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 * userspace 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/notifier.h>
42 #define DRIVER_VERSION "0.2"
43 #define DRIVER_AUTHOR "Alex Williamson <alex.williamson@redhat.com>"
44 #define DRIVER_DESC "Type1 IOMMU driver for VFIO"
46 static bool allow_unsafe_interrupts;
47 module_param_named(allow_unsafe_interrupts,
48 allow_unsafe_interrupts, bool, S_IRUGO | S_IWUSR);
49 MODULE_PARM_DESC(allow_unsafe_interrupts,
50 "Enable VFIO IOMMU support for on platforms without interrupt remapping support.");
52 static bool disable_hugepages;
53 module_param_named(disable_hugepages,
54 disable_hugepages, bool, S_IRUGO | S_IWUSR);
55 MODULE_PARM_DESC(disable_hugepages,
56 "Disable VFIO IOMMU support for IOMMU hugepages.");
58 static unsigned int dma_entry_limit __read_mostly = U16_MAX;
59 module_param_named(dma_entry_limit, dma_entry_limit, uint, 0644);
60 MODULE_PARM_DESC(dma_entry_limit,
61 "Maximum number of user DMA mappings per container (65535).");
64 struct list_head domain_list;
65 struct list_head iova_list;
67 struct rb_root dma_list;
68 struct list_head device_list;
69 struct mutex device_list_lock;
70 unsigned int dma_avail;
71 unsigned int vaddr_invalid_count;
72 uint64_t pgsize_bitmap;
73 uint64_t num_non_pinned_groups;
76 bool dirty_page_tracking;
77 struct list_head emulated_iommu_groups;
81 struct iommu_domain *domain;
82 struct list_head next;
83 struct list_head group_list;
84 bool fgsp : 1; /* Fine-grained super pages */
85 bool enforce_cache_coherency : 1;
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) */
97 struct task_struct *task;
98 struct rb_root pfn_list; /* Ex-user pinned pfn list */
99 unsigned long *bitmap;
100 struct mm_struct *mm;
105 struct page **pages; /* for pin_user_pages_remote */
106 struct page *fallback_page; /* if pages alloc fails */
107 int capacity; /* length of pages array */
108 int size; /* of batch currently */
109 int offset; /* of next entry in pages */
112 struct vfio_iommu_group {
113 struct iommu_group *iommu_group;
114 struct list_head next;
115 bool pinned_page_dirty_scope;
119 struct list_head list;
125 * Guest RAM pinning working set or DMA target
129 dma_addr_t iova; /* Device address */
130 unsigned long pfn; /* Host pfn */
131 unsigned int ref_count;
134 struct vfio_regions {
135 struct list_head list;
141 #define DIRTY_BITMAP_BYTES(n) (ALIGN(n, BITS_PER_TYPE(u64)) / BITS_PER_BYTE)
144 * Input argument of number of bits to bitmap_set() is unsigned integer, which
145 * further casts to signed integer for unaligned multi-bit operation,
147 * Then maximum bitmap size supported is 2^31 bits divided by 2^3 bits/byte,
148 * that is 2^28 (256 MB) which maps to 2^31 * 2^12 = 2^43 (8TB) on 4K page
151 #define DIRTY_BITMAP_PAGES_MAX ((u64)INT_MAX)
152 #define DIRTY_BITMAP_SIZE_MAX DIRTY_BITMAP_BYTES(DIRTY_BITMAP_PAGES_MAX)
154 static int put_pfn(unsigned long pfn, int prot);
156 static struct vfio_iommu_group*
157 vfio_iommu_find_iommu_group(struct vfio_iommu *iommu,
158 struct iommu_group *iommu_group);
161 * This code handles mapping and unmapping of user data buffers
162 * into DMA'ble space using the IOMMU
165 static struct vfio_dma *vfio_find_dma(struct vfio_iommu *iommu,
166 dma_addr_t start, size_t size)
168 struct rb_node *node = iommu->dma_list.rb_node;
171 struct vfio_dma *dma = rb_entry(node, struct vfio_dma, node);
173 if (start + size <= dma->iova)
174 node = node->rb_left;
175 else if (start >= dma->iova + dma->size)
176 node = node->rb_right;
184 static struct rb_node *vfio_find_dma_first_node(struct vfio_iommu *iommu,
185 dma_addr_t start, u64 size)
187 struct rb_node *res = NULL;
188 struct rb_node *node = iommu->dma_list.rb_node;
189 struct vfio_dma *dma_res = NULL;
192 struct vfio_dma *dma = rb_entry(node, struct vfio_dma, node);
194 if (start < dma->iova + dma->size) {
197 if (start >= dma->iova)
199 node = node->rb_left;
201 node = node->rb_right;
204 if (res && size && dma_res->iova >= start + size)
209 static void vfio_link_dma(struct vfio_iommu *iommu, struct vfio_dma *new)
211 struct rb_node **link = &iommu->dma_list.rb_node, *parent = NULL;
212 struct vfio_dma *dma;
216 dma = rb_entry(parent, struct vfio_dma, node);
218 if (new->iova + new->size <= dma->iova)
219 link = &(*link)->rb_left;
221 link = &(*link)->rb_right;
224 rb_link_node(&new->node, parent, link);
225 rb_insert_color(&new->node, &iommu->dma_list);
228 static void vfio_unlink_dma(struct vfio_iommu *iommu, struct vfio_dma *old)
230 rb_erase(&old->node, &iommu->dma_list);
234 static int vfio_dma_bitmap_alloc(struct vfio_dma *dma, size_t pgsize)
236 uint64_t npages = dma->size / pgsize;
238 if (npages > DIRTY_BITMAP_PAGES_MAX)
242 * Allocate extra 64 bits that are used to calculate shift required for
243 * bitmap_shift_left() to manipulate and club unaligned number of pages
244 * in adjacent vfio_dma ranges.
246 dma->bitmap = kvzalloc(DIRTY_BITMAP_BYTES(npages) + sizeof(u64),
254 static void vfio_dma_bitmap_free(struct vfio_dma *dma)
260 static void vfio_dma_populate_bitmap(struct vfio_dma *dma, size_t pgsize)
263 unsigned long pgshift = __ffs(pgsize);
265 for (p = rb_first(&dma->pfn_list); p; p = rb_next(p)) {
266 struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn, node);
268 bitmap_set(dma->bitmap, (vpfn->iova - dma->iova) >> pgshift, 1);
272 static void vfio_iommu_populate_bitmap_full(struct vfio_iommu *iommu)
275 unsigned long pgshift = __ffs(iommu->pgsize_bitmap);
277 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
278 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
280 bitmap_set(dma->bitmap, 0, dma->size >> pgshift);
284 static int vfio_dma_bitmap_alloc_all(struct vfio_iommu *iommu, size_t pgsize)
288 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
289 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
292 ret = vfio_dma_bitmap_alloc(dma, pgsize);
296 for (p = rb_prev(n); p; p = rb_prev(p)) {
297 struct vfio_dma *dma = rb_entry(n,
298 struct vfio_dma, node);
300 vfio_dma_bitmap_free(dma);
304 vfio_dma_populate_bitmap(dma, pgsize);
309 static void vfio_dma_bitmap_free_all(struct vfio_iommu *iommu)
313 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
314 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
316 vfio_dma_bitmap_free(dma);
321 * Helper Functions for host iova-pfn list
323 static struct vfio_pfn *vfio_find_vpfn(struct vfio_dma *dma, dma_addr_t iova)
325 struct vfio_pfn *vpfn;
326 struct rb_node *node = dma->pfn_list.rb_node;
329 vpfn = rb_entry(node, struct vfio_pfn, node);
331 if (iova < vpfn->iova)
332 node = node->rb_left;
333 else if (iova > vpfn->iova)
334 node = node->rb_right;
341 static void vfio_link_pfn(struct vfio_dma *dma,
342 struct vfio_pfn *new)
344 struct rb_node **link, *parent = NULL;
345 struct vfio_pfn *vpfn;
347 link = &dma->pfn_list.rb_node;
350 vpfn = rb_entry(parent, struct vfio_pfn, node);
352 if (new->iova < vpfn->iova)
353 link = &(*link)->rb_left;
355 link = &(*link)->rb_right;
358 rb_link_node(&new->node, parent, link);
359 rb_insert_color(&new->node, &dma->pfn_list);
362 static void vfio_unlink_pfn(struct vfio_dma *dma, struct vfio_pfn *old)
364 rb_erase(&old->node, &dma->pfn_list);
367 static int vfio_add_to_pfn_list(struct vfio_dma *dma, dma_addr_t iova,
370 struct vfio_pfn *vpfn;
372 vpfn = kzalloc(sizeof(*vpfn), GFP_KERNEL);
379 vfio_link_pfn(dma, vpfn);
383 static void vfio_remove_from_pfn_list(struct vfio_dma *dma,
384 struct vfio_pfn *vpfn)
386 vfio_unlink_pfn(dma, vpfn);
390 static struct vfio_pfn *vfio_iova_get_vfio_pfn(struct vfio_dma *dma,
393 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
400 static int vfio_iova_put_vfio_pfn(struct vfio_dma *dma, struct vfio_pfn *vpfn)
405 if (!vpfn->ref_count) {
406 ret = put_pfn(vpfn->pfn, dma->prot);
407 vfio_remove_from_pfn_list(dma, vpfn);
412 static int mm_lock_acct(struct task_struct *task, struct mm_struct *mm,
413 bool lock_cap, long npage)
415 int ret = mmap_write_lock_killable(mm);
420 ret = __account_locked_vm(mm, abs(npage), npage > 0, task, lock_cap);
421 mmap_write_unlock(mm);
425 static int vfio_lock_acct(struct vfio_dma *dma, long npage, bool async)
427 struct mm_struct *mm;
434 if (async && !mmget_not_zero(mm))
435 return -ESRCH; /* process exited */
437 ret = mm_lock_acct(dma->task, mm, dma->lock_cap, npage);
439 dma->locked_vm += npage;
448 * Some mappings aren't backed by a struct page, for example an mmap'd
449 * MMIO range for our own or another device. These use a different
450 * pfn conversion and shouldn't be tracked as locked pages.
451 * For compound pages, any driver that sets the reserved bit in head
452 * page needs to set the reserved bit in all subpages to be safe.
