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;
73 uint64_t num_non_pinned_groups;
76 bool dirty_page_tracking;
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);
152 * This code handles mapping and unmapping of user data buffers
153 * into DMA'ble space using the IOMMU
156 static struct vfio_dma *vfio_find_dma(struct vfio_iommu *iommu,
157 dma_addr_t start, size_t size)
159 struct rb_node *node = iommu->dma_list.rb_node;
162 struct vfio_dma *dma = rb_entry(node, struct vfio_dma, node);
164 if (start + size <= dma->iova)
165 node = node->rb_left;
166 else if (start >= dma->iova + dma->size)
167 node = node->rb_right;
175 static void vfio_link_dma(struct vfio_iommu *iommu, struct vfio_dma *new)
177 struct rb_node **link = &iommu->dma_list.rb_node, *parent = NULL;
178 struct vfio_dma *dma;
182 dma = rb_entry(parent, struct vfio_dma, node);
184 if (new->iova + new->size <= dma->iova)
185 link = &(*link)->rb_left;
187 link = &(*link)->rb_right;
190 rb_link_node(&new->node, parent, link);
191 rb_insert_color(&new->node, &iommu->dma_list);
194 static void vfio_unlink_dma(struct vfio_iommu *iommu, struct vfio_dma *old)
196 rb_erase(&old->node, &iommu->dma_list);
200 static int vfio_dma_bitmap_alloc(struct vfio_dma *dma, size_t pgsize)
202 uint64_t npages = dma->size / pgsize;
204 if (npages > DIRTY_BITMAP_PAGES_MAX)
208 * Allocate extra 64 bits that are used to calculate shift required for
209 * bitmap_shift_left() to manipulate and club unaligned number of pages
210 * in adjacent vfio_dma ranges.
212 dma->bitmap = kvzalloc(DIRTY_BITMAP_BYTES(npages) + sizeof(u64),
220 static void vfio_dma_bitmap_free(struct vfio_dma *dma)
226 static void vfio_dma_populate_bitmap(struct vfio_dma *dma, size_t pgsize)
229 unsigned long pgshift = __ffs(pgsize);
231 for (p = rb_first(&dma->pfn_list); p; p = rb_next(p)) {
232 struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn, node);
234 bitmap_set(dma->bitmap, (vpfn->iova - dma->iova) >> pgshift, 1);
238 static void vfio_iommu_populate_bitmap_full(struct vfio_iommu *iommu)
241 unsigned long pgshift = __ffs(iommu->pgsize_bitmap);
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);
246 bitmap_set(dma->bitmap, 0, dma->size >> pgshift);
250 static int vfio_dma_bitmap_alloc_all(struct vfio_iommu *iommu, size_t pgsize)
254 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
255 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
258 ret = vfio_dma_bitmap_alloc(dma, pgsize);
262 for (p = rb_prev(n); p; p = rb_prev(p)) {
263 struct vfio_dma *dma = rb_entry(n,
264 struct vfio_dma, node);
266 vfio_dma_bitmap_free(dma);
270 vfio_dma_populate_bitmap(dma, pgsize);
275 static void vfio_dma_bitmap_free_all(struct vfio_iommu *iommu)
279 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
280 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
282 vfio_dma_bitmap_free(dma);
287 * Helper Functions for host iova-pfn list
289 static struct vfio_pfn *vfio_find_vpfn(struct vfio_dma *dma, dma_addr_t iova)
291 struct vfio_pfn *vpfn;
292 struct rb_node *node = dma->pfn_list.rb_node;
295 vpfn = rb_entry(node, struct vfio_pfn, node);
297 if (iova < vpfn->iova)
298 node = node->rb_left;
299 else if (iova > vpfn->iova)
300 node = node->rb_right;
307 static void vfio_link_pfn(struct vfio_dma *dma,
308 struct vfio_pfn *new)
310 struct rb_node **link, *parent = NULL;
311 struct vfio_pfn *vpfn;
313 link = &dma->pfn_list.rb_node;
316 vpfn = rb_entry(parent, struct vfio_pfn, node);
318 if (new->iova < vpfn->iova)
319 link = &(*link)->rb_left;
321 link = &(*link)->rb_right;
324 rb_link_node(&new->node, parent, link);
325 rb_insert_color(&new->node, &dma->pfn_list);
328 static void vfio_unlink_pfn(struct vfio_dma *dma, struct vfio_pfn *old)
330 rb_erase(&old->node, &dma->pfn_list);
333 static int vfio_add_to_pfn_list(struct vfio_dma *dma, dma_addr_t iova,
336 struct vfio_pfn *vpfn;
338 vpfn = kzalloc(sizeof(*vpfn), GFP_KERNEL);
345 vfio_link_pfn(dma, vpfn);
349 static void vfio_remove_from_pfn_list(struct vfio_dma *dma,
350 struct vfio_pfn *vpfn)
352 vfio_unlink_pfn(dma, vpfn);
356 static struct vfio_pfn *vfio_iova_get_vfio_pfn(struct vfio_dma *dma,
359 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
366 static int vfio_iova_put_vfio_pfn(struct vfio_dma *dma, struct vfio_pfn *vpfn)
371 if (!vpfn->ref_count) {
372 ret = put_pfn(vpfn->pfn, dma->prot);
373 vfio_remove_from_pfn_list(dma, vpfn);
378 static int vfio_lock_acct(struct vfio_dma *dma, long npage, bool async)
380 struct mm_struct *mm;
386 mm = async ? get_task_mm(dma->task) : dma->task->mm;
388 return -ESRCH; /* process exited */
390 ret = mmap_write_lock_killable(mm);
392 ret = __account_locked_vm(mm, abs(npage), npage > 0, dma->task,
394 mmap_write_unlock(mm);
404 * Some mappings aren't backed by a struct page, for example an mmap'd
405 * MMIO range for our own or another device. These use a different
406 * pfn conversion and shouldn't be tracked as locked pages.
407 * For compound pages, any driver that sets the reserved bit in head
408 * page needs to set the reserved bit in all subpages to be safe.
410 static bool is_invalid_reserved_pfn(unsigned long pfn)
413 return PageReserved(pfn_to_page(pfn));
418 static int put_pfn(unsigned long pfn, int prot)
420 if (!is_invalid_reserved_pfn(pfn)) {
421 struct page *page = pfn_to_page(pfn);
423 unpin_user_pages_dirty_lock(&page, 1, prot & IOMMU_WRITE);
429 static int follow_fault_pfn(struct vm_area_struct *vma, struct mm_struct *mm,
430 unsigned long vaddr, unsigned long *pfn,
435 ret = follow_pfn(vma, vaddr, pfn);
437 bool unlocked = false;
439 ret = fixup_user_fault(mm, vaddr,
441 (write_fault ? FAULT_FLAG_WRITE : 0),
449 ret = follow_pfn(vma, vaddr, pfn);
455 static int vaddr_get_pfn(struct mm_struct *mm, unsigned long vaddr,
456 int prot, unsigned long *pfn)
458 struct page *page[1];
459 struct vm_area_struct *vma;
460 unsigned int flags = 0;
463 if (prot & IOMMU_WRITE)
467 ret = pin_user_pages_remote(mm, vaddr, 1, flags | FOLL_LONGTERM,
470 *pfn = page_to_pfn(page[0]);
475 vaddr = untagged_addr(vaddr);
478 vma = find_vma_intersection(mm, vaddr, vaddr + 1);
480 if (vma && vma->vm_flags & VM_PFNMAP) {
481 ret = follow_fault_pfn(vma, mm, vaddr, pfn, prot & IOMMU_WRITE);
485 if (!ret && !is_invalid_reserved_pfn(*pfn))
489 mmap_read_unlock(mm);
494 * Attempt to pin pages. We really don't want to track all the pfns and
495 * the iommu can only map chunks of consecutive pfns anyway, so get the
496 * first page and all consecutive pages with the same locking.
498 static long vfio_pin_pages_remote(struct vfio_dma *dma, unsigned long vaddr,
499 long npage, unsigned long *pfn_base,
502 unsigned long pfn = 0;
503 long ret, pinned = 0, lock_acct = 0;
505 dma_addr_t iova = vaddr - dma->vaddr + dma->iova;
507 /* This code path is only user initiated */
511 ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, pfn_base);
516 rsvd = is_invalid_reserved_pfn(*pfn_base);
519 * Reserved pages aren't counted against the user, externally pinned
520 * pages are already counted against the user.
