_iova_range *cap_iovas,
2615 struct vfio_info_cap_header *header;
2616 struct vfio_iommu_type1_info_cap_iova_range *iova_cap;
2618 header = vfio_info_cap_add(caps, size,
2619 VFIO_IOMMU_TYPE1_INFO_CAP_IOVA_RANGE, 1);
2621 return PTR_ERR(header);
2623 iova_cap = container_of(header,
2624 struct vfio_iommu_type1_info_cap_iova_range,
2626 iova_cap->nr_iovas = cap_iovas->nr_iovas;
2627 memcpy(iova_cap->iova_ranges, cap_iovas->iova_ranges,
2628 cap_iovas->nr_iovas * sizeof(*cap_iovas->iova_ranges));
2632 static int vfio_iommu_iova_build_caps(struct vfio_iommu *iommu,
2633 struct vfio_info_cap *caps)
2635 struct vfio_iommu_type1_info_cap_iova_range *cap_iovas;
2636 struct vfio_iova *iova;
2638 int iovas = 0, i = 0, ret;
2640 list_for_each_entry(iova, &iommu->iova_list, list)
2645 * Return 0 as a container with a single mdev device
2646 * will have an empty list
2651 size = sizeof(*cap_iovas) + (iovas * sizeof(*cap_iovas->iova_ranges));
2653 cap_iovas = kzalloc(size, GFP_KERNEL);
2657 cap_iovas->nr_iovas = iovas;
2659 list_for_each_entry(iova, &iommu->iova_list, list) {
2660 cap_iovas->iova_ranges[i].start = iova->start;
2661 cap_iovas->iova_ranges[i].end = iova->end;
2665 ret = vfio_iommu_iova_add_cap(caps, cap_iovas, size);
2671 static int vfio_iommu_migration_build_caps(struct vfio_iommu *iommu,
2672 struct vfio_info_cap *caps)
2674 struct vfio_iommu_type1_info_cap_migration cap_mig;
2676 cap_mig.header.id = VFIO_IOMMU_TYPE1_INFO_CAP_MIGRATION;
2677 cap_mig.header.version = 1;
2680 /* support minimum pgsize */
2681 cap_mig.pgsize_bitmap = (size_t)1 << __ffs(iommu->pgsize_bitmap);
2682 cap_mig.max_dirty_bitmap_size = DIRTY_BITMAP_SIZE_MAX;
2684 return vfio_info_add_capability(caps, &cap_mig.header, sizeof(cap_mig));
2687 static int vfio_iommu_dma_avail_build_caps(struct vfio_iommu *iommu,
2688 struct vfio_info_cap *caps)
2690 struct vfio_iommu_type1_info_dma_avail cap_dma_avail;
2692 cap_dma_avail.header.id = VFIO_IOMMU_TYPE1_INFO_DMA_AVAIL;
2693 cap_dma_avail.header.version = 1;
2695 cap_dma_avail.avail = iommu->dma_avail;
2697 return vfio_info_add_capability(caps, &cap_dma_avail.header,
2698 sizeof(cap_dma_avail));
2701 static int vfio_iommu_type1_get_info(struct vfio_iommu *iommu,
2704 struct vfio_iommu_type1_info info;
2705 unsigned long minsz;
2706 struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
2707 unsigned long capsz;
2710 minsz = offsetofend(struct vfio_iommu_type1_info, iova_pgsizes);
2712 /* For backward compatibility, cannot require this */
2713 capsz = offsetofend(struct vfio_iommu_type1_info, cap_offset);
2715 if (copy_from_user(&info, (void __user *)arg, minsz))
2718 if (info.argsz < minsz)
2721 if (info.argsz >= capsz) {
2723 info.cap_offset = 0; /* output, no-recopy necessary */
2726 mutex_lock(&iommu->lock);
2727 info.flags = VFIO_IOMMU_INFO_PGSIZES;
2729 info.iova_pgsizes = iommu->pgsize_bitmap;
2731 ret = vfio_iommu_migration_build_caps(iommu, &caps);
2734 ret = vfio_iommu_dma_avail_build_caps(iommu, &caps);
2737 ret = vfio_iommu_iova_build_caps(iommu, &caps);
2739 mutex_unlock(&iommu->lock);
2745 info.flags |= VFIO_IOMMU_INFO_CAPS;
2747 if (info.argsz < sizeof(info) + caps.size) {
2748 info.argsz = sizeof(info) + caps.size;
2750 vfio_info_cap_shift(&caps, sizeof(info));
2751 if (copy_to_user((void __user *)arg +
2752 sizeof(info), caps.buf,
2757 info.cap_offset = sizeof(info);
2763 return copy_to_user((void __user *)arg, &info, minsz) ?
