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,
570 *pfn = page_to_pfn(pages[0]);
574 vaddr = untagged_addr_remote(mm, vaddr);
577 vma = vma_lookup(mm, vaddr);
579 if (vma && vma->vm_flags & VM_PFNMAP) {
580 ret = follow_fault_pfn(vma, mm, vaddr, pfn, prot & IOMMU_WRITE);
585 if (is_invalid_reserved_pfn(*pfn))
592 mmap_read_unlock(mm);
597 * Attempt to pin pages. We really don't want to track all the pfns and
598 * the iommu can only map chunks of consecutive pfns anyway, so get the
599 * first page and all consecutive pages with the same locking.
601 static long vfio_pin_pages_remote(struct vfio_dma *dma, unsigned long vaddr,
602 long npage, unsigned long *pfn_base,
603 unsigned long limit, struct vfio_batch *batch)
606 struct mm_struct *mm = current->mm;
607 long ret, pinned = 0, lock_acct = 0;
609 dma_addr_t iova = vaddr - dma->vaddr + dma->iova;
611 /* This code path is only user initiated */
616 /* Leftover pages in batch from an earlier call. */
617 *pfn_base = page_to_pfn(batch->pages[batch->offset]);
619 rsvd = is_invalid_reserved_pfn(*pfn_base);
626 /* Empty batch, so refill it. */
627 long req_pages = min_t(long, npage, batch->capacity);
629 ret = vaddr_get_pfns(mm, vaddr, req_pages, dma->prot,
639 rsvd = is_invalid_reserved_pfn(*pfn_base);
644 * pfn is preset for the first iteration of this inner loop and
645 * updated at the end to handle a VM_PFNMAP pfn. In that case,
646 * batch->pages isn't valid (there's no struct page), so allow
647 * batch->pages to be touched only when there's more than one
648 * pfn to check, which guarantees the pfns are from a
652 if (pfn != *pfn_base + pinned ||
653 rsvd != is_invalid_reserved_pfn(pfn))
657 * Reserved pages aren't counted against the user,
658 * externally pinned pages are already counted against
661 if (!rsvd && !vfio_find_vpfn(dma, iova)) {
662 if (!dma->lock_cap &&
663 mm->locked_vm + lock_acct + 1 > limit) {
664 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n",
665 __func__, limit << PAGE_SHIFT);
682 pfn = page_to_pfn(batch->pages[batch->offset]);
685 if (unlikely(disable_hugepages))
690 ret = vfio_lock_acct(dma, lock_acct, false);
693 if (batch->size == 1 && !batch->offset) {
694 /* May be a VM_PFNMAP pfn, which the batch can't remember. */
695 put_pfn(pfn, dma->prot);
700 if (pinned && !rsvd) {
701 for (pfn = *pfn_base ; pinned ; pfn++, pinned--)
702 put_pfn(pfn, dma->prot);
704 vfio_batch_unpin(batch, dma);
712 static long vfio_unpin_pages_remote(struct vfio_dma *dma, dma_addr_t iova,
713 unsigned long pfn, long npage,
716 long unlocked = 0, locked = 0;
719 for (i = 0; i < npage; i++, iova += PAGE_SIZE) {
720 if (put_pfn(pfn++, dma->prot)) {
722 if (vfio_find_vpfn(dma, iova))
728 vfio_lock_acct(dma, locked - unlocked, true);
733 static int vfio_pin_page_external(struct vfio_dma *dma, unsigned long vaddr,
734 unsigned long *pfn_base, bool do_accounting)
736 struct page *pages[1];
737 struct mm_struct *mm;
741 if (!mmget_not_zero(mm))
744 ret = vaddr_get_pfns(mm, vaddr, 1, dma->prot, pfn_base, pages);
750 if (do_accounting && !is_invalid_reserved_pfn(*pfn_base)) {
751 ret = vfio_lock_acct(dma, 1, false);
753 put_pfn(*pfn_base, dma->prot);
755 pr_warn("%s: Task %s (%d) RLIMIT_MEMLOCK "
756 "(%ld) exceeded\n", __func__,
757 dma->task->comm, task_pid_nr(dma->task),
758 task_rlimit(dma->task, RLIMIT_MEMLOCK));
767 static int vfio_unpin_page_external(struct vfio_dma *dma, dma_addr_t iova,
771 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
776 unlocked = vfio_iova_put_vfio_pfn(dma, vpfn);
779 vfio_lock_acct(dma, -unlocked, true);
784 static int vfio_iommu_type1_pin_pages(void *iommu_data,
785 struct iommu_group *iommu_group,
786 dma_addr_t user_iova,
790 struct vfio_iommu *iommu = iommu_data;
791 struct vfio_iommu_group *group;
793 unsigned long remote_vaddr;
794 struct vfio_dma *dma;
797 if (!iommu || !pages)
800 /* Supported for v2 version only */
804 mutex_lock(&iommu->lock);
806 if (WARN_ONCE(iommu->vaddr_invalid_count,
807 "vfio_pin_pages not allowed with VFIO_UPDATE_VADDR\n")) {
812 /* Fail if no dma_umap notifier is registered */
813 if (list_empty(&iommu->device_list)) {
819 * If iommu capable domain exist in the container then all pages are
820 * already pinned and accounted. Accounting should be done if there is no
821 * iommu capable domain in the container.
823 do_accounting = list_empty(&iommu->domain_list);
825 for (i = 0; i < npage; i++) {
826 unsigned long phys_pfn;
828 struct vfio_pfn *vpfn;
830 iova = user_iova + PAGE_SIZE * i;
831 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
837 if ((dma->prot & prot) != prot) {
842 vpfn = vfio_iova_get_vfio_pfn(dma, iova);
844 pages[i] = pfn_to_page(vpfn->pfn);
848 remote_vaddr = dma->vaddr + (iova - dma->iova);
849 ret = vfio_pin_page_external(dma, remote_vaddr, &phys_pfn,
854 if (!pfn_valid(phys_pfn)) {
859 ret = vfio_add_to_pfn_list(dma, iova, phys_pfn);
861 if (put_pfn(phys_pfn, dma->prot) && do_accounting)
862 vfio_lock_acct(dma, -1, true);
866 pages[i] = pfn_to_page(phys_pfn);
868 if (iommu->dirty_page_tracking) {
869 unsigned long pgshift = __ffs(iommu->pgsize_bitmap);
872 * Bitmap populated with the smallest supported page
875 bitmap_set(dma->bitmap,
876 (iova - dma->iova) >> pgshift, 1);
881 group = vfio_iommu_find_iommu_group(iommu, iommu_group);
882 if (!group->pinned_page_dirty_scope) {
883 group->pinned_page_dirty_scope = true;
884 iommu->num_non_pinned_groups--;
891 for (j = 0; j < i; j++) {
894 iova = user_iova + PAGE_SIZE * j;
895 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
896 vfio_unpin_page_external(dma, iova, do_accounting);
900 mutex_unlock(&iommu->lock);
904 static void vfio_iommu_type1_unpin_pages(void *iommu_data,
905 dma_addr_t user_iova, int npage)
907 struct vfio_iommu *iommu = iommu_data;
911 /* Supported for v2 version only */
912 if (WARN_ON(!iommu->v2))
915 mutex_lock(&iommu->lock);
917 do_accounting = list_empty(&iommu->domain_list);
918 for (i = 0; i < npage; i++) {
919 dma_addr_t iova = user_iova + PAGE_SIZE * i;
920 struct vfio_dma *dma;
922 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
926 vfio_unpin_page_external(dma, iova, do_accounting);
929 mutex_unlock(&iommu->lock);
934 static long vfio_sync_unpin(struct vfio_dma *dma, struct vfio_domain *domain,
935 struct list_head *regions,
936 struct iommu_iotlb_gather *iotlb_gather)
939 struct vfio_regions *entry, *next;
941 iommu_iotlb_sync(domain->domain, iotlb_gather);
943 list_for_each_entry_safe(entry, next, regions, list) {
944 unlocked += vfio_unpin_pages_remote(dma,
946 entry->phys >> PAGE_SHIFT,
947 entry->len >> PAGE_SHIFT,
949 list_del(&entry->list);
959 * Generally, VFIO needs to unpin remote pages after each IOTLB flush.
960 * Therefore, when using IOTLB flush sync interface, VFIO need to keep track
961 * of these regions (currently using a list).
963 * This value specifies maximum number of regions for each IOTLB flush sync.
965 #define VFIO_IOMMU_TLB_SYNC_MAX 512
967 static size_t unmap_unpin_fast(struct vfio_domain *domain,
968 struct vfio_dma *dma, dma_addr_t *iova,
969 size_t len, phys_addr_t phys, long *unlocked,
970 struct list_head *unmapped_list,
972 struct iommu_iotlb_gather *iotlb_gather)
975 struct vfio_regions *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
978 unmapped = iommu_unmap_fast(domain->domain, *iova, len,
986 entry->len = unmapped;
987 list_add_tail(&entry->list, unmapped_list);
995 * Sync if the number of fast-unmap regions hits the limit
996 * or in case of errors.