522 if (!rsvd && !vfio_find_vpfn(dma, iova)) {
523 if (!dma->lock_cap && current->mm->locked_vm + 1 > limit) {
524 put_pfn(*pfn_base, dma->prot);
525 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n", __func__,
526 limit << PAGE_SHIFT);
532 if (unlikely(disable_hugepages))
535 /* Lock all the consecutive pages from pfn_base */
536 for (vaddr += PAGE_SIZE, iova += PAGE_SIZE; pinned < npage;
537 pinned++, vaddr += PAGE_SIZE, iova += PAGE_SIZE) {
538 ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, &pfn);
542 if (pfn != *pfn_base + pinned ||
543 rsvd != is_invalid_reserved_pfn(pfn)) {
544 put_pfn(pfn, dma->prot);
548 if (!rsvd && !vfio_find_vpfn(dma, iova)) {
549 if (!dma->lock_cap &&
550 current->mm->locked_vm + lock_acct + 1 > limit) {
551 put_pfn(pfn, dma->prot);
552 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n",
553 __func__, limit << PAGE_SHIFT);
562 ret = vfio_lock_acct(dma, lock_acct, false);
567 for (pfn = *pfn_base ; pinned ; pfn++, pinned--)
568 put_pfn(pfn, dma->prot);
577 static long vfio_unpin_pages_remote(struct vfio_dma *dma, dma_addr_t iova,
578 unsigned long pfn, long npage,
581 long unlocked = 0, locked = 0;
584 for (i = 0; i < npage; i++, iova += PAGE_SIZE) {
585 if (put_pfn(pfn++, dma->prot)) {
587 if (vfio_find_vpfn(dma, iova))
593 vfio_lock_acct(dma, locked - unlocked, true);
598 static int vfio_pin_page_external(struct vfio_dma *dma, unsigned long vaddr,
599 unsigned long *pfn_base, bool do_accounting)
601 struct mm_struct *mm;
604 mm = get_task_mm(dma->task);
608 ret = vaddr_get_pfn(mm, vaddr, dma->prot, pfn_base);
609 if (!ret && do_accounting && !is_invalid_reserved_pfn(*pfn_base)) {
610 ret = vfio_lock_acct(dma, 1, true);
612 put_pfn(*pfn_base, dma->prot);
614 pr_warn("%s: Task %s (%d) RLIMIT_MEMLOCK "
615 "(%ld) exceeded\n", __func__,
616 dma->task->comm, task_pid_nr(dma->task),
617 task_rlimit(dma->task, RLIMIT_MEMLOCK));
625 static int vfio_unpin_page_external(struct vfio_dma *dma, dma_addr_t iova,
629 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
634 unlocked = vfio_iova_put_vfio_pfn(dma, vpfn);
637 vfio_lock_acct(dma, -unlocked, true);
642 static int vfio_iommu_type1_pin_pages(void *iommu_data,
643 struct iommu_group *iommu_group,
644 unsigned long *user_pfn,
646 unsigned long *phys_pfn)
648 struct vfio_iommu *iommu = iommu_data;
649 struct vfio_group *group;
651 unsigned long remote_vaddr;
652 struct vfio_dma *dma;
655 if (!iommu || !user_pfn || !phys_pfn)
658 /* Supported for v2 version only */
662 mutex_lock(&iommu->lock);
664 /* Fail if notifier list is empty */
665 if (!iommu->notifier.head) {
671 * If iommu capable domain exist in the container then all pages are
672 * already pinned and accounted. Accouting should be done if there is no
673 * iommu capable domain in the container.
675 do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu);
677 for (i = 0; i < npage; i++) {
679 struct vfio_pfn *vpfn;
681 iova = user_pfn[i] << PAGE_SHIFT;
682 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
688 if ((dma->prot & prot) != prot) {
693 vpfn = vfio_iova_get_vfio_pfn(dma, iova);
695 phys_pfn[i] = vpfn->pfn;
699 remote_vaddr = dma->vaddr + (iova - dma->iova);
700 ret = vfio_pin_page_external(dma, remote_vaddr, &phys_pfn[i],
705 ret = vfio_add_to_pfn_list(dma, iova, phys_pfn[i]);
707 if (put_pfn(phys_pfn[i], dma->prot) && do_accounting)
708 vfio_lock_acct(dma, -1, true);
712 if (iommu->dirty_page_tracking) {
713 unsigned long pgshift = __ffs(iommu->pgsize_bitmap);
716 * Bitmap populated with the smallest supported page
719 bitmap_set(dma->bitmap,
720 (iova - dma->iova) >> pgshift, 1);
725 group = vfio_iommu_find_iommu_group(iommu, iommu_group);
726 if (!group->pinned_page_dirty_scope) {
727 group->pinned_page_dirty_scope = true;
728 iommu->num_non_pinned_groups--;
735 for (j = 0; j < i; j++) {
738 iova = user_pfn[j] << PAGE_SHIFT;
739 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
740 vfio_unpin_page_external(dma, iova, do_accounting);
744 mutex_unlock(&iommu->lock);
748 static int vfio_iommu_type1_unpin_pages(void *iommu_data,
749 unsigned long *user_pfn,
752 struct vfio_iommu *iommu = iommu_data;
756 if (!iommu || !user_pfn)
759 /* Supported for v2 version only */
763 mutex_lock(&iommu->lock);
765 do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu);
766 for (i = 0; i < npage; i++) {
767 struct vfio_dma *dma;
770 iova = user_pfn[i] << PAGE_SHIFT;
771 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
774 vfio_unpin_page_external(dma, iova, do_accounting);
778 mutex_unlock(&iommu->lock);
779 return i > npage ? npage : (i > 0 ? i : -EINVAL);
782 static long vfio_sync_unpin(struct vfio_dma *dma, struct vfio_domain *domain,
783 struct list_head *regions,
784 struct iommu_iotlb_gather *iotlb_gather)
787 struct vfio_regions *entry, *next;
789 iommu_iotlb_sync(domain->domain, iotlb_gather);
791 list_for_each_entry_safe(entry, next, regions, list) {
792 unlocked += vfio_unpin_pages_remote(dma,
794 entry->phys >> PAGE_SHIFT,
795 entry->len >> PAGE_SHIFT,
797 list_del(&entry->list);
807 * Generally, VFIO needs to unpin remote pages after each IOTLB flush.
808 * Therefore, when using IOTLB flush sync interface, VFIO need to keep track
809 * of these regions (currently using a list).
811 * This value specifies maximum number of regions for each IOTLB flush sync.
813 #define VFIO_IOMMU_TLB_SYNC_MAX 512
815 static size_t unmap_unpin_fast(struct vfio_domain *domain,
816 struct vfio_dma *dma, dma_addr_t *iova,
817 size_t len, phys_addr_t phys, long *unlocked,
818 struct list_head *unmapped_list,
820 struct iommu_iotlb_gather *iotlb_gather)
823 struct vfio_regions *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
826 unmapped = iommu_unmap_fast(domain->domain, *iova, len,
834 entry->len = unmapped;
835 list_add_tail(&entry->list, unmapped_list);
843 * Sync if the number of fast-unmap regions hits the limit
844 * or in case of errors.
846 if (*unmapped_cnt >= VFIO_IOMMU_TLB_SYNC_MAX || !unmapped) {
847 *unlocked += vfio_sync_unpin(dma, domain, unmapped_list,
855 static size_t unmap_unpin_slow(struct vfio_domain *domain,
856 struct vfio_dma *dma, dma_addr_t *iova,
857 size_t len, phys_addr_t phys,
860 size_t unmapped = iommu_unmap(domain->domain, *iova, len);
863 *unlocked += vfio_unpin_pages_remote(dma, *iova,
865 unmapped >> PAGE_SHIFT,
873 static long vfio_unmap_unpin(struct vfio_iommu *iommu, struct vfio_dma *dma,
876 dma_addr_t iova = dma->iova, end = dma->iova + dma->size;
877 struct vfio_domain *domain, *d;
878 LIST_HEAD(unmapped_region_list);
879 struct iommu_iotlb_gather iotlb_gather;
880 int unmapped_region_cnt = 0;
886 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
890 * We use the IOMMU to track the physical addresses, otherwise we'd
891 * need a much more complicated tracking system. Unfortunately that
892 * means we need to use one of the iommu domains to figure out the
893 * pfns to unpin. The rest need to be unmapped in advance so we have
894 * no iommu translations remaining when the pages are unpinned.
896 domain = d = list_first_entry(&iommu->domain_list,
897 struct vfio_domain, next);
899 list_for_each_entry_continue(d, &iommu->domain_list, next) {
900 iommu_unmap(d->domain, dma->iova, dma->size);
904 iommu_iotlb_gather_init(&iotlb_gather);
906 size_t unmapped, len;
907 phys_addr_t phys, next;
909 phys = iommu_iova_to_phys(domain->domain, iova);
910 if (WARN_ON(!phys)) {
916 * To optimize for fewer iommu_unmap() calls, each of which
917 * may require hardware cache flushing, try to find the
918 * largest contiguous physical memory chunk to unmap.
920 for (len = PAGE_SIZE;
921 !domain->fgsp && iova + len < end; len += PAGE_SIZE) {
922 next = iommu_iova_to_phys(domain->domain, iova + len);
923 if (next != phys + len)
928 * First, try to use fast unmap/unpin. In case of failure,
929 * switch to slow unmap/unpin path.
931 unmapped = unmap_unpin_fast(domain, dma, &iova, len, phys,
932 &unlocked, &unmapped_region_list,
933 &unmapped_region_cnt,
936 unmapped = unmap_unpin_slow(domain, dma, &iova, len,
938 if (WARN_ON(!unmapped))
943 dma->iommu_mapped = false;
945 if (unmapped_region_cnt) {
946 unlocked += vfio_sync_unpin(dma, domain, &unmapped_region_list,
951 vfio_lock_acct(dma, -unlocked, true);
957 static void vfio_remove_dma(struct vfio_iommu *iommu, struct vfio_dma *dma)
959 WARN_ON(!RB_EMPTY_ROOT(&dma->pfn_list));
960 vfio_unmap_unpin(iommu, dma, true);
961 vfio_unlink_dma(iommu, dma);
962 put_task_struct(dma->task);
963 vfio_dma_bitmap_free(dma);
968 static void vfio_update_pgsize_bitmap(struct vfio_iommu *iommu)
970 struct vfio_domain *domain;
972 iommu->pgsize_bitmap = ULONG_MAX;
974 list_for_each_entry(domain, &iommu->domain_list, next)
975 iommu->pgsize_bitmap &= domain->domain->pgsize_bitmap;
978 * In case the IOMMU supports page sizes smaller than PAGE_SIZE
979 * we pretend PAGE_SIZE is supported and hide sub-PAGE_SIZE sizes.