454 static bool is_invalid_reserved_pfn(unsigned long pfn)
457 return PageReserved(pfn_to_page(pfn));
462 static int put_pfn(unsigned long pfn, int prot)
464 if (!is_invalid_reserved_pfn(pfn)) {
465 struct page *page = pfn_to_page(pfn);
467 unpin_user_pages_dirty_lock(&page, 1, prot & IOMMU_WRITE);
473 #define VFIO_BATCH_MAX_CAPACITY (PAGE_SIZE / sizeof(struct page *))
475 static void vfio_batch_init(struct vfio_batch *batch)
480 if (unlikely(disable_hugepages))
483 batch->pages = (struct page **) __get_free_page(GFP_KERNEL);
487 batch->capacity = VFIO_BATCH_MAX_CAPACITY;
491 batch->pages = &batch->fallback_page;
495 static void vfio_batch_unpin(struct vfio_batch *batch, struct vfio_dma *dma)
497 while (batch->size) {
498 unsigned long pfn = page_to_pfn(batch->pages[batch->offset]);
500 put_pfn(pfn, dma->prot);
506 static void vfio_batch_fini(struct vfio_batch *batch)
508 if (batch->capacity == VFIO_BATCH_MAX_CAPACITY)
509 free_page((unsigned long)batch->pages);
512 static int follow_fault_pfn(struct vm_area_struct *vma, struct mm_struct *mm,
513 unsigned long vaddr, unsigned long *pfn,
521 ret = follow_pte(vma->vm_mm, vaddr, &ptep, &ptl);
523 bool unlocked = false;
525 ret = fixup_user_fault(mm, vaddr,
527 (write_fault ? FAULT_FLAG_WRITE : 0),
535 ret = follow_pte(vma->vm_mm, vaddr, &ptep, &ptl);
540 pte = ptep_get(ptep);
542 if (write_fault && !pte_write(pte))
547 pte_unmap_unlock(ptep, ptl);
552 * Returns the positive number of pfns successfully obtained or a negative
555 static int vaddr_get_pfns(struct mm_struct *mm, unsigned long vaddr,
556 long npages, int prot, unsigned long *pfn,
559 struct vm_area_struct *vma;
560 unsigned int flags = 0;
563 if (prot & IOMMU_WRITE)
567 ret = pin_user_pages_remote(mm, vaddr, npages, flags | FOLL_LONGTERM,
573 * The zero page is always resident, we don't need to pin it
574 * and it falls into our invalid/reserved test so we don't
575 * unpin in put_pfn(). Unpin all zero pages in the batch here.
577 for (i = 0 ; i < ret; i++) {
578 if (unlikely(is_zero_pfn(page_to_pfn(pages[i]))))
579 unpin_user_page(pages[i]);
582 *pfn = page_to_pfn(pages[0]);
586 vaddr = untagged_addr_remote(mm, vaddr);
589 vma = vma_lookup(mm, vaddr);
591 if (vma && vma->vm_flags & VM_PFNMAP) {
592 ret = follow_fault_pfn(vma, mm, vaddr, pfn, prot & IOMMU_WRITE);
597 if (is_invalid_reserved_pfn(*pfn))
604 mmap_read_unlock(mm);
609 * Attempt to pin pages. We really don't want to track all the pfns and
610 * the iommu can only map chunks of consecutive pfns anyway, so get the
611 * first page and all consecutive pages with the same locking.
613 static long vfio_pin_pages_remote(struct vfio_dma *dma, unsigned long vaddr,
614 long npage, unsigned long *pfn_base,
615 unsigned long limit, struct vfio_batch *batch)
618 struct mm_struct *mm = current->mm;
619 long ret, pinned = 0, lock_acct = 0;
621 dma_addr_t iova = vaddr - dma->vaddr + dma->iova;
623 /* This code path is only user initiated */
628 /* Leftover pages in batch from an earlier call. */
629 *pfn_base = page_to_pfn(batch->pages[batch->offset]);
631 rsvd = is_invalid_reserved_pfn(*pfn_base);
638 /* Empty batch, so refill it. */
639 long req_pages = min_t(long, npage, batch->capacity);
641 ret = vaddr_get_pfns(mm, vaddr, req_pages, dma->prot,
651 rsvd = is_invalid_reserved_pfn(*pfn_base);
656 * pfn is preset for the first iteration of this inner loop and
657 * updated at the end to handle a VM_PFNMAP pfn. In that case,
658 * batch->pages isn't valid (there's no struct page), so allow
659 * batch->pages to be touched only when there's more than one
660 * pfn to check, which guarantees the pfns are from a
664 if (pfn != *pfn_base + pinned ||
665 rsvd != is_invalid_reserved_pfn(pfn))
669 * Reserved pages aren't counted against the user,
670 * externally pinned pages are already counted against
673 if (!rsvd && !vfio_find_vpfn(dma, iova)) {
674 if (!dma->lock_cap &&
675 mm->locked_vm + lock_acct + 1 > limit) {
676 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n",
677 __func__, limit << PAGE_SHIFT);
694 pfn = page_to_pfn(batch->pages[batch->offset]);
697 if (unlikely(disable_hugepages))
702 ret = vfio_lock_acct(dma, lock_acct, false);
705 if (batch->size == 1 && !batch->offset) {
706 /* May be a VM_PFNMAP pfn, which the batch can't remember. */
707 put_pfn(pfn, dma->prot);
712 if (pinned && !rsvd) {
713 for (pfn = *pfn_base ; pinned ; pfn++, pinned--)
714 put_pfn(pfn, dma->prot);
716 vfio_batch_unpin(batch, dma);
724 static long vfio_unpin_pages_remote(struct vfio_dma *dma, dma_addr_t iova,
725 unsigned long pfn, long npage,
728 long unlocked = 0, locked = 0;
731 for (i = 0; i < npage; i++, iova += PAGE_SIZE) {
732 if (put_pfn(pfn++, dma->prot)) {
734 if (vfio_find_vpfn(dma, iova))
740 vfio_lock_acct(dma, locked - unlocked, true);
745 static int vfio_pin_page_external(struct vfio_dma *dma, unsigned long vaddr,
746 unsigned long *pfn_base, bool do_accounting)
748 struct page *pages[1];
749 struct mm_struct *mm;
753 if (!mmget_not_zero(mm))
756 ret = vaddr_get_pfns(mm, vaddr, 1, dma->prot, pfn_base, pages);
762 if (do_accounting && !is_invalid_reserved_pfn(*pfn_base)) {
763 ret = vfio_lock_acct(dma, 1, false);
765 put_pfn(*pfn_base, dma->prot);
767 pr_warn("%s: Task %s (%d) RLIMIT_MEMLOCK "
768 "(%ld) exceeded\n", __func__,
769 dma->task->comm, task_pid_nr(dma->task),
770 task_rlimit(dma->task, RLIMIT_MEMLOCK));
779 static int vfio_unpin_page_external(struct vfio_dma *dma, dma_addr_t iova,
783 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
788 unlocked = vfio_iova_put_vfio_pfn(dma, vpfn);
791 vfio_lock_acct(dma, -unlocked, true);
796 static int vfio_iommu_type1_pin_pages(void *iommu_data,
797 struct iommu_group *iommu_group,
798 dma_addr_t user_iova,
802 struct vfio_iommu *iommu = iommu_data;
803 struct vfio_iommu_group *group;
805 unsigned long remote_vaddr;
806 struct vfio_dma *dma;
809 if (!iommu || !pages)
812 /* Supported for v2 version only */
816 mutex_lock(&iommu->lock);
818 if (WARN_ONCE(iommu->vaddr_invalid_count,
819 "vfio_pin_pages not allowed with VFIO_UPDATE_VADDR\n")) {
824 /* Fail if no dma_umap notifier is registered */
825 if (list_empty(&iommu->device_list)) {
831 * If iommu capable domain exist in the container then all pages are
832 * already pinned and accounted. Accounting should be done if there is no
833 * iommu capable domain in the container.
835 do_accounting = list_empty(&iommu->domain_list);
837 for (i = 0; i < npage; i++) {
838 unsigned long phys_pfn;
840 struct vfio_pfn *vpfn;
842 iova = user_iova + PAGE_SIZE * i;
843 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
849 if ((dma->prot & prot) != prot) {
854 vpfn = vfio_iova_get_vfio_pfn(dma, iova);
856 pages[i] = pfn_to_page(vpfn->pfn);
860 remote_vaddr = dma->vaddr + (iova - dma->iova);
861 ret = vfio_pin_page_external(dma, remote_vaddr, &phys_pfn,
866 if (!pfn_valid(phys_pfn)) {
871 ret = vfio_add_to_pfn_list(dma, iova, phys_pfn);
873 if (put_pfn(phys_pfn, dma->prot) && do_accounting)
874 vfio_lock_acct(dma, -1, true);
878 pages[i] = pfn_to_page(phys_pfn);
880 if (iommu->dirty_page_tracking) {
881 unsigned long pgshift = __ffs(iommu->pgsize_bitmap);
884 * Bitmap populated with the smallest supported page
887 bitmap_set(dma->bitmap,
888 (iova - dma->iova) >> pgshift, 1);
893 group = vfio_iommu_find_iommu_group(iommu, iommu_group);
894 if (!group->pinned_page_dirty_scope) {
895 group->pinned_page_dirty_scope = true;
896 iommu->num_non_pinned_groups--;
903 for (j = 0; j < i; j++) {
906 iova = user_iova + PAGE_SIZE * j;
907 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
908 vfio_unpin_page_external(dma, iova, do_accounting);
912 mutex_unlock(&iommu->lock);
916 static void vfio_iommu_type1_unpin_pages(void *iommu_data,
917 dma_addr_t user_iova, int npage)
919 struct vfio_iommu *iommu = iommu_data;
923 /* Supported for v2 version only */
924 if (WARN_ON(!iommu->v2))
927 mutex_lock(&iommu->lock);
929 do_accounting = list_empty(&iommu->domain_list);
930 for (i = 0; i < npage; i++) {
931 dma_addr_t iova = user_iova + PAGE_SIZE * i;
932 struct vfio_dma *dma;
934 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
938 vfio_unpin_page_external(dma, iova, do_accounting);
941 mutex_unlock(&iommu->lock);
946 static long vfio_sync_unpin(struct vfio_dma *dma, struct vfio_domain *domain,
947 struct list_head *regions,
948 struct iommu_iotlb_gather *iotlb_gather)
951 struct vfio_regions *entry, *next;
953 iommu_iotlb_sync(domain->domain, iotlb_gather);
955 list_for_each_entry_safe(entry, next, regions, list) {
956 unlocked += vfio_unpin_pages_remote(dma,
958 entry->phys >> PAGE_SHIFT,
959 entry->len >> PAGE_SHIFT,
961 list_del(&entry->list);
971 * Generally, VFIO needs to unpin remote pages after each IOTLB flush.