998 if (*unmapped_cnt >= VFIO_IOMMU_TLB_SYNC_MAX || !unmapped) {
999 *unlocked += vfio_sync_unpin(dma, domain, unmapped_list,
1007 static size_t unmap_unpin_slow(struct vfio_domain *domain,
1008 struct vfio_dma *dma, dma_addr_t *iova,
1009 size_t len, phys_addr_t phys,
1012 size_t unmapped = iommu_unmap(domain->domain, *iova, len);
1015 *unlocked += vfio_unpin_pages_remote(dma, *iova,
1017 unmapped >> PAGE_SHIFT,
1025 static long vfio_unmap_unpin(struct vfio_iommu *iommu, struct vfio_dma *dma,
1028 dma_addr_t iova = dma->iova, end = dma->iova + dma->size;
1029 struct vfio_domain *domain, *d;
1030 LIST_HEAD(unmapped_region_list);
1031 struct iommu_iotlb_gather iotlb_gather;
1032 int unmapped_region_cnt = 0;
1038 if (list_empty(&iommu->domain_list))
1042 * We use the IOMMU to track the physical addresses, otherwise we'd
1043 * need a much more complicated tracking system. Unfortunately that
1044 * means we need to use one of the iommu domains to figure out the
1045 * pfns to unpin. The rest need to be unmapped in advance so we have
1046 * no iommu translations remaining when the pages are unpinned.
1048 domain = d = list_first_entry(&iommu->domain_list,
1049 struct vfio_domain, next);
1051 list_for_each_entry_continue(d, &iommu->domain_list, next) {
1052 iommu_unmap(d->domain, dma->iova, dma->size);
1056 iommu_iotlb_gather_init(&iotlb_gather);
1057 while (iova < end) {
1058 size_t unmapped, len;
1059 phys_addr_t phys, next;
1061 phys = iommu_iova_to_phys(domain->domain, iova);
1062 if (WARN_ON(!phys)) {
1068 * To optimize for fewer iommu_unmap() calls, each of which
1069 * may require hardware cache flushing, try to find the
1070 * largest contiguous physical memory chunk to unmap.
1072 for (len = PAGE_SIZE;
1073 !domain->fgsp && iova + len < end; len += PAGE_SIZE) {
1074 next = iommu_iova_to_phys(domain->domain, iova + len);
1075 if (next != phys + len)
1080 * First, try to use fast unmap/unpin. In case of failure,
1081 * switch to slow unmap/unpin path.
1083 unmapped = unmap_unpin_fast(domain, dma, &iova, len, phys,
1084 &unlocked, &unmapped_region_list,
1085 &unmapped_region_cnt,
1088 unmapped = unmap_unpin_slow(domain, dma, &iova, len,
1090 if (WARN_ON(!unmapped))
1095 dma->iommu_mapped = false;
1097 if (unmapped_region_cnt) {
1098 unlocked += vfio_sync_unpin(dma, domain, &unmapped_region_list,
1102 if (do_accounting) {
1103 vfio_lock_acct(dma, -unlocked, true);
1109 static void vfio_remove_dma(struct vfio_iommu *iommu, struct vfio_dma *dma)
1111 WARN_ON(!RB_EMPTY_ROOT(&dma->pfn_list));
1112 vfio_unmap_unpin(iommu, dma, true);
1113 vfio_unlink_dma(iommu, dma);
1114 put_task_struct(dma->task);
1116 vfio_dma_bitmap_free(dma);
1117 if (dma->vaddr_invalid)
1118 iommu->vaddr_invalid_count--;
1123 static void vfio_update_pgsize_bitmap(struct vfio_iommu *iommu)
1125 struct vfio_domain *domain;
1127 iommu->pgsize_bitmap = ULONG_MAX;
1129 list_for_each_entry(domain, &iommu->domain_list, next)
1130 iommu->pgsize_bitmap &= domain->domain->pgsize_bitmap;
1133 * In case the IOMMU supports page sizes smaller than PAGE_SIZE
1134 * we pretend PAGE_SIZE is supported and hide sub-PAGE_SIZE sizes.
1135 * That way the user will be able to map/unmap buffers whose size/
1136 * start address is aligned with PAGE_SIZE. Pinning code uses that
1137 * granularity while iommu driver can use the sub-PAGE_SIZE size
1138 * to map the buffer.
1140 if (iommu->pgsize_bitmap & ~PAGE_MASK) {
1141 iommu->pgsize_bitmap &= PAGE_MASK;
1142 iommu->pgsize_bitmap |= PAGE_SIZE;
1146 static int update_user_bitmap(u64 __user *bitmap, struct vfio_iommu *iommu,
1147 struct vfio_dma *dma, dma_addr_t base_iova,
1150 unsigned long pgshift = __ffs(pgsize);
1151 unsigned long nbits = dma->size >> pgshift;
1152 unsigned long bit_offset = (dma->iova - base_iova) >> pgshift;
1153 unsigned long copy_offset = bit_offset / BITS_PER_LONG;
1154 unsigned long shift = bit_offset % BITS_PER_LONG;
1155 unsigned long leftover;
1158 * mark all pages dirty if any IOMMU capable device is not able
1159 * to report dirty pages and all pages are pinned and mapped.
1161 if (iommu->num_non_pinned_groups && dma->iommu_mapped)
1162 bitmap_set(dma->bitmap, 0, nbits);
1165 bitmap_shift_left(dma->bitmap, dma->bitmap, shift,
1168 if (copy_from_user(&leftover,
1169 (void __user *)(bitmap + copy_offset),
1173 bitmap_or(dma->bitmap, dma->bitmap, &leftover, shift);
1176 if (copy_to_user((void __user *)(bitmap + copy_offset), dma->bitmap,
1177 DIRTY_BITMAP_BYTES(nbits + shift)))
1183 static int vfio_iova_dirty_bitmap(u64 __user *bitmap, struct vfio_iommu *iommu,
1184 dma_addr_t iova, size_t size, size_t pgsize)
1186 struct vfio_dma *dma;
1188 unsigned long pgshift = __ffs(pgsize);
1192 * GET_BITMAP request must fully cover vfio_dma mappings. Multiple
1193 * vfio_dma mappings may be clubbed by specifying large ranges, but
1194 * there must not be any previous mappings bisected by the range.
1195 * An error will be returned if these conditions are not met.
1197 dma = vfio_find_dma(iommu, iova, 1);
1198 if (dma && dma->iova != iova)
1201 dma = vfio_find_dma(iommu, iova + size - 1, 0);
1202 if (dma && dma->iova + dma->size != iova + size)
1205 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
1206 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1208 if (dma->iova < iova)
1211 if (dma->iova > iova + size - 1)
1214 ret = update_user_bitmap(bitmap, iommu, dma, iova, pgsize);
1219 * Re-populate bitmap to include all pinned pages which are
1220 * considered as dirty but exclude pages which are unpinned and
1221 * pages which are marked dirty by vfio_dma_rw()
1223 bitmap_clear(dma->bitmap, 0, dma->size >> pgshift);
1224 vfio_dma_populate_bitmap(dma, pgsize);
1229 static int verify_bitmap_size(uint64_t npages, uint64_t bitmap_size)
1231 if (!npages || !bitmap_size || (bitmap_size > DIRTY_BITMAP_SIZE_MAX) ||
1232 (bitmap_size < DIRTY_BITMAP_BYTES(npages)))
1239 * Notify VFIO drivers using vfio_register_emulated_iommu_dev() to invalidate
1240 * and unmap iovas within the range we're about to unmap. Drivers MUST unpin
1241 * pages in response to an invalidation.
