1 // SPDX-License-Identifier: GPL-2.0-only
3 * VFIO: IOMMU DMA mapping support for Type1 IOMMU
5 * Copyright (C) 2012 Red Hat, Inc. All rights reserved.
6 * Author: Alex Williamson <alex.williamson@redhat.com>
8 * Derived from original vfio:
9 * Copyright 2010 Cisco Systems, Inc. All rights reserved.
10 * Author: Tom Lyon, pugs@cisco.com
12 * We arbitrarily define a Type1 IOMMU as one matching the below code.
13 * It could be called the x86 IOMMU as it's designed for AMD-Vi & Intel
14 * VT-d, but that makes it harder to re-use as theoretically anyone
15 * implementing a similar IOMMU could make use of this. We expect the
16 * IOMMU to support the IOMMU API and have few to no restrictions around
17 * the IOVA range that can be mapped. The Type1 IOMMU is currently
18 * optimized for relatively static mappings of a userspace process with
19 * userpsace pages pinned into memory. We also assume devices and IOMMU
20 * domains are PCI based as the IOMMU API is still centered around a
21 * device/bus interface rather than a group interface.
24 #include <linux/compat.h>
25 #include <linux/device.h>
27 #include <linux/highmem.h>
28 #include <linux/iommu.h>
29 #include <linux/module.h>
31 #include <linux/kthread.h>
32 #include <linux/rbtree.h>
33 #include <linux/sched/signal.h>
34 #include <linux/sched/mm.h>
35 #include <linux/slab.h>
36 #include <linux/uaccess.h>
37 #include <linux/vfio.h>
38 #include <linux/workqueue.h>
39 #include <linux/mdev.h>
40 #include <linux/notifier.h>
41 #include <linux/dma-iommu.h>
42 #include <linux/irqdomain.h>
44 #define DRIVER_VERSION "0.2"
45 #define DRIVER_AUTHOR "Alex Williamson <alex.williamson@redhat.com>"
46 #define DRIVER_DESC "Type1 IOMMU driver for VFIO"
48 static bool allow_unsafe_interrupts;
49 module_param_named(allow_unsafe_interrupts,
50 allow_unsafe_interrupts, bool, S_IRUGO | S_IWUSR);
51 MODULE_PARM_DESC(allow_unsafe_interrupts,
52 "Enable VFIO IOMMU support for on platforms without interrupt remapping support.");
54 static bool disable_hugepages;
55 module_param_named(disable_hugepages,
56 disable_hugepages, bool, S_IRUGO | S_IWUSR);
57 MODULE_PARM_DESC(disable_hugepages,
58 "Disable VFIO IOMMU support for IOMMU hugepages.");
60 static unsigned int dma_entry_limit __read_mostly = U16_MAX;
61 module_param_named(dma_entry_limit, dma_entry_limit, uint, 0644);
62 MODULE_PARM_DESC(dma_entry_limit,
63 "Maximum number of user DMA mappings per container (65535).");
66 struct list_head domain_list;
67 struct list_head iova_list;
68 struct vfio_domain *external_domain; /* domain for external user */
70 struct rb_root dma_list;
71 struct blocking_notifier_head notifier;
72 unsigned int dma_avail;
73 unsigned int vaddr_invalid_count;
74 uint64_t pgsize_bitmap;
75 uint64_t num_non_pinned_groups;
76 wait_queue_head_t vaddr_wait;
79 bool dirty_page_tracking;
80 bool pinned_page_dirty_scope;
85 struct iommu_domain *domain;
86 struct list_head next;
87 struct list_head group_list;
88 int prot; /* IOMMU_CACHE */
89 bool fgsp; /* Fine-grained super pages */
94 dma_addr_t iova; /* Device address */
95 unsigned long vaddr; /* Process virtual addr */
96 size_t size; /* Map size (bytes) */
97 int prot; /* IOMMU_READ/WRITE */
99 bool lock_cap; /* capable(CAP_IPC_LOCK) */
101 struct task_struct *task;
102 struct rb_root pfn_list; /* Ex-user pinned pfn list */
103 unsigned long *bitmap;
107 struct iommu_group *iommu_group;
108 struct list_head next;
109 bool mdev_group; /* An mdev group */
110 bool pinned_page_dirty_scope;
114 struct list_head list;
120 * Guest RAM pinning working set or DMA target
124 dma_addr_t iova; /* Device address */
125 unsigned long pfn; /* Host pfn */
126 unsigned int ref_count;
129 struct vfio_regions {
130 struct list_head list;
136 #define IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu) \
137 (!list_empty(&iommu->domain_list))
139 #define DIRTY_BITMAP_BYTES(n) (ALIGN(n, BITS_PER_TYPE(u64)) / BITS_PER_BYTE)
142 * Input argument of number of bits to bitmap_set() is unsigned integer, which
143 * further casts to signed integer for unaligned multi-bit operation,
145 * Then maximum bitmap size supported is 2^31 bits divided by 2^3 bits/byte,
146 * that is 2^28 (256 MB) which maps to 2^31 * 2^12 = 2^43 (8TB) on 4K page
149 #define DIRTY_BITMAP_PAGES_MAX ((u64)INT_MAX)
150 #define DIRTY_BITMAP_SIZE_MAX DIRTY_BITMAP_BYTES(DIRTY_BITMAP_PAGES_MAX)
154 static int put_pfn(unsigned long pfn, int prot);
156 static struct vfio_group *vfio_iommu_find_iommu_group(struct vfio_iommu *iommu,
157 struct iommu_group *iommu_group);
160 * This code handles mapping and unmapping of user data buffers
161 * into DMA'ble space using the IOMMU
164 static struct vfio_dma *vfio_find_dma(struct vfio_iommu *iommu,
165 dma_addr_t start, size_t size)
167 struct rb_node *node = iommu->dma_list.rb_node;
170 struct vfio_dma *dma = rb_entry(node, struct vfio_dma, node);
172 if (start + size <= dma->iova)
173 node = node->rb_left;
174 else if (start >= dma->iova + dma->size)
175 node = node->rb_right;
183 static struct rb_node *vfio_find_dma_first_node(struct vfio_iommu *iommu,
184 dma_addr_t start, size_t size)
186 struct rb_node *res = NULL;
187 struct rb_node *node = iommu->dma_list.rb_node;
188 struct vfio_dma *dma_res = NULL;
191 struct vfio_dma *dma = rb_entry(node, struct vfio_dma, node);
193 if (start < dma->iova + dma->size) {
196 if (start >= dma->iova)
198 node = node->rb_left;
200 node = node->rb_right;
203 if (res && size && dma_res->iova >= start + size)
208 static void vfio_link_dma(struct vfio_iommu *iommu, struct vfio_dma *new)
210 struct rb_node **link = &iommu->dma_list.rb_node, *parent = NULL;
211 struct vfio_dma *dma;
215 dma = rb_entry(parent, struct vfio_dma, node);
217 if (new->iova + new->size <= dma->iova)
218 link = &(*link)->rb_left;
220 link = &(*link)->rb_right;
223 rb_link_node(&new->node, parent, link);
224 rb_insert_color(&new->node, &iommu->dma_list);
227 static void vfio_unlink_dma(struct vfio_iommu *iommu, struct vfio_dma *old)
229 rb_erase(&old->node, &iommu->dma_list);
233 static int vfio_dma_bitmap_alloc(struct vfio_dma *dma, size_t pgsize)
235 uint64_t npages = dma->size / pgsize;
237 if (npages > DIRTY_BITMAP_PAGES_MAX)
241 * Allocate extra 64 bits that are used to calculate shift required for
242 * bitmap_shift_left() to manipulate and club unaligned number of pages
243 * in adjacent vfio_dma ranges.
245 dma->bitmap = kvzalloc(DIRTY_BITMAP_BYTES(npages) + sizeof(u64),
253 static void vfio_dma_bitmap_free(struct vfio_dma *dma)
259 static void vfio_dma_populate_bitmap(struct vfio_dma *dma, size_t pgsize)
262 unsigned long pgshift = __ffs(pgsize);
264 for (p = rb_first(&dma->pfn_list); p; p = rb_next(p)) {
265 struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn, node);
267 bitmap_set(dma->bitmap, (vpfn->iova - dma->iova) >> pgshift, 1);
271 static void vfio_iommu_populate_bitmap_full(struct vfio_iommu *iommu)
274 unsigned long pgshift = __ffs(iommu->pgsize_bitmap);
276 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
277 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
279 bitmap_set(dma->bitmap, 0, dma->size >> pgshift);
283 static int vfio_dma_bitmap_alloc_all(struct vfio_iommu *iommu, size_t pgsize)
287 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
288 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
291 ret = vfio_dma_bitmap_alloc(dma, pgsize);
295 for (p = rb_prev(n); p; p = rb_prev(p)) {
296 struct vfio_dma *dma = rb_entry(n,
297 struct vfio_dma, node);
299 vfio_dma_bitmap_free(dma);
303 vfio_dma_populate_bitmap(dma, pgsize);
308 static void vfio_dma_bitmap_free_all(struct vfio_iommu *iommu)
312 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
313 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
315 vfio_dma_bitmap_free(dma);
320 * Helper Functions for host iova-pfn list
322 static struct vfio_pfn *vfio_find_vpfn(struct vfio_dma *dma, dma_addr_t iova)
324 struct vfio_pfn *vpfn;
325 struct rb_node *node = dma->pfn_list.rb_node;
328 vpfn = rb_entry(node, struct vfio_pfn, node);
330 if (iova < vpfn->iova)
331 node = node->rb_left;
332 else if (iova > vpfn->iova)
333 node = node->rb_right;
340 static void vfio_link_pfn(struct vfio_dma *dma,
341 struct vfio_pfn *new)
343 struct rb_node **link, *parent = NULL;
344 struct vfio_pfn *vpfn;
346 link = &dma->pfn_list.rb_node;
349 vpfn = rb_entry(parent, struct vfio_pfn, node);
351 if (new->iova < vpfn->iova)
352 link = &(*link)->rb_left;
354 link = &(*link)->rb_right;
357 rb_link_node(&new->node, parent, link);
358 rb_insert_color(&new->node, &dma->pfn_list);
361 static void vfio_unlink_pfn(struct vfio_dma *dma, struct vfio_pfn *old)
363 rb_erase(&old->node, &dma->pfn_list);
366 static int vfio_add_to_pfn_list(struct vfio_dma *dma, dma_addr_t iova,
369 struct vfio_pfn *vpfn;
371 vpfn = kzalloc(sizeof(*vpfn), GFP_KERNEL);
378 vfio_link_pfn(dma, vpfn);
382 static void vfio_remove_from_pfn_list(struct vfio_dma *dma,
383 struct vfio_pfn *vpfn)
385 vfio_unlink_pfn(dma, vpfn);
389 static struct vfio_pfn *vfio_iova_get_vfio_pfn(struct vfio_dma *dma,
392 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
399 static int vfio_iova_put_vfio_pfn(struct vfio_dma *dma, struct vfio_pfn *vpfn)
404 if (!vpfn->ref_count) {
405 ret = put_pfn(vpfn->pfn, dma->prot);
406 vfio_remove_from_pfn_list(dma, vpfn);
411 static int vfio_lock_acct(struct vfio_dma *dma, long npage, bool async)
413 struct mm_struct *mm;
419 mm = async ? get_task_mm(dma->task) : dma->task->mm;
421 return -ESRCH; /* process exited */
423 ret = mmap_write_lock_killable(mm);
425 ret = __account_locked_vm(mm, abs(npage), npage > 0, dma->task,
427 mmap_write_unlock(mm);
437 * Some mappings aren't backed by a struct page, for example an mmap'd
438 * MMIO range for our own or another device. These use a different
439 * pfn conversion and shouldn't be tracked as locked pages.
