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);
498 static int vaddr_get_pfn(struct mm_struct *mm, unsigned long vaddr,
499 int prot, unsigned long *pfn)
501 struct page *page[1];
502 struct vm_area_struct *vma;
503 unsigned int flags = 0;
506 if (prot & IOMMU_WRITE)
510 ret = pin_user_pages_remote(mm, vaddr, 1, flags | FOLL_LONGTERM,
513 *pfn = page_to_pfn(page[0]);
518 vaddr = untagged_addr(vaddr);
521 vma = find_vma_intersection(mm, vaddr, vaddr + 1);
523 if (vma && vma->vm_flags & VM_PFNMAP) {
524 ret = follow_fault_pfn(vma, mm, vaddr, pfn, prot & IOMMU_WRITE);
528 if (!ret && !is_invalid_reserved_pfn(*pfn))
532 mmap_read_unlock(mm);
536 static int vfio_wait(struct vfio_iommu *iommu)
540 prepare_to_wait(&iommu->vaddr_wait, &wait, TASK_KILLABLE);
541 mutex_unlock(&iommu->lock);
543 mutex_lock(&iommu->lock);
544 finish_wait(&iommu->vaddr_wait, &wait);
545 if (kthread_should_stop() || !iommu->container_open ||
546 fatal_signal_pending(current)) {
553 * Find dma struct and wait for its vaddr to be valid. iommu lock is dropped
554 * if the task waits, but is re-locked on return. Return result in *dma_p.
555 * Return 0 on success with no waiting, WAITED on success if waited, and -errno
558 static int vfio_find_dma_valid(struct vfio_iommu *iommu, dma_addr_t start,
559 size_t size, struct vfio_dma **dma_p)
564 *dma_p = vfio_find_dma(iommu, start, size);
567 else if (!(*dma_p)->vaddr_invalid)
570 ret = vfio_wait(iommu);
577 * Wait for all vaddr in the dma_list to become valid. iommu lock is dropped
578 * if the task waits, but is re-locked on return. Return 0 on success with no
579 * waiting, WAITED on success if waited, and -errno on error.
581 static int vfio_wait_all_valid(struct vfio_iommu *iommu)
585 while (iommu->vaddr_invalid_count && ret >= 0)
586 ret = vfio_wait(iommu);
592 * Attempt to pin pages. We really don't want to track all the pfns and
593 * the iommu can only map chunks of consecutive pfns anyway, so get the
594 * first page and all consecutive pages with the same locking.
596 static long vfio_pin_pages_remote(struct vfio_dma *dma, unsigned long vaddr,
597 long npage, unsigned long *pfn_base,
600 unsigned long pfn = 0;
601 long ret, pinned = 0, lock_acct = 0;
603 dma_addr_t iova = vaddr - dma->vaddr + dma->iova;
605 /* This code path is only user initiated */
609 ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, pfn_base);
614 rsvd = is_invalid_reserved_pfn(*pfn_base);
617 * Reserved pages aren't counted against the user, externally pinned
618 * pages are already counted against the user.