1243 static void vfio_notify_dma_unmap(struct vfio_iommu *iommu,
1244 struct vfio_dma *dma)
1246 struct vfio_device *device;
1248 if (list_empty(&iommu->device_list))
1252 * The device is expected to call vfio_unpin_pages() for any IOVA it has
1253 * pinned within the range. Since vfio_unpin_pages() will eventually
1254 * call back down to this code and try to obtain the iommu->lock we must
1257 mutex_lock(&iommu->device_list_lock);
1258 mutex_unlock(&iommu->lock);
1260 list_for_each_entry(device, &iommu->device_list, iommu_entry)
1261 device->ops->dma_unmap(device, dma->iova, dma->size);
1263 mutex_unlock(&iommu->device_list_lock);
1264 mutex_lock(&iommu->lock);
1267 static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
1268 struct vfio_iommu_type1_dma_unmap *unmap,
1269 struct vfio_bitmap *bitmap)
1271 struct vfio_dma *dma, *dma_last = NULL;
1272 size_t unmapped = 0, pgsize;
1273 int ret = -EINVAL, retries = 0;
1274 unsigned long pgshift;
1275 dma_addr_t iova = unmap->iova;
1276 u64 size = unmap->size;
1277 bool unmap_all = unmap->flags & VFIO_DMA_UNMAP_FLAG_ALL;
1278 bool invalidate_vaddr = unmap->flags & VFIO_DMA_UNMAP_FLAG_VADDR;
1279 struct rb_node *n, *first_n;
1281 mutex_lock(&iommu->lock);
1283 /* Cannot update vaddr if mdev is present. */
1284 if (invalidate_vaddr && !list_empty(&iommu->emulated_iommu_groups)) {
1289 pgshift = __ffs(iommu->pgsize_bitmap);
1290 pgsize = (size_t)1 << pgshift;
1292 if (iova & (pgsize - 1))
1299 } else if (!size || size & (pgsize - 1) ||
1300 iova + size - 1 < iova || size > SIZE_MAX) {
1304 /* When dirty tracking is enabled, allow only min supported pgsize */
1305 if ((unmap->flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) &&
1306 (!iommu->dirty_page_tracking || (bitmap->pgsize != pgsize))) {
1310 WARN_ON((pgsize - 1) & PAGE_MASK);
1313 * vfio-iommu-type1 (v1) - User mappings were coalesced together to
1314 * avoid tracking individual mappings. This means that the granularity
1315 * of the original mapping was lost and the user was allowed to attempt
1316 * to unmap any range. Depending on the contiguousness of physical
1317 * memory and page sizes supported by the IOMMU, arbitrary unmaps may
1318 * or may not have worked. We only guaranteed unmap granularity
1319 * matching the original mapping; even though it was untracked here,
1320 * the original mappings are reflected in IOMMU mappings. This
1321 * resulted in a couple unusual behaviors. First, if a range is not
1322 * able to be unmapped, ex. a set of 4k pages that was mapped as a
1323 * 2M hugepage into the IOMMU, the unmap ioctl returns success but with
1324 * a zero sized unmap. Also, if an unmap request overlaps the first
1325 * address of a hugepage, the IOMMU will unmap the entire hugepage.
1326 * This also returns success and the returned unmap size reflects the
1327 * actual size unmapped.
1329 * We attempt to maintain compatibility with this "v1" interface, but
1330 * we take control out of the hands of the IOMMU. Therefore, an unmap
1331 * request offset from the beginning of the original mapping will
1332 * return success with zero sized unmap. And an unmap request covering
1333 * the first iova of mapping will unmap the entire range.
1335 * The v2 version of this interface intends to be more deterministic.
1336 * Unmap requests must fully cover previous mappings. Multiple
1337 * mappings may still be unmaped by specifying large ranges, but there
1338 * must not be any previous mappings bisected by the range. An error
1339 * will be returned if these conditions are not met. The v2 interface
1340 * will only return success and a size of zero if there were no
1341 * mappings within the range.
1343 if (iommu->v2 && !unmap_all) {
1344 dma = vfio_find_dma(iommu, iova, 1);
1345 if (dma && dma->iova != iova)
1348 dma = vfio_find_dma(iommu, iova + size - 1, 0);
1349 if (dma && dma->iova + dma->size != iova + size)
1354 n = first_n = vfio_find_dma_first_node(iommu, iova, size);
1357 dma = rb_entry(n, struct vfio_dma, node);
1358 if (dma->iova >= iova + size)
1361 if (!iommu->v2 && iova > dma->iova)
1364 if (invalidate_vaddr) {
1365 if (dma->vaddr_invalid) {
1366 struct rb_node *last_n = n;
1368 for (n = first_n; n != last_n; n = rb_next(n)) {
1370 struct vfio_dma, node);
1371 dma->vaddr_invalid = false;
1372 iommu->vaddr_invalid_count--;
1378 dma->vaddr_invalid = true;
1379 iommu->vaddr_invalid_count++;
1380 unmapped += dma->size;
1385 if (!RB_EMPTY_ROOT(&dma->pfn_list)) {
1386 if (dma_last == dma) {
1387 BUG_ON(++retries > 10);
1393 vfio_notify_dma_unmap(iommu, dma);
1397 if (unmap->flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) {
1398 ret = update_user_bitmap(bitmap->data, iommu, dma,
1404 unmapped += dma->size;
1406 vfio_remove_dma(iommu, dma);
1410 mutex_unlock(&iommu->lock);
1412 /* Report how much was unmapped */
1413 unmap->size = unmapped;
1418 static int vfio_iommu_map(struct vfio_iommu *iommu, dma_addr_t iova,
1419 unsigned long pfn, long npage, int prot)
1421 struct vfio_domain *d;
1424 list_for_each_entry(d, &iommu->domain_list, next) {
1425 ret = iommu_map(d->domain, iova, (phys_addr_t)pfn << PAGE_SHIFT,
1426 npage << PAGE_SHIFT, prot | IOMMU_CACHE,
1427 GFP_KERNEL_ACCOUNT);
1437 list_for_each_entry_continue_reverse(d, &iommu->domain_list, next) {
1438 iommu_unmap(d->domain, iova, npage << PAGE_SHIFT);
1445 static int vfio_pin_map_dma(struct vfio_iommu *iommu, struct vfio_dma *dma,
1448 dma_addr_t iova = dma->iova;
1449 unsigned long vaddr = dma->vaddr;
1450 struct vfio_batch batch;
1451 size_t size = map_size;
1453 unsigned long pfn, limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1456 vfio_batch_init(&batch);
1459 /* Pin a contiguous chunk of memory */
1460 npage = vfio_pin_pages_remote(dma, vaddr + dma->size,
1461 size >> PAGE_SHIFT, &pfn, limit,
1470 ret = vfio_iommu_map(iommu, iova + dma->size, pfn, npage,
1473 vfio_unpin_pages_remote(dma, iova + dma->size, pfn,
1475 vfio_batch_unpin(&batch, dma);
1479 size -= npage << PAGE_SHIFT;
1480 dma->size += npage << PAGE_SHIFT;
1483 vfio_batch_fini(&batch);
1484 dma->iommu_mapped = true;
1487 vfio_remove_dma(iommu, dma);
1493 * Check dma map request is within a valid iova range
1495 static bool vfio_iommu_iova_dma_valid(struct vfio_iommu *iommu,
1496 dma_addr_t start, dma_addr_t end)
1498 struct list_head *iova = &iommu->iova_list;
1499 struct vfio_iova *node;
1501 list_for_each_entry(node, iova, list) {
1502 if (start >= node->start && end <= node->end)
1507 * Check for list_empty() as well since a container with
1508 * a single mdev device will have an empty list.
1510 return list_empty(iova);
1513 static int vfio_change_dma_owner(struct vfio_dma *dma)
1515 struct task_struct *task = current->group_leader;
1516 struct mm_struct *mm = current->mm;
1517 long npage = dma->locked_vm;
1524 lock_cap = capable(CAP_IPC_LOCK);
1525 ret = mm_lock_acct(task, mm, lock_cap, npage);
1529 if (mmget_not_zero(dma->mm)) {
1530 mm_lock_acct(dma->task, dma->mm, dma->lock_cap, -npage);
1534 if (dma->task != task) {
1535 put_task_struct(dma->task);
1536 dma->task = get_task_struct(task);
1541 dma->lock_cap = lock_cap;
1545 static int vfio_dma_do_map(struct vfio_iommu *iommu,
1546 struct vfio_iommu_type1_dma_map *map)
1548 bool set_vaddr = map->flags & VFIO_DMA_MAP_FLAG_VADDR;
1549 dma_addr_t iova = map->iova;
1550 unsigned long vaddr = map->vaddr;
1551 size_t size = map->size;
1552 int ret = 0, prot = 0;
1554 struct vfio_dma *dma;
1556 /* Verify that none of our __u64 fields overflow */
1557 if (map->size != size || map->vaddr != vaddr || map->iova != iova)
1560 /* READ/WRITE from device perspective */
1561 if (map->flags & VFIO_DMA_MAP_FLAG_WRITE)
1562 prot |= IOMMU_WRITE;
1563 if (map->flags & VFIO_DMA_MAP_FLAG_READ)
1566 if ((prot && set_vaddr) || (!prot && !set_vaddr))
1569 mutex_lock(&iommu->lock);
1571 pgsize = (size_t)1 << __ffs(iommu->pgsize_bitmap);
1573 WARN_ON((pgsize - 1) & PAGE_MASK);
1575 if (!size || (size | iova | vaddr) & (pgsize - 1)) {
1580 /* Don't allow IOVA or virtual address wrap */
1581 if (iova + size - 1 < iova || vaddr + size - 1 < vaddr) {
1586 dma = vfio_find_dma(iommu, iova, size);
1590 } else if (!dma->vaddr_invalid || dma->iova != iova ||
1591 dma->size != size) {
1594 ret = vfio_change_dma_owner(dma);
1598 dma->vaddr_invalid = false;
1599 iommu->vaddr_invalid_count--;
1607 if (!iommu->dma_avail) {
1612 if (!vfio_iommu_iova_dma_valid(iommu, iova, iova + size - 1)) {
1617 dma = kzalloc(sizeof(*dma), GFP_KERNEL);
1629 * We need to be able to both add to a task's locked memory and test
1630 * against the locked memory limit and we need to be able to do both
1631 * outside of this call path as pinning can be asynchronous via the
1632 * external interfaces for mdev devices. RLIMIT_MEMLOCK requires a
1633 * task_struct. Save the group_leader so that all DMA tracking uses
1634 * the same task, to make debugging easier. VM locked pages requires
1635 * an mm_struct, so grab the mm in case the task dies.