440 * For compound pages, any driver that sets the reserved bit in head
441 * page needs to set the reserved bit in all subpages to be safe.
443 static bool is_invalid_reserved_pfn(unsigned long pfn)
446 return PageReserved(pfn_to_page(pfn));
451 static int put_pfn(unsigned long pfn, int prot)
453 if (!is_invalid_reserved_pfn(pfn)) {
454 struct page *page = pfn_to_page(pfn);
456 unpin_user_pages_dirty_lock(&page, 1, prot & IOMMU_WRITE);
462 static int follow_fault_pfn(struct vm_area_struct *vma, struct mm_struct *mm,
463 unsigned long vaddr, unsigned long *pfn,
470 ret = follow_pte(vma->vm_mm, vaddr, &ptep, &ptl);
472 bool unlocked = false;
474 ret = fixup_user_fault(mm, vaddr,
476 (write_fault ? FAULT_FLAG_WRITE : 0),
484 ret = follow_pte(vma->vm_mm, vaddr, &ptep, &ptl);
489 if (write_fault && !pte_write(*ptep))
492 *pfn = pte_pfn(*ptep);
494 pte_unmap_unlock(ptep, ptl);
499 * Returns the positive number of pfns successfully obtained or a negative
502 static int vaddr_get_pfn(struct mm_struct *mm, unsigned long vaddr,
503 int prot, unsigned long *pfn)
505 struct page *page[1];
506 struct vm_area_struct *vma;
507 unsigned int flags = 0;
510 if (prot & IOMMU_WRITE)
514 ret = pin_user_pages_remote(mm, vaddr, 1, flags | FOLL_LONGTERM,
517 *pfn = page_to_pfn(page[0]);
521 vaddr = untagged_addr(vaddr);
524 vma = find_vma_intersection(mm, vaddr, vaddr + 1);
526 if (vma && vma->vm_flags & VM_PFNMAP) {
527 ret = follow_fault_pfn(vma, mm, vaddr, pfn, prot & IOMMU_WRITE);
532 if (is_invalid_reserved_pfn(*pfn))
539 mmap_read_unlock(mm);
543 static int vfio_wait(struct vfio_iommu *iommu)
547 prepare_to_wait(&iommu->vaddr_wait, &wait, TASK_KILLABLE);
548 mutex_unlock(&iommu->lock);
550 mutex_lock(&iommu->lock);
551 finish_wait(&iommu->vaddr_wait, &wait);
552 if (kthread_should_stop() || !iommu->container_open ||
553 fatal_signal_pending(current)) {
560 * Find dma struct and wait for its vaddr to be valid. iommu lock is dropped
561 * if the task waits, but is re-locked on return. Return result in *dma_p.
562 * Return 0 on success with no waiting, WAITED on success if waited, and -errno
565 static int vfio_find_dma_valid(struct vfio_iommu *iommu, dma_addr_t start,
566 size_t size, struct vfio_dma **dma_p)
571 *dma_p = vfio_find_dma(iommu, start, size);
574 else if (!(*dma_p)->vaddr_invalid)
577 ret = vfio_wait(iommu);
584 * Wait for all vaddr in the dma_list to become valid. iommu lock is dropped
585 * if the task waits, but is re-locked on return. Return 0 on success with no
586 * waiting, WAITED on success if waited, and -errno on error.
588 static int vfio_wait_all_valid(struct vfio_iommu *iommu)
592 while (iommu->vaddr_invalid_count && ret >= 0)
593 ret = vfio_wait(iommu);
599 * Attempt to pin pages. We really don't want to track all the pfns and
600 * the iommu can only map chunks of consecutive pfns anyway, so get the
601 * first page and all consecutive pages with the same locking.
603 static long vfio_pin_pages_remote(struct vfio_dma *dma, unsigned long vaddr,
604 long npage, unsigned long *pfn_base,
607 unsigned long pfn = 0;
608 long ret, pinned = 0, lock_acct = 0;
610 dma_addr_t iova = vaddr - dma->vaddr + dma->iova;
612 /* This code path is only user initiated */
616 ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, pfn_base);
621 rsvd = is_invalid_reserved_pfn(*pfn_base);
624 * Reserved pages aren't counted against the user, externally pinned
625 * pages are already counted against the user.
627 if (!rsvd && !vfio_find_vpfn(dma, iova)) {
628 if (!dma->lock_cap && current->mm->locked_vm + 1 > limit) {
629 put_pfn(*pfn_base, dma->prot);
630 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n", __func__,
631 limit << PAGE_SHIFT);
637 if (unlikely(disable_hugepages))
640 /* Lock all the consecutive pages from pfn_base */
641 for (vaddr += PAGE_SIZE, iova += PAGE_SIZE; pinned < npage;
642 pinned++, vaddr += PAGE_SIZE, iova += PAGE_SIZE) {
643 ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, &pfn);
647 if (pfn != *pfn_base + pinned ||
648 rsvd != is_invalid_reserved_pfn(pfn)) {
649 put_pfn(pfn, dma->prot);
653 if (!rsvd && !vfio_find_vpfn(dma, iova)) {
654 if (!dma->lock_cap &&
655 current->mm->locked_vm + lock_acct + 1 > limit) {
656 put_pfn(pfn, dma->prot);
657 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n",
658 __func__, limit << PAGE_SHIFT);
667 ret = vfio_lock_acct(dma, lock_acct, false);
672 for (pfn = *pfn_base ; pinned ; pfn++, pinned--)
673 put_pfn(pfn, dma->prot);
682 static long vfio_unpin_pages_remote(struct vfio_dma *dma, dma_addr_t iova,
683 unsigned long pfn, long npage,
686 long unlocked = 0, locked = 0;
689 for (i = 0; i < npage; i++, iova += PAGE_SIZE) {
690 if (put_pfn(pfn++, dma->prot)) {
692 if (vfio_find_vpfn(dma, iova))
698 vfio_lock_acct(dma, locked - unlocked, true);
703 static int vfio_pin_page_external(struct vfio_dma *dma, unsigned long vaddr,
704 unsigned long *pfn_base, bool do_accounting)
706 struct mm_struct *mm;
709 mm = get_task_mm(dma->task);
713 ret = vaddr_get_pfn(mm, vaddr, dma->prot, pfn_base);
714 if (ret == 1 && do_accounting && !is_invalid_reserved_pfn(*pfn_base)) {
715 ret = vfio_lock_acct(dma, 1, true);
717 put_pfn(*pfn_base, dma->prot);
719 pr_warn("%s: Task %s (%d) RLIMIT_MEMLOCK "
720 "(%ld) exceeded\n", __func__,
721 dma->task->comm, task_pid_nr(dma->task),
722 task_rlimit(dma->task, RLIMIT_MEMLOCK));
730 static int vfio_unpin_page_external(struct vfio_dma *dma, dma_addr_t iova,
734 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
739 unlocked = vfio_iova_put_vfio_pfn(dma, vpfn);
742 vfio_lock_acct(dma, -unlocked, true);
747 static int vfio_iommu_type1_pin_pages(void *iommu_data,
748 struct iommu_group *iommu_group,
749 unsigned long *user_pfn,
751 unsigned long *phys_pfn)
753 struct vfio_iommu *iommu = iommu_data;
754 struct vfio_group *group;
756 unsigned long remote_vaddr;
757 struct vfio_dma *dma;
761 if (!iommu || !user_pfn || !phys_pfn)
764 /* Supported for v2 version only */
768 mutex_lock(&iommu->lock);
771 * Wait for all necessary vaddr's to be valid so they can be used in
772 * the main loop without dropping the lock, to avoid racing vs unmap.
775 if (iommu->vaddr_invalid_count) {
776 for (i = 0; i < npage; i++) {
777 iova = user_pfn[i] << PAGE_SHIFT;
778 ret = vfio_find_dma_valid(iommu, iova, PAGE_SIZE, &dma);
786 /* Fail if notifier list is empty */
787 if (!iommu->notifier.head) {
793 * If iommu capable domain exist in the container then all pages are
794 * already pinned and accounted. Accouting should be done if there is no
795 * iommu capable domain in the container.
797 do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu);
799 for (i = 0; i < npage; i++) {
800 struct vfio_pfn *vpfn;
802 iova = user_pfn[i] << PAGE_SHIFT;
803 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
809 if ((dma->prot & prot) != prot) {
814 vpfn = vfio_iova_get_vfio_pfn(dma, iova);
816 phys_pfn[i] = vpfn->pfn;
820 remote_vaddr = dma->vaddr + (iova - dma->iova);
821 ret = vfio_pin_page_external(dma, remote_vaddr, &phys_pfn[i],
826 ret = vfio_add_to_pfn_list(dma, iova, phys_pfn[i]);
828 if (put_pfn(phys_pfn[i], dma->prot) && do_accounting)
829 vfio_lock_acct(dma, -1, true);
833 if (iommu->dirty_page_tracking) {
834 unsigned long pgshift = __ffs(iommu->pgsize_bitmap);
837 * Bitmap populated with the smallest supported page
840 bitmap_set(dma->bitmap,
841 (iova - dma->iova) >> pgshift, 1);
846 group = vfio_iommu_find_iommu_group(iommu, iommu_group);
847 if (!group->pinned_page_dirty_scope) {
848 group->pinned_page_dirty_scope = true;
849 iommu->num_non_pinned_groups--;
856 for (j = 0; j < i; j++) {
859 iova = user_pfn[j] << PAGE_SHIFT;
860 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
861 vfio_unpin_page_external(dma, iova, do_accounting);
865 mutex_unlock(&iommu->lock);
869 static int vfio_iommu_type1_unpin_pages(void *iommu_data,
870 unsigned long *user_pfn,
873 struct vfio_iommu *iommu = iommu_data;
877 if (!iommu || !user_pfn)
880 /* Supported for v2 version only */
884 mutex_lock(&iommu->lock);
886 do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu);
887 for (i = 0; i < npage; i++) {
888 struct vfio_dma *dma;
891 iova = user_pfn[i] << PAGE_SHIFT;
892 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
895 vfio_unpin_page_external(dma, iova, do_accounting);
899 mutex_unlock(&iommu->lock);
900 return i > npage ? npage : (i > 0 ? i : -EINVAL);
903 static long vfio_sync_unpin(struct vfio_dma *dma, struct vfio_domain *domain,
904 struct list_head *regions,
905 struct iommu_iotlb_gather *iotlb_gather)
908 struct vfio_regions *entry, *next;
910 iommu_iotlb_sync(domain->domain, iotlb_gather);
912 list_for_each_entry_safe(entry, next, regions, list) {
913 unlocked += vfio_unpin_pages_remote(dma,
915 entry->phys >> PAGE_SHIFT,
916 entry->len >> PAGE_SHIFT,
918 list_del(&entry->list);
928 * Generally, VFIO needs to unpin remote pages after each IOTLB flush.
