1 // SPDX-License-Identifier: GPL-2.0-only
3 * VFIO: IOMMU DMA mapping support for Type1 IOMMU
5 * Copyright (C) 2012 Red Hat, Inc. All rights reserved.
6 * Author: Alex Williamson <alex.williamson@redhat.com>
8 * Derived from original vfio:
9 * Copyright 2010 Cisco Systems, Inc. All rights reserved.
10 * Author: Tom Lyon, pugs@cisco.com
12 * We arbitrarily define a Type1 IOMMU as one matching the below code.
13 * It could be called the x86 IOMMU as it's designed for AMD-Vi & Intel
14 * VT-d, but that makes it harder to re-use as theoretically anyone
15 * implementing a similar IOMMU could make use of this. We expect the
16 * IOMMU to support the IOMMU API and have few to no restrictions around
17 * the IOVA range that can be mapped. The Type1 IOMMU is currently
18 * optimized for relatively static mappings of a userspace process with
19 * userpsace pages pinned into memory. We also assume devices and IOMMU
20 * domains are PCI based as the IOMMU API is still centered around a
21 * device/bus interface rather than a group interface.
24 #include <linux/compat.h>
25 #include <linux/device.h>
27 #include <linux/iommu.h>
28 #include <linux/module.h>
30 #include <linux/kthread.h>
31 #include <linux/rbtree.h>
32 #include <linux/sched/signal.h>
33 #include <linux/sched/mm.h>
34 #include <linux/slab.h>
35 #include <linux/uaccess.h>
36 #include <linux/vfio.h>
37 #include <linux/workqueue.h>
38 #include <linux/mdev.h>
39 #include <linux/notifier.h>
40 #include <linux/dma-iommu.h>
41 #include <linux/irqdomain.h>
43 #define DRIVER_VERSION "0.2"
44 #define DRIVER_AUTHOR "Alex Williamson <alex.williamson@redhat.com>"
45 #define DRIVER_DESC "Type1 IOMMU driver for VFIO"
47 static bool allow_unsafe_interrupts;
48 module_param_named(allow_unsafe_interrupts,
49 allow_unsafe_interrupts, bool, S_IRUGO | S_IWUSR);
50 MODULE_PARM_DESC(allow_unsafe_interrupts,
51 "Enable VFIO IOMMU support for on platforms without interrupt remapping support.");
53 static bool disable_hugepages;
54 module_param_named(disable_hugepages,
55 disable_hugepages, bool, S_IRUGO | S_IWUSR);
56 MODULE_PARM_DESC(disable_hugepages,
57 "Disable VFIO IOMMU support for IOMMU hugepages.");
59 static unsigned int dma_entry_limit __read_mostly = U16_MAX;
60 module_param_named(dma_entry_limit, dma_entry_limit, uint, 0644);
61 MODULE_PARM_DESC(dma_entry_limit,
62 "Maximum number of user DMA mappings per container (65535).");
65 struct list_head domain_list;
66 struct list_head iova_list;
67 struct vfio_domain *external_domain; /* domain for external user */
69 struct rb_root dma_list;
70 struct blocking_notifier_head notifier;
71 unsigned int dma_avail;
72 uint64_t pgsize_bitmap;
75 bool dirty_page_tracking;
76 bool pinned_page_dirty_scope;
80 struct iommu_domain *domain;
81 struct list_head next;
82 struct list_head group_list;
83 int prot; /* IOMMU_CACHE */
84 bool fgsp; /* Fine-grained super pages */
89 dma_addr_t iova; /* Device address */
90 unsigned long vaddr; /* Process virtual addr */
91 size_t size; /* Map size (bytes) */
92 int prot; /* IOMMU_READ/WRITE */
94 bool lock_cap; /* capable(CAP_IPC_LOCK) */
95 struct task_struct *task;
96 struct rb_root pfn_list; /* Ex-user pinned pfn list */
97 unsigned long *bitmap;
101 struct iommu_group *iommu_group;
102 struct list_head next;
103 bool mdev_group; /* An mdev group */
104 bool pinned_page_dirty_scope;
108 struct list_head list;
114 * Guest RAM pinning working set or DMA target
118 dma_addr_t iova; /* Device address */
119 unsigned long pfn; /* Host pfn */
120 unsigned int ref_count;
123 struct vfio_regions {
124 struct list_head list;
130 #define IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu) \
131 (!list_empty(&iommu->domain_list))
133 #define DIRTY_BITMAP_BYTES(n) (ALIGN(n, BITS_PER_TYPE(u64)) / BITS_PER_BYTE)
136 * Input argument of number of bits to bitmap_set() is unsigned integer, which
137 * further casts to signed integer for unaligned multi-bit operation,
139 * Then maximum bitmap size supported is 2^31 bits divided by 2^3 bits/byte,
140 * that is 2^28 (256 MB) which maps to 2^31 * 2^12 = 2^43 (8TB) on 4K page
143 #define DIRTY_BITMAP_PAGES_MAX ((u64)INT_MAX)
144 #define DIRTY_BITMAP_SIZE_MAX DIRTY_BITMAP_BYTES(DIRTY_BITMAP_PAGES_MAX)
146 static int put_pfn(unsigned long pfn, int prot);
148 static struct vfio_group *vfio_iommu_find_iommu_group(struct vfio_iommu *iommu,
149 struct iommu_group *iommu_group);
151 static void update_pinned_page_dirty_scope(struct vfio_iommu *iommu);
153 * This code handles mapping and unmapping of user data buffers
154 * into DMA'ble space using the IOMMU
157 static struct vfio_dma *vfio_find_dma(struct vfio_iommu *iommu,
158 dma_addr_t start, size_t size)
160 struct rb_node *node = iommu->dma_list.rb_node;
163 struct vfio_dma *dma = rb_entry(node, struct vfio_dma, node);
165 if (start + size <= dma->iova)
166 node = node->rb_left;
167 else if (start >= dma->iova + dma->size)
168 node = node->rb_right;
176 static void vfio_link_dma(struct vfio_iommu *iommu, struct vfio_dma *new)
178 struct rb_node **link = &iommu->dma_list.rb_node, *parent = NULL;
179 struct vfio_dma *dma;
183 dma = rb_entry(parent, struct vfio_dma, node);
185 if (new->iova + new->size <= dma->iova)
186 link = &(*link)->rb_left;
188 link = &(*link)->rb_right;
191 rb_link_node(&new->node, parent, link);
192 rb_insert_color(&new->node, &iommu->dma_list);
195 static void vfio_unlink_dma(struct vfio_iommu *iommu, struct vfio_dma *old)
197 rb_erase(&old->node, &iommu->dma_list);
201 static int vfio_dma_bitmap_alloc(struct vfio_dma *dma, size_t pgsize)
203 uint64_t npages = dma->size / pgsize;
205 if (npages > DIRTY_BITMAP_PAGES_MAX)
209 * Allocate extra 64 bits that are used to calculate shift required for
210 * bitmap_shift_left() to manipulate and club unaligned number of pages
211 * in adjacent vfio_dma ranges.
213 dma->bitmap = kvzalloc(DIRTY_BITMAP_BYTES(npages) + sizeof(u64),
221 static void vfio_dma_bitmap_free(struct vfio_dma *dma)
227 static void vfio_dma_populate_bitmap(struct vfio_dma *dma, size_t pgsize)
230 unsigned long pgshift = __ffs(pgsize);
232 for (p = rb_first(&dma->pfn_list); p; p = rb_next(p)) {
233 struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn, node);
235 bitmap_set(dma->bitmap, (vpfn->iova - dma->iova) >> pgshift, 1);
239 static int vfio_dma_bitmap_alloc_all(struct vfio_iommu *iommu, size_t pgsize)
243 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
244 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
247 ret = vfio_dma_bitmap_alloc(dma, pgsize);
251 for (p = rb_prev(n); p; p = rb_prev(p)) {
252 struct vfio_dma *dma = rb_entry(n,
253 struct vfio_dma, node);
255 vfio_dma_bitmap_free(dma);
259 vfio_dma_populate_bitmap(dma, pgsize);
264 static void vfio_dma_bitmap_free_all(struct vfio_iommu *iommu)
268 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
269 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
271 vfio_dma_bitmap_free(dma);
276 * Helper Functions for host iova-pfn list
278 static struct vfio_pfn *vfio_find_vpfn(struct vfio_dma *dma, dma_addr_t iova)
280 struct vfio_pfn *vpfn;
281 struct rb_node *node = dma->pfn_list.rb_node;
284 vpfn = rb_entry(node, struct vfio_pfn, node);
286 if (iova < vpfn->iova)
287 node = node->rb_left;
288 else if (iova > vpfn->iova)
289 node = node->rb_right;
296 static void vfio_link_pfn(struct vfio_dma *dma,
297 struct vfio_pfn *new)
299 struct rb_node **link, *parent = NULL;
300 struct vfio_pfn *vpfn;
302 link = &dma->pfn_list.rb_node;
305 vpfn = rb_entry(parent, struct vfio_pfn, node);
307 if (new->iova < vpfn->iova)
308 link = &(*link)->rb_left;
310 link = &(*link)->rb_right;
313 rb_link_node(&new->node, parent, link);
314 rb_insert_color(&new->node, &dma->pfn_list);
317 static void vfio_unlink_pfn(struct vfio_dma *dma, struct vfio_pfn *old)
319 rb_erase(&old->node, &dma->pfn_list);
322 static int vfio_add_to_pfn_list(struct vfio_dma *dma, dma_addr_t iova,
325 struct vfio_pfn *vpfn;
327 vpfn = kzalloc(sizeof(*vpfn), GFP_KERNEL);
334 vfio_link_pfn(dma, vpfn);
338 static void vfio_remove_from_pfn_list(struct vfio_dma *dma,
339 struct vfio_pfn *vpfn)
341 vfio_unlink_pfn(dma, vpfn);
345 static struct vfio_pfn *vfio_iova_get_vfio_pfn(struct vfio_dma *dma,
348 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
355 static int vfio_iova_put_vfio_pfn(struct vfio_dma *dma, struct vfio_pfn *vpfn)
360 if (!vpfn->ref_count) {
361 ret = put_pfn(vpfn->pfn, dma->prot);
362 vfio_remove_from_pfn_list(dma, vpfn);
367 static int vfio_lock_acct(struct vfio_dma *dma, long npage, bool async)
369 struct mm_struct *mm;
375 mm = async ? get_task_mm(dma->task) : dma->task->mm;
377 return -ESRCH; /* process exited */
379 ret = mmap_write_lock_killable(mm);
381 ret = __account_locked_vm(mm, abs(npage), npage > 0, dma->task,
383 mmap_write_unlock(mm);
393 * Some mappings aren't backed by a struct page, for example an mmap'd
394 * MMIO range for our own or another device. These use a different
395 * pfn conversion and shouldn't be tracked as locked pages.