1637 get_task_struct(current->group_leader);
1638 dma->task = current->group_leader;
1639 dma->lock_cap = capable(CAP_IPC_LOCK);
1640 dma->mm = current->mm;
1643 dma->pfn_list = RB_ROOT;
1645 /* Insert zero-sized and grow as we map chunks of it */
1646 vfio_link_dma(iommu, dma);
1648 /* Don't pin and map if container doesn't contain IOMMU capable domain*/
1649 if (list_empty(&iommu->domain_list))
1652 ret = vfio_pin_map_dma(iommu, dma, size);
1654 if (!ret && iommu->dirty_page_tracking) {
1655 ret = vfio_dma_bitmap_alloc(dma, pgsize);
1657 vfio_remove_dma(iommu, dma);
1661 mutex_unlock(&iommu->lock);
1665 static int vfio_iommu_replay(struct vfio_iommu *iommu,
1666 struct vfio_domain *domain)
1668 struct vfio_batch batch;
1669 struct vfio_domain *d = NULL;
1671 unsigned long limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1674 /* Arbitrarily pick the first domain in the list for lookups */
1675 if (!list_empty(&iommu->domain_list))
1676 d = list_first_entry(&iommu->domain_list,
1677 struct vfio_domain, next);
1679 vfio_batch_init(&batch);
1681 n = rb_first(&iommu->dma_list);
1683 for (; n; n = rb_next(n)) {
1684 struct vfio_dma *dma;
1687 dma = rb_entry(n, struct vfio_dma, node);
1690 while (iova < dma->iova + dma->size) {
1694 if (dma->iommu_mapped) {
1698 if (WARN_ON(!d)) { /* mapped w/o a domain?! */
1703 phys = iommu_iova_to_phys(d->domain, iova);
1705 if (WARN_ON(!phys)) {
1713 while (i < dma->iova + dma->size &&
1714 p == iommu_iova_to_phys(d->domain, i)) {
1721 unsigned long vaddr = dma->vaddr +
1723 size_t n = dma->iova + dma->size - iova;
1726 npage = vfio_pin_pages_remote(dma, vaddr,
1736 phys = pfn << PAGE_SHIFT;
1737 size = npage << PAGE_SHIFT;
1740 ret = iommu_map(domain->domain, iova, phys, size,
1741 dma->prot | IOMMU_CACHE,
1742 GFP_KERNEL_ACCOUNT);
1744 if (!dma->iommu_mapped) {
1745 vfio_unpin_pages_remote(dma, iova,
1749 vfio_batch_unpin(&batch, dma);
1758 /* All dmas are now mapped, defer to second tree walk for unwind */
1759 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
1760 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1762 dma->iommu_mapped = true;
1765 vfio_batch_fini(&batch);
1769 for (; n; n = rb_prev(n)) {
1770 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1773 if (dma->iommu_mapped) {
1774 iommu_unmap(domain->domain, dma->iova, dma->size);
1779 while (iova < dma->iova + dma->size) {
1780 phys_addr_t phys, p;
1784 phys = iommu_iova_to_phys(domain->domain, iova);
1793 while (i < dma->iova + dma->size &&
1794 p == iommu_iova_to_phys(domain->domain, i)) {
1800 iommu_unmap(domain->domain, iova, size);
1801 vfio_unpin_pages_remote(dma, iova, phys >> PAGE_SHIFT,
1802 size >> PAGE_SHIFT, true);
1806 vfio_batch_fini(&batch);
1811 * We change our unmap behavior slightly depending on whether the IOMMU
1812 * supports fine-grained superpages. IOMMUs like AMD-Vi will use a superpage
1813 * for practically any contiguous power-of-two mapping we give it. This means
1814 * we don't need to look for contiguous chunks ourselves to make unmapping
1815 * more efficient. On IOMMUs with coarse-grained super pages, like Intel VT-d
1816 * with discrete 2M/1G/512G/1T superpages, identifying contiguous chunks
1817 * significantly boosts non-hugetlbfs mappings and doesn't seem to hurt when
1818 * hugetlbfs is in use.
1820 static void vfio_test_domain_fgsp(struct vfio_domain *domain, struct list_head *regions)
1822 int ret, order = get_order(PAGE_SIZE * 2);
1823 struct vfio_iova *region;
1827 pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order);
1831 list_for_each_entry(region, regions, list) {
1832 start = ALIGN(region->start, PAGE_SIZE * 2);
1833 if (start >= region->end || (region->end - start < PAGE_SIZE * 2))
1836 ret = iommu_map(domain->domain, start, page_to_phys(pages), PAGE_SIZE * 2,
1837 IOMMU_READ | IOMMU_WRITE | IOMMU_CACHE,
1838 GFP_KERNEL_ACCOUNT);
1840 size_t unmapped = iommu_unmap(domain->domain, start, PAGE_SIZE);
1842 if (unmapped == PAGE_SIZE)
1843 iommu_unmap(domain->domain, start + PAGE_SIZE, PAGE_SIZE);
1845 domain->fgsp = true;
1850 __free_pages(pages, order);
1853 static struct vfio_iommu_group *find_iommu_group(struct vfio_domain *domain,
1854 struct iommu_group *iommu_group)
1856 struct vfio_iommu_group *g;
1858 list_for_each_entry(g, &domain->group_list, next) {
1859 if (g->iommu_group == iommu_group)
1866 static struct vfio_iommu_group*
1867 vfio_iommu_find_iommu_group(struct vfio_iommu *iommu,
1868 struct iommu_group *iommu_group)
1870 struct vfio_iommu_group *group;
1871 struct vfio_domain *domain;
1873 list_for_each_entry(domain, &iommu->domain_list, next) {
1874 group = find_iommu_group(domain, iommu_group);
1879 list_for_each_entry(group, &iommu->emulated_iommu_groups, next)
1880 if (group->iommu_group == iommu_group)
1885 static bool vfio_iommu_has_sw_msi(struct list_head *group_resv_regions,
1888 struct iommu_resv_region *region;
1891 list_for_each_entry(region, group_resv_regions, list) {
1893 * The presence of any 'real' MSI regions should take
1894 * precedence over the software-managed one if the
1895 * IOMMU driver happens to advertise both types.
1897 if (region->type == IOMMU_RESV_MSI) {
1902 if (region->type == IOMMU_RESV_SW_MSI) {
1903 *base = region->start;
1912 * This is a helper function to insert an address range to iova list.
1913 * The list is initially created with a single entry corresponding to
1914 * the IOMMU domain geometry to which the device group is attached.
1915 * The list aperture gets modified when a new domain is added to the
1916 * container if the new aperture doesn't conflict with the current one
1917 * or with any existing dma mappings. The list is also modified to
1918 * exclude any reserved regions associated with the device group.
1920 static int vfio_iommu_iova_insert(struct list_head *head,
1921 dma_addr_t start, dma_addr_t end)
1923 struct vfio_iova *region;
1925 region = kmalloc(sizeof(*region), GFP_KERNEL);
1929 INIT_LIST_HEAD(®ion->list);
1930 region->start = start;
1933 list_add_tail(®ion->list, head);
1938 * Check the new iommu aperture conflicts with existing aper or with any
1939 * existing dma mappings.
1941 static bool vfio_iommu_aper_conflict(struct vfio_iommu *iommu,
1942 dma_addr_t start, dma_addr_t end)
1944 struct vfio_iova *first, *last;
1945 struct list_head *iova = &iommu->iova_list;
1947 if (list_empty(iova))
1950 /* Disjoint sets, return conflict */
1951 first = list_first_entry(iova, struct vfio_iova, list);
1952 last = list_last_entry(iova, struct vfio_iova, list);
1953 if (start > last->end || end < first->start)
1956 /* Check for any existing dma mappings below the new start */
1957 if (start > first->start) {
1958 if (vfio_find_dma(iommu, first->start, start - first->start))
1962 /* Check for any existing dma mappings beyond the new end */
1963 if (end < last->end) {
1964 if (vfio_find_dma(iommu, end + 1, last->end - end))
1972 * Resize iommu iova aperture window. This is called only if the new
1973 * aperture has no conflict with existing aperture and dma mappings.