972 * Therefore, when using IOTLB flush sync interface, VFIO need to keep track
973 * of these regions (currently using a list).
975 * This value specifies maximum number of regions for each IOTLB flush sync.
977 #define VFIO_IOMMU_TLB_SYNC_MAX 512
979 static size_t unmap_unpin_fast(struct vfio_domain *domain,
980 struct vfio_dma *dma, dma_addr_t *iova,
981 size_t len, phys_addr_t phys, long *unlocked,
982 struct list_head *unmapped_list,
984 struct iommu_iotlb_gather *iotlb_gather)
987 struct vfio_regions *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
990 unmapped = iommu_unmap_fast(domain->domain, *iova, len,
998 entry->len = unmapped;
999 list_add_tail(&entry->list, unmapped_list);
1007 * Sync if the number of fast-unmap regions hits the limit
1008 * or in case of errors.
1010 if (*unmapped_cnt >= VFIO_IOMMU_TLB_SYNC_MAX || !unmapped) {
1011 *unlocked += vfio_sync_unpin(dma, domain, unmapped_list,
1019 static size_t unmap_unpin_slow(struct vfio_domain *domain,
1020 struct vfio_dma *dma, dma_addr_t *iova,
1021 size_t len, phys_addr_t phys,
1024 size_t unmapped = iommu_unmap(domain->domain, *iova, len);
1027 *unlocked += vfio_unpin_pages_remote(dma, *iova,
1029 unmapped >> PAGE_SHIFT,
1037 static long vfio_unmap_unpin(struct vfio_iommu *iommu, struct vfio_dma *dma,
1040 dma_addr_t iova = dma->iova, end = dma->iova + dma->size;
1041 struct vfio_domain *domain, *d;
1042 LIST_HEAD(unmapped_region_list);
1043 struct iommu_iotlb_gather iotlb_gather;
1044 int unmapped_region_cnt = 0;
1050 if (list_empty(&iommu->domain_list))
1054 * We use the IOMMU to track the physical addresses, otherwise we'd
1055 * need a much more complicated tracking system. Unfortunately that
1056 * means we need to use one of the iommu domains to figure out the
1057 * pfns to unpin. The rest need to be unmapped in advance so we have
1058 * no iommu translations remaining when the pages are unpinned.
1060 domain = d = list_first_entry(&iommu->domain_list,
1061 struct vfio_domain, next);
1063 list_for_each_entry_continue(d, &iommu->domain_list, next) {
1064 iommu_unmap(d->domain, dma->iova, dma->size);
1068 iommu_iotlb_gather_init(&iotlb_gather);
1069 while (iova < end) {
1070 size_t unmapped, len;
1071 phys_addr_t phys, next;
1073 phys = iommu_iova_to_phys(domain->domain, iova);
1074 if (WARN_ON(!phys)) {
1080 * To optimize for fewer iommu_unmap() calls, each of which
1081 * may require hardware cache flushing, try to find the
1082 * largest contiguous physical memory chunk to unmap.
1084 for (len = PAGE_SIZE;
1085 !domain->fgsp && iova + len < end; len += PAGE_SIZE) {
1086 next = iommu_iova_to_phys(domain->domain, iova + len);
1087 if (next != phys + len)
1092 * First, try to use fast unmap/unpin. In case of failure,
1093 * switch to slow unmap/unpin path.
1095 unmapped = unmap_unpin_fast(domain, dma, &iova, len, phys,
1096 &unlocked, &unmapped_region_list,
1097 &unmapped_region_cnt,
1100 unmapped = unmap_unpin_slow(domain, dma, &iova, len,
1102 if (WARN_ON(!unmapped))
1107 dma->iommu_mapped = false;
1109 if (unmapped_region_cnt) {
1110 unlocked += vfio_sync_unpin(dma, domain, &unmapped_region_list,
1114 if (do_accounting) {
1115 vfio_lock_acct(dma, -unlocked, true);
1121 static void vfio_remove_dma(struct vfio_iommu *iommu, struct vfio_dma *dma)
1123 WARN_ON(!RB_EMPTY_ROOT(&dma->pfn_list));
1124 vfio_unmap_unpin(iommu, dma, true);
1125 vfio_unlink_dma(iommu, dma);
1126 put_task_struct(dma->task);
1128 vfio_dma_bitmap_free(dma);
1129 if (dma->vaddr_invalid)
1130 iommu->vaddr_invalid_count--;
1135 static void vfio_update_pgsize_bitmap(struct vfio_iommu *iommu)
1137 struct vfio_domain *domain;
1139 iommu->pgsize_bitmap = ULONG_MAX;
1141 list_for_each_entry(domain, &iommu->domain_list, next)
1142 iommu->pgsize_bitmap &= domain->domain->pgsize_bitmap;
1145 * In case the IOMMU supports page sizes smaller than PAGE_SIZE
1146 * we pretend PAGE_SIZE is supported and hide sub-PAGE_SIZE sizes.
1147 * That way the user will be able to map/unmap buffers whose size/
1148 * start address is aligned with PAGE_SIZE. Pinning code uses that
1149 * granularity while iommu driver can use the sub-PAGE_SIZE size
1150 * to map the buffer.
1152 if (iommu->pgsize_bitmap & ~PAGE_MASK) {
1153 iommu->pgsize_bitmap &= PAGE_MASK;
1154 iommu->pgsize_bitmap |= PAGE_SIZE;
1158 static int update_user_bitmap(u64 __user *bitmap, struct vfio_iommu *iommu,
1159 struct vfio_dma *dma, dma_addr_t base_iova,
1162 unsigned long pgshift = __ffs(pgsize);
1163 unsigned long nbits = dma->size >> pgshift;
1164 unsigned long bit_offset = (dma->iova - base_iova) >> pgshift;
1165 unsigned long copy_offset = bit_offset / BITS_PER_LONG;
1166 unsigned long shift = bit_offset % BITS_PER_LONG;
1167 unsigned long leftover;
1170 * mark all pages dirty if any IOMMU capable device is not able
1171 * to report dirty pages and all pages are pinned and mapped.
1173 if (iommu->num_non_pinned_groups && dma->iommu_mapped)
1174 bitmap_set(dma->bitmap, 0, nbits);
1177 bitmap_shift_left(dma->bitmap, dma->bitmap, shift,
1180 if (copy_from_user(&leftover,
1181 (void __user *)(bitmap + copy_offset),
1185 bitmap_or(dma->bitmap, dma->bitmap, &leftover, shift);
1188 if (copy_to_user((void __user *)(bitmap + copy_offset), dma->bitmap,
1189 DIRTY_BITMAP_BYTES(nbits + shift)))
1195 static int vfio_iova_dirty_bitmap(u64 __user *bitmap, struct vfio_iommu *iommu,
1196 dma_addr_t iova, size_t size, size_t pgsize)
1198 struct vfio_dma *dma;
1200 unsigned long pgshift = __ffs(pgsize);
1204 * GET_BITMAP request must fully cover vfio_dma mappings. Multiple
1205 * vfio_dma mappings may be clubbed by specifying large ranges, but
1206 * there must not be any previous mappings bisected by the range.
1207 * An error will be returned if these conditions are not met.
1209 dma = vfio_find_dma(iommu, iova, 1);
1210 if (dma && dma->iova != iova)
1213 dma = vfio_find_dma(iommu, iova + size - 1, 0);
1214 if (dma && dma->iova + dma->size != iova + size)
1217 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
1218 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1220 if (dma->iova < iova)
1223 if (dma->iova > iova + size - 1)
1226 ret = update_user_bitmap(bitmap, iommu, dma, iova, pgsize);
1231 * Re-populate bitmap to include all pinned pages which are
1232 * considered as dirty but exclude pages which are unpinned and
1233 * pages which are marked dirty by vfio_dma_rw()
1235 bitmap_clear(dma->bitmap, 0, dma->size >> pgshift);
1236 vfio_dma_populate_bitmap(dma, pgsize);
1241 static int verify_bitmap_size(uint64_t npages, uint64_t bitmap_size)
1243 if (!npages || !bitmap_size || (bitmap_size > DIRTY_BITMAP_SIZE_MAX) ||
1244 (bitmap_size < DIRTY_BITMAP_BYTES(npages)))
1251 * Notify VFIO drivers using vfio_register_emulated_iommu_dev() to invalidate
1252 * and unmap iovas within the range we're about to unmap. Drivers MUST unpin
1253 * pages in response to an invalidation.