396 * For compound pages, any driver that sets the reserved bit in head
397 * page needs to set the reserved bit in all subpages to be safe.
399 static bool is_invalid_reserved_pfn(unsigned long pfn)
402 return PageReserved(pfn_to_page(pfn));
407 static int put_pfn(unsigned long pfn, int prot)
409 if (!is_invalid_reserved_pfn(pfn)) {
410 struct page *page = pfn_to_page(pfn);
412 unpin_user_pages_dirty_lock(&page, 1, prot & IOMMU_WRITE);
418 static int follow_fault_pfn(struct vm_area_struct *vma, struct mm_struct *mm,
419 unsigned long vaddr, unsigned long *pfn,
424 ret = follow_pfn(vma, vaddr, pfn);
426 bool unlocked = false;
428 ret = fixup_user_fault(mm, vaddr,
430 (write_fault ? FAULT_FLAG_WRITE : 0),
438 ret = follow_pfn(vma, vaddr, pfn);
444 static int vaddr_get_pfn(struct mm_struct *mm, unsigned long vaddr,
445 int prot, unsigned long *pfn)
447 struct page *page[1];
448 struct vm_area_struct *vma;
449 unsigned int flags = 0;
452 if (prot & IOMMU_WRITE)
456 ret = pin_user_pages_remote(mm, vaddr, 1, flags | FOLL_LONGTERM,
459 *pfn = page_to_pfn(page[0]);
464 vaddr = untagged_addr(vaddr);
467 vma = find_vma_intersection(mm, vaddr, vaddr + 1);
469 if (vma && vma->vm_flags & VM_PFNMAP) {
470 ret = follow_fault_pfn(vma, mm, vaddr, pfn, prot & IOMMU_WRITE);
474 if (!ret && !is_invalid_reserved_pfn(*pfn))
478 mmap_read_unlock(mm);
483 * Attempt to pin pages. We really don't want to track all the pfns and
484 * the iommu can only map chunks of consecutive pfns anyway, so get the
485 * first page and all consecutive pages with the same locking.
487 static long vfio_pin_pages_remote(struct vfio_dma *dma, unsigned long vaddr,
488 long npage, unsigned long *pfn_base,
491 unsigned long pfn = 0;
492 long ret, pinned = 0, lock_acct = 0;
494 dma_addr_t iova = vaddr - dma->vaddr + dma->iova;
496 /* This code path is only user initiated */
500 ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, pfn_base);
505 rsvd = is_invalid_reserved_pfn(*pfn_base);
508 * Reserved pages aren't counted against the user, externally pinned
509 * pages are already counted against the user.
511 if (!rsvd && !vfio_find_vpfn(dma, iova)) {
512 if (!dma->lock_cap && current->mm->locked_vm + 1 > limit) {
513 put_pfn(*pfn_base, dma->prot);
514 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n", __func__,
515 limit << PAGE_SHIFT);
521 if (unlikely(disable_hugepages))
524 /* Lock all the consecutive pages from pfn_base */
525 for (vaddr += PAGE_SIZE, iova += PAGE_SIZE; pinned < npage;
526 pinned++, vaddr += PAGE_SIZE, iova += PAGE_SIZE) {
527 ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, &pfn);
531 if (pfn != *pfn_base + pinned ||
532 rsvd != is_invalid_reserved_pfn(pfn)) {
533 put_pfn(pfn, dma->prot);
537 if (!rsvd && !vfio_find_vpfn(dma, iova)) {
538 if (!dma->lock_cap &&
539 current->mm->locked_vm + lock_acct + 1 > limit) {
540 put_pfn(pfn, dma->prot);
541 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n",
542 __func__, limit << PAGE_SHIFT);
551 ret = vfio_lock_acct(dma, lock_acct, false);
556 for (pfn = *pfn_base ; pinned ; pfn++, pinned--)
557 put_pfn(pfn, dma->prot);
566 static long vfio_unpin_pages_remote(struct vfio_dma *dma, dma_addr_t iova,
567 unsigned long pfn, long npage,
570 long unlocked = 0, locked = 0;
573 for (i = 0; i < npage; i++, iova += PAGE_SIZE) {
574 if (put_pfn(pfn++, dma->prot)) {
576 if (vfio_find_vpfn(dma, iova))
582 vfio_lock_acct(dma, locked - unlocked, true);
587 static int vfio_pin_page_external(struct vfio_dma *dma, unsigned long vaddr,
588 unsigned long *pfn_base, bool do_accounting)
590 struct mm_struct *mm;
593 mm = get_task_mm(dma->task);
597 ret = vaddr_get_pfn(mm, vaddr, dma->prot, pfn_base);
598 if (!ret && do_accounting && !is_invalid_reserved_pfn(*pfn_base)) {
599 ret = vfio_lock_acct(dma, 1, true);
601 put_pfn(*pfn_base, dma->prot);
603 pr_warn("%s: Task %s (%d) RLIMIT_MEMLOCK "
604 "(%ld) exceeded\n", __func__,
605 dma->task->comm, task_pid_nr(dma->task),
606 task_rlimit(dma->task, RLIMIT_MEMLOCK));
614 static int vfio_unpin_page_external(struct vfio_dma *dma, dma_addr_t iova,
618 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
623 unlocked = vfio_iova_put_vfio_pfn(dma, vpfn);
626 vfio_lock_acct(dma, -unlocked, true);
631 static int vfio_iommu_type1_pin_pages(void *iommu_data,
632 struct iommu_group *iommu_group,
633 unsigned long *user_pfn,
635 unsigned long *phys_pfn)
637 struct vfio_iommu *iommu = iommu_data;
638 struct vfio_group *group;
640 unsigned long remote_vaddr;
641 struct vfio_dma *dma;
644 if (!iommu || !user_pfn || !phys_pfn)
647 /* Supported for v2 version only */
651 mutex_lock(&iommu->lock);
653 /* Fail if notifier list is empty */
654 if (!iommu->notifier.head) {
660 * If iommu capable domain exist in the container then all pages are
661 * already pinned and accounted. Accouting should be done if there is no
662 * iommu capable domain in the container.
664 do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu);
666 for (i = 0; i < npage; i++) {
668 struct vfio_pfn *vpfn;
670 iova = user_pfn[i] << PAGE_SHIFT;
671 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
677 if ((dma->prot & prot) != prot) {
682 vpfn = vfio_iova_get_vfio_pfn(dma, iova);
684 phys_pfn[i] = vpfn->pfn;
688 remote_vaddr = dma->vaddr + (iova - dma->iova);
689 ret = vfio_pin_page_external(dma, remote_vaddr, &phys_pfn[i],
694 ret = vfio_add_to_pfn_list(dma, iova, phys_pfn[i]);
696 if (put_pfn(phys_pfn[i], dma->prot) && do_accounting)
697 vfio_lock_acct(dma, -1, true);
701 if (iommu->dirty_page_tracking) {
702 unsigned long pgshift = __ffs(iommu->pgsize_bitmap);
705 * Bitmap populated with the smallest supported page
708 bitmap_set(dma->bitmap,
709 (iova - dma->iova) >> pgshift, 1);
714 group = vfio_iommu_find_iommu_group(iommu, iommu_group);
715 if (!group->pinned_page_dirty_scope) {
716 group->pinned_page_dirty_scope = true;
717 update_pinned_page_dirty_scope(iommu);
724 for (j = 0; j < i; j++) {
727 iova = user_pfn[j] << PAGE_SHIFT;
728 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
729 vfio_unpin_page_external(dma, iova, do_accounting);
733 mutex_unlock(&iommu->lock);
737 static int vfio_iommu_type1_unpin_pages(void *iommu_data,
738 unsigned long *user_pfn,
741 struct vfio_iommu *iommu = iommu_data;
745 if (!iommu || !user_pfn)
748 /* Supported for v2 version only */
752 mutex_lock(&iommu->lock);
754 do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu);
755 for (i = 0; i < npage; i++) {
756 struct vfio_dma *dma;
759 iova = user_pfn[i] << PAGE_SHIFT;
760 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
763 vfio_unpin_page_external(dma, iova, do_accounting);
767 mutex_unlock(&iommu->lock);
768 return i > npage ? npage : (i > 0 ? i : -EINVAL);
771 static long vfio_sync_unpin(struct vfio_dma *dma, struct vfio_domain *domain,
772 struct list_head *regions,
773 struct iommu_iotlb_gather *iotlb_gather)
776 struct vfio_regions *entry, *next;
778 iommu_iotlb_sync(domain->domain, iotlb_gather);
780 list_for_each_entry_safe(entry, next, regions, list) {
781 unlocked += vfio_unpin_pages_remote(dma,
783 entry->phys >> PAGE_SHIFT,
784 entry->len >> PAGE_SHIFT,
786 list_del(&entry->list);
796 * Generally, VFIO needs to unpin remote pages after each IOTLB flush.
797 * Therefore, when using IOTLB flush sync interface, VFIO need to keep track
798 * of these regions (currently using a list).
800 * This value specifies maximum number of regions for each IOTLB flush sync.
802 #define VFIO_IOMMU_TLB_SYNC_MAX 512
804 static size_t unmap_unpin_fast(struct vfio_domain *domain,
805 struct vfio_dma *dma, dma_addr_t *iova,
806 size_t len, phys_addr_t phys, long *unlocked,
807 struct list_head *unmapped_list,
809 struct iommu_iotlb_gather *iotlb_gather)
812 struct vfio_regions *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
815 unmapped = iommu_unmap_fast(domain->domain, *iova, len,
823 entry->len = unmapped;
824 list_add_tail(&entry->list, unmapped_list);
832 * Sync if the number of fast-unmap regions hits the limit
833 * or in case of errors.