980 * That way the user will be able to map/unmap buffers whose size/
981 * start address is aligned with PAGE_SIZE. Pinning code uses that
982 * granularity while iommu driver can use the sub-PAGE_SIZE size
985 if (iommu->pgsize_bitmap & ~PAGE_MASK) {
986 iommu->pgsize_bitmap &= PAGE_MASK;
987 iommu->pgsize_bitmap |= PAGE_SIZE;
991 static int update_user_bitmap(u64 __user *bitmap, struct vfio_iommu *iommu,
992 struct vfio_dma *dma, dma_addr_t base_iova,
995 unsigned long pgshift = __ffs(pgsize);
996 unsigned long nbits = dma->size >> pgshift;
997 unsigned long bit_offset = (dma->iova - base_iova) >> pgshift;
998 unsigned long copy_offset = bit_offset / BITS_PER_LONG;
999 unsigned long shift = bit_offset % BITS_PER_LONG;
1000 unsigned long leftover;
1003 * mark all pages dirty if any IOMMU capable device is not able
1004 * to report dirty pages and all pages are pinned and mapped.
1006 if (iommu->num_non_pinned_groups && dma->iommu_mapped)
1007 bitmap_set(dma->bitmap, 0, nbits);
1010 bitmap_shift_left(dma->bitmap, dma->bitmap, shift,
1013 if (copy_from_user(&leftover,
1014 (void __user *)(bitmap + copy_offset),
1018 bitmap_or(dma->bitmap, dma->bitmap, &leftover, shift);
1021 if (copy_to_user((void __user *)(bitmap + copy_offset), dma->bitmap,
1022 DIRTY_BITMAP_BYTES(nbits + shift)))
1028 static int vfio_iova_dirty_bitmap(u64 __user *bitmap, struct vfio_iommu *iommu,
1029 dma_addr_t iova, size_t size, size_t pgsize)
1031 struct vfio_dma *dma;
1033 unsigned long pgshift = __ffs(pgsize);
1037 * GET_BITMAP request must fully cover vfio_dma mappings. Multiple
1038 * vfio_dma mappings may be clubbed by specifying large ranges, but
1039 * there must not be any previous mappings bisected by the range.
1040 * An error will be returned if these conditions are not met.
1042 dma = vfio_find_dma(iommu, iova, 1);
1043 if (dma && dma->iova != iova)
1046 dma = vfio_find_dma(iommu, iova + size - 1, 0);
1047 if (dma && dma->iova + dma->size != iova + size)
1050 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
1051 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1053 if (dma->iova < iova)
1056 if (dma->iova > iova + size - 1)
1059 ret = update_user_bitmap(bitmap, iommu, dma, iova, pgsize);
1064 * Re-populate bitmap to include all pinned pages which are
1065 * considered as dirty but exclude pages which are unpinned and
1066 * pages which are marked dirty by vfio_dma_rw()
1068 bitmap_clear(dma->bitmap, 0, dma->size >> pgshift);
1069 vfio_dma_populate_bitmap(dma, pgsize);
1074 static int verify_bitmap_size(uint64_t npages, uint64_t bitmap_size)
1076 if (!npages || !bitmap_size || (bitmap_size > DIRTY_BITMAP_SIZE_MAX) ||
1077 (bitmap_size < DIRTY_BITMAP_BYTES(npages)))
1083 static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
1084 struct vfio_iommu_type1_dma_unmap *unmap,
1085 struct vfio_bitmap *bitmap)
1087 struct vfio_dma *dma, *dma_last = NULL;
1088 size_t unmapped = 0, pgsize;
1089 int ret = 0, retries = 0;
1090 unsigned long pgshift;
1092 mutex_lock(&iommu->lock);
1094 pgshift = __ffs(iommu->pgsize_bitmap);
1095 pgsize = (size_t)1 << pgshift;
1097 if (unmap->iova & (pgsize - 1)) {
1102 if (!unmap->size || unmap->size & (pgsize - 1)) {
1107 if (unmap->iova + unmap->size - 1 < unmap->iova ||
1108 unmap->size > SIZE_MAX) {
1113 /* When dirty tracking is enabled, allow only min supported pgsize */
1114 if ((unmap->flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) &&
1115 (!iommu->dirty_page_tracking || (bitmap->pgsize != pgsize))) {
1120 WARN_ON((pgsize - 1) & PAGE_MASK);
1123 * vfio-iommu-type1 (v1) - User mappings were coalesced together to
1124 * avoid tracking individual mappings. This means that the granularity
1125 * of the original mapping was lost and the user was allowed to attempt
1126 * to unmap any range. Depending on the contiguousness of physical
1127 * memory and page sizes supported by the IOMMU, arbitrary unmaps may
1128 * or may not have worked. We only guaranteed unmap granularity
1129 * matching the original mapping; even though it was untracked here,
1130 * the original mappings are reflected in IOMMU mappings. This
1131 * resulted in a couple unusual behaviors. First, if a range is not
1132 * able to be unmapped, ex. a set of 4k pages that was mapped as a
1133 * 2M hugepage into the IOMMU, the unmap ioctl returns success but with
1134 * a zero sized unmap. Also, if an unmap request overlaps the first
1135 * address of a hugepage, the IOMMU will unmap the entire hugepage.
1136 * This also returns success and the returned unmap size reflects the
1137 * actual size unmapped.
1139 * We attempt to maintain compatibility with this "v1" interface, but
1140 * we take control out of the hands of the IOMMU. Therefore, an unmap
1141 * request offset from the beginning of the original mapping will
1142 * return success with zero sized unmap. And an unmap request covering
1143 * the first iova of mapping will unmap the entire range.
1145 * The v2 version of this interface intends to be more deterministic.
1146 * Unmap requests must fully cover previous mappings. Multiple
1147 * mappings may still be unmaped by specifying large ranges, but there
1148 * must not be any previous mappings bisected by the range. An error
1149 * will be returned if these conditions are not met. The v2 interface
1150 * will only return success and a size of zero if there were no
1151 * mappings within the range.
1154 dma = vfio_find_dma(iommu, unmap->iova, 1);
1155 if (dma && dma->iova != unmap->iova) {
1159 dma = vfio_find_dma(iommu, unmap->iova + unmap->size - 1, 0);
1160 if (dma && dma->iova + dma->size != unmap->iova + unmap->size) {
1166 while ((dma = vfio_find_dma(iommu, unmap->iova, unmap->size))) {
1167 if (!iommu->v2 && unmap->iova > dma->iova)
1170 * Task with same address space who mapped this iova range is
1171 * allowed to unmap the iova range.
1173 if (dma->task->mm != current->mm)
1176 if (!RB_EMPTY_ROOT(&dma->pfn_list)) {
1177 struct vfio_iommu_type1_dma_unmap nb_unmap;
1179 if (dma_last == dma) {
1180 BUG_ON(++retries > 10);
1186 nb_unmap.iova = dma->iova;
1187 nb_unmap.size = dma->size;
1190 * Notify anyone (mdev vendor drivers) to invalidate and
1191 * unmap iovas within the range we're about to unmap.
1192 * Vendor drivers MUST unpin pages in response to an
1195 mutex_unlock(&iommu->lock);
1196 blocking_notifier_call_chain(&iommu->notifier,
1197 VFIO_IOMMU_NOTIFY_DMA_UNMAP,
1199 mutex_lock(&iommu->lock);
1203 if (unmap->flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) {
1204 ret = update_user_bitmap(bitmap->data, iommu, dma,
1205 unmap->iova, pgsize);
1210 unmapped += dma->size;
1211 vfio_remove_dma(iommu, dma);
1215 mutex_unlock(&iommu->lock);
1217 /* Report how much was unmapped */
1218 unmap->size = unmapped;
1223 static int vfio_iommu_map(struct vfio_iommu *iommu, dma_addr_t iova,
1224 unsigned long pfn, long npage, int prot)
1226 struct vfio_domain *d;
1229 list_for_each_entry(d, &iommu->domain_list, next) {
1230 ret = iommu_map(d->domain, iova, (phys_addr_t)pfn << PAGE_SHIFT,
1231 npage << PAGE_SHIFT, prot | d->prot);
1241 list_for_each_entry_continue_reverse(d, &iommu->domain_list, next) {
1242 iommu_unmap(d->domain, iova, npage << PAGE_SHIFT);
1249 static int vfio_pin_map_dma(struct vfio_iommu *iommu, struct vfio_dma *dma,
1252 dma_addr_t iova = dma->iova;
1253 unsigned long vaddr = dma->vaddr;
1254 size_t size = map_size;
1256 unsigned long pfn, limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1260 /* Pin a contiguous chunk of memory */
1261 npage = vfio_pin_pages_remote(dma, vaddr + dma->size,
1262 size >> PAGE_SHIFT, &pfn, limit);
1270 ret = vfio_iommu_map(iommu, iova + dma->size, pfn, npage,
1273 vfio_unpin_pages_remote(dma, iova + dma->size, pfn,
1278 size -= npage << PAGE_SHIFT;
1279 dma->size += npage << PAGE_SHIFT;
1282 dma->iommu_mapped = true;
1285 vfio_remove_dma(iommu, dma);
1291 * Check dma map request is within a valid iova range
1293 static bool vfio_iommu_iova_dma_valid(struct vfio_iommu *iommu,
1294 dma_addr_t start, dma_addr_t end)
1296 struct list_head *iova = &iommu->iova_list;
1297 struct vfio_iova *node;
1299 list_for_each_entry(node, iova, list) {
1300 if (start >= node->start && end <= node->end)
1305 * Check for list_empty() as well since a container with
1306 * a single mdev device will have an empty list.