1255 static void vfio_notify_dma_unmap(struct vfio_iommu *iommu,
1256 struct vfio_dma *dma)
1258 struct vfio_device *device;
1260 if (list_empty(&iommu->device_list))
1264 * The device is expected to call vfio_unpin_pages() for any IOVA it has
1265 * pinned within the range. Since vfio_unpin_pages() will eventually
1266 * call back down to this code and try to obtain the iommu->lock we must
1269 mutex_lock(&iommu->device_list_lock);
1270 mutex_unlock(&iommu->lock);
1272 list_for_each_entry(device, &iommu->device_list, iommu_entry)
1273 device->ops->dma_unmap(device, dma->iova, dma->size);
1275 mutex_unlock(&iommu->device_list_lock);
1276 mutex_lock(&iommu->lock);
1279 static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
1280 struct vfio_iommu_type1_dma_unmap *unmap,
1281 struct vfio_bitmap *bitmap)
1283 struct vfio_dma *dma, *dma_last = NULL;
1284 size_t unmapped = 0, pgsize;
1285 int ret = -EINVAL, retries = 0;
1286 unsigned long pgshift;
1287 dma_addr_t iova = unmap->iova;
1288 u64 size = unmap->size;
1289 bool unmap_all = unmap->flags & VFIO_DMA_UNMAP_FLAG_ALL;
1290 bool invalidate_vaddr = unmap->flags & VFIO_DMA_UNMAP_FLAG_VADDR;
1291 struct rb_node *n, *first_n;
1293 mutex_lock(&iommu->lock);
1295 /* Cannot update vaddr if mdev is present. */
1296 if (invalidate_vaddr && !list_empty(&iommu->emulated_iommu_groups)) {
1301 pgshift = __ffs(iommu->pgsize_bitmap);
1302 pgsize = (size_t)1 << pgshift;
1304 if (iova & (pgsize - 1))
1311 } else if (!size || size & (pgsize - 1) ||
1312 iova + size - 1 < iova || size > SIZE_MAX) {
1316 /* When dirty tracking is enabled, allow only min supported pgsize */
1317 if ((unmap->flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) &&
1318 (!iommu->dirty_page_tracking || (bitmap->pgsize != pgsize))) {
1322 WARN_ON((pgsize - 1) & PAGE_MASK);
1325 * vfio-iommu-type1 (v1) - User mappings were coalesced together to
1326 * avoid tracking individual mappings. This means that the granularity
1327 * of the original mapping was lost and the user was allowed to attempt
1328 * to unmap any range. Depending on the contiguousness of physical
1329 * memory and page sizes supported by the IOMMU, arbitrary unmaps may
1330 * or may not have worked. We only guaranteed unmap granularity
1331 * matching the original mapping; even though it was untracked here,
1332 * the original mappings are reflected in IOMMU mappings. This
1333 * resulted in a couple unusual behaviors. First, if a range is not
1334 * able to be unmapped, ex. a set of 4k pages that was mapped as a
1335 * 2M hugepage into the IOMMU, the unmap ioctl returns success but with
1336 * a zero sized unmap. Also, if an unmap request overlaps the first
1337 * address of a hugepage, the IOMMU will unmap the entire hugepage.
1338 * This also returns success and the returned unmap size reflects the
1339 * actual size unmapped.
1341 * We attempt to maintain compatibility with this "v1" interface, but
1342 * we take control out of the hands of the IOMMU. Therefore, an unmap
1343 * request offset from the beginning of the original mapping will
1344 * return success with zero sized unmap. And an unmap request covering
1345 * the first iova of mapping will unmap the entire range.
1347 * The v2 version of this interface intends to be more deterministic.
1348 * Unmap requests must fully cover previous mappings. Multiple
1349 * mappings may still be unmaped by specifying large ranges, but there
1350 * must not be any previous mappings bisected by the range. An error
1351 * will be returned if these conditions are not met. The v2 interface
1352 * will only return success and a size of zero if there were no
1353 * mappings within the range.
1355 if (iommu->v2 && !unmap_all) {
1356 dma = vfio_find_dma(iommu, iova, 1);
1357 if (dma && dma->iova != iova)
1360 dma = vfio_find_dma(iommu, iova + size - 1, 0);
1361 if (dma && dma->iova + dma->size != iova + size)
1366 n = first_n = vfio_find_dma_first_node(iommu, iova, size);
1369 dma = rb_entry(n, struct vfio_dma, node);
1370 if (dma->iova >= iova + size)
1373 if (!iommu->v2 && iova > dma->iova)
1376 if (invalidate_vaddr) {
1377 if (dma->vaddr_invalid) {
1378 struct rb_node *last_n = n;
1380 for (n = first_n; n != last_n; n = rb_next(n)) {
1382 struct vfio_dma, node);
1383 dma->vaddr_invalid = false;
1384 iommu->vaddr_invalid_count--;
1390 dma->vaddr_invalid = true;
1391 iommu->vaddr_invalid_count++;
1392 unmapped += dma->size;
1397 if (!RB_EMPTY_ROOT(&dma->pfn_list)) {
1398 if (dma_last == dma) {
1399 BUG_ON(++retries > 10);
1405 vfio_notify_dma_unmap(iommu, dma);
1409 if (unmap->flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) {
1410 ret = update_user_bitmap(bitmap->data, iommu, dma,
1416 unmapped += dma->size;
1418 vfio_remove_dma(iommu, dma);
1422 mutex_unlock(&iommu->lock);
1424 /* Report how much was unmapped */
1425 unmap->size = unmapped;
1430 static int vfio_iommu_map(struct vfio_iommu *iommu, dma_addr_t iova,
1431 unsigned long pfn, long npage, int prot)
1433 struct vfio_domain *d;
1436 list_for_each_entry(d, &iommu->domain_list, next) {
1437 ret = iommu_map(d->domain, iova, (phys_addr_t)pfn << PAGE_SHIFT,
1438 npage << PAGE_SHIFT, prot | IOMMU_CACHE,
1449 list_for_each_entry_continue_reverse(d, &iommu->domain_list, next) {
1450 iommu_unmap(d->domain, iova, npage << PAGE_SHIFT);
1457 static int vfio_pin_map_dma(struct vfio_iommu *iommu, struct vfio_dma *dma,
1460 dma_addr_t iova = dma->iova;
1461 unsigned long vaddr = dma->vaddr;
1462 struct vfio_batch batch;
1463 size_t size = map_size;
1465 unsigned long pfn, limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1468 vfio_batch_init(&batch);
1471 /* Pin a contiguous chunk of memory */
1472 npage = vfio_pin_pages_remote(dma, vaddr + dma->size,
1473 size >> PAGE_SHIFT, &pfn, limit,
1482 ret = vfio_iommu_map(iommu, iova + dma->size, pfn, npage,
1485 vfio_unpin_pages_remote(dma, iova + dma->size, pfn,
1487 vfio_batch_unpin(&batch, dma);
1491 size -= npage << PAGE_SHIFT;
1492 dma->size += npage << PAGE_SHIFT;
1495 vfio_batch_fini(&batch);
1496 dma->iommu_mapped = true;
1499 vfio_remove_dma(iommu, dma);
1505 * Check dma map request is within a valid iova range
1507 static bool vfio_iommu_iova_dma_valid(struct vfio_iommu *iommu,
1508 dma_addr_t start, dma_addr_t end)
1510 struct list_head *iova = &iommu->iova_list;
1511 struct vfio_iova *node;
1513 list_for_each_entry(node, iova, list) {
1514 if (start >= node->start && end <= node->end)
1519 * Check for list_empty() as well since a container with
1520 * a single mdev device will have an empty list.
1522 return list_empty(iova);
1525 static int vfio_change_dma_owner(struct vfio_dma *dma)
1527 struct task_struct *task = current->group_leader;
1528 struct mm_struct *mm = current->mm;
1529 long npage = dma->locked_vm;
1536 lock_cap = capable(CAP_IPC_LOCK);
1537 ret = mm_lock_acct(task, mm, lock_cap, npage);
1541 if (mmget_not_zero(dma->mm)) {
1542 mm_lock_acct(dma->task, dma->mm, dma->lock_cap, -npage);
1546 if (dma->task != task) {
1547 put_task_struct(dma->task);
1548 dma->task = get_task_struct(task);
1553 dma->lock_cap = lock_cap;
1557 static int vfio_dma_do_map(struct vfio_iommu *iommu,
1558 struct vfio_iommu_type1_dma_map *map)
1560 bool set_vaddr = map->flags & VFIO_DMA_MAP_FLAG_VADDR;
1561 dma_addr_t iova = map->iova;
1562 unsigned long vaddr = map->vaddr;
1563 size_t size = map->size;
1564 int ret = 0, prot = 0;
1566 struct vfio_dma *dma;
1568 /* Verify that none of our __u64 fields overflow */
1569 if (map->size != size || map->vaddr != vaddr || map->iova != iova)
1572 /* READ/WRITE from device perspective */
1573 if (map->flags & VFIO_DMA_MAP_FLAG_WRITE)
1574 prot |= IOMMU_WRITE;
1575 if (map->flags & VFIO_DMA_MAP_FLAG_READ)
1578 if ((prot && set_vaddr) || (!prot && !set_vaddr))
1581 mutex_lock(&iommu->lock);
1583 pgsize = (size_t)1 << __ffs(iommu->pgsize_bitmap);
1585 WARN_ON((pgsize - 1) & PAGE_MASK);
1587 if (!size || (size | iova | vaddr) & (pgsize - 1)) {
1592 /* Don't allow IOVA or virtual address wrap */
1593 if (iova + size - 1 < iova || vaddr + size - 1 < vaddr) {
1598 dma = vfio_find_dma(iommu, iova, size);
1602 } else if (!dma->vaddr_invalid || dma->iova != iova ||
1603 dma->size != size) {
1606 ret = vfio_change_dma_owner(dma);
1610 dma->vaddr_invalid = false;
1611 iommu->vaddr_invalid_count--;
1619 if (!iommu->dma_avail) {
1624 if (!vfio_iommu_iova_dma_valid(iommu, iova, iova + size - 1)) {
1629 dma = kzalloc(sizeof(*dma), GFP_KERNEL);
1641 * We need to be able to both add to a task's locked memory and test
1642 * against the locked memory limit and we need to be able to do both
1643 * outside of this call path as pinning can be asynchronous via the
1644 * external interfaces for mdev devices. RLIMIT_MEMLOCK requires a
1645 * task_struct. Save the group_leader so that all DMA tracking uses
1646 * the same task, to make debugging easier. VM locked pages requires
1647 * an mm_struct, so grab the mm in case the task dies.