835 if (*unmapped_cnt >= VFIO_IOMMU_TLB_SYNC_MAX || !unmapped) {
836 *unlocked += vfio_sync_unpin(dma, domain, unmapped_list,
844 static size_t unmap_unpin_slow(struct vfio_domain *domain,
845 struct vfio_dma *dma, dma_addr_t *iova,
846 size_t len, phys_addr_t phys,
849 size_t unmapped = iommu_unmap(domain->domain, *iova, len);
852 *unlocked += vfio_unpin_pages_remote(dma, *iova,
854 unmapped >> PAGE_SHIFT,
862 static long vfio_unmap_unpin(struct vfio_iommu *iommu, struct vfio_dma *dma,
865 dma_addr_t iova = dma->iova, end = dma->iova + dma->size;
866 struct vfio_domain *domain, *d;
867 LIST_HEAD(unmapped_region_list);
868 struct iommu_iotlb_gather iotlb_gather;
869 int unmapped_region_cnt = 0;
875 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
879 * We use the IOMMU to track the physical addresses, otherwise we'd
880 * need a much more complicated tracking system. Unfortunately that
881 * means we need to use one of the iommu domains to figure out the
882 * pfns to unpin. The rest need to be unmapped in advance so we have
883 * no iommu translations remaining when the pages are unpinned.
885 domain = d = list_first_entry(&iommu->domain_list,
886 struct vfio_domain, next);
888 list_for_each_entry_continue(d, &iommu->domain_list, next) {
889 iommu_unmap(d->domain, dma->iova, dma->size);
893 iommu_iotlb_gather_init(&iotlb_gather);
895 size_t unmapped, len;
896 phys_addr_t phys, next;
898 phys = iommu_iova_to_phys(domain->domain, iova);
899 if (WARN_ON(!phys)) {
905 * To optimize for fewer iommu_unmap() calls, each of which
906 * may require hardware cache flushing, try to find the
907 * largest contiguous physical memory chunk to unmap.
909 for (len = PAGE_SIZE;
910 !domain->fgsp && iova + len < end; len += PAGE_SIZE) {
911 next = iommu_iova_to_phys(domain->domain, iova + len);
912 if (next != phys + len)
917 * First, try to use fast unmap/unpin. In case of failure,
918 * switch to slow unmap/unpin path.
920 unmapped = unmap_unpin_fast(domain, dma, &iova, len, phys,
921 &unlocked, &unmapped_region_list,
922 &unmapped_region_cnt,
925 unmapped = unmap_unpin_slow(domain, dma, &iova, len,
927 if (WARN_ON(!unmapped))
932 dma->iommu_mapped = false;
934 if (unmapped_region_cnt) {
935 unlocked += vfio_sync_unpin(dma, domain, &unmapped_region_list,
940 vfio_lock_acct(dma, -unlocked, true);
946 static void vfio_remove_dma(struct vfio_iommu *iommu, struct vfio_dma *dma)
948 vfio_unmap_unpin(iommu, dma, true);
949 vfio_unlink_dma(iommu, dma);
950 put_task_struct(dma->task);
951 vfio_dma_bitmap_free(dma);
956 static void vfio_update_pgsize_bitmap(struct vfio_iommu *iommu)
958 struct vfio_domain *domain;
960 iommu->pgsize_bitmap = ULONG_MAX;
962 list_for_each_entry(domain, &iommu->domain_list, next)
963 iommu->pgsize_bitmap &= domain->domain->pgsize_bitmap;
966 * In case the IOMMU supports page sizes smaller than PAGE_SIZE
967 * we pretend PAGE_SIZE is supported and hide sub-PAGE_SIZE sizes.
968 * That way the user will be able to map/unmap buffers whose size/
969 * start address is aligned with PAGE_SIZE. Pinning code uses that
970 * granularity while iommu driver can use the sub-PAGE_SIZE size
973 if (iommu->pgsize_bitmap & ~PAGE_MASK) {
974 iommu->pgsize_bitmap &= PAGE_MASK;
975 iommu->pgsize_bitmap |= PAGE_SIZE;
979 static int update_user_bitmap(u64 __user *bitmap, struct vfio_iommu *iommu,
980 struct vfio_dma *dma, dma_addr_t base_iova,
983 unsigned long pgshift = __ffs(pgsize);
984 unsigned long nbits = dma->size >> pgshift;
985 unsigned long bit_offset = (dma->iova - base_iova) >> pgshift;
986 unsigned long copy_offset = bit_offset / BITS_PER_LONG;
987 unsigned long shift = bit_offset % BITS_PER_LONG;
988 unsigned long leftover;
991 * mark all pages dirty if any IOMMU capable device is not able
992 * to report dirty pages and all pages are pinned and mapped.
994 if (!iommu->pinned_page_dirty_scope && dma->iommu_mapped)
995 bitmap_set(dma->bitmap, 0, nbits);
998 bitmap_shift_left(dma->bitmap, dma->bitmap, shift,
1001 if (copy_from_user(&leftover,
1002 (void __user *)(bitmap + copy_offset),
1006 bitmap_or(dma->bitmap, dma->bitmap, &leftover, shift);
1009 if (copy_to_user((void __user *)(bitmap + copy_offset), dma->bitmap,
1010 DIRTY_BITMAP_BYTES(nbits + shift)))
1016 static int vfio_iova_dirty_bitmap(u64 __user *bitmap, struct vfio_iommu *iommu,
1017 dma_addr_t iova, size_t size, size_t pgsize)
1019 struct vfio_dma *dma;
1021 unsigned long pgshift = __ffs(pgsize);
1025 * GET_BITMAP request must fully cover vfio_dma mappings. Multiple
1026 * vfio_dma mappings may be clubbed by specifying large ranges, but
1027 * there must not be any previous mappings bisected by the range.
1028 * An error will be returned if these conditions are not met.
1030 dma = vfio_find_dma(iommu, iova, 1);
1031 if (dma && dma->iova != iova)
1034 dma = vfio_find_dma(iommu, iova + size - 1, 0);
1035 if (dma && dma->iova + dma->size != iova + size)
1038 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
1039 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1041 if (dma->iova < iova)
1044 if (dma->iova > iova + size - 1)
1047 ret = update_user_bitmap(bitmap, iommu, dma, iova, pgsize);
1052 * Re-populate bitmap to include all pinned pages which are
1053 * considered as dirty but exclude pages which are unpinned and
1054 * pages which are marked dirty by vfio_dma_rw()
1056 bitmap_clear(dma->bitmap, 0, dma->size >> pgshift);
1057 vfio_dma_populate_bitmap(dma, pgsize);
1062 static int verify_bitmap_size(uint64_t npages, uint64_t bitmap_size)
1064 if (!npages || !bitmap_size || (bitmap_size > DIRTY_BITMAP_SIZE_MAX) ||
1065 (bitmap_size < DIRTY_BITMAP_BYTES(npages)))
1071 static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
1072 struct vfio_iommu_type1_dma_unmap *unmap,
1073 struct vfio_bitmap *bitmap)
1075 struct vfio_dma *dma, *dma_last = NULL;
1076 size_t unmapped = 0, pgsize;
1077 int ret = -EINVAL, retries = 0;
1078 unsigned long pgshift;
1079 dma_addr_t iova = unmap->iova;
1080 unsigned long size = unmap->size;
1081 bool unmap_all = unmap->flags & VFIO_DMA_UNMAP_FLAG_ALL;
1083 mutex_lock(&iommu->lock);
1085 pgshift = __ffs(iommu->pgsize_bitmap);
1086 pgsize = (size_t)1 << pgshift;
1088 if (iova & (pgsize - 1))
1095 } else if (!size || size & (pgsize - 1)) {
1099 if (iova + size - 1 < iova || size > SIZE_MAX)
1102 /* When dirty tracking is enabled, allow only min supported pgsize */
1103 if ((unmap->flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) &&
1104 (!iommu->dirty_page_tracking || (bitmap->pgsize != pgsize))) {
1108 WARN_ON((pgsize - 1) & PAGE_MASK);
1111 * vfio-iommu-type1 (v1) - User mappings were coalesced together to
1112 * avoid tracking individual mappings. This means that the granularity
1113 * of the original mapping was lost and the user was allowed to attempt
1114 * to unmap any range. Depending on the contiguousness of physical
1115 * memory and page sizes supported by the IOMMU, arbitrary unmaps may
1116 * or may not have worked. We only guaranteed unmap granularity
1117 * matching the original mapping; even though it was untracked here,
1118 * the original mappings are reflected in IOMMU mappings. This
1119 * resulted in a couple unusual behaviors. First, if a range is not
1120 * able to be unmapped, ex. a set of 4k pages that was mapped as a
1121 * 2M hugepage into the IOMMU, the unmap ioctl returns success but with
1122 * a zero sized unmap. Also, if an unmap request overlaps the first
1123 * address of a hugepage, the IOMMU will unmap the entire hugepage.
1124 * This also returns success and the returned unmap size reflects the
1125 * actual size unmapped.
1127 * We attempt to maintain compatibility with this "v1" interface, but
1128 * we take control out of the hands of the IOMMU. Therefore, an unmap
1129 * request offset from the beginning of the original mapping will
1130 * return success with zero sized unmap. And an unmap request covering
1131 * the first iova of mapping will unmap the entire range.
1133 * The v2 version of this interface intends to be more deterministic.
1134 * Unmap requests must fully cover previous mappings. Multiple
1135 * mappings may still be unmaped by specifying large ranges, but there
1136 * must not be any previous mappings bisected by the range. An error
1137 * will be returned if these conditions are not met. The v2 interface
1138 * will only return success and a size of zero if there were no
1139 * mappings within the range.