1308 return list_empty(iova);
1311 static int vfio_dma_do_map(struct vfio_iommu *iommu,
1312 struct vfio_iommu_type1_dma_map *map)
1314 dma_addr_t iova = map->iova;
1315 unsigned long vaddr = map->vaddr;
1316 size_t size = map->size;
1317 int ret = 0, prot = 0;
1319 struct vfio_dma *dma;
1321 /* Verify that none of our __u64 fields overflow */
1322 if (map->size != size || map->vaddr != vaddr || map->iova != iova)
1325 /* READ/WRITE from device perspective */
1326 if (map->flags & VFIO_DMA_MAP_FLAG_WRITE)
1327 prot |= IOMMU_WRITE;
1328 if (map->flags & VFIO_DMA_MAP_FLAG_READ)
1331 mutex_lock(&iommu->lock);
1333 pgsize = (size_t)1 << __ffs(iommu->pgsize_bitmap);
1335 WARN_ON((pgsize - 1) & PAGE_MASK);
1337 if (!prot || !size || (size | iova | vaddr) & (pgsize - 1)) {
1342 /* Don't allow IOVA or virtual address wrap */
1343 if (iova + size - 1 < iova || vaddr + size - 1 < vaddr) {
1348 if (vfio_find_dma(iommu, iova, size)) {
1353 if (!iommu->dma_avail) {
1358 if (!vfio_iommu_iova_dma_valid(iommu, iova, iova + size - 1)) {
1363 dma = kzalloc(sizeof(*dma), GFP_KERNEL);
1375 * We need to be able to both add to a task's locked memory and test
1376 * against the locked memory limit and we need to be able to do both
1377 * outside of this call path as pinning can be asynchronous via the
1378 * external interfaces for mdev devices. RLIMIT_MEMLOCK requires a
1379 * task_struct and VM locked pages requires an mm_struct, however
1380 * holding an indefinite mm reference is not recommended, therefore we
1381 * only hold a reference to a task. We could hold a reference to
1382 * current, however QEMU uses this call path through vCPU threads,
1383 * which can be killed resulting in a NULL mm and failure in the unmap
1384 * path when called via a different thread. Avoid this problem by
1385 * using the group_leader as threads within the same group require
1386 * both CLONE_THREAD and CLONE_VM and will therefore use the same
1389 * Previously we also used the task for testing CAP_IPC_LOCK at the
1390 * time of pinning and accounting, however has_capability() makes use
1391 * of real_cred, a copy-on-write field, so we can't guarantee that it
1392 * matches group_leader, or in fact that it might not change by the
1393 * time it's evaluated. If a process were to call MAP_DMA with
1394 * CAP_IPC_LOCK but later drop it, it doesn't make sense that they
1395 * possibly see different results for an iommu_mapped vfio_dma vs
1396 * externally mapped. Therefore track CAP_IPC_LOCK in vfio_dma at the
1397 * time of calling MAP_DMA.
1399 get_task_struct(current->group_leader);
1400 dma->task = current->group_leader;
1401 dma->lock_cap = capable(CAP_IPC_LOCK);
1403 dma->pfn_list = RB_ROOT;
1405 /* Insert zero-sized and grow as we map chunks of it */
1406 vfio_link_dma(iommu, dma);
1408 /* Don't pin and map if container doesn't contain IOMMU capable domain*/
1409 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
1412 ret = vfio_pin_map_dma(iommu, dma, size);
1414 if (!ret && iommu->dirty_page_tracking) {
1415 ret = vfio_dma_bitmap_alloc(dma, pgsize);
1417 vfio_remove_dma(iommu, dma);
1421 mutex_unlock(&iommu->lock);
1425 static int vfio_bus_type(struct device *dev, void *data)
1427 struct bus_type **bus = data;
1429 if (*bus && *bus != dev->bus)
1437 static int vfio_iommu_replay(struct vfio_iommu *iommu,
1438 struct vfio_domain *domain)
1440 struct vfio_domain *d = NULL;
1442 unsigned long limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1445 /* Arbitrarily pick the first domain in the list for lookups */
1446 if (!list_empty(&iommu->domain_list))
1447 d = list_first_entry(&iommu->domain_list,
1448 struct vfio_domain, next);
1450 n = rb_first(&iommu->dma_list);
1452 for (; n; n = rb_next(n)) {
1453 struct vfio_dma *dma;
1456 dma = rb_entry(n, struct vfio_dma, node);
1459 while (iova < dma->iova + dma->size) {
1463 if (dma->iommu_mapped) {
1467 if (WARN_ON(!d)) { /* mapped w/o a domain?! */
1472 phys = iommu_iova_to_phys(d->domain, iova);
1474 if (WARN_ON(!phys)) {
1482 while (i < dma->iova + dma->size &&
1483 p == iommu_iova_to_phys(d->domain, i)) {
1490 unsigned long vaddr = dma->vaddr +
1492 size_t n = dma->iova + dma->size - iova;
1495 npage = vfio_pin_pages_remote(dma, vaddr,
1504 phys = pfn << PAGE_SHIFT;
1505 size = npage << PAGE_SHIFT;
1508 ret = iommu_map(domain->domain, iova, phys,
1509 size, dma->prot | domain->prot);
1511 if (!dma->iommu_mapped)
1512 vfio_unpin_pages_remote(dma, iova,
1523 /* All dmas are now mapped, defer to second tree walk for unwind */
1524 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
1525 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1527 dma->iommu_mapped = true;
1533 for (; n; n = rb_prev(n)) {
1534 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1537 if (dma->iommu_mapped) {
1538 iommu_unmap(domain->domain, dma->iova, dma->size);
1543 while (iova < dma->iova + dma->size) {
1544 phys_addr_t phys, p;
1548 phys = iommu_iova_to_phys(domain->domain, iova);
1557 while (i < dma->iova + dma->size &&
1558 p == iommu_iova_to_phys(domain->domain, i)) {
1564 iommu_unmap(domain->domain, iova, size);
1565 vfio_unpin_pages_remote(dma, iova, phys >> PAGE_SHIFT,
1566 size >> PAGE_SHIFT, true);
1574 * We change our unmap behavior slightly depending on whether the IOMMU
1575 * supports fine-grained superpages. IOMMUs like AMD-Vi will use a superpage
1576 * for practically any contiguous power-of-two mapping we give it. This means
1577 * we don't need to look for contiguous chunks ourselves to make unmapping
1578 * more efficient. On IOMMUs with coarse-grained super pages, like Intel VT-d
1579 * with discrete 2M/1G/512G/1T superpages, identifying contiguous chunks
1580 * significantly boosts non-hugetlbfs mappings and doesn't seem to hurt when
1581 * hugetlbfs is in use.
1583 static void vfio_test_domain_fgsp(struct vfio_domain *domain)
1586 int ret, order = get_order(PAGE_SIZE * 2);
1588 pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order);
1592 ret = iommu_map(domain->domain, 0, page_to_phys(pages), PAGE_SIZE * 2,
1593 IOMMU_READ | IOMMU_WRITE | domain->prot);
1595 size_t unmapped = iommu_unmap(domain->domain, 0, PAGE_SIZE);
1597 if (unmapped == PAGE_SIZE)
1598 iommu_unmap(domain->domain, PAGE_SIZE, PAGE_SIZE);
1600 domain->fgsp = true;
1603 __free_pages(pages, order);
1606 static struct vfio_group *find_iommu_group(struct vfio_domain *domain,
1607 struct iommu_group *iommu_group)
1609 struct vfio_group *g;
1611 list_for_each_entry(g, &domain->group_list, next) {
1612 if (g->iommu_group == iommu_group)
1619 static struct vfio_group *vfio_iommu_find_iommu_group(struct vfio_iommu *iommu,
1620 struct iommu_group *iommu_group)
1622 struct vfio_domain *domain;
1623 struct vfio_group *group = NULL;
1625 list_for_each_entry(domain, &iommu->domain_list, next) {
1626 group = find_iommu_group(domain, iommu_group);
1631 if (iommu->external_domain)
1632 group = find_iommu_group(iommu->external_domain, iommu_group);
1637 static bool vfio_iommu_has_sw_msi(struct list_head *group_resv_regions,
1640 struct iommu_resv_region *region;
1643 list_for_each_entry(region, group_resv_regions, list) {
1645 * The presence of any 'real' MSI regions should take
1646 * precedence over the software-managed one if the
1647 * IOMMU driver happens to advertise both types.