929 * Therefore, when using IOTLB flush sync interface, VFIO need to keep track
930 * of these regions (currently using a list).
932 * This value specifies maximum number of regions for each IOTLB flush sync.
934 #define VFIO_IOMMU_TLB_SYNC_MAX 512
936 static size_t unmap_unpin_fast(struct vfio_domain *domain,
937 struct vfio_dma *dma, dma_addr_t *iova,
938 size_t len, phys_addr_t phys, long *unlocked,
939 struct list_head *unmapped_list,
941 struct iommu_iotlb_gather *iotlb_gather)
944 struct vfio_regions *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
947 unmapped = iommu_unmap_fast(domain->domain, *iova, len,
955 entry->len = unmapped;
956 list_add_tail(&entry->list, unmapped_list);
964 * Sync if the number of fast-unmap regions hits the limit
965 * or in case of errors.
967 if (*unmapped_cnt >= VFIO_IOMMU_TLB_SYNC_MAX || !unmapped) {
968 *unlocked += vfio_sync_unpin(dma, domain, unmapped_list,
976 static size_t unmap_unpin_slow(struct vfio_domain *domain,
977 struct vfio_dma *dma, dma_addr_t *iova,
978 size_t len, phys_addr_t phys,
981 size_t unmapped = iommu_unmap(domain->domain, *iova, len);
984 *unlocked += vfio_unpin_pages_remote(dma, *iova,
986 unmapped >> PAGE_SHIFT,
994 static long vfio_unmap_unpin(struct vfio_iommu *iommu, struct vfio_dma *dma,
997 dma_addr_t iova = dma->iova, end = dma->iova + dma->size;
998 struct vfio_domain *domain, *d;
999 LIST_HEAD(unmapped_region_list);
1000 struct iommu_iotlb_gather iotlb_gather;
1001 int unmapped_region_cnt = 0;
1007 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
1011 * We use the IOMMU to track the physical addresses, otherwise we'd
1012 * need a much more complicated tracking system. Unfortunately that
1013 * means we need to use one of the iommu domains to figure out the
1014 * pfns to unpin. The rest need to be unmapped in advance so we have
1015 * no iommu translations remaining when the pages are unpinned.
1017 domain = d = list_first_entry(&iommu->domain_list,
1018 struct vfio_domain, next);
1020 list_for_each_entry_continue(d, &iommu->domain_list, next) {
1021 iommu_unmap(d->domain, dma->iova, dma->size);
1025 iommu_iotlb_gather_init(&iotlb_gather);
1026 while (iova < end) {
1027 size_t unmapped, len;
1028 phys_addr_t phys, next;
1030 phys = iommu_iova_to_phys(domain->domain, iova);
1031 if (WARN_ON(!phys)) {
1037 * To optimize for fewer iommu_unmap() calls, each of which
1038 * may require hardware cache flushing, try to find the
1039 * largest contiguous physical memory chunk to unmap.
1041 for (len = PAGE_SIZE;
1042 !domain->fgsp && iova + len < end; len += PAGE_SIZE) {
1043 next = iommu_iova_to_phys(domain->domain, iova + len);
1044 if (next != phys + len)
1049 * First, try to use fast unmap/unpin. In case of failure,
1050 * switch to slow unmap/unpin path.
1052 unmapped = unmap_unpin_fast(domain, dma, &iova, len, phys,
1053 &unlocked, &unmapped_region_list,
1054 &unmapped_region_cnt,
1057 unmapped = unmap_unpin_slow(domain, dma, &iova, len,
1059 if (WARN_ON(!unmapped))
1064 dma->iommu_mapped = false;
1066 if (unmapped_region_cnt) {
1067 unlocked += vfio_sync_unpin(dma, domain, &unmapped_region_list,
1071 if (do_accounting) {
1072 vfio_lock_acct(dma, -unlocked, true);
1078 static void vfio_remove_dma(struct vfio_iommu *iommu, struct vfio_dma *dma)
1080 WARN_ON(!RB_EMPTY_ROOT(&dma->pfn_list));
1081 vfio_unmap_unpin(iommu, dma, true);
1082 vfio_unlink_dma(iommu, dma);
1083 put_task_struct(dma->task);
1084 vfio_dma_bitmap_free(dma);
1085 if (dma->vaddr_invalid) {
1086 iommu->vaddr_invalid_count--;
1087 wake_up_all(&iommu->vaddr_wait);
1093 static void vfio_update_pgsize_bitmap(struct vfio_iommu *iommu)
1095 struct vfio_domain *domain;
1097 iommu->pgsize_bitmap = ULONG_MAX;
1099 list_for_each_entry(domain, &iommu->domain_list, next)
1100 iommu->pgsize_bitmap &= domain->domain->pgsize_bitmap;
1103 * In case the IOMMU supports page sizes smaller than PAGE_SIZE
1104 * we pretend PAGE_SIZE is supported and hide sub-PAGE_SIZE sizes.
1105 * That way the user will be able to map/unmap buffers whose size/
1106 * start address is aligned with PAGE_SIZE. Pinning code uses that
1107 * granularity while iommu driver can use the sub-PAGE_SIZE size
1108 * to map the buffer.
1110 if (iommu->pgsize_bitmap & ~PAGE_MASK) {
1111 iommu->pgsize_bitmap &= PAGE_MASK;
1112 iommu->pgsize_bitmap |= PAGE_SIZE;
1116 static int update_user_bitmap(u64 __user *bitmap, struct vfio_iommu *iommu,
1117 struct vfio_dma *dma, dma_addr_t base_iova,
1120 unsigned long pgshift = __ffs(pgsize);
1121 unsigned long nbits = dma->size >> pgshift;
1122 unsigned long bit_offset = (dma->iova - base_iova) >> pgshift;
1123 unsigned long copy_offset = bit_offset / BITS_PER_LONG;
1124 unsigned long shift = bit_offset % BITS_PER_LONG;
1125 unsigned long leftover;
1128 * mark all pages dirty if any IOMMU capable device is not able
1129 * to report dirty pages and all pages are pinned and mapped.
1131 if (iommu->num_non_pinned_groups && dma->iommu_mapped)
1132 bitmap_set(dma->bitmap, 0, nbits);
1135 bitmap_shift_left(dma->bitmap, dma->bitmap, shift,
1138 if (copy_from_user(&leftover,
1139 (void __user *)(bitmap + copy_offset),
1143 bitmap_or(dma->bitmap, dma->bitmap, &leftover, shift);
1146 if (copy_to_user((void __user *)(bitmap + copy_offset), dma->bitmap,
1147 DIRTY_BITMAP_BYTES(nbits + shift)))
1153 static int vfio_iova_dirty_bitmap(u64 __user *bitmap, struct vfio_iommu *iommu,
1154 dma_addr_t iova, size_t size, size_t pgsize)
1156 struct vfio_dma *dma;
1158 unsigned long pgshift = __ffs(pgsize);
1162 * GET_BITMAP request must fully cover vfio_dma mappings. Multiple
1163 * vfio_dma mappings may be clubbed by specifying large ranges, but
1164 * there must not be any previous mappings bisected by the range.
1165 * An error will be returned if these conditions are not met.
1167 dma = vfio_find_dma(iommu, iova, 1);
1168 if (dma && dma->iova != iova)
1171 dma = vfio_find_dma(iommu, iova + size - 1, 0);
1172 if (dma && dma->iova + dma->size != iova + size)
1175 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
1176 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1178 if (dma->iova < iova)
1181 if (dma->iova > iova + size - 1)
1184 ret = update_user_bitmap(bitmap, iommu, dma, iova, pgsize);
1189 * Re-populate bitmap to include all pinned pages which are
1190 * considered as dirty but exclude pages which are unpinned and
1191 * pages which are marked dirty by vfio_dma_rw()
1193 bitmap_clear(dma->bitmap, 0, dma->size >> pgshift);
1194 vfio_dma_populate_bitmap(dma, pgsize);
1199 static int verify_bitmap_size(uint64_t npages, uint64_t bitmap_size)
1201 if (!npages || !bitmap_size || (bitmap_size > DIRTY_BITMAP_SIZE_MAX) ||
1202 (bitmap_size < DIRTY_BITMAP_BYTES(npages)))
1208 static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
1209 struct vfio_iommu_type1_dma_unmap *unmap,
1210 struct vfio_bitmap *bitmap)
1212 struct vfio_dma *dma, *dma_last = NULL;
1213 size_t unmapped = 0, pgsize;
1214 int ret = -EINVAL, retries = 0;
1215 unsigned long pgshift;
1216 dma_addr_t iova = unmap->iova;
1217 unsigned long size = unmap->size;
1218 bool unmap_all = unmap->flags & VFIO_DMA_UNMAP_FLAG_ALL;
1219 bool invalidate_vaddr = unmap->flags & VFIO_DMA_UNMAP_FLAG_VADDR;
1220 struct rb_node *n, *first_n;
1222 mutex_lock(&iommu->lock);
1224 pgshift = __ffs(iommu->pgsize_bitmap);
1225 pgsize = (size_t)1 << pgshift;
1227 if (iova & (pgsize - 1))
1234 } else if (!size || size & (pgsize - 1)) {
1238 if (iova + size - 1 < iova || size > SIZE_MAX)
1241 /* When dirty tracking is enabled, allow only min supported pgsize */
1242 if ((unmap->flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) &&
1243 (!iommu->dirty_page_tracking || (bitmap->pgsize != pgsize))) {
1247 WARN_ON((pgsize - 1) & PAGE_MASK);
1250 * vfio-iommu-type1 (v1) - User mappings were coalesced together to
1251 * avoid tracking individual mappings. This means that the granularity
1252 * of the original mapping was lost and the user was allowed to attempt
1253 * to unmap any range. Depending on the contiguousness of physical
1254 * memory and page sizes supported by the IOMMU, arbitrary unmaps may
1255 * or may not have worked. We only guaranteed unmap granularity
1256 * matching the original mapping; even though it was untracked here,
1257 * the original mappings are reflected in IOMMU mappings. This
1258 * resulted in a couple unusual behaviors. First, if a range is not
1259 * able to be unmapped, ex. a set of 4k pages that was mapped as a
1260 * 2M hugepage into the IOMMU, the unmap ioctl returns success but with
1261 * a zero sized unmap. Also, if an unmap request overlaps the first
1262 * address of a hugepage, the IOMMU will unmap the entire hugepage.