1141 if (iommu->v2 && !unmap_all) {
1142 dma = vfio_find_dma(iommu, iova, 1);
1143 if (dma && dma->iova != iova)
1146 dma = vfio_find_dma(iommu, iova + size - 1, 0);
1147 if (dma && dma->iova + dma->size != iova + size)
1152 while ((dma = vfio_find_dma(iommu, iova, size))) {
1153 if (!iommu->v2 && iova > dma->iova)
1156 * Task with same address space who mapped this iova range is
1157 * allowed to unmap the iova range.
1159 if (dma->task->mm != current->mm)
1162 if (!RB_EMPTY_ROOT(&dma->pfn_list)) {
1163 struct vfio_iommu_type1_dma_unmap nb_unmap;
1165 if (dma_last == dma) {
1166 BUG_ON(++retries > 10);
1172 nb_unmap.iova = dma->iova;
1173 nb_unmap.size = dma->size;
1176 * Notify anyone (mdev vendor drivers) to invalidate and
1177 * unmap iovas within the range we're about to unmap.
1178 * Vendor drivers MUST unpin pages in response to an
1181 mutex_unlock(&iommu->lock);
1182 blocking_notifier_call_chain(&iommu->notifier,
1183 VFIO_IOMMU_NOTIFY_DMA_UNMAP,
1185 mutex_lock(&iommu->lock);
1189 if (unmap->flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) {
1190 ret = update_user_bitmap(bitmap->data, iommu, dma,
1196 unmapped += dma->size;
1197 vfio_remove_dma(iommu, dma);
1201 mutex_unlock(&iommu->lock);
1203 /* Report how much was unmapped */
1204 unmap->size = unmapped;
1209 static int vfio_iommu_map(struct vfio_iommu *iommu, dma_addr_t iova,
1210 unsigned long pfn, long npage, int prot)
1212 struct vfio_domain *d;
1215 list_for_each_entry(d, &iommu->domain_list, next) {
1216 ret = iommu_map(d->domain, iova, (phys_addr_t)pfn << PAGE_SHIFT,
1217 npage << PAGE_SHIFT, prot | d->prot);
1227 list_for_each_entry_continue_reverse(d, &iommu->domain_list, next) {
1228 iommu_unmap(d->domain, iova, npage << PAGE_SHIFT);
1235 static int vfio_pin_map_dma(struct vfio_iommu *iommu, struct vfio_dma *dma,
1238 dma_addr_t iova = dma->iova;
1239 unsigned long vaddr = dma->vaddr;
1240 size_t size = map_size;
1242 unsigned long pfn, limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1246 /* Pin a contiguous chunk of memory */
1247 npage = vfio_pin_pages_remote(dma, vaddr + dma->size,
1248 size >> PAGE_SHIFT, &pfn, limit);
1256 ret = vfio_iommu_map(iommu, iova + dma->size, pfn, npage,
1259 vfio_unpin_pages_remote(dma, iova + dma->size, pfn,
1264 size -= npage << PAGE_SHIFT;
1265 dma->size += npage << PAGE_SHIFT;
1268 dma->iommu_mapped = true;
1271 vfio_remove_dma(iommu, dma);
1277 * Check dma map request is within a valid iova range
1279 static bool vfio_iommu_iova_dma_valid(struct vfio_iommu *iommu,
1280 dma_addr_t start, dma_addr_t end)
1282 struct list_head *iova = &iommu->iova_list;
1283 struct vfio_iova *node;
1285 list_for_each_entry(node, iova, list) {
1286 if (start >= node->start && end <= node->end)
1291 * Check for list_empty() as well since a container with
1292 * a single mdev device will have an empty list.
1294 return list_empty(iova);
1297 static int vfio_dma_do_map(struct vfio_iommu *iommu,
1298 struct vfio_iommu_type1_dma_map *map)
1300 dma_addr_t iova = map->iova;
1301 unsigned long vaddr = map->vaddr;
1302 size_t size = map->size;
1303 int ret = 0, prot = 0;
1305 struct vfio_dma *dma;
1307 /* Verify that none of our __u64 fields overflow */
1308 if (map->size != size || map->vaddr != vaddr || map->iova != iova)
1311 /* READ/WRITE from device perspective */
1312 if (map->flags & VFIO_DMA_MAP_FLAG_WRITE)
1313 prot |= IOMMU_WRITE;
1314 if (map->flags & VFIO_DMA_MAP_FLAG_READ)
1317 mutex_lock(&iommu->lock);
1319 pgsize = (size_t)1 << __ffs(iommu->pgsize_bitmap);
1321 WARN_ON((pgsize - 1) & PAGE_MASK);
1323 if (!prot || !size || (size | iova | vaddr) & (pgsize - 1)) {
1328 /* Don't allow IOVA or virtual address wrap */
1329 if (iova + size - 1 < iova || vaddr + size - 1 < vaddr) {
1334 if (vfio_find_dma(iommu, iova, size)) {
1339 if (!iommu->dma_avail) {
1344 if (!vfio_iommu_iova_dma_valid(iommu, iova, iova + size - 1)) {
1349 dma = kzalloc(sizeof(*dma), GFP_KERNEL);
1361 * We need to be able to both add to a task's locked memory and test
1362 * against the locked memory limit and we need to be able to do both
1363 * outside of this call path as pinning can be asynchronous via the
1364 * external interfaces for mdev devices. RLIMIT_MEMLOCK requires a
1365 * task_struct and VM locked pages requires an mm_struct, however
1366 * holding an indefinite mm reference is not recommended, therefore we
1367 * only hold a reference to a task. We could hold a reference to
1368 * current, however QEMU uses this call path through vCPU threads,
1369 * which can be killed resulting in a NULL mm and failure in the unmap
1370 * path when called via a different thread. Avoid this problem by
1371 * using the group_leader as threads within the same group require
1372 * both CLONE_THREAD and CLONE_VM and will therefore use the same
1375 * Previously we also used the task for testing CAP_IPC_LOCK at the
1376 * time of pinning and accounting, however has_capability() makes use
1377 * of real_cred, a copy-on-write field, so we can't guarantee that it
1378 * matches group_leader, or in fact that it might not change by the
1379 * time it's evaluated. If a process were to call MAP_DMA with
1380 * CAP_IPC_LOCK but later drop it, it doesn't make sense that they
1381 * possibly see different results for an iommu_mapped vfio_dma vs
1382 * externally mapped. Therefore track CAP_IPC_LOCK in vfio_dma at the
1383 * time of calling MAP_DMA.
1385 get_task_struct(current->group_leader);
1386 dma->task = current->group_leader;
1387 dma->lock_cap = capable(CAP_IPC_LOCK);
1389 dma->pfn_list = RB_ROOT;
1391 /* Insert zero-sized and grow as we map chunks of it */
1392 vfio_link_dma(iommu, dma);
1394 /* Don't pin and map if container doesn't contain IOMMU capable domain*/
1395 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
1398 ret = vfio_pin_map_dma(iommu, dma, size);
1400 if (!ret && iommu->dirty_page_tracking) {
1401 ret = vfio_dma_bitmap_alloc(dma, pgsize);
1403 vfio_remove_dma(iommu, dma);
1407 mutex_unlock(&iommu->lock);
1411 static int vfio_bus_type(struct device *dev, void *data)
1413 struct bus_type **bus = data;
1415 if (*bus && *bus != dev->bus)
1423 static int vfio_iommu_replay(struct vfio_iommu *iommu,
1424 struct vfio_domain *domain)
1426 struct vfio_domain *d = NULL;
1428 unsigned long limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1431 /* Arbitrarily pick the first domain in the list for lookups */
1432 if (!list_empty(&iommu->domain_list))
1433 d = list_first_entry(&iommu->domain_list,
1434 struct vfio_domain, next);
1436 n = rb_first(&iommu->dma_list);
1438 for (; n; n = rb_next(n)) {
1439 struct vfio_dma *dma;
1442 dma = rb_entry(n, struct vfio_dma, node);
1445 while (iova < dma->iova + dma->size) {
1449 if (dma->iommu_mapped) {
1453 if (WARN_ON(!d)) { /* mapped w/o a domain?! */
1458 phys = iommu_iova_to_phys(d->domain, iova);
1460 if (WARN_ON(!phys)) {
1468 while (i < dma->iova + dma->size &&
1469 p == iommu_iova_to_phys(d->domain, i)) {
1476 unsigned long vaddr = dma->vaddr +
1478 size_t n = dma->iova + dma->size - iova;
1481 npage = vfio_pin_pages_remote(dma, vaddr,
1490 phys = pfn << PAGE_SHIFT;
1491 size = npage << PAGE_SHIFT;
1494 ret = iommu_map(domain->domain, iova, phys,
1495 size, dma->prot | domain->prot);
1497 if (!dma->iommu_mapped)
1498 vfio_unpin_pages_remote(dma, iova,
1509 /* All dmas are now mapped, defer to second tree walk for unwind */
1510 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
1511 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1513 dma->iommu_mapped = true;
1519 for (; n; n = rb_prev(n)) {
1520 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1523 if (dma->iommu_mapped) {
1524 iommu_unmap(domain->domain, dma->iova, dma->size);
1529 while (iova < dma->iova + dma->size) {
1530 phys_addr_t phys, p;
1534 phys = iommu_iova_to_phys(domain->domain, iova);
1543 while (i < dma->iova + dma->size &&
1544 p == iommu_iova_to_phys(domain->domain, i)) {
1550 iommu_unmap(domain->domain, iova, size);
1551 vfio_unpin_pages_remote(dma, iova, phys >> PAGE_SHIFT,
1552 size >> PAGE_SHIFT, true);
1560 * We change our unmap behavior slightly depending on whether the IOMMU
1561 * supports fine-grained superpages. IOMMUs like AMD-Vi will use a superpage
1562 * for practically any contiguous power-of-two mapping we give it. This means
1563 * we don't need to look for contiguous chunks ourselves to make unmapping
1564 * more efficient. On IOMMUs with coarse-grained super pages, like Intel VT-d
1565 * with discrete 2M/1G/512G/1T superpages, identifying contiguous chunks
1566 * significantly boosts non-hugetlbfs mappings and doesn't seem to hurt when
1567 * hugetlbfs is in use.