1649 if (region->type == IOMMU_RESV_MSI) {
1654 if (region->type == IOMMU_RESV_SW_MSI) {
1655 *base = region->start;
1663 static struct device *vfio_mdev_get_iommu_device(struct device *dev)
1665 struct device *(*fn)(struct device *dev);
1666 struct device *iommu_device;
1668 fn = symbol_get(mdev_get_iommu_device);
1670 iommu_device = fn(dev);
1671 symbol_put(mdev_get_iommu_device);
1673 return iommu_device;
1679 static int vfio_mdev_attach_domain(struct device *dev, void *data)
1681 struct iommu_domain *domain = data;
1682 struct device *iommu_device;
1684 iommu_device = vfio_mdev_get_iommu_device(dev);
1686 if (iommu_dev_feature_enabled(iommu_device, IOMMU_DEV_FEAT_AUX))
1687 return iommu_aux_attach_device(domain, iommu_device);
1689 return iommu_attach_device(domain, iommu_device);
1695 static int vfio_mdev_detach_domain(struct device *dev, void *data)
1697 struct iommu_domain *domain = data;
1698 struct device *iommu_device;
1700 iommu_device = vfio_mdev_get_iommu_device(dev);
1702 if (iommu_dev_feature_enabled(iommu_device, IOMMU_DEV_FEAT_AUX))
1703 iommu_aux_detach_device(domain, iommu_device);
1705 iommu_detach_device(domain, iommu_device);
1711 static int vfio_iommu_attach_group(struct vfio_domain *domain,
1712 struct vfio_group *group)
1714 if (group->mdev_group)
1715 return iommu_group_for_each_dev(group->iommu_group,
1717 vfio_mdev_attach_domain);
1719 return iommu_attach_group(domain->domain, group->iommu_group);
1722 static void vfio_iommu_detach_group(struct vfio_domain *domain,
1723 struct vfio_group *group)
1725 if (group->mdev_group)
1726 iommu_group_for_each_dev(group->iommu_group, domain->domain,
1727 vfio_mdev_detach_domain);
1729 iommu_detach_group(domain->domain, group->iommu_group);
1732 static bool vfio_bus_is_mdev(struct bus_type *bus)
1734 struct bus_type *mdev_bus;
1737 mdev_bus = symbol_get(mdev_bus_type);
1739 ret = (bus == mdev_bus);
1740 symbol_put(mdev_bus_type);
1746 static int vfio_mdev_iommu_device(struct device *dev, void *data)
1748 struct device **old = data, *new;
1750 new = vfio_mdev_get_iommu_device(dev);
1751 if (!new || (*old && *old != new))
1760 * This is a helper function to insert an address range to iova list.
1761 * The list is initially created with a single entry corresponding to
1762 * the IOMMU domain geometry to which the device group is attached.
1763 * The list aperture gets modified when a new domain is added to the
1764 * container if the new aperture doesn't conflict with the current one
1765 * or with any existing dma mappings. The list is also modified to
1766 * exclude any reserved regions associated with the device group.
1768 static int vfio_iommu_iova_insert(struct list_head *head,
1769 dma_addr_t start, dma_addr_t end)
1771 struct vfio_iova *region;
1773 region = kmalloc(sizeof(*region), GFP_KERNEL);
1777 INIT_LIST_HEAD(®ion->list);
1778 region->start = start;
1781 list_add_tail(®ion->list, head);
1786 * Check the new iommu aperture conflicts with existing aper or with any
1787 * existing dma mappings.
1789 static bool vfio_iommu_aper_conflict(struct vfio_iommu *iommu,
1790 dma_addr_t start, dma_addr_t end)
1792 struct vfio_iova *first, *last;
1793 struct list_head *iova = &iommu->iova_list;
1795 if (list_empty(iova))
1798 /* Disjoint sets, return conflict */
1799 first = list_first_entry(iova, struct vfio_iova, list);
1800 last = list_last_entry(iova, struct vfio_iova, list);
1801 if (start > last->end || end < first->start)
1804 /* Check for any existing dma mappings below the new start */
1805 if (start > first->start) {
1806 if (vfio_find_dma(iommu, first->start, start - first->start))
1810 /* Check for any existing dma mappings beyond the new end */
1811 if (end < last->end) {
1812 if (vfio_find_dma(iommu, end + 1, last->end - end))
1820 * Resize iommu iova aperture window. This is called only if the new
1821 * aperture has no conflict with existing aperture and dma mappings.
1823 static int vfio_iommu_aper_resize(struct list_head *iova,
1824 dma_addr_t start, dma_addr_t end)
1826 struct vfio_iova *node, *next;
1828 if (list_empty(iova))
1829 return vfio_iommu_iova_insert(iova, start, end);
1831 /* Adjust iova list start */
1832 list_for_each_entry_safe(node, next, iova, list) {
1833 if (start < node->start)
1835 if (start >= node->start && start < node->end) {
1836 node->start = start;
1839 /* Delete nodes before new start */
1840 list_del(&node->list);
1844 /* Adjust iova list end */
1845 list_for_each_entry_safe(node, next, iova, list) {
1846 if (end > node->end)
1848 if (end > node->start && end <= node->end) {
1852 /* Delete nodes after new end */
1853 list_del(&node->list);
1861 * Check reserved region conflicts with existing dma mappings
1863 static bool vfio_iommu_resv_conflict(struct vfio_iommu *iommu,
1864 struct list_head *resv_regions)
1866 struct iommu_resv_region *region;
1868 /* Check for conflict with existing dma mappings */
1869 list_for_each_entry(region, resv_regions, list) {
1870 if (region->type == IOMMU_RESV_DIRECT_RELAXABLE)
1873 if (vfio_find_dma(iommu, region->start, region->length))
1881 * Check iova region overlap with reserved regions and
1882 * exclude them from the iommu iova range
1884 static int vfio_iommu_resv_exclude(struct list_head *iova,
1885 struct list_head *resv_regions)
1887 struct iommu_resv_region *resv;
1888 struct vfio_iova *n, *next;
1890 list_for_each_entry(resv, resv_regions, list) {
1891 phys_addr_t start, end;
1893 if (resv->type == IOMMU_RESV_DIRECT_RELAXABLE)
1896 start = resv->start;
1897 end = resv->start + resv->length - 1;
1899 list_for_each_entry_safe(n, next, iova, list) {
1903 if (start > n->end || end < n->start)
1906 * Insert a new node if current node overlaps with the
1907 * reserve region to exlude that from valid iova range.
1908 * Note that, new node is inserted before the current
1909 * node and finally the current node is deleted keeping
1910 * the list updated and sorted.
1912 if (start > n->start)
1913 ret = vfio_iommu_iova_insert(&n->list, n->start,
1915 if (!ret && end < n->end)
1916 ret = vfio_iommu_iova_insert(&n->list, end + 1,
1926 if (list_empty(iova))
1932 static void vfio_iommu_resv_free(struct list_head *resv_regions)
1934 struct iommu_resv_region *n, *next;
1936 list_for_each_entry_safe(n, next, resv_regions, list) {
1942 static void vfio_iommu_iova_free(struct list_head *iova)
1944 struct vfio_iova *n, *next;
1946 list_for_each_entry_safe(n, next, iova, list) {
1952 static int vfio_iommu_iova_get_copy(struct vfio_iommu *iommu,
1953 struct list_head *iova_copy)
1955 struct list_head *iova = &iommu->iova_list;
1956 struct vfio_iova *n;
1959 list_for_each_entry(n, iova, list) {
1960 ret = vfio_iommu_iova_insert(iova_copy, n->start, n->end);
1968 vfio_iommu_iova_free(iova_copy);
1972 static void vfio_iommu_iova_insert_copy(struct vfio_iommu *iommu,
1973 struct list_head *iova_copy)
1975 struct list_head *iova = &iommu->iova_list;
1977 vfio_iommu_iova_free(iova);
1979 list_splice_tail(iova_copy, iova);
1982 static int vfio_iommu_type1_attach_group(void *iommu_data,
1983 struct iommu_group *iommu_group)
1985 struct vfio_iommu *iommu = iommu_data;
1986 struct vfio_group *group;
1987 struct vfio_domain *domain, *d;
1988 struct bus_type *bus = NULL;
1990 bool resv_msi, msi_remap;
1991 phys_addr_t resv_msi_base = 0;
1992 struct iommu_domain_geometry geo;
1993 LIST_HEAD(iova_copy);
1994 LIST_HEAD(group_resv_regions);
1996 mutex_lock(&iommu->lock);
1998 /* Check for duplicates */
1999 if (vfio_iommu_find_iommu_group(iommu, iommu_group)) {
2000 mutex_unlock(&iommu->lock);
2004 group = kzalloc(sizeof(*group), GFP_KERNEL);
2005 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2006 if (!group || !domain) {
2011 group->iommu_group = iommu_group;
2013 /* Determine bus_type in order to allocate a domain */
2014 ret = iommu_group_for_each_dev(iommu_group, &bus, vfio_bus_type);
2018 if (vfio_bus_is_mdev(bus)) {
2019 struct device *iommu_device = NULL;
2021 group->mdev_group = true;
2023 /* Determine the isolation type */
2024 ret = iommu_group_for_each_dev(iommu_group, &iommu_device,
2025 vfio_mdev_iommu_device);
2026 if (ret || !iommu_device) {
2027 if (!iommu->external_domain) {
2028 INIT_LIST_HEAD(&domain->group_list);
2029 iommu->external_domain = domain;
2030 vfio_update_pgsize_bitmap(iommu);
2035 list_add(&group->next,
2036 &iommu->external_domain->group_list);
2038 * Non-iommu backed group cannot dirty memory directly,
2039 * it can only use interfaces that provide dirty
2041 * The iommu scope can only be promoted with the
2042 * addition of a dirty tracking group.