1649 get_task_struct(current->group_leader);
1650 dma->task = current->group_leader;
1651 dma->lock_cap = capable(CAP_IPC_LOCK);
1652 dma->mm = current->mm;
1655 dma->pfn_list = RB_ROOT;
1657 /* Insert zero-sized and grow as we map chunks of it */
1658 vfio_link_dma(iommu, dma);
1660 /* Don't pin and map if container doesn't contain IOMMU capable domain*/
1661 if (list_empty(&iommu->domain_list))
1664 ret = vfio_pin_map_dma(iommu, dma, size);
1666 if (!ret && iommu->dirty_page_tracking) {
1667 ret = vfio_dma_bitmap_alloc(dma, pgsize);
1669 vfio_remove_dma(iommu, dma);
1673 mutex_unlock(&iommu->lock);
1677 static int vfio_iommu_replay(struct vfio_iommu *iommu,
1678 struct vfio_domain *domain)
1680 struct vfio_batch batch;
1681 struct vfio_domain *d = NULL;
1683 unsigned long limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1686 /* Arbitrarily pick the first domain in the list for lookups */
1687 if (!list_empty(&iommu->domain_list))
1688 d = list_first_entry(&iommu->domain_list,
1689 struct vfio_domain, next);
1691 vfio_batch_init(&batch);
1693 n = rb_first(&iommu->dma_list);
1695 for (; n; n = rb_next(n)) {
1696 struct vfio_dma *dma;
1699 dma = rb_entry(n, struct vfio_dma, node);
1702 while (iova < dma->iova + dma->size) {
1706 if (dma->iommu_mapped) {
1710 if (WARN_ON(!d)) { /* mapped w/o a domain?! */
1715 phys = iommu_iova_to_phys(d->domain, iova);
1717 if (WARN_ON(!phys)) {
1725 while (i < dma->iova + dma->size &&
1726 p == iommu_iova_to_phys(d->domain, i)) {
1733 unsigned long vaddr = dma->vaddr +
1735 size_t n = dma->iova + dma->size - iova;
1738 npage = vfio_pin_pages_remote(dma, vaddr,
1748 phys = pfn << PAGE_SHIFT;
1749 size = npage << PAGE_SHIFT;
1752 ret = iommu_map(domain->domain, iova, phys, size,
1753 dma->prot | IOMMU_CACHE, GFP_KERNEL);
1755 if (!dma->iommu_mapped) {
1756 vfio_unpin_pages_remote(dma, iova,
1760 vfio_batch_unpin(&batch, dma);
1769 /* All dmas are now mapped, defer to second tree walk for unwind */
1770 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
1771 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1773 dma->iommu_mapped = true;
1776 vfio_batch_fini(&batch);
1780 for (; n; n = rb_prev(n)) {
1781 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1784 if (dma->iommu_mapped) {
1785 iommu_unmap(domain->domain, dma->iova, dma->size);
1790 while (iova < dma->iova + dma->size) {
1791 phys_addr_t phys, p;
1795 phys = iommu_iova_to_phys(domain->domain, iova);
1804 while (i < dma->iova + dma->size &&
1805 p == iommu_iova_to_phys(domain->domain, i)) {
1811 iommu_unmap(domain->domain, iova, size);
1812 vfio_unpin_pages_remote(dma, iova, phys >> PAGE_SHIFT,
1813 size >> PAGE_SHIFT, true);
1817 vfio_batch_fini(&batch);
1822 * We change our unmap behavior slightly depending on whether the IOMMU
1823 * supports fine-grained superpages. IOMMUs like AMD-Vi will use a superpage
1824 * for practically any contiguous power-of-two mapping we give it. This means
1825 * we don't need to look for contiguous chunks ourselves to make unmapping
1826 * more efficient. On IOMMUs with coarse-grained super pages, like Intel VT-d
1827 * with discrete 2M/1G/512G/1T superpages, identifying contiguous chunks
1828 * significantly boosts non-hugetlbfs mappings and doesn't seem to hurt when
1829 * hugetlbfs is in use.
1831 static void vfio_test_domain_fgsp(struct vfio_domain *domain, struct list_head *regions)
1833 int ret, order = get_order(PAGE_SIZE * 2);
1834 struct vfio_iova *region;
1838 pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order);
1842 list_for_each_entry(region, regions, list) {
1843 start = ALIGN(region->start, PAGE_SIZE * 2);
1844 if (start >= region->end || (region->end - start < PAGE_SIZE * 2))
1847 ret = iommu_map(domain->domain, start, page_to_phys(pages), PAGE_SIZE * 2,
1848 IOMMU_READ | IOMMU_WRITE | IOMMU_CACHE, GFP_KERNEL);
1850 size_t unmapped = iommu_unmap(domain->domain, start, PAGE_SIZE);
1852 if (unmapped == PAGE_SIZE)
1853 iommu_unmap(domain->domain, start + PAGE_SIZE, PAGE_SIZE);
1855 domain->fgsp = true;
1860 __free_pages(pages, order);
1863 static struct vfio_iommu_group *find_iommu_group(struct vfio_domain *domain,
1864 struct iommu_group *iommu_group)
1866 struct vfio_iommu_group *g;
1868 list_for_each_entry(g, &domain->group_list, next) {
1869 if (g->iommu_group == iommu_group)
1876 static struct vfio_iommu_group*
1877 vfio_iommu_find_iommu_group(struct vfio_iommu *iommu,
1878 struct iommu_group *iommu_group)
1880 struct vfio_iommu_group *group;
1881 struct vfio_domain *domain;
1883 list_for_each_entry(domain, &iommu->domain_list, next) {
1884 group = find_iommu_group(domain, iommu_group);
1889 list_for_each_entry(group, &iommu->emulated_iommu_groups, next)
1890 if (group->iommu_group == iommu_group)
1895 static bool vfio_iommu_has_sw_msi(struct list_head *group_resv_regions,
1898 struct iommu_resv_region *region;
1901 list_for_each_entry(region, group_resv_regions, list) {
1903 * The presence of any 'real' MSI regions should take
1904 * precedence over the software-managed one if the
1905 * IOMMU driver happens to advertise both types.
1907 if (region->type == IOMMU_RESV_MSI) {
1912 if (region->type == IOMMU_RESV_SW_MSI) {
1913 *base = region->start;
1922 * This is a helper function to insert an address range to iova list.
1923 * The list is initially created with a single entry corresponding to
1924 * the IOMMU domain geometry to which the device group is attached.
1925 * The list aperture gets modified when a new domain is added to the
1926 * container if the new aperture doesn't conflict with the current one
1927 * or with any existing dma mappings. The list is also modified to
1928 * exclude any reserved regions associated with the device group.
1930 static int vfio_iommu_iova_insert(struct list_head *head,
1931 dma_addr_t start, dma_addr_t end)
1933 struct vfio_iova *region;
1935 region = kmalloc(sizeof(*region), GFP_KERNEL);
1939 INIT_LIST_HEAD(®ion->list);
1940 region->start = start;
1943 list_add_tail(®ion->list, head);
1948 * Check the new iommu aperture conflicts with existing aper or with any
1949 * existing dma mappings.
1951 static bool vfio_iommu_aper_conflict(struct vfio_iommu *iommu,
1952 dma_addr_t start, dma_addr_t end)
1954 struct vfio_iova *first, *last;
1955 struct list_head *iova = &iommu->iova_list;
1957 if (list_empty(iova))
1960 /* Disjoint sets, return conflict */
1961 first = list_first_entry(iova, struct vfio_iova, list);
1962 last = list_last_entry(iova, struct vfio_iova, list);
1963 if (start > last->end || end < first->start)
1966 /* Check for any existing dma mappings below the new start */
1967 if (start > first->start) {
1968 if (vfio_find_dma(iommu, first->start, start - first->start))
1972 /* Check for any existing dma mappings beyond the new end */
1973 if (end < last->end) {
1974 if (vfio_find_dma(iommu, end + 1, last->end - end))
1982 * Resize iommu iova aperture window. This is called only if the new
1983 * aperture has no conflict with existing aperture and dma mappings.
1985 static int vfio_iommu_aper_resize(struct list_head *iova,
1986 dma_addr_t start, dma_addr_t end)
1988 struct vfio_iova *node, *next;
1990 if (list_empty(iova))
1991 return vfio_iommu_iova_insert(iova, start, end);
1993 /* Adjust iova list start */
1994 list_for_each_entry_safe(node, next, iova, list) {
1995 if (start < node->start)
1997 if (start >= node->start && start < node->end) {
1998 node->start = start;
2001 /* Delete nodes before new start */
2002 list_del(&node->list);
2006 /* Adjust iova list end */
2007 list_for_each_entry_safe(node, next, iova, list) {
2008 if (end > node->end)
2010 if (end > node->start && end <= node->end) {
2014 /* Delete nodes after new end */
2015 list_del(&node->list);
2023 * Check reserved region conflicts with existing dma mappings
2025 static bool vfio_iommu_resv_conflict(struct vfio_iommu *iommu,
2026 struct list_head *resv_regions)
2028 struct iommu_resv_region *region;
2030 /* Check for conflict with existing dma mappings */
2031 list_for_each_entry(region, resv_regions, list) {
2032 if (region->type == IOMMU_RESV_DIRECT_RELAXABLE)
2035 if (vfio_find_dma(iommu, region->start, region->length))
2043 * Check iova region overlap with reserved regions and
2044 * exclude them from the iommu iova range
2046 static int vfio_iommu_resv_exclude(struct list_head *iova,
2047 struct list_head *resv_regions)
2049 struct iommu_resv_region *resv;
2050 struct vfio_iova *n, *next;
2052 list_for_each_entry(resv, resv_regions, list) {
2053 phys_addr_t start, end;
2055 if (resv->type == IOMMU_RESV_DIRECT_RELAXABLE)
2058 start = resv->start;
2059 end = resv->start + resv->length - 1;
2061 list_for_each_entry_safe(n, next, iova, list) {
2065 if (start > n->end || end < n->start)
2068 * Insert a new node if current node overlaps with the
2069 * reserve region to exclude that from valid iova range.
2070 * Note that, new node is inserted before the current
2071 * node and finally the current node is deleted keeping
2072 * the list updated and sorted.