1569 static void vfio_test_domain_fgsp(struct vfio_domain *domain)
1572 int ret, order = get_order(PAGE_SIZE * 2);
1574 pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order);
1578 ret = iommu_map(domain->domain, 0, page_to_phys(pages), PAGE_SIZE * 2,
1579 IOMMU_READ | IOMMU_WRITE | domain->prot);
1581 size_t unmapped = iommu_unmap(domain->domain, 0, PAGE_SIZE);
1583 if (unmapped == PAGE_SIZE)
1584 iommu_unmap(domain->domain, PAGE_SIZE, PAGE_SIZE);
1586 domain->fgsp = true;
1589 __free_pages(pages, order);
1592 static struct vfio_group *find_iommu_group(struct vfio_domain *domain,
1593 struct iommu_group *iommu_group)
1595 struct vfio_group *g;
1597 list_for_each_entry(g, &domain->group_list, next) {
1598 if (g->iommu_group == iommu_group)
1605 static struct vfio_group *vfio_iommu_find_iommu_group(struct vfio_iommu *iommu,
1606 struct iommu_group *iommu_group)
1608 struct vfio_domain *domain;
1609 struct vfio_group *group = NULL;
1611 list_for_each_entry(domain, &iommu->domain_list, next) {
1612 group = find_iommu_group(domain, iommu_group);
1617 if (iommu->external_domain)
1618 group = find_iommu_group(iommu->external_domain, iommu_group);
1623 static void update_pinned_page_dirty_scope(struct vfio_iommu *iommu)
1625 struct vfio_domain *domain;
1626 struct vfio_group *group;
1628 list_for_each_entry(domain, &iommu->domain_list, next) {
1629 list_for_each_entry(group, &domain->group_list, next) {
1630 if (!group->pinned_page_dirty_scope) {
1631 iommu->pinned_page_dirty_scope = false;
1637 if (iommu->external_domain) {
1638 domain = iommu->external_domain;
1639 list_for_each_entry(group, &domain->group_list, next) {
1640 if (!group->pinned_page_dirty_scope) {
1641 iommu->pinned_page_dirty_scope = false;
1647 iommu->pinned_page_dirty_scope = true;
1650 static bool vfio_iommu_has_sw_msi(struct list_head *group_resv_regions,
1653 struct iommu_resv_region *region;
1656 list_for_each_entry(region, group_resv_regions, list) {
1658 * The presence of any 'real' MSI regions should take
1659 * precedence over the software-managed one if the
1660 * IOMMU driver happens to advertise both types.
1662 if (region->type == IOMMU_RESV_MSI) {
1667 if (region->type == IOMMU_RESV_SW_MSI) {
1668 *base = region->start;
1676 static struct device *vfio_mdev_get_iommu_device(struct device *dev)
1678 struct device *(*fn)(struct device *dev);
1679 struct device *iommu_device;
1681 fn = symbol_get(mdev_get_iommu_device);
1683 iommu_device = fn(dev);
1684 symbol_put(mdev_get_iommu_device);
1686 return iommu_device;
1692 static int vfio_mdev_attach_domain(struct device *dev, void *data)
1694 struct iommu_domain *domain = data;
1695 struct device *iommu_device;
1697 iommu_device = vfio_mdev_get_iommu_device(dev);
1699 if (iommu_dev_feature_enabled(iommu_device, IOMMU_DEV_FEAT_AUX))
1700 return iommu_aux_attach_device(domain, iommu_device);
1702 return iommu_attach_device(domain, iommu_device);
1708 static int vfio_mdev_detach_domain(struct device *dev, void *data)
1710 struct iommu_domain *domain = data;
1711 struct device *iommu_device;
1713 iommu_device = vfio_mdev_get_iommu_device(dev);
1715 if (iommu_dev_feature_enabled(iommu_device, IOMMU_DEV_FEAT_AUX))
1716 iommu_aux_detach_device(domain, iommu_device);
1718 iommu_detach_device(domain, iommu_device);
1724 static int vfio_iommu_attach_group(struct vfio_domain *domain,
1725 struct vfio_group *group)
1727 if (group->mdev_group)
1728 return iommu_group_for_each_dev(group->iommu_group,
1730 vfio_mdev_attach_domain);
1732 return iommu_attach_group(domain->domain, group->iommu_group);
1735 static void vfio_iommu_detach_group(struct vfio_domain *domain,
1736 struct vfio_group *group)
1738 if (group->mdev_group)
1739 iommu_group_for_each_dev(group->iommu_group, domain->domain,
1740 vfio_mdev_detach_domain);
1742 iommu_detach_group(domain->domain, group->iommu_group);
1745 static bool vfio_bus_is_mdev(struct bus_type *bus)
1747 struct bus_type *mdev_bus;
1750 mdev_bus = symbol_get(mdev_bus_type);
1752 ret = (bus == mdev_bus);
1753 symbol_put(mdev_bus_type);
1759 static int vfio_mdev_iommu_device(struct device *dev, void *data)
1761 struct device **old = data, *new;
1763 new = vfio_mdev_get_iommu_device(dev);
1764 if (!new || (*old && *old != new))
1773 * This is a helper function to insert an address range to iova list.
1774 * The list is initially created with a single entry corresponding to
1775 * the IOMMU domain geometry to which the device group is attached.
1776 * The list aperture gets modified when a new domain is added to the
1777 * container if the new aperture doesn't conflict with the current one
1778 * or with any existing dma mappings. The list is also modified to
1779 * exclude any reserved regions associated with the device group.
1781 static int vfio_iommu_iova_insert(struct list_head *head,
1782 dma_addr_t start, dma_addr_t end)
1784 struct vfio_iova *region;
1786 region = kmalloc(sizeof(*region), GFP_KERNEL);
1790 INIT_LIST_HEAD(®ion->list);
1791 region->start = start;
1794 list_add_tail(®ion->list, head);
1799 * Check the new iommu aperture conflicts with existing aper or with any
1800 * existing dma mappings.
1802 static bool vfio_iommu_aper_conflict(struct vfio_iommu *iommu,
1803 dma_addr_t start, dma_addr_t end)
1805 struct vfio_iova *first, *last;
1806 struct list_head *iova = &iommu->iova_list;
1808 if (list_empty(iova))
1811 /* Disjoint sets, return conflict */
1812 first = list_first_entry(iova, struct vfio_iova, list);
1813 last = list_last_entry(iova, struct vfio_iova, list);
1814 if (start > last->end || end < first->start)
1817 /* Check for any existing dma mappings below the new start */
1818 if (start > first->start) {
1819 if (vfio_find_dma(iommu, first->start, start - first->start))
1823 /* Check for any existing dma mappings beyond the new end */
1824 if (end < last->end) {
1825 if (vfio_find_dma(iommu, end + 1, last->end - end))
1833 * Resize iommu iova aperture window. This is called only if the new
1834 * aperture has no conflict with existing aperture and dma mappings.
1836 static int vfio_iommu_aper_resize(struct list_head *iova,
1837 dma_addr_t start, dma_addr_t end)
1839 struct vfio_iova *node, *next;
1841 if (list_empty(iova))
1842 return vfio_iommu_iova_insert(iova, start, end);
1844 /* Adjust iova list start */
1845 list_for_each_entry_safe(node, next, iova, list) {
1846 if (start < node->start)
1848 if (start >= node->start && start < node->end) {
1849 node->start = start;
1852 /* Delete nodes before new start */
1853 list_del(&node->list);
1857 /* Adjust iova list end */
1858 list_for_each_entry_safe(node, next, iova, list) {
1859 if (end > node->end)
1861 if (end > node->start && end <= node->end) {
1865 /* Delete nodes after new end */
1866 list_del(&node->list);
1874 * Check reserved region conflicts with existing dma mappings
1876 static bool vfio_iommu_resv_conflict(struct vfio_iommu *iommu,
1877 struct list_head *resv_regions)
1879 struct iommu_resv_region *region;
1881 /* Check for conflict with existing dma mappings */
1882 list_for_each_entry(region, resv_regions, list) {
1883 if (region->type == IOMMU_RESV_DIRECT_RELAXABLE)
1886 if (vfio_find_dma(iommu, region->start, region->length))
1894 * Check iova region overlap with reserved regions and
1895 * exclude them from the iommu iova range
1897 static int vfio_iommu_resv_exclude(struct list_head *iova,
1898 struct list_head *resv_regions)
1900 struct iommu_resv_region *resv;
1901 struct vfio_iova *n, *next;
1903 list_for_each_entry(resv, resv_regions, list) {
1904 phys_addr_t start, end;
1906 if (resv->type == IOMMU_RESV_DIRECT_RELAXABLE)
1909 start = resv->start;
1910 end = resv->start + resv->length - 1;
1912 list_for_each_entry_safe(n, next, iova, list) {
1916 if (start > n->end || end < n->start)
1919 * Insert a new node if current node overlaps with the
1920 * reserve region to exlude that from valid iova range.
1921 * Note that, new node is inserted before the current
1922 * node and finally the current node is deleted keeping
1923 * the list updated and sorted.