2044 group->pinned_page_dirty_scope = true;
2045 mutex_unlock(&iommu->lock);
2050 bus = iommu_device->bus;
2053 domain->domain = iommu_domain_alloc(bus);
2054 if (!domain->domain) {
2059 if (iommu->nesting) {
2062 ret = iommu_domain_set_attr(domain->domain, DOMAIN_ATTR_NESTING,
2068 ret = vfio_iommu_attach_group(domain, group);
2072 /* Get aperture info */
2073 iommu_domain_get_attr(domain->domain, DOMAIN_ATTR_GEOMETRY, &geo);
2075 if (vfio_iommu_aper_conflict(iommu, geo.aperture_start,
2076 geo.aperture_end)) {
2081 ret = iommu_get_group_resv_regions(iommu_group, &group_resv_regions);
2085 if (vfio_iommu_resv_conflict(iommu, &group_resv_regions)) {
2091 * We don't want to work on the original iova list as the list
2092 * gets modified and in case of failure we have to retain the
2093 * original list. Get a copy here.
2095 ret = vfio_iommu_iova_get_copy(iommu, &iova_copy);
2099 ret = vfio_iommu_aper_resize(&iova_copy, geo.aperture_start,
2104 ret = vfio_iommu_resv_exclude(&iova_copy, &group_resv_regions);
2108 resv_msi = vfio_iommu_has_sw_msi(&group_resv_regions, &resv_msi_base);
2110 INIT_LIST_HEAD(&domain->group_list);
2111 list_add(&group->next, &domain->group_list);
2113 msi_remap = irq_domain_check_msi_remap() ||
2114 iommu_capable(bus, IOMMU_CAP_INTR_REMAP);
2116 if (!allow_unsafe_interrupts && !msi_remap) {
2117 pr_warn("%s: No interrupt remapping support. Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n",
2123 if (iommu_capable(bus, IOMMU_CAP_CACHE_COHERENCY))
2124 domain->prot |= IOMMU_CACHE;
2127 * Try to match an existing compatible domain. We don't want to
2128 * preclude an IOMMU driver supporting multiple bus_types and being
2129 * able to include different bus_types in the same IOMMU domain, so
2130 * we test whether the domains use the same iommu_ops rather than
2131 * testing if they're on the same bus_type.
2133 list_for_each_entry(d, &iommu->domain_list, next) {
2134 if (d->domain->ops == domain->domain->ops &&
2135 d->prot == domain->prot) {
2136 vfio_iommu_detach_group(domain, group);
2137 if (!vfio_iommu_attach_group(d, group)) {
2138 list_add(&group->next, &d->group_list);
2139 iommu_domain_free(domain->domain);
2144 ret = vfio_iommu_attach_group(domain, group);
2150 vfio_test_domain_fgsp(domain);
2152 /* replay mappings on new domains */
2153 ret = vfio_iommu_replay(iommu, domain);
2158 ret = iommu_get_msi_cookie(domain->domain, resv_msi_base);
2159 if (ret && ret != -ENODEV)
2163 list_add(&domain->next, &iommu->domain_list);
2164 vfio_update_pgsize_bitmap(iommu);
2166 /* Delete the old one and insert new iova list */
2167 vfio_iommu_iova_insert_copy(iommu, &iova_copy);
2170 * An iommu backed group can dirty memory directly and therefore
2171 * demotes the iommu scope until it declares itself dirty tracking
2172 * capable via the page pinning interface.
2174 iommu->num_non_pinned_groups++;
2175 mutex_unlock(&iommu->lock);
2176 vfio_iommu_resv_free(&group_resv_regions);
2181 vfio_iommu_detach_group(domain, group);
2183 iommu_domain_free(domain->domain);
2184 vfio_iommu_iova_free(&iova_copy);
2185 vfio_iommu_resv_free(&group_resv_regions);
2189 mutex_unlock(&iommu->lock);
2193 static void vfio_iommu_unmap_unpin_all(struct vfio_iommu *iommu)
2195 struct rb_node *node;
2197 while ((node = rb_first(&iommu->dma_list)))
2198 vfio_remove_dma(iommu, rb_entry(node, struct vfio_dma, node));
2201 static void vfio_iommu_unmap_unpin_reaccount(struct vfio_iommu *iommu)
2203 struct rb_node *n, *p;
2205 n = rb_first(&iommu->dma_list);
2206 for (; n; n = rb_next(n)) {
2207 struct vfio_dma *dma;
2208 long locked = 0, unlocked = 0;
2210 dma = rb_entry(n, struct vfio_dma, node);
2211 unlocked += vfio_unmap_unpin(iommu, dma, false);
2212 p = rb_first(&dma->pfn_list);
2213 for (; p; p = rb_next(p)) {
2214 struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn,
2217 if (!is_invalid_reserved_pfn(vpfn->pfn))
2220 vfio_lock_acct(dma, locked - unlocked, true);
2225 * Called when a domain is removed in detach. It is possible that
2226 * the removed domain decided the iova aperture window. Modify the
2227 * iova aperture with the smallest window among existing domains.
2229 static void vfio_iommu_aper_expand(struct vfio_iommu *iommu,
2230 struct list_head *iova_copy)
2232 struct vfio_domain *domain;
2233 struct iommu_domain_geometry geo;
2234 struct vfio_iova *node;
2235 dma_addr_t start = 0;
2236 dma_addr_t end = (dma_addr_t)~0;
2238 if (list_empty(iova_copy))
2241 list_for_each_entry(domain, &iommu->domain_list, next) {
2242 iommu_domain_get_attr(domain->domain, DOMAIN_ATTR_GEOMETRY,
2244 if (geo.aperture_start > start)
2245 start = geo.aperture_start;
2246 if (geo.aperture_end < end)
2247 end = geo.aperture_end;
2250 /* Modify aperture limits. The new aper is either same or bigger */
2251 node = list_first_entry(iova_copy, struct vfio_iova, list);
2252 node->start = start;
2253 node = list_last_entry(iova_copy, struct vfio_iova, list);
2258 * Called when a group is detached. The reserved regions for that
2259 * group can be part of valid iova now. But since reserved regions
2260 * may be duplicated among groups, populate the iova valid regions
2263 static int vfio_iommu_resv_refresh(struct vfio_iommu *iommu,
2264 struct list_head *iova_copy)
2266 struct vfio_domain *d;
2267 struct vfio_group *g;
2268 struct vfio_iova *node;
2269 dma_addr_t start, end;
2270 LIST_HEAD(resv_regions);
2273 if (list_empty(iova_copy))
2276 list_for_each_entry(d, &iommu->domain_list, next) {
2277 list_for_each_entry(g, &d->group_list, next) {
2278 ret = iommu_get_group_resv_regions(g->iommu_group,
2285 node = list_first_entry(iova_copy, struct vfio_iova, list);
2286 start = node->start;
2287 node = list_last_entry(iova_copy, struct vfio_iova, list);
2290 /* purge the iova list and create new one */
2291 vfio_iommu_iova_free(iova_copy);
2293 ret = vfio_iommu_aper_resize(iova_copy, start, end);
2297 /* Exclude current reserved regions from iova ranges */
2298 ret = vfio_iommu_resv_exclude(iova_copy, &resv_regions);
2300 vfio_iommu_resv_free(&resv_regions);
2304 static void vfio_iommu_type1_detach_group(void *iommu_data,
2305 struct iommu_group *iommu_group)
2307 struct vfio_iommu *iommu = iommu_data;
2308 struct vfio_domain *domain;
2309 struct vfio_group *group;
2310 bool update_dirty_scope = false;
2311 LIST_HEAD(iova_copy);
2313 mutex_lock(&iommu->lock);
2315 if (iommu->external_domain) {
2316 group = find_iommu_group(iommu->external_domain, iommu_group);
2318 update_dirty_scope = !group->pinned_page_dirty_scope;
2319 list_del(&group->next);
2322 if (list_empty(&iommu->external_domain->group_list)) {
2323 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu)) {
2324 WARN_ON(iommu->notifier.head);
2325 vfio_iommu_unmap_unpin_all(iommu);
2328 kfree(iommu->external_domain);
2329 iommu->external_domain = NULL;
2331 goto detach_group_done;
2336 * Get a copy of iova list. This will be used to update
2337 * and to replace the current one later. Please note that
2338 * we will leave the original list as it is if update fails.