620 if (!rsvd && !vfio_find_vpfn(dma, iova)) {
621 if (!dma->lock_cap && current->mm->locked_vm + 1 > limit) {
622 put_pfn(*pfn_base, dma->prot);
623 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n", __func__,
624 limit << PAGE_SHIFT);
630 if (unlikely(disable_hugepages))
633 /* Lock all the consecutive pages from pfn_base */
634 for (vaddr += PAGE_SIZE, iova += PAGE_SIZE; pinned < npage;
635 pinned++, vaddr += PAGE_SIZE, iova += PAGE_SIZE) {
636 ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, &pfn);
640 if (pfn != *pfn_base + pinned ||
641 rsvd != is_invalid_reserved_pfn(pfn)) {
642 put_pfn(pfn, dma->prot);
646 if (!rsvd && !vfio_find_vpfn(dma, iova)) {
647 if (!dma->lock_cap &&
648 current->mm->locked_vm + lock_acct + 1 > limit) {
649 put_pfn(pfn, dma->prot);
650 pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n",
651 __func__, limit << PAGE_SHIFT);
660 ret = vfio_lock_acct(dma, lock_acct, false);
665 for (pfn = *pfn_base ; pinned ; pfn++, pinned--)
666 put_pfn(pfn, dma->prot);
675 static long vfio_unpin_pages_remote(struct vfio_dma *dma, dma_addr_t iova,
676 unsigned long pfn, long npage,
679 long unlocked = 0, locked = 0;
682 for (i = 0; i < npage; i++, iova += PAGE_SIZE) {
683 if (put_pfn(pfn++, dma->prot)) {
685 if (vfio_find_vpfn(dma, iova))
691 vfio_lock_acct(dma, locked - unlocked, true);
696 static int vfio_pin_page_external(struct vfio_dma *dma, unsigned long vaddr,
697 unsigned long *pfn_base, bool do_accounting)
699 struct mm_struct *mm;
702 mm = get_task_mm(dma->task);
706 ret = vaddr_get_pfn(mm, vaddr, dma->prot, pfn_base);
707 if (!ret && do_accounting && !is_invalid_reserved_pfn(*pfn_base)) {
708 ret = vfio_lock_acct(dma, 1, true);
710 put_pfn(*pfn_base, dma->prot);
712 pr_warn("%s: Task %s (%d) RLIMIT_MEMLOCK "
713 "(%ld) exceeded\n", __func__,
714 dma->task->comm, task_pid_nr(dma->task),
715 task_rlimit(dma->task, RLIMIT_MEMLOCK));
723 static int vfio_unpin_page_external(struct vfio_dma *dma, dma_addr_t iova,
727 struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
732 unlocked = vfio_iova_put_vfio_pfn(dma, vpfn);
735 vfio_lock_acct(dma, -unlocked, true);
740 static int vfio_iommu_type1_pin_pages(void *iommu_data,
741 struct iommu_group *iommu_group,
742 unsigned long *user_pfn,
744 unsigned long *phys_pfn)
746 struct vfio_iommu *iommu = iommu_data;
747 struct vfio_group *group;
749 unsigned long remote_vaddr;
750 struct vfio_dma *dma;
754 if (!iommu || !user_pfn || !phys_pfn)
757 /* Supported for v2 version only */
761 mutex_lock(&iommu->lock);
764 * Wait for all necessary vaddr's to be valid so they can be used in
765 * the main loop without dropping the lock, to avoid racing vs unmap.
768 if (iommu->vaddr_invalid_count) {
769 for (i = 0; i < npage; i++) {
770 iova = user_pfn[i] << PAGE_SHIFT;
771 ret = vfio_find_dma_valid(iommu, iova, PAGE_SIZE, &dma);
779 /* Fail if notifier list is empty */
780 if (!iommu->notifier.head) {
786 * If iommu capable domain exist in the container then all pages are
787 * already pinned and accounted. Accouting should be done if there is no
788 * iommu capable domain in the container.