1263 * This also returns success and the returned unmap size reflects the
1264 * actual size unmapped.
1266 * We attempt to maintain compatibility with this "v1" interface, but
1267 * we take control out of the hands of the IOMMU. Therefore, an unmap
1268 * request offset from the beginning of the original mapping will
1269 * return success with zero sized unmap. And an unmap request covering
1270 * the first iova of mapping will unmap the entire range.
1272 * The v2 version of this interface intends to be more deterministic.
1273 * Unmap requests must fully cover previous mappings. Multiple
1274 * mappings may still be unmaped by specifying large ranges, but there
1275 * must not be any previous mappings bisected by the range. An error
1276 * will be returned if these conditions are not met. The v2 interface
1277 * will only return success and a size of zero if there were no
1278 * mappings within the range.
1280 if (iommu->v2 && !unmap_all) {
1281 dma = vfio_find_dma(iommu, iova, 1);
1282 if (dma && dma->iova != iova)
1285 dma = vfio_find_dma(iommu, iova + size - 1, 0);
1286 if (dma && dma->iova + dma->size != iova + size)
1291 n = first_n = vfio_find_dma_first_node(iommu, iova, size);
1294 dma = rb_entry(n, struct vfio_dma, node);
1295 if (dma->iova >= iova + size)
1298 if (!iommu->v2 && iova > dma->iova)
1301 * Task with same address space who mapped this iova range is
1302 * allowed to unmap the iova range.
1304 if (dma->task->mm != current->mm)
1307 if (invalidate_vaddr) {
1308 if (dma->vaddr_invalid) {
1309 struct rb_node *last_n = n;
1311 for (n = first_n; n != last_n; n = rb_next(n)) {
1313 struct vfio_dma, node);
1314 dma->vaddr_invalid = false;
1315 iommu->vaddr_invalid_count--;
1321 dma->vaddr_invalid = true;
1322 iommu->vaddr_invalid_count++;
1323 unmapped += dma->size;
1328 if (!RB_EMPTY_ROOT(&dma->pfn_list)) {
1329 struct vfio_iommu_type1_dma_unmap nb_unmap;
1331 if (dma_last == dma) {
1332 BUG_ON(++retries > 10);
1338 nb_unmap.iova = dma->iova;
1339 nb_unmap.size = dma->size;
1342 * Notify anyone (mdev vendor drivers) to invalidate and
1343 * unmap iovas within the range we're about to unmap.
1344 * Vendor drivers MUST unpin pages in response to an
1347 mutex_unlock(&iommu->lock);
1348 blocking_notifier_call_chain(&iommu->notifier,
1349 VFIO_IOMMU_NOTIFY_DMA_UNMAP,
1351 mutex_lock(&iommu->lock);
1355 if (unmap->flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) {
1356 ret = update_user_bitmap(bitmap->data, iommu, dma,
1362 unmapped += dma->size;
1364 vfio_remove_dma(iommu, dma);
1368 mutex_unlock(&iommu->lock);
1370 /* Report how much was unmapped */
1371 unmap->size = unmapped;
1376 static int vfio_iommu_map(struct vfio_iommu *iommu, dma_addr_t iova,
1377 unsigned long pfn, long npage, int prot)
1379 struct vfio_domain *d;
1382 list_for_each_entry(d, &iommu->domain_list, next) {
1383 ret = iommu_map(d->domain, iova, (phys_addr_t)pfn << PAGE_SHIFT,
1384 npage << PAGE_SHIFT, prot | d->prot);
1394 list_for_each_entry_continue_reverse(d, &iommu->domain_list, next) {
1395 iommu_unmap(d->domain, iova, npage << PAGE_SHIFT);
1402 static int vfio_pin_map_dma(struct vfio_iommu *iommu, struct vfio_dma *dma,
1405 dma_addr_t iova = dma->iova;
1406 unsigned long vaddr = dma->vaddr;
1407 size_t size = map_size;
1409 unsigned long pfn, limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1413 /* Pin a contiguous chunk of memory */
1414 npage = vfio_pin_pages_remote(dma, vaddr + dma->size,
1415 size >> PAGE_SHIFT, &pfn, limit);
1423 ret = vfio_iommu_map(iommu, iova + dma->size, pfn, npage,
1426 vfio_unpin_pages_remote(dma, iova + dma->size, pfn,
1431 size -= npage << PAGE_SHIFT;
1432 dma->size += npage << PAGE_SHIFT;
1435 dma->iommu_mapped = true;
1438 vfio_remove_dma(iommu, dma);
1444 * Check dma map request is within a valid iova range
1446 static bool vfio_iommu_iova_dma_valid(struct vfio_iommu *iommu,
1447 dma_addr_t start, dma_addr_t end)
1449 struct list_head *iova = &iommu->iova_list;
1450 struct vfio_iova *node;
1452 list_for_each_entry(node, iova, list) {
1453 if (start >= node->start && end <= node->end)
1458 * Check for list_empty() as well since a container with
1459 * a single mdev device will have an empty list.
1461 return list_empty(iova);
1464 static int vfio_dma_do_map(struct vfio_iommu *iommu,
1465 struct vfio_iommu_type1_dma_map *map)
1467 bool set_vaddr = map->flags & VFIO_DMA_MAP_FLAG_VADDR;
1468 dma_addr_t iova = map->iova;
1469 unsigned long vaddr = map->vaddr;
1470 size_t size = map->size;
1471 int ret = 0, prot = 0;
1473 struct vfio_dma *dma;
1475 /* Verify that none of our __u64 fields overflow */
1476 if (map->size != size || map->vaddr != vaddr || map->iova != iova)
1479 /* READ/WRITE from device perspective */
1480 if (map->flags & VFIO_DMA_MAP_FLAG_WRITE)
1481 prot |= IOMMU_WRITE;
1482 if (map->flags & VFIO_DMA_MAP_FLAG_READ)
1485 if ((prot && set_vaddr) || (!prot && !set_vaddr))
1488 mutex_lock(&iommu->lock);
1490 pgsize = (size_t)1 << __ffs(iommu->pgsize_bitmap);
1492 WARN_ON((pgsize - 1) & PAGE_MASK);
1494 if (!size || (size | iova | vaddr) & (pgsize - 1)) {
1499 /* Don't allow IOVA or virtual address wrap */
1500 if (iova + size - 1 < iova || vaddr + size - 1 < vaddr) {
1505 dma = vfio_find_dma(iommu, iova, size);
1509 } else if (!dma->vaddr_invalid || dma->iova != iova ||
1510 dma->size != size) {
1514 dma->vaddr_invalid = false;
1515 iommu->vaddr_invalid_count--;
1516 wake_up_all(&iommu->vaddr_wait);
1524 if (!iommu->dma_avail) {
1529 if (!vfio_iommu_iova_dma_valid(iommu, iova, iova + size - 1)) {
1534 dma = kzalloc(sizeof(*dma), GFP_KERNEL);
1546 * We need to be able to both add to a task's locked memory and test
1547 * against the locked memory limit and we need to be able to do both
1548 * outside of this call path as pinning can be asynchronous via the
1549 * external interfaces for mdev devices. RLIMIT_MEMLOCK requires a
1550 * task_struct and VM locked pages requires an mm_struct, however
1551 * holding an indefinite mm reference is not recommended, therefore we
1552 * only hold a reference to a task. We could hold a reference to
1553 * current, however QEMU uses this call path through vCPU threads,
1554 * which can be killed resulting in a NULL mm and failure in the unmap
1555 * path when called via a different thread. Avoid this problem by
1556 * using the group_leader as threads within the same group require
1557 * both CLONE_THREAD and CLONE_VM and will therefore use the same
1560 * Previously we also used the task for testing CAP_IPC_LOCK at the
1561 * time of pinning and accounting, however has_capability() makes use
1562 * of real_cred, a copy-on-write field, so we can't guarantee that it
1563 * matches group_leader, or in fact that it might not change by the
1564 * time it's evaluated. If a process were to call MAP_DMA with
1565 * CAP_IPC_LOCK but later drop it, it doesn't make sense that they
1566 * possibly see different results for an iommu_mapped vfio_dma vs
1567 * externally mapped. Therefore track CAP_IPC_LOCK in vfio_dma at the
1568 * time of calling MAP_DMA.