2767 static int vfio_iommu_type1_map_dma(struct vfio_iommu *iommu,
2770 struct vfio_iommu_type1_dma_map map;
2771 unsigned long minsz;
2772 uint32_t mask = VFIO_DMA_MAP_FLAG_READ | VFIO_DMA_MAP_FLAG_WRITE;
2774 minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
2776 if (copy_from_user(&map, (void __user *)arg, minsz))
2779 if (map.argsz < minsz || map.flags & ~mask)
2782 return vfio_dma_do_map(iommu, &map);
2785 static int vfio_iommu_type1_unmap_dma(struct vfio_iommu *iommu,
2788 struct vfio_iommu_type1_dma_unmap unmap;
2789 struct vfio_bitmap bitmap = { 0 };
2790 unsigned long minsz;
2793 minsz = offsetofend(struct vfio_iommu_type1_dma_unmap, size);
2795 if (copy_from_user(&unmap, (void __user *)arg, minsz))
2798 if (unmap.argsz < minsz ||
2799 unmap.flags & ~VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP)
2802 if (unmap.flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) {
2803 unsigned long pgshift;
2805 if (unmap.argsz < (minsz + sizeof(bitmap)))
2808 if (copy_from_user(&bitmap,
2809 (void __user *)(arg + minsz),
2813 if (!access_ok((void __user *)bitmap.data, bitmap.size))
2816 pgshift = __ffs(bitmap.pgsize);
2817 ret = verify_bitmap_size(unmap.size >> pgshift,
2823 ret = vfio_dma_do_unmap(iommu, &unmap, &bitmap);
2827 return copy_to_user((void __user *)arg, &unmap, minsz) ?
2831 static int vfio_iommu_type1_dirty_pages(struct vfio_iommu *iommu,
2834 struct vfio_iommu_type1_dirty_bitmap dirty;
2835 uint32_t mask = VFIO_IOMMU_DIRTY_PAGES_FLAG_START |
2836 VFIO_IOMMU_DIRTY_PAGES_FLAG_STOP |
2837 VFIO_IOMMU_DIRTY_PAGES_FLAG_GET_BITMAP;
2838 unsigned long minsz;
2844 minsz = offsetofend(struct vfio_iommu_type1_dirty_bitmap, flags);
2846 if (copy_from_user(&dirty, (void __user *)arg, minsz))
2849 if (dirty.argsz < minsz || dirty.flags & ~mask)
2852 /* only one flag should be set at a time */
2853 if (__ffs(dirty.flags) != __fls(dirty.flags))
2856 if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_START) {
2859 mutex_lock(&iommu->lock);
2860 pgsize = 1 << __ffs(iommu->pgsize_bitmap);
2861 if (!iommu->dirty_page_tracking) {
2862 ret = vfio_dma_bitmap_alloc_all(iommu, pgsize);
2864 iommu->dirty_page_tracking = true;
2866 mutex_unlock(&iommu->lock);
2868 } else if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_STOP) {
2869 mutex_lock(&iommu->lock);
2870 if (iommu->dirty_page_tracking) {
2871 iommu->dirty_page_tracking = false;
2872 vfio_dma_bitmap_free_all(iommu);
2874 mutex_unlock(&iommu->lock);
2876 } else if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_GET_BITMAP) {
2877 struct vfio_iommu_type1_dirty_bitmap_get range;
2878 unsigned long pgshift;
2879 size_t data_size = dirty.argsz - minsz;
2880 size_t iommu_pgsize;
2882 if (!data_size || data_size < sizeof(range))
2885 if (copy_from_user(&range, (void __user *)(arg + minsz),
2889 if (range.iova + range.size < range.iova)
2891 if (!access_ok((void __user *)range.bitmap.data,
2895 pgshift = __ffs(range.bitmap.pgsize);
2896 ret = verify_bitmap_size(range.size >> pgshift,
2901 mutex_lock(&iommu->lock);
2903 iommu_pgsize = (size_t)1 << __ffs(iommu->pgsize_bitmap);
2905 /* allow only smallest supported pgsize */
2906 if (range.bitmap.pgsize != iommu_pgsize) {
2910 if (range.iova & (iommu_pgsize - 1)) {
2914 if (!range.size || range.size & (iommu_pgsize - 1)) {
2919 if (iommu->dirty_page_tracking)
2920 ret = vfio_iova_dirty_bitmap(range.bitmap.data,
2923 range.bitmap.pgsize);
2927 mutex_unlock(&iommu->lock);
2935 static long vfio_iommu_type1_ioctl(void *iommu_data,
2936 unsigned int cmd, unsigned long arg)