2340 vfio_iommu_iova_get_copy(iommu, &iova_copy);
2342 list_for_each_entry(domain, &iommu->domain_list, next) {
2343 group = find_iommu_group(domain, iommu_group);
2347 vfio_iommu_detach_group(domain, group);
2348 update_dirty_scope = !group->pinned_page_dirty_scope;
2349 list_del(&group->next);
2352 * Group ownership provides privilege, if the group list is
2353 * empty, the domain goes away. If it's the last domain with
2354 * iommu and external domain doesn't exist, then all the
2355 * mappings go away too. If it's the last domain with iommu and
2356 * external domain exist, update accounting
2358 if (list_empty(&domain->group_list)) {
2359 if (list_is_singular(&iommu->domain_list)) {
2360 if (!iommu->external_domain) {
2361 WARN_ON(iommu->notifier.head);
2362 vfio_iommu_unmap_unpin_all(iommu);
2364 vfio_iommu_unmap_unpin_reaccount(iommu);
2367 iommu_domain_free(domain->domain);
2368 list_del(&domain->next);
2370 vfio_iommu_aper_expand(iommu, &iova_copy);
2371 vfio_update_pgsize_bitmap(iommu);
2376 if (!vfio_iommu_resv_refresh(iommu, &iova_copy))
2377 vfio_iommu_iova_insert_copy(iommu, &iova_copy);
2379 vfio_iommu_iova_free(&iova_copy);
2383 * Removal of a group without dirty tracking may allow the iommu scope
2386 if (update_dirty_scope) {
2387 iommu->num_non_pinned_groups--;
2388 if (iommu->dirty_page_tracking)
2389 vfio_iommu_populate_bitmap_full(iommu);
2391 mutex_unlock(&iommu->lock);
2394 static void *vfio_iommu_type1_open(unsigned long arg)
2396 struct vfio_iommu *iommu;
2398 iommu = kzalloc(sizeof(*iommu), GFP_KERNEL);
2400 return ERR_PTR(-ENOMEM);
2403 case VFIO_TYPE1_IOMMU:
2405 case VFIO_TYPE1_NESTING_IOMMU:
2406 iommu->nesting = true;
2408 case VFIO_TYPE1v2_IOMMU:
2413 return ERR_PTR(-EINVAL);
2416 INIT_LIST_HEAD(&iommu->domain_list);
2417 INIT_LIST_HEAD(&iommu->iova_list);
2418 iommu->dma_list = RB_ROOT;
2419 iommu->dma_avail = dma_entry_limit;
2420 mutex_init(&iommu->lock);
2421 BLOCKING_INIT_NOTIFIER_HEAD(&iommu->notifier);
2426 static void vfio_release_domain(struct vfio_domain *domain, bool external)
2428 struct vfio_group *group, *group_tmp;
2430 list_for_each_entry_safe(group, group_tmp,
2431 &domain->group_list, next) {
2433 vfio_iommu_detach_group(domain, group);
2434 list_del(&group->next);
2439 iommu_domain_free(domain->domain);
2442 static void vfio_iommu_type1_release(void *iommu_data)
2444 struct vfio_iommu *iommu = iommu_data;
2445 struct vfio_domain *domain, *domain_tmp;
2447 if (iommu->external_domain) {
2448 vfio_release_domain(iommu->external_domain, true);
2449 kfree(iommu->external_domain);
2452 vfio_iommu_unmap_unpin_all(iommu);
2454 list_for_each_entry_safe(domain, domain_tmp,
2455 &iommu->domain_list, next) {
2456 vfio_release_domain(domain, false);
2457 list_del(&domain->next);
2461 vfio_iommu_iova_free(&iommu->iova_list);
2466 static int vfio_domains_have_iommu_cache(struct vfio_iommu *iommu)
2468 struct vfio_domain *domain;
2471 mutex_lock(&iommu->lock);
2472 list_for_each_entry(domain, &iommu->domain_list, next) {
2473 if (!(domain->prot & IOMMU_CACHE)) {
2478 mutex_unlock(&iommu->lock);
2483 static int vfio_iommu_type1_check_extension(struct vfio_iommu *iommu,
2487 case VFIO_TYPE1_IOMMU:
2488 case VFIO_TYPE1v2_IOMMU:
2489 case VFIO_TYPE1_NESTING_IOMMU:
2491 case VFIO_DMA_CC_IOMMU:
2494 return vfio_domains_have_iommu_cache(iommu);
2500 static int vfio_iommu_iova_add_cap(struct vfio_info_cap *caps,
2501 struct vfio_iommu_type1_info_cap_iova_range *cap_iovas,
2504 struct vfio_info_cap_header *header;
2505 struct vfio_iommu_type1_info_cap_iova_range *iova_cap;
2507 header = vfio_info_cap_add(caps, size,
2508 VFIO_IOMMU_TYPE1_INFO_CAP_IOVA_RANGE, 1);
2510 return PTR_ERR(header);
2512 iova_cap = container_of(header,
2513 struct vfio_iommu_type1_info_cap_iova_range,
2515 iova_cap->nr_iovas = cap_iovas->nr_iovas;
2516 memcpy(iova_cap->iova_ranges, cap_iovas->iova_ranges,
2517 cap_iovas->nr_iovas * sizeof(*cap_iovas->iova_ranges));
2521 static int vfio_iommu_iova_build_caps(struct vfio_iommu *iommu,
2522 struct vfio_info_cap *caps)
2524 struct vfio_iommu_type1_info_cap_iova_range *cap_iovas;
2525 struct vfio_iova *iova;
2527 int iovas = 0, i = 0, ret;
2529 list_for_each_entry(iova, &iommu->iova_list, list)
2534 * Return 0 as a container with a single mdev device
2535 * will have an empty list
2540 size = sizeof(*cap_iovas) + (iovas * sizeof(*cap_iovas->iova_ranges));
2542 cap_iovas = kzalloc(size, GFP_KERNEL);
2546 cap_iovas->nr_iovas = iovas;
2548 list_for_each_entry(iova, &iommu->iova_list, list) {
2549 cap_iovas->iova_ranges[i].start = iova->start;
2550 cap_iovas->iova_ranges[i].end = iova->end;
2554 ret = vfio_iommu_iova_add_cap(caps, cap_iovas, size);
2560 static int vfio_iommu_migration_build_caps(struct vfio_iommu *iommu,
2561 struct vfio_info_cap *caps)
2563 struct vfio_iommu_type1_info_cap_migration cap_mig;
2565 cap_mig.header.id = VFIO_IOMMU_TYPE1_INFO_CAP_MIGRATION;
2566 cap_mig.header.version = 1;
2569 /* support minimum pgsize */
2570 cap_mig.pgsize_bitmap = (size_t)1 << __ffs(iommu->pgsize_bitmap);
2571 cap_mig.max_dirty_bitmap_size = DIRTY_BITMAP_SIZE_MAX;
2573 return vfio_info_add_capability(caps, &cap_mig.header, sizeof(cap_mig));
2576 static int vfio_iommu_dma_avail_build_caps(struct vfio_iommu *iommu,
2577 struct vfio_info_cap *caps)
2579 struct vfio_iommu_type1_info_dma_avail cap_dma_avail;
2581 cap_dma_avail.header.id = VFIO_IOMMU_TYPE1_INFO_DMA_AVAIL;
2582 cap_dma_avail.header.version = 1;
2584 cap_dma_avail.avail = iommu->dma_avail;
2586 return vfio_info_add_capability(caps, &cap_dma_avail.header,
2587 sizeof(cap_dma_avail));
2590 static int vfio_iommu_type1_get_info(struct vfio_iommu *iommu,
2593 struct vfio_iommu_type1_info info;
2594 unsigned long minsz;
2595 struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
2596 unsigned long capsz;
2599 minsz = offsetofend(struct vfio_iommu_type1_info, iova_pgsizes);
2601 /* For backward compatibility, cannot require this */
2602 capsz = offsetofend(struct vfio_iommu_type1_info, cap_offset);
2604 if (copy_from_user(&info, (void __user *)arg, minsz))
2607 if (info.argsz < minsz)
2610 if (info.argsz >= capsz) {
2612 info.cap_offset = 0; /* output, no-recopy necessary */
2615 mutex_lock(&iommu->lock);
2616 info.flags = VFIO_IOMMU_INFO_PGSIZES;
2618 info.iova_pgsizes = iommu->pgsize_bitmap;
2620 ret = vfio_iommu_migration_build_caps(iommu, &caps);
2623 ret = vfio_iommu_dma_avail_build_caps(iommu, &caps);
2626 ret = vfio_iommu_iova_build_caps(iommu, &caps);
2628 mutex_unlock(&iommu->lock);
2634 info.flags |= VFIO_IOMMU_INFO_CAPS;
2636 if (info.argsz < sizeof(info) + caps.size) {
2637 info.argsz = sizeof(info) + caps.size;
2639 vfio_info_cap_shift(&caps, sizeof(info));
2640 if (copy_to_user((void __user *)arg +
2641 sizeof(info), caps.buf,
2646 info.cap_offset = sizeof(info);
2652 return copy_to_user((void __user *)arg, &info, minsz) ?