1570 get_task_struct(current->group_leader);
1571 dma->task = current->group_leader;
1572 dma->lock_cap = capable(CAP_IPC_LOCK);
1574 dma->pfn_list = RB_ROOT;
1576 /* Insert zero-sized and grow as we map chunks of it */
1577 vfio_link_dma(iommu, dma);
1579 /* Don't pin and map if container doesn't contain IOMMU capable domain*/
1580 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
1583 ret = vfio_pin_map_dma(iommu, dma, size);
1585 if (!ret && iommu->dirty_page_tracking) {
1586 ret = vfio_dma_bitmap_alloc(dma, pgsize);
1588 vfio_remove_dma(iommu, dma);
1592 mutex_unlock(&iommu->lock);
1596 static int vfio_bus_type(struct device *dev, void *data)
1598 struct bus_type **bus = data;
1600 if (*bus && *bus != dev->bus)
1608 static int vfio_iommu_replay(struct vfio_iommu *iommu,
1609 struct vfio_domain *domain)
1611 struct vfio_domain *d = NULL;
1613 unsigned long limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1616 ret = vfio_wait_all_valid(iommu);
1620 /* Arbitrarily pick the first domain in the list for lookups */
1621 if (!list_empty(&iommu->domain_list))
1622 d = list_first_entry(&iommu->domain_list,
1623 struct vfio_domain, next);
1625 n = rb_first(&iommu->dma_list);
1627 for (; n; n = rb_next(n)) {
1628 struct vfio_dma *dma;
1631 dma = rb_entry(n, struct vfio_dma, node);
1634 while (iova < dma->iova + dma->size) {
1638 if (dma->iommu_mapped) {
1642 if (WARN_ON(!d)) { /* mapped w/o a domain?! */
1647 phys = iommu_iova_to_phys(d->domain, iova);
1649 if (WARN_ON(!phys)) {
1657 while (i < dma->iova + dma->size &&
1658 p == iommu_iova_to_phys(d->domain, i)) {
1665 unsigned long vaddr = dma->vaddr +
1667 size_t n = dma->iova + dma->size - iova;
1670 npage = vfio_pin_pages_remote(dma, vaddr,
1679 phys = pfn << PAGE_SHIFT;
1680 size = npage << PAGE_SHIFT;
1683 ret = iommu_map(domain->domain, iova, phys,
1684 size, dma->prot | domain->prot);
1686 if (!dma->iommu_mapped)
1687 vfio_unpin_pages_remote(dma, iova,
1698 /* All dmas are now mapped, defer to second tree walk for unwind */
1699 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
1700 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1702 dma->iommu_mapped = true;
1708 for (; n; n = rb_prev(n)) {
1709 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1712 if (dma->iommu_mapped) {
1713 iommu_unmap(domain->domain, dma->iova, dma->size);
1718 while (iova < dma->iova + dma->size) {
1719 phys_addr_t phys, p;
1723 phys = iommu_iova_to_phys(domain->domain, iova);
1732 while (i < dma->iova + dma->size &&
1733 p == iommu_iova_to_phys(domain->domain, i)) {
1739 iommu_unmap(domain->domain, iova, size);
1740 vfio_unpin_pages_remote(dma, iova, phys >> PAGE_SHIFT,
1741 size >> PAGE_SHIFT, true);
1749 * We change our unmap behavior slightly depending on whether the IOMMU
1750 * supports fine-grained superpages. IOMMUs like AMD-Vi will use a superpage
1751 * for practically any contiguous power-of-two mapping we give it. This means
1752 * we don't need to look for contiguous chunks ourselves to make unmapping
1753 * more efficient. On IOMMUs with coarse-grained super pages, like Intel VT-d
1754 * with discrete 2M/1G/512G/1T superpages, identifying contiguous chunks
1755 * significantly boosts non-hugetlbfs mappings and doesn't seem to hurt when
1756 * hugetlbfs is in use.
1758 static void vfio_test_domain_fgsp(struct vfio_domain *domain)
1761 int ret, order = get_order(PAGE_SIZE * 2);
1763 pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order);
1767 ret = iommu_map(domain->domain, 0, page_to_phys(pages), PAGE_SIZE * 2,
1768 IOMMU_READ | IOMMU_WRITE | domain->prot);
1770 size_t unmapped = iommu_unmap(domain->domain, 0, PAGE_SIZE);
1772 if (unmapped == PAGE_SIZE)
1773 iommu_unmap(domain->domain, PAGE_SIZE, PAGE_SIZE);
1775 domain->fgsp = true;
1778 __free_pages(pages, order);
1781 static struct vfio_group *find_iommu_group(struct vfio_domain *domain,
1782 struct iommu_group *iommu_group)
1784 struct vfio_group *g;
1786 list_for_each_entry(g, &domain->group_list, next) {
1787 if (g->iommu_group == iommu_group)
1794 static struct vfio_group *vfio_iommu_find_iommu_group(struct vfio_iommu *iommu,
1795 struct iommu_group *iommu_group)
1797 struct vfio_domain *domain;
1798 struct vfio_group *group = NULL;
1800 list_for_each_entry(domain, &iommu->domain_list, next) {
1801 group = find_iommu_group(domain, iommu_group);
1806 if (iommu->external_domain)
1807 group = find_iommu_group(iommu->external_domain, iommu_group);
1812 static bool vfio_iommu_has_sw_msi(struct list_head *group_resv_regions,
1815 struct iommu_resv_region *region;
1818 list_for_each_entry(region, group_resv_regions, list) {
1820 * The presence of any 'real' MSI regions should take
1821 * precedence over the software-managed one if the
1822 * IOMMU driver happens to advertise both types.
1824 if (region->type == IOMMU_RESV_MSI) {
1829 if (region->type == IOMMU_RESV_SW_MSI) {
1830 *base = region->start;
1838 static struct device *vfio_mdev_get_iommu_device(struct device *dev)
1840 struct device *(*fn)(struct device *dev);
1841 struct device *iommu_device;
1843 fn = symbol_get(mdev_get_iommu_device);
1845 iommu_device = fn(dev);
1846 symbol_put(mdev_get_iommu_device);
1848 return iommu_device;
1854 static int vfio_mdev_attach_domain(struct device *dev, void *data)
1856 struct iommu_domain *domain = data;
1857 struct device *iommu_device;
1859 iommu_device = vfio_mdev_get_iommu_device(dev);
1861 if (iommu_dev_feature_enabled(iommu_device, IOMMU_DEV_FEAT_AUX))
1862 return iommu_aux_attach_device(domain, iommu_device);
1864 return iommu_attach_device(domain, iommu_device);
1870 static int vfio_mdev_detach_domain(struct device *dev, void *data)
1872 struct iommu_domain *domain = data;
1873 struct device *iommu_device;
1875 iommu_device = vfio_mdev_get_iommu_device(dev);
1877 if (iommu_dev_feature_enabled(iommu_device, IOMMU_DEV_FEAT_AUX))
1878 iommu_aux_detach_device(domain, iommu_device);
1880 iommu_detach_device(domain, iommu_device);
1886 static int vfio_iommu_attach_group(struct vfio_domain *domain,
1887 struct vfio_group *group)
1889 if (group->mdev_group)
1890 return iommu_group_for_each_dev(group->iommu_group,
1892 vfio_mdev_attach_domain);
1894 return iommu_attach_group(domain->domain, group->iommu_group);
1897 static void vfio_iommu_detach_group(struct vfio_domain *domain,
1898 struct vfio_group *group)
1900 if (group->mdev_group)
1901 iommu_group_for_each_dev(group->iommu_group, domain->domain,
1902 vfio_mdev_detach_domain);
1904 iommu_detach_group(domain->domain, group->iommu_group);
1907 static bool vfio_bus_is_mdev(struct bus_type *bus)
1909 struct bus_type *mdev_bus;
1912 mdev_bus = symbol_get(mdev_bus_type);
1914 ret = (bus == mdev_bus);
1915 symbol_put(mdev_bus_type);
1921 static int vfio_mdev_iommu_device(struct device *dev, void *data)
1923 struct device **old = data, *new;
1925 new = vfio_mdev_get_iommu_device(dev);
1926 if (!new || (*old && *old != new))
1935 * This is a helper function to insert an address range to iova list.
1936 * The list is initially created with a single entry corresponding to
1937 * the IOMMU domain geometry to which the device group is attached.
1938 * The list aperture gets modified when a new domain is added to the
1939 * container if the new aperture doesn't conflict with the current one
1940 * or with any existing dma mappings. The list is also modified to
1941 * exclude any reserved regions associated with the device group.
1943 static int vfio_iommu_iova_insert(struct list_head *head,
1944 dma_addr_t start, dma_addr_t end)
1946 struct vfio_iova *region;
1948 region = kmalloc(sizeof(*region), GFP_KERNEL);
1952 INIT_LIST_HEAD(®ion->list);
1953 region->start = start;
1956 list_add_tail(®ion->list, head);
1961 * Check the new iommu aperture conflicts with existing aper or with any
1962 * existing dma mappings.
1964 static bool vfio_iommu_aper_conflict(struct vfio_iommu *iommu,
1965 dma_addr_t start, dma_addr_t end)
1967 struct vfio_iova *first, *last;
1968 struct list_head *iova = &iommu->iova_list;
1970 if (list_empty(iova))
1973 /* Disjoint sets, return conflict */
1974 first = list_first_entry(iova, struct vfio_iova, list);
1975 last = list_last_entry(iova, struct vfio_iova, list);
1976 if (start > last->end || end < first->start)
1979 /* Check for any existing dma mappings below the new start */
1980 if (start > first->start) {
1981 if (vfio_find_dma(iommu, first->start, start - first->start))
1985 /* Check for any existing dma mappings beyond the new end */
1986 if (end < last->end) {
1987 if (vfio_find_dma(iommu, end + 1, last->end - end))
1995 * Resize iommu iova aperture window. This is called only if the new
1996 * aperture has no conflict with existing aperture and dma mappings.
1998 static int vfio_iommu_aper_resize(struct list_head *iova,
1999 dma_addr_t start, dma_addr_t end)
2001 struct vfio_iova *node, *next;
2003 if (list_empty(iova))
2004 return vfio_iommu_iova_insert(iova, start, end);
2006 /* Adjust iova list start */
2007 list_for_each_entry_safe(node, next, iova, list) {
2008 if (start < node->start)
2010 if (start >= node->start && start < node->end) {
2011 node->start = start;
2014 /* Delete nodes before new start */
2015 list_del(&node->list);
2019 /* Adjust iova list end */
2020 list_for_each_entry_safe(node, next, iova, list) {
2021 if (end > node->end)
2023 if (end > node->start && end <= node->end) {
2027 /* Delete nodes after new end */
2028 list_del(&node->list);
2036 * Check reserved region conflicts with existing dma mappings
2038 static bool vfio_iommu_resv_conflict(struct vfio_iommu *iommu,
2039 struct list_head *resv_regions)
2041 struct iommu_resv_region *region;
2043 /* Check for conflict with existing dma mappings */
2044 list_for_each_entry(region, resv_regions, list) {
2045 if (region->type == IOMMU_RESV_DIRECT_RELAXABLE)
2048 if (vfio_find_dma(iommu, region->start, region->length))
2056 * Check iova region overlap with reserved regions and
2057 * exclude them from the iommu iova range
2059 static int vfio_iommu_resv_exclude(struct list_head *iova,
2060 struct list_head *resv_regions)
2062 struct iommu_resv_region *resv;
2063 struct vfio_iova *n, *next;
2065 list_for_each_entry(resv, resv_regions, list) {
2066 phys_addr_t start, end;
2068 if (resv->type == IOMMU_RESV_DIRECT_RELAXABLE)
2071 start = resv->start;
2072 end = resv->start + resv->length - 1;
2074 list_for_each_entry_safe(n, next, iova, list) {
2078 if (start > n->end || end < n->start)
2081 * Insert a new node if current node overlaps with the
2082 * reserve region to exlude that from valid iova range.