2938 struct vfio_iommu *iommu = iommu_data;
2941 case VFIO_CHECK_EXTENSION:
2942 return vfio_iommu_type1_check_extension(iommu, arg);
2943 case VFIO_IOMMU_GET_INFO:
2944 return vfio_iommu_type1_get_info(iommu, arg);
2945 case VFIO_IOMMU_MAP_DMA:
2946 return vfio_iommu_type1_map_dma(iommu, arg);
2947 case VFIO_IOMMU_UNMAP_DMA:
2948 return vfio_iommu_type1_unmap_dma(iommu, arg);
2949 case VFIO_IOMMU_DIRTY_PAGES:
2950 return vfio_iommu_type1_dirty_pages(iommu, arg);
2956 static int vfio_iommu_type1_register_notifier(void *iommu_data,
2957 unsigned long *events,
2958 struct notifier_block *nb)
2960 struct vfio_iommu *iommu = iommu_data;
2962 /* clear known events */
2963 *events &= ~VFIO_IOMMU_NOTIFY_DMA_UNMAP;
2965 /* refuse to register if still events remaining */
2969 return blocking_notifier_chain_register(&iommu->notifier, nb);
2972 static int vfio_iommu_type1_unregister_notifier(void *iommu_data,
2973 struct notifier_block *nb)
2975 struct vfio_iommu *iommu = iommu_data;
2977 return blocking_notifier_chain_unregister(&iommu->notifier, nb);
2980 static int vfio_iommu_type1_dma_rw_chunk(struct vfio_iommu *iommu,
2981 dma_addr_t user_iova, void *data,
2982 size_t count, bool write,
2985 struct mm_struct *mm;
2986 unsigned long vaddr;
2987 struct vfio_dma *dma;
2988 bool kthread = current->mm == NULL;
2993 dma = vfio_find_dma(iommu, user_iova, 1);
2997 if ((write && !(dma->prot & IOMMU_WRITE)) ||
2998 !(dma->prot & IOMMU_READ))
3001 mm = get_task_mm(dma->task);
3008 else if (current->mm != mm)
3011 offset = user_iova - dma->iova;
3013 if (count > dma->size - offset)
3014 count = dma->size - offset;
3016 vaddr = dma->vaddr + offset;
3019 *copied = copy_to_user((void __user *)vaddr, data,
3021 if (*copied && iommu->dirty_page_tracking) {
3022 unsigned long pgshift = __ffs(iommu->pgsize_bitmap);
3024 * Bitmap populated with the smallest supported page
3027 bitmap_set(dma->bitmap, offset >> pgshift,
3028 ((offset + *copied - 1) >> pgshift) -
3029 (offset >> pgshift) + 1);
3032 *copied = copy_from_user(data, (void __user *)vaddr,
3035 kthread_unuse_mm(mm);
3038 return *copied ? 0 : -EFAULT;
3041 static int vfio_iommu_type1_dma_rw(void *iommu_data, dma_addr_t user_iova,
3042 void *data, size_t count, bool write)
3044 struct vfio_iommu *iommu = iommu_data;
3048 mutex_lock(&iommu->lock);
3050 ret = vfio_iommu_type1_dma_rw_chunk(iommu, user_iova, data,
3051 count, write, &done);
3060 mutex_unlock(&iommu->lock);
3064 static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = {
3065 .name = "vfio-iommu-type1",
3066 .owner = THIS_MODULE,
3067 .open = vfio_iommu_type1_open,
3068 .release = vfio_iommu_type1_release,
3069 .ioctl = vfio_iommu_type1_ioctl,
3070 .attach_group = vfio_iommu_type1_attach_group,
3071 .detach_group = vfio_iommu_type1_detach_group,
3072 .pin_pages = vfio_iommu_type1_pin_pages,
3073 .unpin_pages = vfio_iommu_type1_unpin_pages,
3074 .register_notifier = vfio_iommu_type1_register_notifier,
3075 .unregister_notifier = vfio_iommu_type1_unregister_notifier,
3076 .dma_rw = vfio_iommu_type1_dma_rw,
3079 static int __init vfio_iommu_type1_init(void)
3081 return vfio_register_iommu_driver(&vfio_iommu_driver_ops_type1);
3084 static void __exit vfio_iommu_type1_cleanup(void)
3086 vfio_unregister_iommu_driver(&vfio_iommu_driver_ops_type1);
3089 module_init(vfio_iommu_type1_init);
3090 module_exit(vfio_iommu_type1_cleanup);
3092 MODULE_VERSION(DRIVER_VERSION);
3093 MODULE_LICENSE("GPL v2");
3094 MODULE_AUTHOR(DRIVER_AUTHOR);
3095 MODULE_DESCRIPTION(DRIVER_DESC);