2656 static int vfio_iommu_type1_map_dma(struct vfio_iommu *iommu,
2659 struct vfio_iommu_type1_dma_map map;
2660 unsigned long minsz;
2661 uint32_t mask = VFIO_DMA_MAP_FLAG_READ | VFIO_DMA_MAP_FLAG_WRITE;
2663 minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
2665 if (copy_from_user(&map, (void __user *)arg, minsz))
2668 if (map.argsz < minsz || map.flags & ~mask)
2671 return vfio_dma_do_map(iommu, &map);
2674 static int vfio_iommu_type1_unmap_dma(struct vfio_iommu *iommu,
2677 struct vfio_iommu_type1_dma_unmap unmap;
2678 struct vfio_bitmap bitmap = { 0 };
2679 unsigned long minsz;
2682 minsz = offsetofend(struct vfio_iommu_type1_dma_unmap, size);
2684 if (copy_from_user(&unmap, (void __user *)arg, minsz))
2687 if (unmap.argsz < minsz ||
2688 unmap.flags & ~VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP)
2691 if (unmap.flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) {
2692 unsigned long pgshift;
2694 if (unmap.argsz < (minsz + sizeof(bitmap)))
2697 if (copy_from_user(&bitmap,
2698 (void __user *)(arg + minsz),
2702 if (!access_ok((void __user *)bitmap.data, bitmap.size))
2705 pgshift = __ffs(bitmap.pgsize);
2706 ret = verify_bitmap_size(unmap.size >> pgshift,
2712 ret = vfio_dma_do_unmap(iommu, &unmap, &bitmap);
2716 return copy_to_user((void __user *)arg, &unmap, minsz) ?
2720 static int vfio_iommu_type1_dirty_pages(struct vfio_iommu *iommu,
2723 struct vfio_iommu_type1_dirty_bitmap dirty;
2724 uint32_t mask = VFIO_IOMMU_DIRTY_PAGES_FLAG_START |
2725 VFIO_IOMMU_DIRTY_PAGES_FLAG_STOP |
2726 VFIO_IOMMU_DIRTY_PAGES_FLAG_GET_BITMAP;
2727 unsigned long minsz;
2733 minsz = offsetofend(struct vfio_iommu_type1_dirty_bitmap, flags);
2735 if (copy_from_user(&dirty, (void __user *)arg, minsz))
2738 if (dirty.argsz < minsz || dirty.flags & ~mask)
2741 /* only one flag should be set at a time */
2742 if (__ffs(dirty.flags) != __fls(dirty.flags))
2745 if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_START) {
2748 mutex_lock(&iommu->lock);
2749 pgsize = 1 << __ffs(iommu->pgsize_bitmap);
2750 if (!iommu->dirty_page_tracking) {
2751 ret = vfio_dma_bitmap_alloc_all(iommu, pgsize);
2753 iommu->dirty_page_tracking = true;
2755 mutex_unlock(&iommu->lock);
2757 } else if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_STOP) {
2758 mutex_lock(&iommu->lock);
2759 if (iommu->dirty_page_tracking) {
2760 iommu->dirty_page_tracking = false;
2761 vfio_dma_bitmap_free_all(iommu);
2763 mutex_unlock(&iommu->lock);
2765 } else if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_GET_BITMAP) {
2766 struct vfio_iommu_type1_dirty_bitmap_get range;
2767 unsigned long pgshift;
2768 size_t data_size = dirty.argsz - minsz;
2769 size_t iommu_pgsize;
2771 if (!data_size || data_size < sizeof(range))
2774 if (copy_from_user(&range, (void __user *)(arg + minsz),
2778 if (range.iova + range.size < range.iova)
2780 if (!access_ok((void __user *)range.bitmap.data,
2784 pgshift = __ffs(range.bitmap.pgsize);
2785 ret = verify_bitmap_size(range.size >> pgshift,
2790 mutex_lock(&iommu->lock);
2792 iommu_pgsize = (size_t)1 << __ffs(iommu->pgsize_bitmap);
2794 /* allow only smallest supported pgsize */
2795 if (range.bitmap.pgsize != iommu_pgsize) {
2799 if (range.iova & (iommu_pgsize - 1)) {
2803 if (!range.size || range.size & (iommu_pgsize - 1)) {
2808 if (iommu->dirty_page_tracking)
2809 ret = vfio_iova_dirty_bitmap(range.bitmap.data,
2812 range.bitmap.pgsize);
2816 mutex_unlock(&iommu->lock);
2824 static long vfio_iommu_type1_ioctl(void *iommu_data,
2825 unsigned int cmd, unsigned long arg)
2827 struct vfio_iommu *iommu = iommu_data;
2830 case VFIO_CHECK_EXTENSION:
2831 return vfio_iommu_type1_check_extension(iommu, arg);
2832 case VFIO_IOMMU_GET_INFO:
2833 return vfio_iommu_type1_get_info(iommu, arg);
2834 case VFIO_IOMMU_MAP_DMA:
2835 return vfio_iommu_type1_map_dma(iommu, arg);
2836 case VFIO_IOMMU_UNMAP_DMA:
2837 return vfio_iommu_type1_unmap_dma(iommu, arg);
2838 case VFIO_IOMMU_DIRTY_PAGES:
2839 return vfio_iommu_type1_dirty_pages(iommu, arg);
2845 static int vfio_iommu_type1_register_notifier(void *iommu_data,
2846 unsigned long *events,
2847 struct notifier_block *nb)
2849 struct vfio_iommu *iommu = iommu_data;
2851 /* clear known events */
2852 *events &= ~VFIO_IOMMU_NOTIFY_DMA_UNMAP;
2854 /* refuse to register if still events remaining */
2858 return blocking_notifier_chain_register(&iommu->notifier, nb);
2861 static int vfio_iommu_type1_unregister_notifier(void *iommu_data,
2862 struct notifier_block *nb)
2864 struct vfio_iommu *iommu = iommu_data;
2866 return blocking_notifier_chain_unregister(&iommu->notifier, nb);
2869 static int vfio_iommu_type1_dma_rw_chunk(struct vfio_iommu *iommu,
2870 dma_addr_t user_iova, void *data,
2871 size_t count, bool write,
2874 struct mm_struct *mm;
2875 unsigned long vaddr;
2876 struct vfio_dma *dma;
2877 bool kthread = current->mm == NULL;
2882 dma = vfio_find_dma(iommu, user_iova, 1);
2886 if ((write && !(dma->prot & IOMMU_WRITE)) ||
2887 !(dma->prot & IOMMU_READ))
2890 mm = get_task_mm(dma->task);
2897 else if (current->mm != mm)
2900 offset = user_iova - dma->iova;
2902 if (count > dma->size - offset)
2903 count = dma->size - offset;
2905 vaddr = dma->vaddr + offset;
2908 *copied = copy_to_user((void __user *)vaddr, data,
2910 if (*copied && iommu->dirty_page_tracking) {
2911 unsigned long pgshift = __ffs(iommu->pgsize_bitmap);
2913 * Bitmap populated with the smallest supported page
2916 bitmap_set(dma->bitmap, offset >> pgshift,
2917 ((offset + *copied - 1) >> pgshift) -
2918 (offset >> pgshift) + 1);
2921 *copied = copy_from_user(data, (void __user *)vaddr,
2924 kthread_unuse_mm(mm);
2927 return *copied ? 0 : -EFAULT;
2930 static int vfio_iommu_type1_dma_rw(void *iommu_data, dma_addr_t user_iova,
2931 void *data, size_t count, bool write)
2933 struct vfio_iommu *iommu = iommu_data;
2937 mutex_lock(&iommu->lock);
2939 ret = vfio_iommu_type1_dma_rw_chunk(iommu, user_iova, data,
2940 count, write, &done);
2949 mutex_unlock(&iommu->lock);
2953 static struct iommu_domain *
2954 vfio_iommu_type1_group_iommu_domain(void *iommu_data,
2955 struct iommu_group *iommu_group)
2957 struct iommu_domain *domain = ERR_PTR(-ENODEV);
2958 struct vfio_iommu *iommu = iommu_data;
2959 struct vfio_domain *d;
2961 if (!iommu || !iommu_group)
2962 return ERR_PTR(-EINVAL);
2964 mutex_lock(&iommu->lock);
2965 list_for_each_entry(d, &iommu->domain_list, next) {
2966 if (find_iommu_group(d, iommu_group)) {
2971 mutex_unlock(&iommu->lock);
2976 static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = {
2977 .name = "vfio-iommu-type1",
2978 .owner = THIS_MODULE,
2979 .open = vfio_iommu_type1_open,
2980 .release = vfio_iommu_type1_release,
2981 .ioctl = vfio_iommu_type1_ioctl,
2982 .attach_group = vfio_iommu_type1_attach_group,
2983 .detach_group = vfio_iommu_type1_detach_group,
2984 .pin_pages = vfio_iommu_type1_pin_pages,
2985 .unpin_pages = vfio_iommu_type1_unpin_pages,
2986 .register_notifier = vfio_iommu_type1_register_notifier,
2987 .unregister_notifier = vfio_iommu_type1_unregister_notifier,
2988 .dma_rw = vfio_iommu_type1_dma_rw,
2989 .group_iommu_domain = vfio_iommu_type1_group_iommu_domain,
2992 static int __init vfio_iommu_type1_init(void)
2994 return vfio_register_iommu_driver(&vfio_iommu_driver_ops_type1);
2997 static void __exit vfio_iommu_type1_cleanup(void)
2999 vfio_unregister_iommu_driver(&vfio_iommu_driver_ops_type1);
3002 module_init(vfio_iommu_type1_init);
3003 module_exit(vfio_iommu_type1_cleanup);
3005 MODULE_VERSION(DRIVER_VERSION);
3006 MODULE_LICENSE("GPL v2");
3007 MODULE_AUTHOR(DRIVER_AUTHOR);
3008 MODULE_DESCRIPTION(DRIVER_DESC);