2074 if (start > n->start)
2075 ret = vfio_iommu_iova_insert(&n->list, n->start,
2077 if (!ret && end < n->end)
2078 ret = vfio_iommu_iova_insert(&n->list, end + 1,
2088 if (list_empty(iova))
2094 static void vfio_iommu_resv_free(struct list_head *resv_regions)
2096 struct iommu_resv_region *n, *next;
2098 list_for_each_entry_safe(n, next, resv_regions, list) {
2104 static void vfio_iommu_iova_free(struct list_head *iova)
2106 struct vfio_iova *n, *next;
2108 list_for_each_entry_safe(n, next, iova, list) {
2114 static int vfio_iommu_iova_get_copy(struct vfio_iommu *iommu,
2115 struct list_head *iova_copy)
2117 struct list_head *iova = &iommu->iova_list;
2118 struct vfio_iova *n;
2121 list_for_each_entry(n, iova, list) {
2122 ret = vfio_iommu_iova_insert(iova_copy, n->start, n->end);
2130 vfio_iommu_iova_free(iova_copy);
2134 static void vfio_iommu_iova_insert_copy(struct vfio_iommu *iommu,
2135 struct list_head *iova_copy)
2137 struct list_head *iova = &iommu->iova_list;
2139 vfio_iommu_iova_free(iova);
2141 list_splice_tail(iova_copy, iova);
2144 static int vfio_iommu_domain_alloc(struct device *dev, void *data)
2146 struct iommu_domain **domain = data;
2148 *domain = iommu_domain_alloc(dev->bus);
2149 return 1; /* Don't iterate */
2152 static int vfio_iommu_type1_attach_group(void *iommu_data,
2153 struct iommu_group *iommu_group, enum vfio_group_type type)
2155 struct vfio_iommu *iommu = iommu_data;
2156 struct vfio_iommu_group *group;
2157 struct vfio_domain *domain, *d;
2159 phys_addr_t resv_msi_base = 0;
2160 struct iommu_domain_geometry *geo;
2161 LIST_HEAD(iova_copy);
2162 LIST_HEAD(group_resv_regions);
2165 mutex_lock(&iommu->lock);
2167 /* Attach could require pinning, so disallow while vaddr is invalid. */
2168 if (iommu->vaddr_invalid_count)
2171 /* Check for duplicates */
2173 if (vfio_iommu_find_iommu_group(iommu, iommu_group))
2177 group = kzalloc(sizeof(*group), GFP_KERNEL);
2180 group->iommu_group = iommu_group;
2182 if (type == VFIO_EMULATED_IOMMU) {
2183 list_add(&group->next, &iommu->emulated_iommu_groups);
2185 * An emulated IOMMU group cannot dirty memory directly, it can
2186 * only use interfaces that provide dirty tracking.
2187 * The iommu scope can only be promoted with the addition of a
2188 * dirty tracking group.
2190 group->pinned_page_dirty_scope = true;
2196 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2198 goto out_free_group;
2201 * Going via the iommu_group iterator avoids races, and trivially gives
2202 * us a representative device for the IOMMU API call. We don't actually
2203 * want to iterate beyond the first device (if any).
2206 iommu_group_for_each_dev(iommu_group, &domain->domain,
2207 vfio_iommu_domain_alloc);
2208 if (!domain->domain)
2209 goto out_free_domain;
2211 if (iommu->nesting) {
2212 ret = iommu_enable_nesting(domain->domain);
2217 ret = iommu_attach_group(domain->domain, group->iommu_group);
2221 /* Get aperture info */
2222 geo = &domain->domain->geometry;
2223 if (vfio_iommu_aper_conflict(iommu, geo->aperture_start,
2224 geo->aperture_end)) {
2229 ret = iommu_get_group_resv_regions(iommu_group, &group_resv_regions);
2233 if (vfio_iommu_resv_conflict(iommu, &group_resv_regions)) {
2239 * We don't want to work on the original iova list as the list
2240 * gets modified and in case of failure we have to retain the
2241 * original list. Get a copy here.
2243 ret = vfio_iommu_iova_get_copy(iommu, &iova_copy);
2247 ret = vfio_iommu_aper_resize(&iova_copy, geo->aperture_start,
2252 ret = vfio_iommu_resv_exclude(&iova_copy, &group_resv_regions);
2256 resv_msi = vfio_iommu_has_sw_msi(&group_resv_regions, &resv_msi_base);
2258 INIT_LIST_HEAD(&domain->group_list);
2259 list_add(&group->next, &domain->group_list);
2261 if (!allow_unsafe_interrupts &&
2262 !iommu_group_has_isolated_msi(iommu_group)) {
2263 pr_warn("%s: No interrupt remapping support. Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n",
2270 * If the IOMMU can block non-coherent operations (ie PCIe TLPs with
2271 * no-snoop set) then VFIO always turns this feature on because on Intel
2272 * platforms it optimizes KVM to disable wbinvd emulation.
2274 if (domain->domain->ops->enforce_cache_coherency)
2275 domain->enforce_cache_coherency =
2276 domain->domain->ops->enforce_cache_coherency(
2280 * Try to match an existing compatible domain. We don't want to
2281 * preclude an IOMMU driver supporting multiple bus_types and being
2282 * able to include different bus_types in the same IOMMU domain, so
2283 * we test whether the domains use the same iommu_ops rather than
2284 * testing if they're on the same bus_type.
2286 list_for_each_entry(d, &iommu->domain_list, next) {
2287 if (d->domain->ops == domain->domain->ops &&
2288 d->enforce_cache_coherency ==
2289 domain->enforce_cache_coherency) {
2290 iommu_detach_group(domain->domain, group->iommu_group);
2291 if (!iommu_attach_group(d->domain,
2292 group->iommu_group)) {
2293 list_add(&group->next, &d->group_list);
2294 iommu_domain_free(domain->domain);
2299 ret = iommu_attach_group(domain->domain,
2300 group->iommu_group);
2306 vfio_test_domain_fgsp(domain, &iova_copy);
2308 /* replay mappings on new domains */
2309 ret = vfio_iommu_replay(iommu, domain);
2314 ret = iommu_get_msi_cookie(domain->domain, resv_msi_base);
2315 if (ret && ret != -ENODEV)
2319 list_add(&domain->next, &iommu->domain_list);
2320 vfio_update_pgsize_bitmap(iommu);
2322 /* Delete the old one and insert new iova list */
2323 vfio_iommu_iova_insert_copy(iommu, &iova_copy);
2326 * An iommu backed group can dirty memory directly and therefore
2327 * demotes the iommu scope until it declares itself dirty tracking
2328 * capable via the page pinning interface.
2330 iommu->num_non_pinned_groups++;
2331 mutex_unlock(&iommu->lock);
2332 vfio_iommu_resv_free(&group_resv_regions);
2337 iommu_detach_group(domain->domain, group->iommu_group);
2339 iommu_domain_free(domain->domain);
2340 vfio_iommu_iova_free(&iova_copy);
2341 vfio_iommu_resv_free(&group_resv_regions);
2347 mutex_unlock(&iommu->lock);
2351 static void vfio_iommu_unmap_unpin_all(struct vfio_iommu *iommu)
2353 struct rb_node *node;
2355 while ((node = rb_first(&iommu->dma_list)))
2356 vfio_remove_dma(iommu, rb_entry(node, struct vfio_dma, node));
2359 static void vfio_iommu_unmap_unpin_reaccount(struct vfio_iommu *iommu)
2361 struct rb_node *n, *p;
2363 n = rb_first(&iommu->dma_list);
2364 for (; n; n = rb_next(n)) {
2365 struct vfio_dma *dma;
2366 long locked = 0, unlocked = 0;
2368 dma = rb_entry(n, struct vfio_dma, node);
2369 unlocked += vfio_unmap_unpin(iommu, dma, false);
2370 p = rb_first(&dma->pfn_list);
2371 for (; p; p = rb_next(p)) {
2372 struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn,
2375 if (!is_invalid_reserved_pfn(vpfn->pfn))
2378 vfio_lock_acct(dma, locked - unlocked, true);
2383 * Called when a domain is removed in detach. It is possible that
2384 * the removed domain decided the iova aperture window. Modify the
2385 * iova aperture with the smallest window among existing domains.
2387 static void vfio_iommu_aper_expand(struct vfio_iommu *iommu,
2388 struct list_head *iova_copy)
2390 struct vfio_domain *domain;
2391 struct vfio_iova *node;
2392 dma_addr_t start = 0;
2393 dma_addr_t end = (dma_addr_t)~0;
2395 if (list_empty(iova_copy))
2398 list_for_each_entry(domain, &iommu->domain_list, next) {
2399 struct iommu_domain_geometry *geo = &domain->domain->geometry;
2401 if (geo->aperture_start > start)
2402 start = geo->aperture_start;
2403 if (geo->aperture_end < end)
2404 end = geo->aperture_end;
2407 /* Modify aperture limits. The new aper is either same or bigger */
2408 node = list_first_entry(iova_copy, struct vfio_iova, list);
2409 node->start = start;
2410 node = list_last_entry(iova_copy, struct vfio_iova, list);
2415 * Called when a group is detached. The reserved regions for that
2416 * group can be part of valid iova now. But since reserved regions
2417 * may be duplicated among groups, populate the iova valid regions
2420 static int vfio_iommu_resv_refresh(struct vfio_iommu *iommu,
2421 struct list_head *iova_copy)
2423 struct vfio_domain *d;
2424 struct vfio_iommu_group *g;
2425 struct vfio_iova *node;
2426 dma_addr_t start, end;
2427 LIST_HEAD(resv_regions);
2430 if (list_empty(iova_copy))
2433 list_for_each_entry(d, &iommu->domain_list, next) {
2434 list_for_each_entry(g, &d->group_list, next) {
2435 ret = iommu_get_group_resv_regions(g->iommu_group,
2442 node = list_first_entry(iova_copy, struct vfio_iova, list);
2443 start = node->start;
2444 node = list_last_entry(iova_copy, struct vfio_iova, list);
2447 /* purge the iova list and create new one */
2448 vfio_iommu_iova_free(iova_copy);
2450 ret = vfio_iommu_aper_resize(iova_copy, start, end);
2454 /* Exclude current reserved regions from iova ranges */
2455 ret = vfio_iommu_resv_exclude(iova_copy, &resv_regions);
2457 vfio_iommu_resv_free(&resv_regions);
2461 static void vfio_iommu_type1_detach_group(void *iommu_data,
2462 struct iommu_group *iommu_group)
2464 struct vfio_iommu *iommu = iommu_data;
2465 struct vfio_domain *domain;
2466 struct vfio_iommu_group *group;
2467 bool update_dirty_scope = false;
2468 LIST_HEAD(iova_copy);
2470 mutex_lock(&iommu->lock);
2471 list_for_each_entry(group, &iommu->emulated_iommu_groups, next) {
2472 if (group->iommu_group != iommu_group)
2474 update_dirty_scope = !group->pinned_page_dirty_scope;
2475 list_del(&group->next);
2478 if (list_empty(&iommu->emulated_iommu_groups) &&
2479 list_empty(&iommu->domain_list)) {
2480 WARN_ON(!list_empty(&iommu->device_list));
2481 vfio_iommu_unmap_unpin_all(iommu);
2483 goto detach_group_done;
2487 * Get a copy of iova list. This will be used to update
2488 * and to replace the current one later. Please note that
2489 * we will leave the original list as it is if update fails.