790 do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu);
792 for (i = 0; i < npage; i++) {
793 struct vfio_pfn *vpfn;
795 iova = user_pfn[i] << PAGE_SHIFT;
796 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
802 if ((dma->prot & prot) != prot) {
807 vpfn = vfio_iova_get_vfio_pfn(dma, iova);
809 phys_pfn[i] = vpfn->pfn;
813 remote_vaddr = dma->vaddr + (iova - dma->iova);
814 ret = vfio_pin_page_external(dma, remote_vaddr, &phys_pfn[i],
819 ret = vfio_add_to_pfn_list(dma, iova, phys_pfn[i]);
821 if (put_pfn(phys_pfn[i], dma->prot) && do_accounting)
822 vfio_lock_acct(dma, -1, true);
826 if (iommu->dirty_page_tracking) {
827 unsigned long pgshift = __ffs(iommu->pgsize_bitmap);
830 * Bitmap populated with the smallest supported page
833 bitmap_set(dma->bitmap,
834 (iova - dma->iova) >> pgshift, 1);
839 group = vfio_iommu_find_iommu_group(iommu, iommu_group);
840 if (!group->pinned_page_dirty_scope) {
841 group->pinned_page_dirty_scope = true;
842 iommu->num_non_pinned_groups--;
849 for (j = 0; j < i; j++) {
852 iova = user_pfn[j] << PAGE_SHIFT;
853 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
854 vfio_unpin_page_external(dma, iova, do_accounting);
858 mutex_unlock(&iommu->lock);
862 static int vfio_iommu_type1_unpin_pages(void *iommu_data,
863 unsigned long *user_pfn,
866 struct vfio_iommu *iommu = iommu_data;
870 if (!iommu || !user_pfn)
873 /* Supported for v2 version only */
877 mutex_lock(&iommu->lock);
879 do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu);
880 for (i = 0; i < npage; i++) {
881 struct vfio_dma *dma;
884 iova = user_pfn[i] << PAGE_SHIFT;
885 dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
888 vfio_unpin_page_external(dma, iova, do_accounting);
892 mutex_unlock(&iommu->lock);
893 return i > npage ? npage : (i > 0 ? i : -EINVAL);
896 static long vfio_sync_unpin(struct vfio_dma *dma, struct vfio_domain *domain,
897 struct list_head *regions,
898 struct iommu_iotlb_gather *iotlb_gather)
901 struct vfio_regions *entry, *next;
903 iommu_iotlb_sync(domain->domain, iotlb_gather);
905 list_for_each_entry_safe(entry, next, regions, list) {
906 unlocked += vfio_unpin_pages_remote(dma,
908 entry->phys >> PAGE_SHIFT,
909 entry->len >> PAGE_SHIFT,
911 list_del(&entry->list);
921 * Generally, VFIO needs to unpin remote pages after each IOTLB flush.
922 * Therefore, when using IOTLB flush sync interface, VFIO need to keep track
923 * of these regions (currently using a list).
925 * This value specifies maximum number of regions for each IOTLB flush sync.
927 #define VFIO_IOMMU_TLB_SYNC_MAX 512
929 static size_t unmap_unpin_fast(struct vfio_domain *domain,
930 struct vfio_dma *dma, dma_addr_t *iova,
931 size_t len, phys_addr_t phys, long *unlocked,
932 struct list_head *unmapped_list,
934 struct iommu_iotlb_gather *iotlb_gather)
937 struct vfio_regions *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
940 unmapped = iommu_unmap_fast(domain->domain, *iova, len,
948 entry->len = unmapped;
949 list_add_tail(&entry->list, unmapped_list);
957 * Sync if the number of fast-unmap regions hits the limit
958 * or in case of errors.
960 if (*unmapped_cnt >= VFIO_IOMMU_TLB_SYNC_MAX || !unmapped) {
961 *unlocked += vfio_sync_unpin(dma, domain, unmapped_list,
969 static size_t unmap_unpin_slow(struct vfio_domain *domain,
970 struct vfio_dma *dma, dma_addr_t *iova,
971 size_t len, phys_addr_t phys,
974 size_t unmapped = iommu_unmap(domain->domain, *iova, len);
977 *unlocked += vfio_unpin_pages_remote(dma, *iova,
979 unmapped >> PAGE_SHIFT,
987 static long vfio_unmap_unpin(struct vfio_iommu *iommu, struct vfio_dma *dma,
990 dma_addr_t iova = dma->iova, end = dma->iova + dma->size;
991 struct vfio_domain *domain, *d;
992 LIST_HEAD(unmapped_region_list);
993 struct iommu_iotlb_gather iotlb_gather;
994 int unmapped_region_cnt = 0;
1000 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
1004 * We use the IOMMU to track the physical addresses, otherwise we'd
1005 * need a much more complicated tracking system. Unfortunately that
1006 * means we need to use one of the iommu domains to figure out the
1007 * pfns to unpin. The rest need to be unmapped in advance so we have
1008 * no iommu translations remaining when the pages are unpinned.