2083 * Note that, new node is inserted before the current
2084 * node and finally the current node is deleted keeping
2085 * the list updated and sorted.
2087 if (start > n->start)
2088 ret = vfio_iommu_iova_insert(&n->list, n->start,
2090 if (!ret && end < n->end)
2091 ret = vfio_iommu_iova_insert(&n->list, end + 1,
2101 if (list_empty(iova))
2107 static void vfio_iommu_resv_free(struct list_head *resv_regions)
2109 struct iommu_resv_region *n, *next;
2111 list_for_each_entry_safe(n, next, resv_regions, list) {
2117 static void vfio_iommu_iova_free(struct list_head *iova)
2119 struct vfio_iova *n, *next;
2121 list_for_each_entry_safe(n, next, iova, list) {
2127 static int vfio_iommu_iova_get_copy(struct vfio_iommu *iommu,
2128 struct list_head *iova_copy)
2130 struct list_head *iova = &iommu->iova_list;
2131 struct vfio_iova *n;
2134 list_for_each_entry(n, iova, list) {
2135 ret = vfio_iommu_iova_insert(iova_copy, n->start, n->end);
2143 vfio_iommu_iova_free(iova_copy);
2147 static void vfio_iommu_iova_insert_copy(struct vfio_iommu *iommu,
2148 struct list_head *iova_copy)
2150 struct list_head *iova = &iommu->iova_list;
2152 vfio_iommu_iova_free(iova);
2154 list_splice_tail(iova_copy, iova);
2157 static int vfio_iommu_type1_attach_group(void *iommu_data,
2158 struct iommu_group *iommu_group)
2160 struct vfio_iommu *iommu = iommu_data;
2161 struct vfio_group *group;
2162 struct vfio_domain *domain, *d;
2163 struct bus_type *bus = NULL;
2165 bool resv_msi, msi_remap;
2166 phys_addr_t resv_msi_base = 0;
2167 struct iommu_domain_geometry geo;
2168 LIST_HEAD(iova_copy);
2169 LIST_HEAD(group_resv_regions);
2171 mutex_lock(&iommu->lock);
2173 /* Check for duplicates */
2174 if (vfio_iommu_find_iommu_group(iommu, iommu_group)) {
2175 mutex_unlock(&iommu->lock);
2179 group = kzalloc(sizeof(*group), GFP_KERNEL);
2180 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2181 if (!group || !domain) {
2186 group->iommu_group = iommu_group;
2188 /* Determine bus_type in order to allocate a domain */
2189 ret = iommu_group_for_each_dev(iommu_group, &bus, vfio_bus_type);
2193 if (vfio_bus_is_mdev(bus)) {
2194 struct device *iommu_device = NULL;
2196 group->mdev_group = true;
2198 /* Determine the isolation type */
2199 ret = iommu_group_for_each_dev(iommu_group, &iommu_device,
2200 vfio_mdev_iommu_device);
2201 if (ret || !iommu_device) {
2202 if (!iommu->external_domain) {
2203 INIT_LIST_HEAD(&domain->group_list);
2204 iommu->external_domain = domain;
2205 vfio_update_pgsize_bitmap(iommu);
2210 list_add(&group->next,
2211 &iommu->external_domain->group_list);
2213 * Non-iommu backed group cannot dirty memory directly,
2214 * it can only use interfaces that provide dirty
2216 * The iommu scope can only be promoted with the
2217 * addition of a dirty tracking group.
2219 group->pinned_page_dirty_scope = true;
2220 mutex_unlock(&iommu->lock);
2225 bus = iommu_device->bus;
2228 domain->domain = iommu_domain_alloc(bus);
2229 if (!domain->domain) {
2234 if (iommu->nesting) {
2237 ret = iommu_domain_set_attr(domain->domain, DOMAIN_ATTR_NESTING,
2243 ret = vfio_iommu_attach_group(domain, group);
2247 /* Get aperture info */
2248 iommu_domain_get_attr(domain->domain, DOMAIN_ATTR_GEOMETRY, &geo);
2250 if (vfio_iommu_aper_conflict(iommu, geo.aperture_start,
2251 geo.aperture_end)) {
2256 ret = iommu_get_group_resv_regions(iommu_group, &group_resv_regions);
2260 if (vfio_iommu_resv_conflict(iommu, &group_resv_regions)) {
2266 * We don't want to work on the original iova list as the list
2267 * gets modified and in case of failure we have to retain the
2268 * original list. Get a copy here.
2270 ret = vfio_iommu_iova_get_copy(iommu, &iova_copy);
2274 ret = vfio_iommu_aper_resize(&iova_copy, geo.aperture_start,
2279 ret = vfio_iommu_resv_exclude(&iova_copy, &group_resv_regions);
2283 resv_msi = vfio_iommu_has_sw_msi(&group_resv_regions, &resv_msi_base);
2285 INIT_LIST_HEAD(&domain->group_list);
2286 list_add(&group->next, &domain->group_list);
2288 msi_remap = irq_domain_check_msi_remap() ||
2289 iommu_capable(bus, IOMMU_CAP_INTR_REMAP);
2291 if (!allow_unsafe_interrupts && !msi_remap) {
2292 pr_warn("%s: No interrupt remapping support. Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n",
2298 if (iommu_capable(bus, IOMMU_CAP_CACHE_COHERENCY))
2299 domain->prot |= IOMMU_CACHE;
2302 * Try to match an existing compatible domain. We don't want to
2303 * preclude an IOMMU driver supporting multiple bus_types and being
2304 * able to include different bus_types in the same IOMMU domain, so
2305 * we test whether the domains use the same iommu_ops rather than
2306 * testing if they're on the same bus_type.
2308 list_for_each_entry(d, &iommu->domain_list, next) {
2309 if (d->domain->ops == domain->domain->ops &&
2310 d->prot == domain->prot) {
2311 vfio_iommu_detach_group(domain, group);
2312 if (!vfio_iommu_attach_group(d, group)) {
2313 list_add(&group->next, &d->group_list);
2314 iommu_domain_free(domain->domain);
2319 ret = vfio_iommu_attach_group(domain, group);
2325 vfio_test_domain_fgsp(domain);
2327 /* replay mappings on new domains */
2328 ret = vfio_iommu_replay(iommu, domain);
2333 ret = iommu_get_msi_cookie(domain->domain, resv_msi_base);
2334 if (ret && ret != -ENODEV)
2338 list_add(&domain->next, &iommu->domain_list);
2339 vfio_update_pgsize_bitmap(iommu);
2341 /* Delete the old one and insert new iova list */
2342 vfio_iommu_iova_insert_copy(iommu, &iova_copy);
2345 * An iommu backed group can dirty memory directly and therefore
2346 * demotes the iommu scope until it declares itself dirty tracking
2347 * capable via the page pinning interface.
2349 iommu->num_non_pinned_groups++;
2350 mutex_unlock(&iommu->lock);
2351 vfio_iommu_resv_free(&group_resv_regions);
2356 vfio_iommu_detach_group(domain, group);
2358 iommu_domain_free(domain->domain);
2359 vfio_iommu_iova_free(&iova_copy);
2360 vfio_iommu_resv_free(&group_resv_regions);
2364 mutex_unlock(&iommu->lock);
2368 static void vfio_iommu_unmap_unpin_all(struct vfio_iommu *iommu)
2370 struct rb_node *node;
2372 while ((node = rb_first(&iommu->dma_list)))
2373 vfio_remove_dma(iommu, rb_entry(node, struct vfio_dma, node));
2376 static void vfio_iommu_unmap_unpin_reaccount(struct vfio_iommu *iommu)
2378 struct rb_node *n, *p;
2380 n = rb_first(&iommu->dma_list);
2381 for (; n; n = rb_next(n)) {
2382 struct vfio_dma *dma;
2383 long locked = 0, unlocked = 0;
2385 dma = rb_entry(n, struct vfio_dma, node);
2386 unlocked += vfio_unmap_unpin(iommu, dma, false);
2387 p = rb_first(&dma->pfn_list);
2388 for (; p; p = rb_next(p)) {
2389 struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn,
2392 if (!is_invalid_reserved_pfn(vpfn->pfn))
2395 vfio_lock_acct(dma, locked - unlocked, true);
2400 * Called when a domain is removed in detach. It is possible that
2401 * the removed domain decided the iova aperture window. Modify the
2402 * iova aperture with the smallest window among existing domains.
2404 static void vfio_iommu_aper_expand(struct vfio_iommu *iommu,
2405 struct list_head *iova_copy)
2407 struct vfio_domain *domain;
2408 struct iommu_domain_geometry geo;
2409 struct vfio_iova *node;
2410 dma_addr_t start = 0;
2411 dma_addr_t end = (dma_addr_t)~0;
2413 if (list_empty(iova_copy))
2416 list_for_each_entry(domain, &iommu->domain_list, next) {
2417 iommu_domain_get_attr(domain->domain, DOMAIN_ATTR_GEOMETRY,
2419 if (geo.aperture_start > start)
2420 start = geo.aperture_start;
2421 if (geo.aperture_end < end)
2422 end = geo.aperture_end;
2425 /* Modify aperture limits. The new aper is either same or bigger */
2426 node = list_first_entry(iova_copy, struct vfio_iova, list);
2427 node->start = start;
2428 node = list_last_entry(iova_copy, struct vfio_iova, list);
2433 * Called when a group is detached. The reserved regions for that
2434 * group can be part of valid iova now. But since reserved regions
2435 * may be duplicated among groups, populate the iova valid regions
2438 static int vfio_iommu_resv_refresh(struct vfio_iommu *iommu,
2439 struct list_head *iova_copy)
2441 struct vfio_domain *d;
2442 struct vfio_group *g;
2443 struct vfio_iova *node;
2444 dma_addr_t start, end;
2445 LIST_HEAD(resv_regions);
2448 if (list_empty(iova_copy))
2451 list_for_each_entry(d, &iommu->domain_list, next) {
2452 list_for_each_entry(g, &d->group_list, next) {
2453 ret = iommu_get_group_resv_regions(g->iommu_group,
2460 node = list_first_entry(iova_copy, struct vfio_iova, list);
2461 start = node->start;
2462 node = list_last_entry(iova_copy, struct vfio_iova, list);
2465 /* purge the iova list and create new one */
2466 vfio_iommu_iova_free(iova_copy);
2468 ret = vfio_iommu_aper_resize(iova_copy, start, end);
2472 /* Exclude current reserved regions from iova ranges */
2473 ret = vfio_iommu_resv_exclude(iova_copy, &resv_regions);
2475 vfio_iommu_resv_free(&resv_regions);
2479 static void vfio_iommu_type1_detach_group(void *iommu_data,
2480 struct iommu_group *iommu_group)
2482 struct vfio_iommu *iommu = iommu_data;
2483 struct vfio_domain *domain;
2484 struct vfio_group *group;
2485 bool update_dirty_scope = false;
2486 LIST_HEAD(iova_copy);
2488 mutex_lock(&iommu->lock);
2490 if (iommu->external_domain) {
2491 group = find_iommu_group(iommu->external_domain, iommu_group);
2493 update_dirty_scope = !group->pinned_page_dirty_scope;
2494 list_del(&group->next);
2497 if (list_empty(&iommu->external_domain->group_list)) {
2498 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu)) {
2499 WARN_ON(iommu->notifier.head);
2500 vfio_iommu_unmap_unpin_all(iommu);
2503 kfree(iommu->external_domain);
2504 iommu->external_domain = NULL;
2506 goto detach_group_done;
2511 * Get a copy of iova list. This will be used to update
2512 * and to replace the current one later. Please note that
2513 * we will leave the original list as it is if update fails.