2491 vfio_iommu_iova_get_copy(iommu, &iova_copy);
2493 list_for_each_entry(domain, &iommu->domain_list, next) {
2494 group = find_iommu_group(domain, iommu_group);
2498 iommu_detach_group(domain->domain, group->iommu_group);
2499 update_dirty_scope = !group->pinned_page_dirty_scope;
2500 list_del(&group->next);
2503 * Group ownership provides privilege, if the group list is
2504 * empty, the domain goes away. If it's the last domain with
2505 * iommu and external domain doesn't exist, then all the
2506 * mappings go away too. If it's the last domain with iommu and
2507 * external domain exist, update accounting
2509 if (list_empty(&domain->group_list)) {
2510 if (list_is_singular(&iommu->domain_list)) {
2511 if (list_empty(&iommu->emulated_iommu_groups)) {
2512 WARN_ON(!list_empty(
2513 &iommu->device_list));
2514 vfio_iommu_unmap_unpin_all(iommu);
2516 vfio_iommu_unmap_unpin_reaccount(iommu);
2519 iommu_domain_free(domain->domain);
2520 list_del(&domain->next);
2522 vfio_iommu_aper_expand(iommu, &iova_copy);
2523 vfio_update_pgsize_bitmap(iommu);
2528 if (!vfio_iommu_resv_refresh(iommu, &iova_copy))
2529 vfio_iommu_iova_insert_copy(iommu, &iova_copy);
2531 vfio_iommu_iova_free(&iova_copy);
2535 * Removal of a group without dirty tracking may allow the iommu scope
2538 if (update_dirty_scope) {
2539 iommu->num_non_pinned_groups--;
2540 if (iommu->dirty_page_tracking)
2541 vfio_iommu_populate_bitmap_full(iommu);
2543 mutex_unlock(&iommu->lock);
2546 static void *vfio_iommu_type1_open(unsigned long arg)
2548 struct vfio_iommu *iommu;
2550 iommu = kzalloc(sizeof(*iommu), GFP_KERNEL);
2552 return ERR_PTR(-ENOMEM);
2555 case VFIO_TYPE1_IOMMU:
2557 case VFIO_TYPE1_NESTING_IOMMU:
2558 iommu->nesting = true;
2560 case VFIO_TYPE1v2_IOMMU:
2565 return ERR_PTR(-EINVAL);
2568 INIT_LIST_HEAD(&iommu->domain_list);
2569 INIT_LIST_HEAD(&iommu->iova_list);
2570 iommu->dma_list = RB_ROOT;
2571 iommu->dma_avail = dma_entry_limit;
2572 mutex_init(&iommu->lock);
2573 mutex_init(&iommu->device_list_lock);
2574 INIT_LIST_HEAD(&iommu->device_list);
2575 iommu->pgsize_bitmap = PAGE_MASK;
2576 INIT_LIST_HEAD(&iommu->emulated_iommu_groups);
2581 static void vfio_release_domain(struct vfio_domain *domain)
2583 struct vfio_iommu_group *group, *group_tmp;
2585 list_for_each_entry_safe(group, group_tmp,
2586 &domain->group_list, next) {
2587 iommu_detach_group(domain->domain, group->iommu_group);
2588 list_del(&group->next);
2592 iommu_domain_free(domain->domain);
2595 static void vfio_iommu_type1_release(void *iommu_data)
2597 struct vfio_iommu *iommu = iommu_data;
2598 struct vfio_domain *domain, *domain_tmp;
2599 struct vfio_iommu_group *group, *next_group;
2601 list_for_each_entry_safe(group, next_group,
2602 &iommu->emulated_iommu_groups, next) {
2603 list_del(&group->next);
2607 vfio_iommu_unmap_unpin_all(iommu);
2609 list_for_each_entry_safe(domain, domain_tmp,
2610 &iommu->domain_list, next) {
2611 vfio_release_domain(domain);
2612 list_del(&domain->next);
2616 vfio_iommu_iova_free(&iommu->iova_list);
2621 static int vfio_domains_have_enforce_cache_coherency(struct vfio_iommu *iommu)
2623 struct vfio_domain *domain;
2626 mutex_lock(&iommu->lock);
2627 list_for_each_entry(domain, &iommu->domain_list, next) {
2628 if (!(domain->enforce_cache_coherency)) {
2633 mutex_unlock(&iommu->lock);
2638 static bool vfio_iommu_has_emulated(struct vfio_iommu *iommu)
2642 mutex_lock(&iommu->lock);
2643 ret = !list_empty(&iommu->emulated_iommu_groups);
2644 mutex_unlock(&iommu->lock);
2648 static int vfio_iommu_type1_check_extension(struct vfio_iommu *iommu,
2652 case VFIO_TYPE1_IOMMU:
2653 case VFIO_TYPE1v2_IOMMU:
2654 case VFIO_TYPE1_NESTING_IOMMU:
2655 case VFIO_UNMAP_ALL:
2657 case VFIO_UPDATE_VADDR:
2659 * Disable this feature if mdevs are present. They cannot
2660 * safely pin/unpin/rw while vaddrs are being updated.
2662 return iommu && !vfio_iommu_has_emulated(iommu);
2663 case VFIO_DMA_CC_IOMMU:
2666 return vfio_domains_have_enforce_cache_coherency(iommu);
2672 static int vfio_iommu_iova_add_cap(struct vfio_info_cap *caps,
2673 struct vfio_iommu_type1_info_cap_iova_range *cap_iovas,
2676 struct vfio_info_cap_header *header;
2677 struct vfio_iommu_type1_info_cap_iova_range *iova_cap;
2679 header = vfio_info_cap_add(caps, size,
2680 VFIO_IOMMU_TYPE1_INFO_CAP_IOVA_RANGE, 1);
2682 return PTR_ERR(header);
2684 iova_cap = container_of(header,
2685 struct vfio_iommu_type1_info_cap_iova_range,
2687 iova_cap->nr_iovas = cap_iovas->nr_iovas;
2688 memcpy(iova_cap->iova_ranges, cap_iovas->iova_ranges,
2689 cap_iovas->nr_iovas * sizeof(*cap_iovas->iova_ranges));
2693 static int vfio_iommu_iova_build_caps(struct vfio_iommu *iommu,
2694 struct vfio_info_cap *caps)
2696 struct vfio_iommu_type1_info_cap_iova_range *cap_iovas;
2697 struct vfio_iova *iova;
2699 int iovas = 0, i = 0, ret;
2701 list_for_each_entry(iova, &iommu->iova_list, list)
2706 * Return 0 as a container with a single mdev device
2707 * will have an empty list
2712 size = struct_size(cap_iovas, iova_ranges, iovas);
2714 cap_iovas = kzalloc(size, GFP_KERNEL);
2718 cap_iovas->nr_iovas = iovas;
2720 list_for_each_entry(iova, &iommu->iova_list, list) {
2721 cap_iovas->iova_ranges[i].start = iova->start;
2722 cap_iovas->iova_ranges[i].end = iova->end;
2726 ret = vfio_iommu_iova_add_cap(caps, cap_iovas, size);
2732 static int vfio_iommu_migration_build_caps(struct vfio_iommu *iommu,
2733 struct vfio_info_cap *caps)
2735 struct vfio_iommu_type1_info_cap_migration cap_mig = {};
2737 cap_mig.header.id = VFIO_IOMMU_TYPE1_INFO_CAP_MIGRATION;
2738 cap_mig.header.version = 1;
2741 /* support minimum pgsize */
2742 cap_mig.pgsize_bitmap = (size_t)1 << __ffs(iommu->pgsize_bitmap);
2743 cap_mig.max_dirty_bitmap_size = DIRTY_BITMAP_SIZE_MAX;
2745 return vfio_info_add_capability(caps, &cap_mig.header, sizeof(cap_mig));
2748 static int vfio_iommu_dma_avail_build_caps(struct vfio_iommu *iommu,
2749 struct vfio_info_cap *caps)
2751 struct vfio_iommu_type1_info_dma_avail cap_dma_avail;
2753 cap_dma_avail.header.id = VFIO_IOMMU_TYPE1_INFO_DMA_AVAIL;
2754 cap_dma_avail.header.version = 1;
2756 cap_dma_avail.avail = iommu->dma_avail;
2758 return vfio_info_add_capability(caps, &cap_dma_avail.header,
2759 sizeof(cap_dma_avail));
2762 static int vfio_iommu_type1_get_info(struct vfio_iommu *iommu,
2765 struct vfio_iommu_type1_info info = {};
2766 unsigned long minsz;
2767 struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
2770 minsz = offsetofend(struct vfio_iommu_type1_info, iova_pgsizes);
2772 if (copy_from_user(&info, (void __user *)arg, minsz))
2775 if (info.argsz < minsz)
2778 minsz = min_t(size_t, info.argsz, sizeof(info));
2780 mutex_lock(&iommu->lock);
2781 info.flags = VFIO_IOMMU_INFO_PGSIZES;
2783 info.iova_pgsizes = iommu->pgsize_bitmap;
2785 ret = vfio_iommu_migration_build_caps(iommu, &caps);
2788 ret = vfio_iommu_dma_avail_build_caps(iommu, &caps);
2791 ret = vfio_iommu_iova_build_caps(iommu, &caps);
2793 mutex_unlock(&iommu->lock);
2799 info.flags |= VFIO_IOMMU_INFO_CAPS;
2801 if (info.argsz < sizeof(info) + caps.size) {
2802 info.argsz = sizeof(info) + caps.size;
2804 vfio_info_cap_shift(&caps, sizeof(info));
2805 if (copy_to_user((void __user *)arg +
2806 sizeof(info), caps.buf,
2811 info.cap_offset = sizeof(info);
2817 return copy_to_user((void __user *)arg, &info, minsz) ?