1925 if (start > n->start)
1926 ret = vfio_iommu_iova_insert(&n->list, n->start,
1928 if (!ret && end < n->end)
1929 ret = vfio_iommu_iova_insert(&n->list, end + 1,
1939 if (list_empty(iova))
1945 static void vfio_iommu_resv_free(struct list_head *resv_regions)
1947 struct iommu_resv_region *n, *next;
1949 list_for_each_entry_safe(n, next, resv_regions, list) {
1955 static void vfio_iommu_iova_free(struct list_head *iova)
1957 struct vfio_iova *n, *next;
1959 list_for_each_entry_safe(n, next, iova, list) {
1965 static int vfio_iommu_iova_get_copy(struct vfio_iommu *iommu,
1966 struct list_head *iova_copy)
1968 struct list_head *iova = &iommu->iova_list;
1969 struct vfio_iova *n;
1972 list_for_each_entry(n, iova, list) {
1973 ret = vfio_iommu_iova_insert(iova_copy, n->start, n->end);
1981 vfio_iommu_iova_free(iova_copy);
1985 static void vfio_iommu_iova_insert_copy(struct vfio_iommu *iommu,
1986 struct list_head *iova_copy)
1988 struct list_head *iova = &iommu->iova_list;
1990 vfio_iommu_iova_free(iova);
1992 list_splice_tail(iova_copy, iova);
1995 static int vfio_iommu_type1_attach_group(void *iommu_data,
1996 struct iommu_group *iommu_group)
1998 struct vfio_iommu *iommu = iommu_data;
1999 struct vfio_group *group;
2000 struct vfio_domain *domain, *d;
2001 struct bus_type *bus = NULL;
2003 bool resv_msi, msi_remap;
2004 phys_addr_t resv_msi_base = 0;
2005 struct iommu_domain_geometry geo;
2006 LIST_HEAD(iova_copy);
2007 LIST_HEAD(group_resv_regions);
2009 mutex_lock(&iommu->lock);
2011 /* Check for duplicates */
2012 if (vfio_iommu_find_iommu_group(iommu, iommu_group)) {
2013 mutex_unlock(&iommu->lock);
2017 group = kzalloc(sizeof(*group), GFP_KERNEL);
2018 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2019 if (!group || !domain) {
2024 group->iommu_group = iommu_group;
2026 /* Determine bus_type in order to allocate a domain */
2027 ret = iommu_group_for_each_dev(iommu_group, &bus, vfio_bus_type);
2031 if (vfio_bus_is_mdev(bus)) {
2032 struct device *iommu_device = NULL;
2034 group->mdev_group = true;
2036 /* Determine the isolation type */
2037 ret = iommu_group_for_each_dev(iommu_group, &iommu_device,
2038 vfio_mdev_iommu_device);
2039 if (ret || !iommu_device) {
2040 if (!iommu->external_domain) {
2041 INIT_LIST_HEAD(&domain->group_list);
2042 iommu->external_domain = domain;
2043 vfio_update_pgsize_bitmap(iommu);
2048 list_add(&group->next,
2049 &iommu->external_domain->group_list);
2051 * Non-iommu backed group cannot dirty memory directly,
2052 * it can only use interfaces that provide dirty
2054 * The iommu scope can only be promoted with the
2055 * addition of a dirty tracking group.
2057 group->pinned_page_dirty_scope = true;
2058 if (!iommu->pinned_page_dirty_scope)
2059 update_pinned_page_dirty_scope(iommu);
2060 mutex_unlock(&iommu->lock);
2065 bus = iommu_device->bus;
2068 domain->domain = iommu_domain_alloc(bus);
2069 if (!domain->domain) {
2074 if (iommu->nesting) {
2077 ret = iommu_domain_set_attr(domain->domain, DOMAIN_ATTR_NESTING,
2083 ret = vfio_iommu_attach_group(domain, group);
2087 /* Get aperture info */
2088 iommu_domain_get_attr(domain->domain, DOMAIN_ATTR_GEOMETRY, &geo);
2090 if (vfio_iommu_aper_conflict(iommu, geo.aperture_start,
2091 geo.aperture_end)) {
2096 ret = iommu_get_group_resv_regions(iommu_group, &group_resv_regions);
2100 if (vfio_iommu_resv_conflict(iommu, &group_resv_regions)) {
2106 * We don't want to work on the original iova list as the list
2107 * gets modified and in case of failure we have to retain the
2108 * original list. Get a copy here.
2110 ret = vfio_iommu_iova_get_copy(iommu, &iova_copy);
2114 ret = vfio_iommu_aper_resize(&iova_copy, geo.aperture_start,
2119 ret = vfio_iommu_resv_exclude(&iova_copy, &group_resv_regions);
2123 resv_msi = vfio_iommu_has_sw_msi(&group_resv_regions, &resv_msi_base);
2125 INIT_LIST_HEAD(&domain->group_list);
2126 list_add(&group->next, &domain->group_list);
2128 msi_remap = irq_domain_check_msi_remap() ||
2129 iommu_capable(bus, IOMMU_CAP_INTR_REMAP);
2131 if (!allow_unsafe_interrupts && !msi_remap) {
2132 pr_warn("%s: No interrupt remapping support. Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n",
2138 if (iommu_capable(bus, IOMMU_CAP_CACHE_COHERENCY))
2139 domain->prot |= IOMMU_CACHE;
2142 * Try to match an existing compatible domain. We don't want to
2143 * preclude an IOMMU driver supporting multiple bus_types and being
2144 * able to include different bus_types in the same IOMMU domain, so
2145 * we test whether the domains use the same iommu_ops rather than
2146 * testing if they're on the same bus_type.
2148 list_for_each_entry(d, &iommu->domain_list, next) {
2149 if (d->domain->ops == domain->domain->ops &&
2150 d->prot == domain->prot) {
2151 vfio_iommu_detach_group(domain, group);
2152 if (!vfio_iommu_attach_group(d, group)) {
2153 list_add(&group->next, &d->group_list);
2154 iommu_domain_free(domain->domain);
2159 ret = vfio_iommu_attach_group(domain, group);
2165 vfio_test_domain_fgsp(domain);
2167 /* replay mappings on new domains */
2168 ret = vfio_iommu_replay(iommu, domain);
2173 ret = iommu_get_msi_cookie(domain->domain, resv_msi_base);
2174 if (ret && ret != -ENODEV)
2178 list_add(&domain->next, &iommu->domain_list);
2179 vfio_update_pgsize_bitmap(iommu);
2181 /* Delete the old one and insert new iova list */
2182 vfio_iommu_iova_insert_copy(iommu, &iova_copy);
2185 * An iommu backed group can dirty memory directly and therefore
2186 * demotes the iommu scope until it declares itself dirty tracking
2187 * capable via the page pinning interface.
2189 iommu->pinned_page_dirty_scope = false;
2190 mutex_unlock(&iommu->lock);
2191 vfio_iommu_resv_free(&group_resv_regions);
2196 vfio_iommu_detach_group(domain, group);
2198 iommu_domain_free(domain->domain);
2199 vfio_iommu_iova_free(&iova_copy);
2200 vfio_iommu_resv_free(&group_resv_regions);
2204 mutex_unlock(&iommu->lock);
2208 static void vfio_iommu_unmap_unpin_all(struct vfio_iommu *iommu)
2210 struct rb_node *node;
2212 while ((node = rb_first(&iommu->dma_list)))
2213 vfio_remove_dma(iommu, rb_entry(node, struct vfio_dma, node));
2216 static void vfio_iommu_unmap_unpin_reaccount(struct vfio_iommu *iommu)
2218 struct rb_node *n, *p;
2220 n = rb_first(&iommu->dma_list);
2221 for (; n; n = rb_next(n)) {
2222 struct vfio_dma *dma;
2223 long locked = 0, unlocked = 0;
2225 dma = rb_entry(n, struct vfio_dma, node);
2226 unlocked += vfio_unmap_unpin(iommu, dma, false);
2227 p = rb_first(&dma->pfn_list);
2228 for (; p; p = rb_next(p)) {
2229 struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn,
2232 if (!is_invalid_reserved_pfn(vpfn->pfn))
2235 vfio_lock_acct(dma, locked - unlocked, true);
2239 static void vfio_sanity_check_pfn_list(struct vfio_iommu *iommu)
2243 n = rb_first(&iommu->dma_list);
2244 for (; n; n = rb_next(n)) {
2245 struct vfio_dma *dma;
2247 dma = rb_entry(n, struct vfio_dma, node);
2249 if (WARN_ON(!RB_EMPTY_ROOT(&dma->pfn_list)))
2252 /* mdev vendor driver must unregister notifier */
2253 WARN_ON(iommu->notifier.head);
2257 * Called when a domain is removed in detach. It is possible that
2258 * the removed domain decided the iova aperture window. Modify the
2259 * iova aperture with the smallest window among existing domains.
2261 static void vfio_iommu_aper_expand(struct vfio_iommu *iommu,
2262 struct list_head *iova_copy)
2264 struct vfio_domain *domain;
2265 struct iommu_domain_geometry geo;
2266 struct vfio_iova *node;
2267 dma_addr_t start = 0;
2268 dma_addr_t end = (dma_addr_t)~0;
2270 if (list_empty(iova_copy))
2273 list_for_each_entry(domain, &iommu->domain_list, next) {
2274 iommu_domain_get_attr(domain->domain, DOMAIN_ATTR_GEOMETRY,
2276 if (geo.aperture_start > start)
2277 start = geo.aperture_start;
2278 if (geo.aperture_end < end)
2279 end = geo.aperture_end;
2282 /* Modify aperture limits. The new aper is either same or bigger */
2283 node = list_first_entry(iova_copy, struct vfio_iova, list);
2284 node->start = start;
2285 node = list_last_entry(iova_copy, struct vfio_iova, list);
2290 * Called when a group is detached. The reserved regions for that
2291 * group can be part of valid iova now. But since reserved regions
2292 * may be duplicated among groups, populate the iova valid regions
2295 static int vfio_iommu_resv_refresh(struct vfio_iommu *iommu,
2296 struct list_head *iova_copy)
2298 struct vfio_domain *d;
2299 struct vfio_group *g;
2300 struct vfio_iova *node;
2301 dma_addr_t start, end;
2302 LIST_HEAD(resv_regions);
2305 if (list_empty(iova_copy))
2308 list_for_each_entry(d, &iommu->domain_list, next) {
2309 list_for_each_entry(g, &d->group_list, next) {
2310 ret = iommu_get_group_resv_regions(g->iommu_group,
2317 node = list_first_entry(iova_copy, struct vfio_iova, list);
2318 start = node->start;
2319 node = list_last_entry(iova_copy, struct vfio_iova, list);
2322 /* purge the iova list and create new one */
2323 vfio_iommu_iova_free(iova_copy);
2325 ret = vfio_iommu_aper_resize(iova_copy, start, end);
2329 /* Exclude current reserved regions from iova ranges */
2330 ret = vfio_iommu_resv_exclude(iova_copy, &resv_regions);
2332 vfio_iommu_resv_free(&resv_regions);
2336 static void vfio_iommu_type1_detach_group(void *iommu_data,
2337 struct iommu_group *iommu_group)
2339 struct vfio_iommu *iommu = iommu_data;
2340 struct vfio_domain *domain;
2341 struct vfio_group *group;
2342 bool update_dirty_scope = false;
2343 LIST_HEAD(iova_copy);
2345 mutex_lock(&iommu->lock);
2347 if (iommu->external_domain) {
2348 group = find_iommu_group(iommu->external_domain, iommu_group);
2350 update_dirty_scope = !group->pinned_page_dirty_scope;
2351 list_del(&group->next);
2354 if (list_empty(&iommu->external_domain->group_list)) {
2355 vfio_sanity_check_pfn_list(iommu);
2357 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
2358 vfio_iommu_unmap_unpin_all(iommu);
2360 kfree(iommu->external_domain);
2361 iommu->external_domain = NULL;
2363 goto detach_group_done;
2368 * Get a copy of iova list. This will be used to update
2369 * and to replace the current one later. Please note that
2370 * we will leave the original list as it is if update fails.