1010 domain = d = list_first_entry(&iommu->domain_list,
1011 struct vfio_domain, next);
1013 list_for_each_entry_continue(d, &iommu->domain_list, next) {
1014 iommu_unmap(d->domain, dma->iova, dma->size);
1018 iommu_iotlb_gather_init(&iotlb_gather);
1019 while (iova < end) {
1020 size_t unmapped, len;
1021 phys_addr_t phys, next;
1023 phys = iommu_iova_to_phys(domain->domain, iova);
1024 if (WARN_ON(!phys)) {
1030 * To optimize for fewer iommu_unmap() calls, each of which
1031 * may require hardware cache flushing, try to find the
1032 * largest contiguous physical memory chunk to unmap.
1034 for (len = PAGE_SIZE;
1035 !domain->fgsp && iova + len < end; len += PAGE_SIZE) {
1036 next = iommu_iova_to_phys(domain->domain, iova + len);
1037 if (next != phys + len)
1042 * First, try to use fast unmap/unpin. In case of failure,
1043 * switch to slow unmap/unpin path.
1045 unmapped = unmap_unpin_fast(domain, dma, &iova, len, phys,
1046 &unlocked, &unmapped_region_list,
1047 &unmapped_region_cnt,
1050 unmapped = unmap_unpin_slow(domain, dma, &iova, len,
1052 if (WARN_ON(!unmapped))
1057 dma->iommu_mapped = false;
1059 if (unmapped_region_cnt) {
1060 unlocked += vfio_sync_unpin(dma, domain, &unmapped_region_list,
1064 if (do_accounting) {
1065 vfio_lock_acct(dma, -unlocked, true);
1071 static void vfio_remove_dma(struct vfio_iommu *iommu, struct vfio_dma *dma)
1073 WARN_ON(!RB_EMPTY_ROOT(&dma->pfn_list));
1074 vfio_unmap_unpin(iommu, dma, true);
1075 vfio_unlink_dma(iommu, dma);
1076 put_task_struct(dma->task);
1077 vfio_dma_bitmap_free(dma);
1078 if (dma->vaddr_invalid) {
1079 iommu->vaddr_invalid_count--;
1080 wake_up_all(&iommu->vaddr_wait);
1086 static void vfio_update_pgsize_bitmap(struct vfio_iommu *iommu)
1088 struct vfio_domain *domain;
1090 iommu->pgsize_bitmap = ULONG_MAX;
1092 list_for_each_entry(domain, &iommu->domain_list, next)
1093 iommu->pgsize_bitmap &= domain->domain->pgsize_bitmap;
1096 * In case the IOMMU supports page sizes smaller than PAGE_SIZE
1097 * we pretend PAGE_SIZE is supported and hide sub-PAGE_SIZE sizes.
1098 * That way the user will be able to map/unmap buffers whose size/
1099 * start address is aligned with PAGE_SIZE. Pinning code uses that
1100 * granularity while iommu driver can use the sub-PAGE_SIZE size
1101 * to map the buffer.
1103 if (iommu->pgsize_bitmap & ~PAGE_MASK) {
1104 iommu->pgsize_bitmap &= PAGE_MASK;
1105 iommu->pgsize_bitmap |= PAGE_SIZE;
1109 static int update_user_bitmap(u64 __user *bitmap, struct vfio_iommu *iommu,
1110 struct vfio_dma *dma, dma_addr_t base_iova,
1113 unsigned long pgshift = __ffs(pgsize);
1114 unsigned long nbits = dma->size >> pgshift;
1115 unsigned long bit_offset = (dma->iova - base_iova) >> pgshift;
1116 unsigned long copy_offset = bit_offset / BITS_PER_LONG;
1117 unsigned long shift = bit_offset % BITS_PER_LONG;
1118 unsigned long leftover;
1121 * mark all pages dirty if any IOMMU capable device is not able
1122 * to report dirty pages and all pages are pinned and mapped.