2821 static int vfio_iommu_type1_map_dma(struct vfio_iommu *iommu,
2824 struct vfio_iommu_type1_dma_map map;
2825 unsigned long minsz;
2826 uint32_t mask = VFIO_DMA_MAP_FLAG_READ | VFIO_DMA_MAP_FLAG_WRITE |
2827 VFIO_DMA_MAP_FLAG_VADDR;
2829 minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
2831 if (copy_from_user(&map, (void __user *)arg, minsz))
2834 if (map.argsz < minsz || map.flags & ~mask)
2837 return vfio_dma_do_map(iommu, &map);
2840 static int vfio_iommu_type1_unmap_dma(struct vfio_iommu *iommu,
2843 struct vfio_iommu_type1_dma_unmap unmap;
2844 struct vfio_bitmap bitmap = { 0 };
2845 uint32_t mask = VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP |
2846 VFIO_DMA_UNMAP_FLAG_VADDR |
2847 VFIO_DMA_UNMAP_FLAG_ALL;
2848 unsigned long minsz;
2851 minsz = offsetofend(struct vfio_iommu_type1_dma_unmap, size);
2853 if (copy_from_user(&unmap, (void __user *)arg, minsz))
2856 if (unmap.argsz < minsz || unmap.flags & ~mask)
2859 if ((unmap.flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) &&
2860 (unmap.flags & (VFIO_DMA_UNMAP_FLAG_ALL |
2861 VFIO_DMA_UNMAP_FLAG_VADDR)))
2864 if (unmap.flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) {
2865 unsigned long pgshift;
2867 if (unmap.argsz < (minsz + sizeof(bitmap)))
2870 if (copy_from_user(&bitmap,
2871 (void __user *)(arg + minsz),
2875 if (!access_ok((void __user *)bitmap.data, bitmap.size))
2878 pgshift = __ffs(bitmap.pgsize);
2879 ret = verify_bitmap_size(unmap.size >> pgshift,
2885 ret = vfio_dma_do_unmap(iommu, &unmap, &bitmap);
2889 return copy_to_user((void __user *)arg, &unmap, minsz) ?
2893 static int vfio_iommu_type1_dirty_pages(struct vfio_iommu *iommu,
2896 struct vfio_iommu_type1_dirty_bitmap dirty;
2897 uint32_t mask = VFIO_IOMMU_DIRTY_PAGES_FLAG_START |
2898 VFIO_IOMMU_DIRTY_PAGES_FLAG_STOP |
2899 VFIO_IOMMU_DIRTY_PAGES_FLAG_GET_BITMAP;
2900 unsigned long minsz;
2906 minsz = offsetofend(struct vfio_iommu_type1_dirty_bitmap, flags);
2908 if (copy_from_user(&dirty, (void __user *)arg, minsz))
2911 if (dirty.argsz < minsz || dirty.flags & ~mask)
2914 /* only one flag should be set at a time */
2915 if (__ffs(dirty.flags) != __fls(dirty.flags))
2918 if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_START) {
2921 mutex_lock(&iommu->lock);
2922 pgsize = 1 << __ffs(iommu->pgsize_bitmap);
2923 if (!iommu->dirty_page_tracking) {
2924 ret = vfio_dma_bitmap_alloc_all(iommu, pgsize);
2926 iommu->dirty_page_tracking = true;
2928 mutex_unlock(&iommu->lock);
2930 } else if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_STOP) {
2931 mutex_lock(&iommu->lock);
2932 if (iommu->dirty_page_tracking) {
2933 iommu->dirty_page_tracking = false;
2934 vfio_dma_bitmap_free_all(iommu);
2936 mutex_unlock(&iommu->lock);
2938 } else if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_GET_BITMAP) {
2939 struct vfio_iommu_type1_dirty_bitmap_get range;
2940 unsigned long pgshift;
2941 size_t data_size = dirty.argsz - minsz;
2942 size_t iommu_pgsize;
2944 if (!data_size || data_size < sizeof(range))
2947 if (copy_from_user(&range, (void __user *)(arg + minsz),
2951 if (range.iova + range.size < range.iova)
2953 if (!access_ok((void __user *)range.bitmap.data,
2957 pgshift = __ffs(range.bitmap.pgsize);
2958 ret = verify_bitmap_size(range.size >> pgshift,
2963 mutex_lock(&iommu->lock);
2965 iommu_pgsize = (size_t)1 << __ffs(iommu->pgsize_bitmap);
2967 /* allow only smallest supported pgsize */
2968 if (range.bitmap.pgsize != iommu_pgsize) {
2972 if (range.iova & (iommu_pgsize - 1)) {
2976 if (!range.size || range.size & (iommu_pgsize - 1)) {
2981 if (iommu->dirty_page_tracking)
2982 ret = vfio_iova_dirty_bitmap(range.bitmap.data,
2985 range.bitmap.pgsize);
2989 mutex_unlock(&iommu->lock);
2997 static long vfio_iommu_type1_ioctl(void *iommu_data,
2998 unsigned int cmd, unsigned long arg)
3000 struct vfio_iommu *iommu = iommu_data;
3003 case VFIO_CHECK_EXTENSION:
3004 return vfio_iommu_type1_check_extension(iommu, arg);
3005 case VFIO_IOMMU_GET_INFO:
3006 return vfio_iommu_type1_get_info(iommu, arg);
3007 case VFIO_IOMMU_MAP_DMA:
3008 return vfio_iommu_type1_map_dma(iommu, arg);
3009 case VFIO_IOMMU_UNMAP_DMA:
3010 return vfio_iommu_type1_unmap_dma(iommu, arg);
3011 case VFIO_IOMMU_DIRTY_PAGES:
3012 return vfio_iommu_type1_dirty_pages(iommu, arg);
3018 static void vfio_iommu_type1_register_device(void *iommu_data,
3019 struct vfio_device *vdev)
3021 struct vfio_iommu *iommu = iommu_data;
3023 if (!vdev->ops->dma_unmap)
3027 * list_empty(&iommu->device_list) is tested under the iommu->lock while
3028 * iteration for dma_unmap must be done under the device_list_lock.
3029 * Holding both locks here allows avoiding the device_list_lock in
3030 * several fast paths. See vfio_notify_dma_unmap()
3032 mutex_lock(&iommu->lock);
3033 mutex_lock(&iommu->device_list_lock);
3034 list_add(&vdev->iommu_entry, &iommu->device_list);
3035 mutex_unlock(&iommu->device_list_lock);
3036 mutex_unlock(&iommu->lock);
3039 static void vfio_iommu_type1_unregister_device(void *iommu_data,
3040 struct vfio_device *vdev)
3042 struct vfio_iommu *iommu = iommu_data;
3044 if (!vdev->ops->dma_unmap)
3047 mutex_lock(&iommu->lock);
3048 mutex_lock(&iommu->device_list_lock);
3049 list_del(&vdev->iommu_entry);
3050 mutex_unlock(&iommu->device_list_lock);
3051 mutex_unlock(&iommu->lock);
3054 static int vfio_iommu_type1_dma_rw_chunk(struct vfio_iommu *iommu,
3055 dma_addr_t user_iova, void *data,
3056 size_t count, bool write,
3059 struct mm_struct *mm;
3060 unsigned long vaddr;
3061 struct vfio_dma *dma;
3062 bool kthread = current->mm == NULL;
3067 dma = vfio_find_dma(iommu, user_iova, 1);
3071 if ((write && !(dma->prot & IOMMU_WRITE)) ||
3072 !(dma->prot & IOMMU_READ))
3076 if (!mmget_not_zero(mm))
3081 else if (current->mm != mm)
3084 offset = user_iova - dma->iova;
3086 if (count > dma->size - offset)
3087 count = dma->size - offset;
3089 vaddr = dma->vaddr + offset;
3092 *copied = copy_to_user((void __user *)vaddr, data,
3094 if (*copied && iommu->dirty_page_tracking) {
3095 unsigned long pgshift = __ffs(iommu->pgsize_bitmap);
3097 * Bitmap populated with the smallest supported page
3100 bitmap_set(dma->bitmap, offset >> pgshift,
3101 ((offset + *copied - 1) >> pgshift) -
3102 (offset >> pgshift) + 1);
3105 *copied = copy_from_user(data, (void __user *)vaddr,
3108 kthread_unuse_mm(mm);
3111 return *copied ? 0 : -EFAULT;
3114 static int vfio_iommu_type1_dma_rw(void *iommu_data, dma_addr_t user_iova,
3115 void *data, size_t count, bool write)
3117 struct vfio_iommu *iommu = iommu_data;
3121 mutex_lock(&iommu->lock);
3123 if (WARN_ONCE(iommu->vaddr_invalid_count,
3124 "vfio_dma_rw not allowed with VFIO_UPDATE_VADDR\n")) {
3130 ret = vfio_iommu_type1_dma_rw_chunk(iommu, user_iova, data,
3131 count, write, &done);
3141 mutex_unlock(&iommu->lock);
3145 static struct iommu_domain *
3146 vfio_iommu_type1_group_iommu_domain(void *iommu_data,
3147 struct iommu_group *iommu_group)
3149 struct iommu_domain *domain = ERR_PTR(-ENODEV);
3150 struct vfio_iommu *iommu = iommu_data;
3151 struct vfio_domain *d;
3153 if (!iommu || !iommu_group)
3154 return ERR_PTR(-EINVAL);
3156 mutex_lock(&iommu->lock);
3157 list_for_each_entry(d, &iommu->domain_list, next) {
3158 if (find_iommu_group(d, iommu_group)) {
3163 mutex_unlock(&iommu->lock);
3168 static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = {
3169 .name = "vfio-iommu-type1",
3170 .owner = THIS_MODULE,
3171 .open = vfio_iommu_type1_open,
3172 .release = vfio_iommu_type1_release,
3173 .ioctl = vfio_iommu_type1_ioctl,
3174 .attach_group = vfio_iommu_type1_attach_group,
3175 .detach_group = vfio_iommu_type1_detach_group,
3176 .pin_pages = vfio_iommu_type1_pin_pages,
3177 .unpin_pages = vfio_iommu_type1_unpin_pages,
3178 .register_device = vfio_iommu_type1_register_device,
3179 .unregister_device = vfio_iommu_type1_unregister_device,
3180 .dma_rw = vfio_iommu_type1_dma_rw,
3181 .group_iommu_domain = vfio_iommu_type1_group_iommu_domain,
3184 static int __init vfio_iommu_type1_init(void)
3186 return vfio_register_iommu_driver(&vfio_iommu_driver_ops_type1);
3189 static void __exit vfio_iommu_type1_cleanup(void)
3191 vfio_unregister_iommu_driver(&vfio_iommu_driver_ops_type1);
3194 module_init(vfio_iommu_type1_init);
3195 module_exit(vfio_iommu_type1_cleanup);
3197 MODULE_VERSION(DRIVER_VERSION);
3198 MODULE_LICENSE("GPL v2");
3199 MODULE_AUTHOR(DRIVER_AUTHOR);
3200 MODULE_DESCRIPTION(DRIVER_DESC);