2515 vfio_iommu_iova_get_copy(iommu, &iova_copy);
2517 list_for_each_entry(domain, &iommu->domain_list, next) {
2518 group = find_iommu_group(domain, iommu_group);
2522 vfio_iommu_detach_group(domain, group);
2523 update_dirty_scope = !group->pinned_page_dirty_scope;
2524 list_del(&group->next);
2527 * Group ownership provides privilege, if the group list is
2528 * empty, the domain goes away. If it's the last domain with
2529 * iommu and external domain doesn't exist, then all the
2530 * mappings go away too. If it's the last domain with iommu and
2531 * external domain exist, update accounting
2533 if (list_empty(&domain->group_list)) {
2534 if (list_is_singular(&iommu->domain_list)) {
2535 if (!iommu->external_domain) {
2536 WARN_ON(iommu->notifier.head);
2537 vfio_iommu_unmap_unpin_all(iommu);
2539 vfio_iommu_unmap_unpin_reaccount(iommu);
2542 iommu_domain_free(domain->domain);
2543 list_del(&domain->next);
2545 vfio_iommu_aper_expand(iommu, &iova_copy);
2546 vfio_update_pgsize_bitmap(iommu);
2551 if (!vfio_iommu_resv_refresh(iommu, &iova_copy))
2552 vfio_iommu_iova_insert_copy(iommu, &iova_copy);
2554 vfio_iommu_iova_free(&iova_copy);
2558 * Removal of a group without dirty tracking may allow the iommu scope
2561 if (update_dirty_scope) {
2562 iommu->num_non_pinned_groups--;
2563 if (iommu->dirty_page_tracking)
2564 vfio_iommu_populate_bitmap_full(iommu);
2566 mutex_unlock(&iommu->lock);
2569 static void *vfio_iommu_type1_open(unsigned long arg)
2571 struct vfio_iommu *iommu;
2573 iommu = kzalloc(sizeof(*iommu), GFP_KERNEL);
2575 return ERR_PTR(-ENOMEM);
2578 case VFIO_TYPE1_IOMMU:
2580 case VFIO_TYPE1_NESTING_IOMMU:
2581 iommu->nesting = true;
2583 case VFIO_TYPE1v2_IOMMU:
2588 return ERR_PTR(-EINVAL);
2591 INIT_LIST_HEAD(&iommu->domain_list);
2592 INIT_LIST_HEAD(&iommu->iova_list);
2593 iommu->dma_list = RB_ROOT;
2594 iommu->dma_avail = dma_entry_limit;
2595 iommu->container_open = true;
2596 mutex_init(&iommu->lock);
2597 BLOCKING_INIT_NOTIFIER_HEAD(&iommu->notifier);
2598 init_waitqueue_head(&iommu->vaddr_wait);
2603 static void vfio_release_domain(struct vfio_domain *domain, bool external)
2605 struct vfio_group *group, *group_tmp;
2607 list_for_each_entry_safe(group, group_tmp,
2608 &domain->group_list, next) {
2610 vfio_iommu_detach_group(domain, group);
2611 list_del(&group->next);
2616 iommu_domain_free(domain->domain);
2619 static void vfio_iommu_type1_release(void *iommu_data)
2621 struct vfio_iommu *iommu = iommu_data;
2622 struct vfio_domain *domain, *domain_tmp;
2624 if (iommu->external_domain) {
2625 vfio_release_domain(iommu->external_domain, true);
2626 kfree(iommu->external_domain);
2629 vfio_iommu_unmap_unpin_all(iommu);
2631 list_for_each_entry_safe(domain, domain_tmp,
2632 &iommu->domain_list, next) {
2633 vfio_release_domain(domain, false);
2634 list_del(&domain->next);
2638 vfio_iommu_iova_free(&iommu->iova_list);
2643 static int vfio_domains_have_iommu_cache(struct vfio_iommu *iommu)
2645 struct vfio_domain *domain;
2648 mutex_lock(&iommu->lock);
2649 list_for_each_entry(domain, &iommu->domain_list, next) {
2650 if (!(domain->prot & IOMMU_CACHE)) {
2655 mutex_unlock(&iommu->lock);
2660 static int vfio_iommu_type1_check_extension(struct vfio_iommu *iommu,
2664 case VFIO_TYPE1_IOMMU:
2665 case VFIO_TYPE1v2_IOMMU:
2666 case VFIO_TYPE1_NESTING_IOMMU:
2667 case VFIO_UNMAP_ALL:
2668 case VFIO_UPDATE_VADDR:
2670 case VFIO_DMA_CC_IOMMU:
2673 return vfio_domains_have_iommu_cache(iommu);
2679 static int vfio_iommu_iova_add_cap(struct vfio_info_cap *caps,
2680 struct vfio_iommu_type1_info_cap_iova_range *cap_iovas,
2683 struct vfio_info_cap_header *header;
2684 struct vfio_iommu_type1_info_cap_iova_range *iova_cap;
2686 header = vfio_info_cap_add(caps, size,
2687 VFIO_IOMMU_TYPE1_INFO_CAP_IOVA_RANGE, 1);
2689 return PTR_ERR(header);
2691 iova_cap = container_of(header,
2692 struct vfio_iommu_type1_info_cap_iova_range,
2694 iova_cap->nr_iovas = cap_iovas->nr_iovas;
2695 memcpy(iova_cap->iova_ranges, cap_iovas->iova_ranges,
2696 cap_iovas->nr_iovas * sizeof(*cap_iovas->iova_ranges));
2700 static int vfio_iommu_iova_build_caps(struct vfio_iommu *iommu,
2701 struct vfio_info_cap *caps)
2703 struct vfio_iommu_type1_info_cap_iova_range *cap_iovas;
2704 struct vfio_iova *iova;
2706 int iovas = 0, i = 0, ret;
2708 list_for_each_entry(iova, &iommu->iova_list, list)
2713 * Return 0 as a container with a single mdev device
2714 * will have an empty list
2719 size = sizeof(*cap_iovas) + (iovas * sizeof(*cap_iovas->iova_ranges));
2721 cap_iovas = kzalloc(size, GFP_KERNEL);
2725 cap_iovas->nr_iovas = iovas;
2727 list_for_each_entry(iova, &iommu->iova_list, list) {
2728 cap_iovas->iova_ranges[i].start = iova->start;
2729 cap_iovas->iova_ranges[i].end = iova->end;
2733 ret = vfio_iommu_iova_add_cap(caps, cap_iovas, size);
2739 static int vfio_iommu_migration_build_caps(struct vfio_iommu *iommu,
2740 struct vfio_info_cap *caps)
2742 struct vfio_iommu_type1_info_cap_migration cap_mig;
2744 cap_mig.header.id = VFIO_IOMMU_TYPE1_INFO_CAP_MIGRATION;
2745 cap_mig.header.version = 1;
2748 /* support minimum pgsize */
2749 cap_mig.pgsize_bitmap = (size_t)1 << __ffs(iommu->pgsize_bitmap);
2750 cap_mig.max_dirty_bitmap_size = DIRTY_BITMAP_SIZE_MAX;
2752 return vfio_info_add_capability(caps, &cap_mig.header, sizeof(cap_mig));
2755 static int vfio_iommu_dma_avail_build_caps(struct vfio_iommu *iommu,
2756 struct vfio_info_cap *caps)
2758 struct vfio_iommu_type1_info_dma_avail cap_dma_avail;
2760 cap_dma_avail.header.id = VFIO_IOMMU_TYPE1_INFO_DMA_AVAIL;
2761 cap_dma_avail.header.version = 1;
2763 cap_dma_avail.avail = iommu->dma_avail;
2765 return vfio_info_add_capability(caps, &cap_dma_avail.header,
2766 sizeof(cap_dma_avail));
2769 static int vfio_iommu_type1_get_info(struct vfio_iommu *iommu,
2772 struct vfio_iommu_type1_info info;
2773 unsigned long minsz;
2774 struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
2775 unsigned long capsz;
2778 minsz = offsetofend(struct vfio_iommu_type1_info, iova_pgsizes);
2780 /* For backward compatibility, cannot require this */
2781 capsz = offsetofend(struct vfio_iommu_type1_info, cap_offset);
2783 if (copy_from_user(&info, (void __user *)arg, minsz))
2786 if (info.argsz < minsz)
2789 if (info.argsz >= capsz) {
2791 info.cap_offset = 0; /* output, no-recopy necessary */
2794 mutex_lock(&iommu->lock);
2795 info.flags = VFIO_IOMMU_INFO_PGSIZES;
2797 info.iova_pgsizes = iommu->pgsize_bitmap;
2799 ret = vfio_iommu_migration_build_caps(iommu, &caps);
2802 ret = vfio_iommu_dma_avail_build_caps(iommu, &caps);
2805 ret = vfio_iommu_iova_build_caps(iommu, &caps);
2807 mutex_unlock(&iommu->lock);
2813 info.flags |= VFIO_IOMMU_INFO_CAPS;
2815 if (info.argsz < sizeof(info) + caps.size) {
2816 info.argsz = sizeof(info) + caps.size;
2818 vfio_info_cap_shift(&caps, sizeof(info));
2819 if (copy_to_user((void __user *)arg +
2820 sizeof(info), caps.buf,
2825 info.cap_offset = sizeof(info);
2831 return copy_to_user((void __user *)arg, &info, minsz) ?