2372 vfio_iommu_iova_get_copy(iommu, &iova_copy);
2374 list_for_each_entry(domain, &iommu->domain_list, next) {
2375 group = find_iommu_group(domain, iommu_group);
2379 vfio_iommu_detach_group(domain, group);
2380 update_dirty_scope = !group->pinned_page_dirty_scope;
2381 list_del(&group->next);
2384 * Group ownership provides privilege, if the group list is
2385 * empty, the domain goes away. If it's the last domain with
2386 * iommu and external domain doesn't exist, then all the
2387 * mappings go away too. If it's the last domain with iommu and
2388 * external domain exist, update accounting
2390 if (list_empty(&domain->group_list)) {
2391 if (list_is_singular(&iommu->domain_list)) {
2392 if (!iommu->external_domain)
2393 vfio_iommu_unmap_unpin_all(iommu);
2395 vfio_iommu_unmap_unpin_reaccount(iommu);
2397 iommu_domain_free(domain->domain);
2398 list_del(&domain->next);
2400 vfio_iommu_aper_expand(iommu, &iova_copy);
2401 vfio_update_pgsize_bitmap(iommu);
2406 if (!vfio_iommu_resv_refresh(iommu, &iova_copy))
2407 vfio_iommu_iova_insert_copy(iommu, &iova_copy);
2409 vfio_iommu_iova_free(&iova_copy);
2413 * Removal of a group without dirty tracking may allow the iommu scope
2416 if (update_dirty_scope)
2417 update_pinned_page_dirty_scope(iommu);
2418 mutex_unlock(&iommu->lock);
2421 static void *vfio_iommu_type1_open(unsigned long arg)
2423 struct vfio_iommu *iommu;
2425 iommu = kzalloc(sizeof(*iommu), GFP_KERNEL);
2427 return ERR_PTR(-ENOMEM);
2430 case VFIO_TYPE1_IOMMU:
2432 case VFIO_TYPE1_NESTING_IOMMU:
2433 iommu->nesting = true;
2435 case VFIO_TYPE1v2_IOMMU:
2440 return ERR_PTR(-EINVAL);
2443 INIT_LIST_HEAD(&iommu->domain_list);
2444 INIT_LIST_HEAD(&iommu->iova_list);
2445 iommu->dma_list = RB_ROOT;
2446 iommu->dma_avail = dma_entry_limit;
2447 mutex_init(&iommu->lock);
2448 BLOCKING_INIT_NOTIFIER_HEAD(&iommu->notifier);
2453 static void vfio_release_domain(struct vfio_domain *domain, bool external)
2455 struct vfio_group *group, *group_tmp;
2457 list_for_each_entry_safe(group, group_tmp,
2458 &domain->group_list, next) {
2460 vfio_iommu_detach_group(domain, group);
2461 list_del(&group->next);
2466 iommu_domain_free(domain->domain);
2469 static void vfio_iommu_type1_release(void *iommu_data)
2471 struct vfio_iommu *iommu = iommu_data;
2472 struct vfio_domain *domain, *domain_tmp;
2474 if (iommu->external_domain) {
2475 vfio_release_domain(iommu->external_domain, true);
2476 vfio_sanity_check_pfn_list(iommu);
2477 kfree(iommu->external_domain);
2480 vfio_iommu_unmap_unpin_all(iommu);
2482 list_for_each_entry_safe(domain, domain_tmp,
2483 &iommu->domain_list, next) {
2484 vfio_release_domain(domain, false);
2485 list_del(&domain->next);
2489 vfio_iommu_iova_free(&iommu->iova_list);
2494 static int vfio_domains_have_iommu_cache(struct vfio_iommu *iommu)
2496 struct vfio_domain *domain;
2499 mutex_lock(&iommu->lock);
2500 list_for_each_entry(domain, &iommu->domain_list, next) {
2501 if (!(domain->prot & IOMMU_CACHE)) {
2506 mutex_unlock(&iommu->lock);
2511 static int vfio_iommu_type1_check_extension(struct vfio_iommu *iommu,
2515 case VFIO_TYPE1_IOMMU:
2516 case VFIO_TYPE1v2_IOMMU:
2517 case VFIO_TYPE1_NESTING_IOMMU:
2518 case VFIO_UNMAP_ALL:
2520 case VFIO_DMA_CC_IOMMU:
2523 return vfio_domains_have_iommu_cache(iommu);
2529 static int vfio_iommu_iova_add_cap(struct vfio_info_cap *caps,
2530 struct vfio_iommu_type1_info_cap_iova_range *cap_iovas,
2533 struct vfio_info_cap_header *header;
2534 struct vfio_iommu_type1_info_cap_iova_range *iova_cap;
2536 header = vfio_info_cap_add(caps, size,
2537 VFIO_IOMMU_TYPE1_INFO_CAP_IOVA_RANGE, 1);
2539 return PTR_ERR(header);
2541 iova_cap = container_of(header,
2542 struct vfio_iommu_type1_info_cap_iova_range,
2544 iova_cap->nr_iovas = cap_iovas->nr_iovas;
2545 memcpy(iova_cap->iova_ranges, cap_iovas->iova_ranges,
2546 cap_iovas->nr_iovas * sizeof(*cap_iovas->iova_ranges));
2550 static int vfio_iommu_iova_build_caps(struct vfio_iommu *iommu,
2551 struct vfio_info_cap *caps)
2553 struct vfio_iommu_type1_info_cap_iova_range *cap_iovas;
2554 struct vfio_iova *iova;
2556 int iovas = 0, i = 0, ret;
2558 list_for_each_entry(iova, &iommu->iova_list, list)
2563 * Return 0 as a container with a single mdev device
2564 * will have an empty list
2569 size = sizeof(*cap_iovas) + (iovas * sizeof(*cap_iovas->iova_ranges));
2571 cap_iovas = kzalloc(size, GFP_KERNEL);
2575 cap_iovas->nr_iovas = iovas;
2577 list_for_each_entry(iova, &iommu->iova_list, list) {
2578 cap_iovas->iova_ranges[i].start = iova->start;
2579 cap_iovas->iova_ranges[i].end = iova->end;
2583 ret = vfio_iommu_iova_add_cap(caps, cap_iovas, size);
2589 static int vfio_iommu_migration_build_caps(struct vfio_iommu *iommu,
2590 struct vfio_info_cap *caps)
2592 struct vfio_iommu_type1_info_cap_migration cap_mig;
2594 cap_mig.header.id = VFIO_IOMMU_TYPE1_INFO_CAP_MIGRATION;
2595 cap_mig.header.version = 1;
2598 /* support minimum pgsize */
2599 cap_mig.pgsize_bitmap = (size_t)1 << __ffs(iommu->pgsize_bitmap);
2600 cap_mig.max_dirty_bitmap_size = DIRTY_BITMAP_SIZE_MAX;
2602 return vfio_info_add_capability(caps, &cap_mig.header, sizeof(cap_mig));
2605 static int vfio_iommu_dma_avail_build_caps(struct vfio_iommu *iommu,
2606 struct vfio_info_cap *caps)
2608 struct vfio_iommu_type1_info_dma_avail cap_dma_avail;
2610 cap_dma_avail.header.id = VFIO_IOMMU_TYPE1_INFO_DMA_AVAIL;
2611 cap_dma_avail.header.version = 1;
2613 cap_dma_avail.avail = iommu->dma_avail;
2615 return vfio_info_add_capability(caps, &cap_dma_avail.header,
2616 sizeof(cap_dma_avail));
2619 static int vfio_iommu_type1_get_info(struct vfio_iommu *iommu,
2622 struct vfio_iommu_type1_info info;
2623 unsigned long minsz;
2624 struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
2625 unsigned long capsz;
2628 minsz = offsetofend(struct vfio_iommu_type1_info, iova_pgsizes);
2630 /* For backward compatibility, cannot require this */
2631 capsz = offsetofend(struct vfio_iommu_type1_info, cap_offset);
2633 if (copy_from_user(&info, (void __user *)arg, minsz))
2636 if (info.argsz < minsz)
2639 if (info.argsz >= capsz) {
2641 info.cap_offset = 0; /* output, no-recopy necessary */
2644 mutex_lock(&iommu->lock);
2645 info.flags = VFIO_IOMMU_INFO_PGSIZES;
2647 info.iova_pgsizes = iommu->pgsize_bitmap;
2649 ret = vfio_iommu_migration_build_caps(iommu, &caps);
2652 ret = vfio_iommu_dma_avail_build_caps(iommu, &caps);
2655 ret = vfio_iommu_iova_build_caps(iommu, &caps);
2657 mutex_unlock(&iommu->lock);
2663 info.flags |= VFIO_IOMMU_INFO_CAPS;
2665 if (info.argsz < sizeof(info) + caps.size) {
2666 info.argsz = sizeof(info) + caps.size;
2668 vfio_info_cap_shift(&caps, sizeof(info));
2669 if (copy_to_user((void __user *)arg +
2670 sizeof(info), caps.buf,
2675 info.cap_offset = sizeof(info);
2681 return copy_to_user((void __user *)arg, &info, minsz) ?