1975 static int vfio_iommu_aper_resize(struct list_head *iova,
1976 dma_addr_t start, dma_addr_t end)
1978 struct vfio_iova *node, *next;
1980 if (list_empty(iova))
1981 return vfio_iommu_iova_insert(iova, start, end);
1983 /* Adjust iova list start */
1984 list_for_each_entry_safe(node, next, iova, list) {
1985 if (start < node->start)
1987 if (start >= node->start && start < node->end) {
1988 node->start = start;
1991 /* Delete nodes before new start */
1992 list_del(&node->list);
1996 /* Adjust iova list end */
1997 list_for_each_entry_safe(node, next, iova, list) {
1998 if (end > node->end)
2000 if (end > node->start && end <= node->end) {
2004 /* Delete nodes after new end */
2005 list_del(&node->list);
2013 * Check reserved region conflicts with existing dma mappings
2015 static bool vfio_iommu_resv_conflict(struct vfio_iommu *iommu,
2016 struct list_head *resv_regions)
2018 struct iommu_resv_region *region;
2020 /* Check for conflict with existing dma mappings */
2021 list_for_each_entry(region, resv_regions, list) {
2022 if (region->type == IOMMU_RESV_DIRECT_RELAXABLE)
2025 if (vfio_find_dma(iommu, region->start, region->length))
2033 * Check iova region overlap with reserved regions and
2034 * exclude them from the iommu iova range
2036 static int vfio_iommu_resv_exclude(struct list_head *iova,
2037 struct list_head *resv_regions)
2039 struct iommu_resv_region *resv;
2040 struct vfio_iova *n, *next;
2042 list_for_each_entry(resv, resv_regions, list) {
2043 phys_addr_t start, end;
2045 if (resv->type == IOMMU_RESV_DIRECT_RELAXABLE)
2048 start = resv->start;
2049 end = resv->start + resv->length - 1;
2051 list_for_each_entry_safe(n, next, iova, list) {
2055 if (start > n->end || end < n->start)
2058 * Insert a new node if current node overlaps with the
2059 * reserve region to exclude that from valid iova range.
2060 * Note that, new node is inserted before the current
2061 * node and finally the current node is deleted keeping
2062 * the list updated and sorted.
2064 if (start > n->start)
2065 ret = vfio_iommu_iova_insert(&n->list, n->start,
2067 if (!ret && end < n->end)
2068 ret = vfio_iommu_iova_insert(&n->list, end + 1,
2078 if (list_empty(iova))
2084 static void vfio_iommu_resv_free(struct list_head *resv_regions)
2086 struct iommu_resv_region *n, *next;
2088 list_for_each_entry_safe(n, next, resv_regions, list) {
2094 static void vfio_iommu_iova_free(struct list_head *iova)
2096 struct vfio_iova *n, *next;
2098 list_for_each_entry_safe(n, next, iova, list) {
2104 static int vfio_iommu_iova_get_copy(struct vfio_iommu *iommu,
2105 struct list_head *iova_copy)
2107 struct list_head *iova = &iommu->iova_list;
2108 struct vfio_iova *n;
2111 list_for_each_entry(n, iova, list) {
2112 ret = vfio_iommu_iova_insert(iova_copy, n->start, n->end);
2120 vfio_iommu_iova_free(iova_copy);
2124 static void vfio_iommu_iova_insert_copy(struct vfio_iommu *iommu,
2125 struct list_head *iova_copy)
2127 struct list_head *iova = &iommu->iova_list;
2129 vfio_iommu_iova_free(iova);
2131 list_splice_tail(iova_copy, iova);
2134 static int vfio_iommu_domain_alloc(struct device *dev, void *data)
2136 struct iommu_domain **domain = data;
2138 *domain = iommu_domain_alloc(dev->bus);
2139 return 1; /* Don't iterate */
2142 static int vfio_iommu_type1_attach_group(void *iommu_data,
2143 struct iommu_group *iommu_group, enum vfio_group_type type)
2145 struct vfio_iommu *iommu = iommu_data;
2146 struct vfio_iommu_group *group;
2147 struct vfio_domain *domain, *d;
2149 phys_addr_t resv_msi_base = 0;
2150 struct iommu_domain_geometry *geo;
2151 LIST_HEAD(iova_copy);
2152 LIST_HEAD(group_resv_regions);
2155 mutex_lock(&iommu->lock);
2157 /* Attach could require pinning, so disallow while vaddr is invalid. */
2158 if (iommu->vaddr_invalid_count)
2161 /* Check for duplicates */
2163 if (vfio_iommu_find_iommu_group(iommu, iommu_group))
2167 group = kzalloc(sizeof(*group), GFP_KERNEL);
2170 group->iommu_group = iommu_group;
2172 if (type == VFIO_EMULATED_IOMMU) {
2173 list_add(&group->next, &iommu->emulated_iommu_groups);
2175 * An emulated IOMMU group cannot dirty memory directly, it can
2176 * only use interfaces that provide dirty tracking.
2177 * The iommu scope can only be promoted with the addition of a
2178 * dirty tracking group.
2180 group->pinned_page_dirty_scope = true;
2186 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2188 goto out_free_group;
2191 * Going via the iommu_group iterator avoids races, and trivially gives
2192 * us a representative device for the IOMMU API call. We don't actually
2193 * want to iterate beyond the first device (if any).
2196 iommu_group_for_each_dev(iommu_group, &domain->domain,
2197 vfio_iommu_domain_alloc);
2198 if (!domain->domain)
2199 goto out_free_domain;
2201 if (iommu->nesting) {
2202 ret = iommu_enable_nesting(domain->domain);
2207 ret = iommu_attach_group(domain->domain, group->iommu_group);
2211 /* Get aperture info */
2212 geo = &domain->domain->geometry;
2213 if (vfio_iommu_aper_conflict(iommu, geo->aperture_start,
2214 geo->aperture_end)) {
2219 ret = iommu_get_group_resv_regions(iommu_group, &group_resv_regions);
2223 if (vfio_iommu_resv_conflict(iommu, &group_resv_regions)) {
2229 * We don't want to work on the original iova list as the list
2230 * gets modified and in case of failure we have to retain the
2231 * original list. Get a copy here.
2233 ret = vfio_iommu_iova_get_copy(iommu, &iova_copy);
2237 ret = vfio_iommu_aper_resize(&iova_copy, geo->aperture_start,
2242 ret = vfio_iommu_resv_exclude(&iova_copy, &group_resv_regions);
2246 resv_msi = vfio_iommu_has_sw_msi(&group_resv_regions, &resv_msi_base);
2248 INIT_LIST_HEAD(&domain->group_list);
2249 list_add(&group->next, &domain->group_list);
2251 if (!allow_unsafe_interrupts &&
2252 !iommu_group_has_isolated_msi(iommu_group)) {
2253 pr_warn("%s: No interrupt remapping support. Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n",
2260 * If the IOMMU can block non-coherent operations (ie PCIe TLPs with
2261 * no-snoop set) then VFIO always turns this feature on because on Intel
2262 * platforms it optimizes KVM to disable wbinvd emulation.
2264 if (domain->domain->ops->enforce_cache_coherency)
2265 domain->enforce_cache_coherency =
2266 domain->domain->ops->enforce_cache_coherency(
2270 * Try to match an existing compatible domain. We don't want to
2271 * preclude an IOMMU driver supporting multiple bus_types and being
2272 * able to include different bus_types in the same IOMMU domain, so
2273 * we test whether the domains use the same iommu_ops rather than
2274 * testing if they're on the same bus_type.
2276 list_for_each_entry(d, &iommu->domain_list, next) {
2277 if (d->domain->ops == domain->domain->ops &&
2278 d->enforce_cache_coherency ==
2279 domain->enforce_cache_coherency) {
2280 iommu_detach_group(domain->domain, group->iommu_group);
2281 if (!iommu_attach_group(d->domain,
2282 group->iommu_group)) {
2283 list_add(&group->next, &d->group_list);
2284 iommu_domain_free(domain->domain);
2289 ret = iommu_attach_group(domain->domain,
2290 group->iommu_group);
2296 vfio_test_domain_fgsp(domain, &iova_copy);
2298 /* replay mappings on new domains */
2299 ret = vfio_iommu_replay(iommu, domain);
2304 ret = iommu_get_msi_cookie(domain->domain, resv_msi_base);
2305 if (ret && ret != -ENODEV)
2309 list_add(&domain->next, &iommu->domain_list);
2310 vfio_update_pgsize_bitmap(iommu);
2312 /* Delete the old one and insert new iova list */
2313 vfio_iommu_iova_insert_copy(iommu, &iova_copy);
2316 * An iommu backed group can dirty memory directly and therefore
2317 * demotes the iommu scope until it declares itself dirty tracking
2318 * capable via the page pinning interface.