2835 static int vfio_iommu_type1_map_dma(struct vfio_iommu *iommu,
2838 struct vfio_iommu_type1_dma_map map;
2839 unsigned long minsz;
2840 uint32_t mask = VFIO_DMA_MAP_FLAG_READ | VFIO_DMA_MAP_FLAG_WRITE |
2841 VFIO_DMA_MAP_FLAG_VADDR;
2843 minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
2845 if (copy_from_user(&map, (void __user *)arg, minsz))
2848 if (map.argsz < minsz || map.flags & ~mask)
2851 return vfio_dma_do_map(iommu, &map);
2854 static int vfio_iommu_type1_unmap_dma(struct vfio_iommu *iommu,
2857 struct vfio_iommu_type1_dma_unmap unmap;
2858 struct vfio_bitmap bitmap = { 0 };
2859 uint32_t mask = VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP |
2860 VFIO_DMA_UNMAP_FLAG_VADDR |
2861 VFIO_DMA_UNMAP_FLAG_ALL;
2862 unsigned long minsz;
2865 minsz = offsetofend(struct vfio_iommu_type1_dma_unmap, size);
2867 if (copy_from_user(&unmap, (void __user *)arg, minsz))
2870 if (unmap.argsz < minsz || unmap.flags & ~mask)
2873 if ((unmap.flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) &&
2874 (unmap.flags & (VFIO_DMA_UNMAP_FLAG_ALL |
2875 VFIO_DMA_UNMAP_FLAG_VADDR)))
2878 if (unmap.flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) {
2879 unsigned long pgshift;
2881 if (unmap.argsz < (minsz + sizeof(bitmap)))
2884 if (copy_from_user(&bitmap,
2885 (void __user *)(arg + minsz),
2889 if (!access_ok((void __user *)bitmap.data, bitmap.size))
2892 pgshift = __ffs(bitmap.pgsize);
2893 ret = verify_bitmap_size(unmap.size >> pgshift,
2899 ret = vfio_dma_do_unmap(iommu, &unmap, &bitmap);
2903 return copy_to_user((void __user *)arg, &unmap, minsz) ?
2907 static int vfio_iommu_type1_dirty_pages(struct vfio_iommu *iommu,
2910 struct vfio_iommu_type1_dirty_bitmap dirty;
2911 uint32_t mask = VFIO_IOMMU_DIRTY_PAGES_FLAG_START |
2912 VFIO_IOMMU_DIRTY_PAGES_FLAG_STOP |
2913 VFIO_IOMMU_DIRTY_PAGES_FLAG_GET_BITMAP;
2914 unsigned long minsz;
2920 minsz = offsetofend(struct vfio_iommu_type1_dirty_bitmap, flags);
2922 if (copy_from_user(&dirty, (void __user *)arg, minsz))
2925 if (dirty.argsz < minsz || dirty.flags & ~mask)
2928 /* only one flag should be set at a time */
2929 if (__ffs(dirty.flags) != __fls(dirty.flags))
2932 if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_START) {
2935 mutex_lock(&iommu->lock);
2936 pgsize = 1 << __ffs(iommu->pgsize_bitmap);
2937 if (!iommu->dirty_page_tracking) {
2938 ret = vfio_dma_bitmap_alloc_all(iommu, pgsize);
2940 iommu->dirty_page_tracking = true;
2942 mutex_unlock(&iommu->lock);
2944 } else if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_STOP) {
2945 mutex_lock(&iommu->lock);
2946 if (iommu->dirty_page_tracking) {
2947 iommu->dirty_page_tracking = false;
2948 vfio_dma_bitmap_free_all(iommu);
2950 mutex_unlock(&iommu->lock);
2952 } else if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_GET_BITMAP) {
2953 struct vfio_iommu_type1_dirty_bitmap_get range;
2954 unsigned long pgshift;
2955 size_t data_size = dirty.argsz - minsz;
2956 size_t iommu_pgsize;
2958 if (!data_size || data_size < sizeof(range))
2961 if (copy_from_user(&range, (void __user *)(arg + minsz),
2965 if (range.iova + range.size < range.iova)
2967 if (!access_ok((void __user *)range.bitmap.data,
2971 pgshift = __ffs(range.bitmap.pgsize);
2972 ret = verify_bitmap_size(range.size >> pgshift,
2977 mutex_lock(&iommu->lock);
2979 iommu_pgsize = (size_t)1 << __ffs(iommu->pgsize_bitmap);
2981 /* allow only smallest supported pgsize */
2982 if (range.bitmap.pgsize != iommu_pgsize) {
2986 if (range.iova & (iommu_pgsize - 1)) {
2990 if (!range.size || range.size & (iommu_pgsize - 1)) {
2995 if (iommu->dirty_page_tracking)
2996 ret = vfio_iova_dirty_bitmap(range.bitmap.data,
2999 range.bitmap.pgsize);
3003 mutex_unlock(&iommu->lock);
3011 static long vfio_iommu_type1_ioctl(void *iommu_data,
3012 unsigned int cmd, unsigned long arg)
3014 struct vfio_iommu *iommu = iommu_data;
3017 case VFIO_CHECK_EXTENSION:
3018 return vfio_iommu_type1_check_extension(iommu, arg);
3019 case VFIO_IOMMU_GET_INFO:
3020 return vfio_iommu_type1_get_info(iommu, arg);
3021 case VFIO_IOMMU_MAP_DMA:
3022 return vfio_iommu_type1_map_dma(iommu, arg);
3023 case VFIO_IOMMU_UNMAP_DMA:
3024 return vfio_iommu_type1_unmap_dma(iommu, arg);
3025 case VFIO_IOMMU_DIRTY_PAGES:
3026 return vfio_iommu_type1_dirty_pages(iommu, arg);
3032 static int vfio_iommu_type1_register_notifier(void *iommu_data,
3033 unsigned long *events,
3034 struct notifier_block *nb)
3036 struct vfio_iommu *iommu = iommu_data;
3038 /* clear known events */
3039 *events &= ~VFIO_IOMMU_NOTIFY_DMA_UNMAP;
3041 /* refuse to register if still events remaining */
3045 return blocking_notifier_chain_register(&iommu->notifier, nb);
3048 static int vfio_iommu_type1_unregister_notifier(void *iommu_data,
3049 struct notifier_block *nb)
3051 struct vfio_iommu *iommu = iommu_data;
3053 return blocking_notifier_chain_unregister(&iommu->notifier, nb);
3056 static int vfio_iommu_type1_dma_rw_chunk(struct vfio_iommu *iommu,
3057 dma_addr_t user_iova, void *data,
3058 size_t count, bool write,
3061 struct mm_struct *mm;
3062 unsigned long vaddr;
3063 struct vfio_dma *dma;
3064 bool kthread = current->mm == NULL;
3070 ret = vfio_find_dma_valid(iommu, user_iova, 1, &dma);
3074 if ((write && !(dma->prot & IOMMU_WRITE)) ||
3075 !(dma->prot & IOMMU_READ))
3078 mm = get_task_mm(dma->task);
3085 else if (current->mm != mm)
3088 offset = user_iova - dma->iova;
3090 if (count > dma->size - offset)
3091 count = dma->size - offset;
3093 vaddr = dma->vaddr + offset;
3096 *copied = copy_to_user((void __user *)vaddr, data,
3098 if (*copied && iommu->dirty_page_tracking) {
3099 unsigned long pgshift = __ffs(iommu->pgsize_bitmap);
3101 * Bitmap populated with the smallest supported page
3104 bitmap_set(dma->bitmap, offset >> pgshift,
3105 ((offset + *copied - 1) >> pgshift) -
3106 (offset >> pgshift) + 1);
3109 *copied = copy_from_user(data, (void __user *)vaddr,
3112 kthread_unuse_mm(mm);
3115 return *copied ? 0 : -EFAULT;
3118 static int vfio_iommu_type1_dma_rw(void *iommu_data, dma_addr_t user_iova,
3119 void *data, size_t count, bool write)
3121 struct vfio_iommu *iommu = iommu_data;
3125 mutex_lock(&iommu->lock);
3127 ret = vfio_iommu_type1_dma_rw_chunk(iommu, user_iova, data,
3128 count, write, &done);
3137 mutex_unlock(&iommu->lock);
3141 static struct iommu_domain *
3142 vfio_iommu_type1_group_iommu_domain(void *iommu_data,
3143 struct iommu_group *iommu_group)
3145 struct iommu_domain *domain = ERR_PTR(-ENODEV);
3146 struct vfio_iommu *iommu = iommu_data;
3147 struct vfio_domain *d;
3149 if (!iommu || !iommu_group)
3150 return ERR_PTR(-EINVAL);
3152 mutex_lock(&iommu->lock);
3153 list_for_each_entry(d, &iommu->domain_list, next) {
3154 if (find_iommu_group(d, iommu_group)) {
3159 mutex_unlock(&iommu->lock);
3164 static void vfio_iommu_type1_notify(void *iommu_data,
3165 enum vfio_iommu_notify_type event)
3167 struct vfio_iommu *iommu = iommu_data;
3169 if (event != VFIO_IOMMU_CONTAINER_CLOSE)
3171 mutex_lock(&iommu->lock);
3172 iommu->container_open = false;
3173 mutex_unlock(&iommu->lock);
3174 wake_up_all(&iommu->vaddr_wait);
3177 static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = {
3178 .name = "vfio-iommu-type1",
3179 .owner = THIS_MODULE,
3180 .open = vfio_iommu_type1_open,
3181 .release = vfio_iommu_type1_release,
3182 .ioctl = vfio_iommu_type1_ioctl,
3183 .attach_group = vfio_iommu_type1_attach_group,
3184 .detach_group = vfio_iommu_type1_detach_group,
3185 .pin_pages = vfio_iommu_type1_pin_pages,
3186 .unpin_pages = vfio_iommu_type1_unpin_pages,
3187 .register_notifier = vfio_iommu_type1_register_notifier,
3188 .unregister_notifier = vfio_iommu_type1_unregister_notifier,
3189 .dma_rw = vfio_iommu_type1_dma_rw,
3190 .group_iommu_domain = vfio_iommu_type1_group_iommu_domain,
3191 .notify = vfio_iommu_type1_notify,
3194 static int __init vfio_iommu_type1_init(void)
3196 return vfio_register_iommu_driver(&vfio_iommu_driver_ops_type1);
3199 static void __exit vfio_iommu_type1_cleanup(void)
3201 vfio_unregister_iommu_driver(&vfio_iommu_driver_ops_type1);
3204 module_init(vfio_iommu_type1_init);
3205 module_exit(vfio_iommu_type1_cleanup);
3207 MODULE_VERSION(DRIVER_VERSION);
3208 MODULE_LICENSE("GPL v2");
3209 MODULE_AUTHOR(DRIVER_AUTHOR);
3210 MODULE_DESCRIPTION(DRIVER_DESC);