2685 static int vfio_iommu_type1_map_dma(struct vfio_iommu *iommu,
2688 struct vfio_iommu_type1_dma_map map;
2689 unsigned long minsz;
2690 uint32_t mask = VFIO_DMA_MAP_FLAG_READ | VFIO_DMA_MAP_FLAG_WRITE;
2692 minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
2694 if (copy_from_user(&map, (void __user *)arg, minsz))
2697 if (map.argsz < minsz || map.flags & ~mask)
2700 return vfio_dma_do_map(iommu, &map);
2703 static int vfio_iommu_type1_unmap_dma(struct vfio_iommu *iommu,
2706 struct vfio_iommu_type1_dma_unmap unmap;
2707 struct vfio_bitmap bitmap = { 0 };
2708 uint32_t mask = VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP |
2709 VFIO_DMA_UNMAP_FLAG_ALL;
2710 unsigned long minsz;
2713 minsz = offsetofend(struct vfio_iommu_type1_dma_unmap, size);
2715 if (copy_from_user(&unmap, (void __user *)arg, minsz))
2718 if (unmap.argsz < minsz || unmap.flags & ~mask)
2721 if ((unmap.flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) &&
2722 (unmap.flags & VFIO_DMA_UNMAP_FLAG_ALL))
2725 if (unmap.flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) {
2726 unsigned long pgshift;
2728 if (unmap.argsz < (minsz + sizeof(bitmap)))
2731 if (copy_from_user(&bitmap,
2732 (void __user *)(arg + minsz),
2736 if (!access_ok((void __user *)bitmap.data, bitmap.size))
2739 pgshift = __ffs(bitmap.pgsize);
2740 ret = verify_bitmap_size(unmap.size >> pgshift,
2746 ret = vfio_dma_do_unmap(iommu, &unmap, &bitmap);
2750 return copy_to_user((void __user *)arg, &unmap, minsz) ?
2754 static int vfio_iommu_type1_dirty_pages(struct vfio_iommu *iommu,
2757 struct vfio_iommu_type1_dirty_bitmap dirty;
2758 uint32_t mask = VFIO_IOMMU_DIRTY_PAGES_FLAG_START |
2759 VFIO_IOMMU_DIRTY_PAGES_FLAG_STOP |
2760 VFIO_IOMMU_DIRTY_PAGES_FLAG_GET_BITMAP;
2761 unsigned long minsz;
2767 minsz = offsetofend(struct vfio_iommu_type1_dirty_bitmap, flags);
2769 if (copy_from_user(&dirty, (void __user *)arg, minsz))
2772 if (dirty.argsz < minsz || dirty.flags & ~mask)
2775 /* only one flag should be set at a time */
2776 if (__ffs(dirty.flags) != __fls(dirty.flags))
2779 if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_START) {
2782 mutex_lock(&iommu->lock);
2783 pgsize = 1 << __ffs(iommu->pgsize_bitmap);
2784 if (!iommu->dirty_page_tracking) {
2785 ret = vfio_dma_bitmap_alloc_all(iommu, pgsize);
2787 iommu->dirty_page_tracking = true;
2789 mutex_unlock(&iommu->lock);
2791 } else if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_STOP) {
2792 mutex_lock(&iommu->lock);
2793 if (iommu->dirty_page_tracking) {
2794 iommu->dirty_page_tracking = false;
2795 vfio_dma_bitmap_free_all(iommu);
2797 mutex_unlock(&iommu->lock);
2799 } else if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_GET_BITMAP) {
2800 struct vfio_iommu_type1_dirty_bitmap_get range;
2801 unsigned long pgshift;
2802 size_t data_size = dirty.argsz - minsz;
2803 size_t iommu_pgsize;
2805 if (!data_size || data_size < sizeof(range))
2808 if (copy_from_user(&range, (void __user *)(arg + minsz),
2812 if (range.iova + range.size < range.iova)
2814 if (!access_ok((void __user *)range.bitmap.data,
2818 pgshift = __ffs(range.bitmap.pgsize);
2819 ret = verify_bitmap_size(range.size >> pgshift,
2824 mutex_lock(&iommu->lock);
2826 iommu_pgsize = (size_t)1 << __ffs(iommu->pgsize_bitmap);
2828 /* allow only smallest supported pgsize */
2829 if (range.bitmap.pgsize != iommu_pgsize) {
2833 if (range.iova & (iommu_pgsize - 1)) {
2837 if (!range.size || range.size & (iommu_pgsize - 1)) {
2842 if (iommu->dirty_page_tracking)
2843 ret = vfio_iova_dirty_bitmap(range.bitmap.data,
2846 range.bitmap.pgsize);
2850 mutex_unlock(&iommu->lock);
2858 static long vfio_iommu_type1_ioctl(void *iommu_data,
2859 unsigned int cmd, unsigned long arg)
2861 struct vfio_iommu *iommu = iommu_data;
2864 case VFIO_CHECK_EXTENSION:
2865 return vfio_iommu_type1_check_extension(iommu, arg);
2866 case VFIO_IOMMU_GET_INFO:
2867 return vfio_iommu_type1_get_info(iommu, arg);
2868 case VFIO_IOMMU_MAP_DMA:
2869 return vfio_iommu_type1_map_dma(iommu, arg);
2870 case VFIO_IOMMU_UNMAP_DMA:
2871 return vfio_iommu_type1_unmap_dma(iommu, arg);
2872 case VFIO_IOMMU_DIRTY_PAGES:
2873 return vfio_iommu_type1_dirty_pages(iommu, arg);
2879 static int vfio_iommu_type1_register_notifier(void *iommu_data,
2880 unsigned long *events,
2881 struct notifier_block *nb)
2883 struct vfio_iommu *iommu = iommu_data;
2885 /* clear known events */
2886 *events &= ~VFIO_IOMMU_NOTIFY_DMA_UNMAP;
2888 /* refuse to register if still events remaining */
2892 return blocking_notifier_chain_register(&iommu->notifier, nb);
2895 static int vfio_iommu_type1_unregister_notifier(void *iommu_data,
2896 struct notifier_block *nb)
2898 struct vfio_iommu *iommu = iommu_data;
2900 return blocking_notifier_chain_unregister(&iommu->notifier, nb);
2903 static int vfio_iommu_type1_dma_rw_chunk(struct vfio_iommu *iommu,
2904 dma_addr_t user_iova, void *data,
2905 size_t count, bool write,
2908 struct mm_struct *mm;
2909 unsigned long vaddr;
2910 struct vfio_dma *dma;
2911 bool kthread = current->mm == NULL;
2916 dma = vfio_find_dma(iommu, user_iova, 1);
2920 if ((write && !(dma->prot & IOMMU_WRITE)) ||
2921 !(dma->prot & IOMMU_READ))
2924 mm = get_task_mm(dma->task);
2931 else if (current->mm != mm)
2934 offset = user_iova - dma->iova;
2936 if (count > dma->size - offset)
2937 count = dma->size - offset;
2939 vaddr = dma->vaddr + offset;
2942 *copied = copy_to_user((void __user *)vaddr, data,
2944 if (*copied && iommu->dirty_page_tracking) {
2945 unsigned long pgshift = __ffs(iommu->pgsize_bitmap);
2947 * Bitmap populated with the smallest supported page
2950 bitmap_set(dma->bitmap, offset >> pgshift,
2951 ((offset + *copied - 1) >> pgshift) -
2952 (offset >> pgshift) + 1);
2955 *copied = copy_from_user(data, (void __user *)vaddr,
2958 kthread_unuse_mm(mm);
2961 return *copied ? 0 : -EFAULT;
2964 static int vfio_iommu_type1_dma_rw(void *iommu_data, dma_addr_t user_iova,
2965 void *data, size_t count, bool write)
2967 struct vfio_iommu *iommu = iommu_data;
2971 mutex_lock(&iommu->lock);
2973 ret = vfio_iommu_type1_dma_rw_chunk(iommu, user_iova, data,
2974 count, write, &done);
2983 mutex_unlock(&iommu->lock);
2987 static struct iommu_domain *
2988 vfio_iommu_type1_group_iommu_domain(void *iommu_data,
2989 struct iommu_group *iommu_group)
2991 struct iommu_domain *domain = ERR_PTR(-ENODEV);
2992 struct vfio_iommu *iommu = iommu_data;
2993 struct vfio_domain *d;
2995 if (!iommu || !iommu_group)
2996 return ERR_PTR(-EINVAL);
2998 mutex_lock(&iommu->lock);
2999 list_for_each_entry(d, &iommu->domain_list, next) {
3000 if (find_iommu_group(d, iommu_group)) {
3005 mutex_unlock(&iommu->lock);
3010 static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = {
3011 .name = "vfio-iommu-type1",
3012 .owner = THIS_MODULE,
3013 .open = vfio_iommu_type1_open,
3014 .release = vfio_iommu_type1_release,
3015 .ioctl = vfio_iommu_type1_ioctl,
3016 .attach_group = vfio_iommu_type1_attach_group,
3017 .detach_group = vfio_iommu_type1_detach_group,
3018 .pin_pages = vfio_iommu_type1_pin_pages,
3019 .unpin_pages = vfio_iommu_type1_unpin_pages,
3020 .register_notifier = vfio_iommu_type1_register_notifier,
3021 .unregister_notifier = vfio_iommu_type1_unregister_notifier,
3022 .dma_rw = vfio_iommu_type1_dma_rw,
3023 .group_iommu_domain = vfio_iommu_type1_group_iommu_domain,
3026 static int __init vfio_iommu_type1_init(void)
3028 return vfio_register_iommu_driver(&vfio_iommu_driver_ops_type1);
3031 static void __exit vfio_iommu_type1_cleanup(void)
3033 vfio_unregister_iommu_driver(&vfio_iommu_driver_ops_type1);
3036 module_init(vfio_iommu_type1_init);
3037 module_exit(vfio_iommu_type1_cleanup);
3039 MODULE_VERSION(DRIVER_VERSION);
3040 MODULE_LICENSE("GPL v2");
3041 MODULE_AUTHOR(DRIVER_AUTHOR);
3042 MODULE_DESCRIPTION(DRIVER_DESC);