2320 iommu->num_non_pinned_groups++;
2321 mutex_unlock(&iommu->lock);
2322 vfio_iommu_resv_free(&group_resv_regions);
2327 iommu_detach_group(domain->domain, group->iommu_group);
2329 iommu_domain_free(domain->domain);
2330 vfio_iommu_iova_free(&iova_copy);
2331 vfio_iommu_resv_free(&group_resv_regions);
2337 mutex_unlock(&iommu->lock);
2341 static void vfio_iommu_unmap_unpin_all(struct vfio_iommu *iommu)
2343 struct rb_node *node;
2345 while ((node = rb_first(&iommu->dma_list)))
2346 vfio_remove_dma(iommu, rb_entry(node, struct vfio_dma, node));
2349 static void vfio_iommu_unmap_unpin_reaccount(struct vfio_iommu *iommu)
2351 struct rb_node *n, *p;
2353 n = rb_first(&iommu->dma_list);
2354 for (; n; n = rb_next(n)) {
2355 struct vfio_dma *dma;
2356 long locked = 0, unlocked = 0;
2358 dma = rb_entry(n, struct vfio_dma, node);
2359 unlocked += vfio_unmap_unpin(iommu, dma, false);
2360 p = rb_first(&dma->pfn_list);
2361 for (; p; p = rb_next(p)) {
2362 struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn,
2365 if (!is_invalid_reserved_pfn(vpfn->pfn))
2368 vfio_lock_acct(dma, locked - unlocked, true);
2373 * Called when a domain is removed in detach. It is possible that
2374 * the removed domain decided the iova aperture window. Modify the
2375 * iova aperture with the smallest window among existing domains.
2377 static void vfio_iommu_aper_expand(struct vfio_iommu *iommu,
2378 struct list_head *iova_copy)
2380 struct vfio_domain *domain;
2381 struct vfio_iova *node;
2382 dma_addr_t start = 0;
2383 dma_addr_t end = (dma_addr_t)~0;
2385 if (list_empty(iova_copy))
2388 list_for_each_entry(domain, &iommu->domain_list, next) {
2389 struct iommu_domain_geometry *geo = &domain->domain->geometry;
2391 if (geo->aperture_start > start)
2392 start = geo->aperture_start;
2393 if (geo->aperture_end < end)
2394 end = geo->aperture_end;
2397 /* Modify aperture limits. The new aper is either same or bigger */
2398 node = list_first_entry(iova_copy, struct vfio_iova, list);
2399 node->start = start;
2400 node = list_last_entry(iova_copy, struct vfio_iova, list);
2405 * Called when a group is detached. The reserved regions for that
2406 * group can be part of valid iova now. But since reserved regions
2407 * may be duplicated among groups, populate the iova valid regions
2410 static int vfio_iommu_resv_refresh(struct vfio_iommu *iommu,
2411 struct list_head *iova_copy)
2413 struct vfio_domain *d;
2414 struct vfio_iommu_group *g;
2415 struct vfio_iova *node;
2416 dma_addr_t start, end;
2417 LIST_HEAD(resv_regions);
2420 if (list_empty(iova_copy))
2423 list_for_each_entry(d, &iommu->domain_list, next) {
2424 list_for_each_entry(g, &d->group_list, next) {
2425 ret = iommu_get_group_resv_regions(g->iommu_group,
2432 node = list_first_entry(iova_copy, struct vfio_iova, list);
2433 start = node->start;
2434 node = list_last_entry(iova_copy, struct vfio_iova, list);
2437 /* purge the iova list and create new one */
2438 vfio_iommu_iova_free(iova_copy);
2440 ret = vfio_iommu_aper_resize(iova_copy, start, end);
2444 /* Exclude current reserved regions from iova ranges */
2445 ret = vfio_iommu_resv_exclude(iova_copy, &resv_regions);
2447 vfio_iommu_resv_free(&resv_regions);
2451 static void vfio_iommu_type1_detach_group(void *iommu_data,
2452 struct iommu_group *iommu_group)
2454 struct vfio_iommu *iommu = iommu_data;
2455 struct vfio_domain *domain;
2456 struct vfio_iommu_group *group;
2457 bool update_dirty_scope = false;
2458 LIST_HEAD(iova_copy);
2460 mutex_lock(&iommu->lock);
2461 list_for_each_entry(group, &iommu->emulated_iommu_groups, next) {
2462 if (group->iommu_group != iommu_group)
2464 update_dirty_scope = !group->pinned_page_dirty_scope;
2465 list_del(&group->next);
2468 if (list_empty(&iommu->emulated_iommu_groups) &&
2469 list_empty(&iommu->domain_list)) {
2470 WARN_ON(!list_empty(&iommu->device_list));
2471 vfio_iommu_unmap_unpin_all(iommu);
2473 goto detach_group_done;
2477 * Get a copy of iova list. This will be used to update
2478 * and to replace the current one later. Please note that
2479 * we will leave the original list as it is if update fails.
2481 vfio_iommu_iova_get_copy(iommu, &iova_copy);
2483 list_for_each_entry(domain, &iommu->domain_list, next) {
2484 group = find_iommu_group(domain, iommu_group);
2488 iommu_detach_group(domain->domain, group->iommu_group);
2489 update_dirty_scope = !group->pinned_page_dirty_scope;
2490 list_del(&group->next);
2493 * Group ownership provides privilege, if the group list is
2494 * empty, the domain goes away. If it's the last domain with
2495 * iommu and external domain doesn't exist, then all the
2496 * mappings go away too. If it's the last domain with iommu and
2497 * external domain exist, update accounting
2499 if (list_empty(&domain->group_list)) {
2500 if (list_is_singular(&iommu->domain_list)) {
2501 if (list_empty(&iommu->emulated_iommu_groups)) {
2502 WARN_ON(!list_empty(
2503 &iommu->device_list));
2504 vfio_iommu_unmap_unpin_all(iommu);
2506 vfio_iommu_unmap_unpin_reaccount(iommu);
2509 iommu_domain_free(domain->domain);
2510 list_del(&domain->next);
2512 vfio_iommu_aper_expand(iommu, &iova_copy);
2513 vfio_update_pgsize_bitmap(iommu);
2518 if (!vfio_iommu_resv_refresh(iommu, &iova_copy))
2519 vfio_iommu_iova_insert_copy(iommu, &iova_copy);
2521 vfio_iommu_iova_free(&iova_copy);
2525 * Removal of a group without dirty tracking may allow the iommu scope
2528 if (update_dirty_scope) {
2529 iommu->num_non_pinned_groups--;
2530 if (iommu->dirty_page_tracking)
2531 vfio_iommu_populate_bitmap_full(iommu);
2533 mutex_unlock(&iommu->lock);
2536 static void *vfio_iommu_type1_open(unsigned long arg)
2538 struct vfio_iommu *iommu;
2540 iommu = kzalloc(sizeof(*iommu), GFP_KERNEL);
2542 return ERR_PTR(-ENOMEM);
2545 case VFIO_TYPE1_IOMMU:
2547 case VFIO_TYPE1_NESTING_IOMMU:
2548 iommu->nesting = true;
2550 case VFIO_TYPE1v2_IOMMU:
2555 return ERR_PTR(-EINVAL);
2558 INIT_LIST_HEAD(&iommu->domain_list);
2559 INIT_LIST_HEAD(&iommu->iova_list);
2560 iommu->dma_list = RB_ROOT;
2561 iommu->dma_avail = dma_entry_limit;
2562 mutex_init(&iommu->lock);
2563 mutex_init(&iommu->device_list_lock);
2564 INIT_LIST_HEAD(&iommu->device_list);
2565 iommu->pgsize_bitmap = PAGE_MASK;
2566 INIT_LIST_HEAD(&iommu->emulated_iommu_groups);
2571 static void vfio_release_domain(struct vfio_domain *domain)
2573 struct vfio_iommu_group *group, *group_tmp;
2575 list_for_each_entry_safe(group, group_tmp,
2576 &domain->group_list, next) {
2577 iommu_detach_group(domain->domain, group->iommu_group);
2578 list_del(&group->next);
2582 iommu_domain_free(domain->domain);
2585 static void vfio_iommu_type1_release(void *iommu_data)
2587 struct vfio_iommu *iommu = iommu_data;
2588 struct vfio_domain *domain, *domain_tmp;
2589 struct vfio_iommu_group *group, *next_group;
2591 list_for_each_entry_safe(group, next_group,
2592 &iommu->emulated_iommu_groups, next) {
2593 list_del(&group->next);
2597 vfio_iommu_unmap_unpin_all(iommu);
2599 list_for_each_entry_safe(domain, domain_tmp,
2600 &iommu->domain_list, next) {
2601 vfio_release_domain(domain);
2602 list_del(&domain->next);
2606 vfio_iommu_iova_free(&iommu->iova_list);
2611 static int vfio_domains_have_enforce_cache_coherency(struct vfio_iommu *iommu)
2613 struct vfio_domain *domain;
2616 mutex_lock(&iommu->lock);
2617 list_for_each_entry(domain, &iommu->domain_list, next) {
2618 if (!(domain->enforce_cache_coherency)) {
2623 mutex_unlock(&iommu->lock);
2628 static bool vfio_iommu_has_emulated(struct vfio_iommu *iommu)
2632 mutex_lock(&iommu->lock);
2633 ret = !list_empty(&iommu->emulated_iommu_groups);
2634 mutex_unlock(&iommu->lock);
2638 static int vfio_iommu_type1_check_extension(struct vfio_iommu *iommu,
2642 case VFIO_TYPE1_IOMMU:
2643 case VFIO_TYPE1v2_IOMMU:
2644 case VFIO_TYPE1_NESTING_IOMMU:
2645 case VFIO_UNMAP_ALL:
2647 case VFIO_UPDATE_VADDR:
2649 * Disable this feature if mdevs are present. They cannot
2650 * safely pin/unpin/rw while vaddrs are being updated.
2652 return iommu && !vfio_iommu_has_emulated(iommu);
2653 case VFIO_DMA_CC_IOMMU:
2656 return vfio_domains_have_enforce_cache_coherency(iommu);
2662 static int vfio_iommu_iova_add_cap(struct vfio_info_cap *caps,
2663 struct vfio_iommu_type1_info_cap_iova_range *cap_iovas,
2666 struct vfio_info_cap_header *header;
2667 struct vfio_iommu_type1_info_cap_iova_range *iova_cap;
2669 header = vfio_info_cap_add(caps, size,
2670 VFIO_IOMMU_TYPE1_INFO_CAP_IOVA_RANGE, 1);
2672 return PTR_ERR(header);
2674 iova_cap = container_of(header,
2675 struct vfio_iommu_type1_info_cap_iova_range,
2677 iova_cap->nr_iovas = cap_iovas->nr_iovas;
2678 memcpy(iova_cap->iova_ranges, cap_iovas->iova_ranges,
2679 cap_iovas->nr_iovas * sizeof(*cap_iovas->iova_ranges));
2683 static int vfio_iommu_iova_build_caps(struct vfio_iommu *iommu,
2684 struct vfio_info_cap *caps)
2686 struct vfio_iommu_type1_info_cap_iova_range *cap_iovas;
2687 struct vfio_iova *iova;
2689 int iovas = 0, i = 0, ret;
2691 list_for_each_entry(iova, &iommu->iova_list, list)
2696 * Return 0 as a container with a single mdev device
2697 * will have an empty list
2702 size = struct_size(cap_iovas, iova_ranges, iovas);
2704 cap_iovas = kzalloc(size, GFP_KERNEL);
2708 cap_iovas->nr_iovas = iovas;
2710 list_for_each_entry(iova, &iommu->iova_list, list) {
2711 cap_iovas->iova_ranges[i].start = iova->start;
2712 cap_iovas->iova_ranges[i].end = iova->end;
2716 ret = vfio_iommu_iova_add_cap(caps, cap_iovas, size);
2722 static int vfio_iommu_migration_build_caps(struct vfio_iommu *iommu,
2723 struct vfio_info_cap *caps)
2725 struct vfio_iommu_type1_info_cap_migration cap_mig = {};
2727 cap_mig.header.id = VFIO_IOMMU_TYPE1_INFO_CAP_MIGRATION;
2728 cap_mig.header.version = 1;
2731 /* support minimum pgsize */
2732 cap_mig.pgsize_bitmap = (size_t)1 << __ffs(iommu->pgsize_bitmap);
2733 cap_mig.max_dirty_bitmap_size = DIRTY_BITMAP_SIZE_MAX;
2735 return vfio_info_add_capability(caps, &cap_mig.header, sizeof(cap_mig));
2738 static int vfio_iommu_dma_avail_build_caps(struct vfio_iommu *iommu,
2739 struct vfio_info_cap *caps)
2741 struct vfio_iommu_type1_info_dma_avail cap_dma_avail;
2743 cap_dma_avail.header.id = VFIO_IOMMU_TYPE1_INFO_DMA_AVAIL;
2744 cap_dma_avail.header.version = 1;
2746 cap_dma_avail.avail = iommu->dma_avail;
2748 return vfio_info_add_capability(caps, &cap_dma_avail.header,
2749 sizeof(cap_dma_avail));
2752 static int vfio_iommu_type1_get_info(struct vfio_iommu *iommu,
2755 struct vfio_iommu_type1_info info = {};
2756 unsigned long minsz;
2757 struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
2760 minsz = offsetofend(struct vfio_iommu_type1_info, iova_pgsizes);
2762 if (copy_from_user(&info, (void __user *)arg, minsz))
2765 if (info.argsz < minsz)
2768 minsz = min_t(size_t, info.argsz, sizeof(info));
2770 mutex_lock(&iommu->lock);
2771 info.flags = VFIO_IOMMU_INFO_PGSIZES;
2773 info.iova_pgsizes = iommu->pgsize_bitmap;
2775 ret = vfio_iommu_migration_build_caps(iommu, &caps);
2778 ret = vfio_iommu_dma_avail_build_caps(iommu, &caps);
2781 ret = vfio_iommu_iova_build_caps(iommu, &caps);
2783 mutex_unlock(&iommu->lock);
2789 info.flags |= VFIO_IOMMU_INFO_CAPS;
2791 if (info.argsz < sizeof(info) + caps.size) {
2792 info.argsz = sizeof(info) + caps.size;
2794 vfio_info_cap_shift(&caps, sizeof(info));
2795 if (copy_to_user((void __user *)arg +
2796 sizeof(info), caps.buf,
2801 info.cap_offset = sizeof(info);
2807 return copy_to_user((void __user *)arg, &info, minsz) ?
2811 static int vfio_iommu_type1_map_dma(struct vfio_iommu *iommu,
2814 struct vfio_iommu_type1_dma_map map;
2815 unsigned long minsz;
2816 uint32_t mask = VFIO_DMA_MAP_FLAG_READ | VFIO_DMA_MAP_FLAG_WRITE |
2817 VFIO_DMA_MAP_FLAG_VADDR;
2819 minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
2821 if (copy_from_user(&map, (void __user *)arg, minsz))
2824 if (map.argsz < minsz || map.flags & ~mask)
2827 return vfio_dma_do_map(iommu, &map);
2830 static int vfio_iommu_type1_unmap_dma(struct vfio_iommu *iommu,
2833 struct vfio_iommu_type1_dma_unmap unmap;
2834 struct vfio_bitmap bitmap = { 0 };
2835 uint32_t mask = VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP |
2836 VFIO_DMA_UNMAP_FLAG_VADDR |
2837 VFIO_DMA_UNMAP_FLAG_ALL;
2838 unsigned long minsz;
2841 minsz = offsetofend(struct vfio_iommu_type1_dma_unmap, size);
2843 if (copy_from_user(&unmap, (void __user *)arg, minsz))
2846 if (unmap.argsz < minsz || unmap.flags & ~mask)
2849 if ((unmap.flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) &&
2850 (unmap.flags & (VFIO_DMA_UNMAP_FLAG_ALL |
2851 VFIO_DMA_UNMAP_FLAG_VADDR)))
2854 if (unmap.flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) {
2855 unsigned long pgshift;
2857 if (unmap.argsz < (minsz + sizeof(bitmap)))
2860 if (copy_from_user(&bitmap,
2861 (void __user *)(arg + minsz),
2865 if (!access_ok((void __user *)bitmap.data, bitmap.size))
2868 pgshift = __ffs(bitmap.pgsize);
2869 ret = verify_bitmap_size(unmap.size >> pgshift,
2875 ret = vfio_dma_do_unmap(iommu, &unmap, &bitmap);
2879 return copy_to_user((void __user *)arg, &unmap, minsz) ?
2883 static int vfio_iommu_type1_dirty_pages(struct vfio_iommu *iommu,
2886 struct vfio_iommu_type1_dirty_bitmap dirty;
2887 uint32_t mask = VFIO_IOMMU_DIRTY_PAGES_FLAG_START |
2888 VFIO_IOMMU_DIRTY_PAGES_FLAG_STOP |
2889 VFIO_IOMMU_DIRTY_PAGES_FLAG_GET_BITMAP;
2890 unsigned long minsz;
2896 minsz = offsetofend(struct vfio_iommu_type1_dirty_bitmap, flags);
2898 if (copy_from_user(&dirty, (void __user *)arg, minsz))
2901 if (dirty.argsz < minsz || dirty.flags & ~mask)
2904 /* only one flag should be set at a time */
2905 if (__ffs(dirty.flags) != __fls(dirty.flags))
2908 if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_START) {
2911 mutex_lock(&iommu->lock);
2912 pgsize = 1 << __ffs(iommu->pgsize_bitmap);
2913 if (!iommu->dirty_page_tracking) {
2914 ret = vfio_dma_bitmap_alloc_all(iommu, pgsize);
2916 iommu->dirty_page_tracking = true;
2918 mutex_unlock(&iommu->lock);
2920 } else if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_STOP) {
2921 mutex_lock(&iommu->lock);
2922 if (iommu->dirty_page_tracking) {
2923 iommu->dirty_page_tracking = false;
2924 vfio_dma_bitmap_free_all(iommu);
2926 mutex_unlock(&iommu->lock);
2928 } else if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_GET_BITMAP) {
2929 struct vfio_iommu_type1_dirty_bitmap_get range;
2930 unsigned long pgshift;
2931 size_t data_size = dirty.argsz - minsz;
2932 size_t iommu_pgsize;
2934 if (!data_size || data_size < sizeof(range))
2937 if (copy_from_user(&range, (void __user *)(arg + minsz),
2941 if (range.iova + range.size < range.iova)
2943 if (!access_ok((void __user *)range.bitmap.data,
2947 pgshift = __ffs(range.bitmap.pgsize);
2948 ret = verify_bitmap_size(range.size >> pgshift,
2953 mutex_lock(&iommu->lock);
2955 iommu_pgsize = (size_t)1 << __ffs(iommu->pgsize_bitmap);
2957 /* allow only smallest supported pgsize */
2958 if (range.bitmap.pgsize != iommu_pgsize) {
2962 if (range.iova & (iommu_pgsize - 1)) {
2966 if (!range.size || range.size & (iommu_pgsize - 1)) {
2971 if (iommu->dirty_page_tracking)
2972 ret = vfio_iova_dirty_bitmap(range.bitmap.data,
2975 range.bitmap.pgsize);
2979 mutex_unlock(&iommu->lock);
2987 static long vfio_iommu_type1_ioctl(void *iommu_data,
2988 unsigned int cmd, unsigned long arg)
2990 struct vfio_iommu *iommu = iommu_data;
2993 case VFIO_CHECK_EXTENSION:
2994 return vfio_iommu_type1_check_extension(iommu, arg);
2995 case VFIO_IOMMU_GET_INFO:
2996 return vfio_iommu_type1_get_info(iommu, arg);
2997 case VFIO_IOMMU_MAP_DMA:
2998 return vfio_iommu_type1_map_dma(iommu, arg);
2999 case VFIO_IOMMU_UNMAP_DMA:
3000 return vfio_iommu_type1_unmap_dma(iommu, arg);
3001 case VFIO_IOMMU_DIRTY_PAGES:
3002 return vfio_iommu_type1_dirty_pages(iommu, arg);
3008 static void vfio_iommu_type1_register_device(void *iommu_data,
3009 struct vfio_device *vdev)
3011 struct vfio_iommu *iommu = iommu_data;
3013 if (!vdev->ops->dma_unmap)
3017 * list_empty(&iommu->device_list) is tested under the iommu->lock while
3018 * iteration for dma_unmap must be done under the device_list_lock.
3019 * Holding both locks here allows avoiding the device_list_lock in
3020 * several fast paths. See vfio_notify_dma_unmap()
3022 mutex_lock(&iommu->lock);
3023 mutex_lock(&iommu->device_list_lock);
3024 list_add(&vdev->iommu_entry, &iommu->device_list);
3025 mutex_unlock(&iommu->device_list_lock);
3026 mutex_unlock(&iommu->lock);
3029 static void vfio_iommu_type1_unregister_device(void *iommu_data,
3030 struct vfio_device *vdev)
3032 struct vfio_iommu *iommu = iommu_data;
3034 if (!vdev->ops->dma_unmap)
3037 mutex_lock(&iommu->lock);
3038 mutex_lock(&iommu->device_list_lock);
3039 list_del(&vdev->iommu_entry);
3040 mutex_unlock(&iommu->device_list_lock);
3041 mutex_unlock(&iommu->lock);
3044 static int vfio_iommu_type1_dma_rw_chunk(struct vfio_iommu *iommu,
3045 dma_addr_t user_iova, void *data,
3046 size_t count, bool write,
3049 struct mm_struct *mm;
3050 unsigned long vaddr;
3051 struct vfio_dma *dma;
3052 bool kthread = current->mm == NULL;
3057 dma = vfio_find_dma(iommu, user_iova, 1);
3061 if ((write && !(dma->prot & IOMMU_WRITE)) ||
3062 !(dma->prot & IOMMU_READ))
3066 if (!mmget_not_zero(mm))
3071 else if (current->mm != mm)
3074 offset = user_iova - dma->iova;
3076 if (count > dma->size - offset)
3077 count = dma->size - offset;
3079 vaddr = dma->vaddr + offset;
3082 *copied = copy_to_user((void __user *)vaddr, data,
3084 if (*copied && iommu->dirty_page_tracking) {
3085 unsigned long pgshift = __ffs(iommu->pgsize_bitmap);
3087 * Bitmap populated with the smallest supported page
3090 bitmap_set(dma->bitmap, offset >> pgshift,
3091 ((offset + *copied - 1) >> pgshift) -
3092 (offset >> pgshift) + 1);
3095 *copied = copy_from_user(data, (void __user *)vaddr,
3098 kthread_unuse_mm(mm);
3101 return *copied ? 0 : -EFAULT;
3104 static int vfio_iommu_type1_dma_rw(void *iommu_data, dma_addr_t user_iova,
3105 void *data, size_t count, bool write)
3107 struct vfio_iommu *iommu = iommu_data;
3111 mutex_lock(&iommu->lock);
3113 if (WARN_ONCE(iommu->vaddr_invalid_count,
3114 "vfio_dma_rw not allowed with VFIO_UPDATE_VADDR\n")) {
3120 ret = vfio_iommu_type1_dma_rw_chunk(iommu, user_iova, data,
3121 count, write, &done);
3131 mutex_unlock(&iommu->lock);
3135 static struct iommu_domain *
3136 vfio_iommu_type1_group_iommu_domain(void *iommu_data,
3137 struct iommu_group *iommu_group)
3139 struct iommu_domain *domain = ERR_PTR(-ENODEV);
3140 struct vfio_iommu *iommu = iommu_data;
3141 struct vfio_domain *d;
3143 if (!iommu || !iommu_group)
3144 return ERR_PTR(-EINVAL);
3146 mutex_lock(&iommu->lock);
3147 list_for_each_entry(d, &iommu->domain_list, next) {
3148 if (find_iommu_group(d, iommu_group)) {
3153 mutex_unlock(&iommu->lock);
3158 static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = {
3159 .name = "vfio-iommu-type1",
3160 .owner = THIS_MODULE,
3161 .open = vfio_iommu_type1_open,
3162 .release = vfio_iommu_type1_release,
3163 .ioctl = vfio_iommu_type1_ioctl,
3164 .attach_group = vfio_iommu_type1_attach_group,
3165 .detach_group = vfio_iommu_type1_detach_group,
3166 .pin_pages = vfio_iommu_type1_pin_pages,
3167 .unpin_pages = vfio_iommu_type1_unpin_pages,
3168 .register_device = vfio_iommu_type1_register_device,
3169 .unregister_device = vfio_iommu_type1_unregister_device,
3170 .dma_rw = vfio_iommu_type1_dma_rw,
3171 .group_iommu_domain = vfio_iommu_type1_group_iommu_domain,
3174 static int __init vfio_iommu_type1_init(void)
3176 return vfio_register_iommu_driver(&vfio_iommu_driver_ops_type1);
3179 static void __exit vfio_iommu_type1_cleanup(void)
3181 vfio_unregister_iommu_driver(&vfio_iommu_driver_ops_type1);
3184 module_init(vfio_iommu_type1_init);
3185 module_exit(vfio_iommu_type1_cleanup);
3187 MODULE_VERSION(DRIVER_VERSION);
3188 MODULE_LICENSE("GPL v2");
3189 MODULE_AUTHOR(DRIVER_AUTHOR);
3190 MODULE_DESCRIPTION(DRIVER_DESC);