1124 if (iommu->num_non_pinned_groups && dma->iommu_mapped)
1125 bitmap_set(dma->bitmap, 0, nbits);
1128 bitmap_shift_left(dma->bitmap, dma->bitmap, shift,
1131 if (copy_from_user(&leftover,
1132 (void __user *)(bitmap + copy_offset),
1136 bitmap_or(dma->bitmap, dma->bitmap, &leftover, shift);
1139 if (copy_to_user((void __user *)(bitmap + copy_offset), dma->bitmap,
1140 DIRTY_BITMAP_BYTES(nbits + shift)))
1146 static int vfio_iova_dirty_bitmap(u64 __user *bitmap, struct vfio_iommu *iommu,
1147 dma_addr_t iova, size_t size, size_t pgsize)
1149 struct vfio_dma *dma;
1151 unsigned long pgshift = __ffs(pgsize);
1155 * GET_BITMAP request must fully cover vfio_dma mappings. Multiple
1156 * vfio_dma mappings may be clubbed by specifying large ranges, but
1157 * there must not be any previous mappings bisected by the range.
1158 * An error will be returned if these conditions are not met.
1160 dma = vfio_find_dma(iommu, iova, 1);
1161 if (dma && dma->iova != iova)
1164 dma = vfio_find_dma(iommu, iova + size - 1, 0);
1165 if (dma && dma->iova + dma->size != iova + size)
1168 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
1169 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1171 if (dma->iova < iova)
1174 if (dma->iova > iova + size - 1)
1177 ret = update_user_bitmap(bitmap, iommu, dma, iova, pgsize);
1182 * Re-populate bitmap to include all pinned pages which are
1183 * considered as dirty but exclude pages which are unpinned and
1184 * pages which are marked dirty by vfio_dma_rw()
1186 bitmap_clear(dma->bitmap, 0, dma->size >> pgshift);
1187 vfio_dma_populate_bitmap(dma, pgsize);
1192 static int verify_bitmap_size(uint64_t npages, uint64_t bitmap_size)
1194 if (!npages || !bitmap_size || (bitmap_size > DIRTY_BITMAP_SIZE_MAX) ||
1195 (bitmap_size < DIRTY_BITMAP_BYTES(npages)))
1201 static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
1202 struct vfio_iommu_type1_dma_unmap *unmap,
1203 struct vfio_bitmap *bitmap)
1205 struct vfio_dma *dma, *dma_last = NULL;
1206 size_t unmapped = 0, pgsize;
1207 int ret = -EINVAL, retries = 0;
1208 unsigned long pgshift;
1209 dma_addr_t iova = unmap->iova;
1210 unsigned long size = unmap->size;
1211 bool unmap_all = unmap->flags & VFIO_DMA_UNMAP_FLAG_ALL;
1212 bool invalidate_vaddr = unmap->flags & VFIO_DMA_UNMAP_FLAG_VADDR;
1213 struct rb_node *n, *first_n;
1215 mutex_lock(&iommu->lock);
1217 pgshift = __ffs(iommu->pgsize_bitmap);
1218 pgsize = (size_t)1 << pgshift;
1220 if (iova & (pgsize - 1))
1227 } else if (!size || size & (pgsize - 1)) {
1231 if (iova + size - 1 < iova || size > SIZE_MAX)
1234 /* When dirty tracking is enabled, allow only min supported pgsize */
1235 if ((unmap->flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) &&
1236 (!iommu->dirty_page_tracking || (bitmap->pgsize != pgsize))) {
1240 WARN_ON((pgsize - 1) & PAGE_MASK);
1243 * vfio-iommu-type1 (v1) - User mappings were coalesced together to
1244 * avoid tracking individual mappings. This means that the granularity
1245 * of the original mapping was lost and the user was allowed to attempt
1246 * to unmap any range. Depending on the contiguousness of physical
1247 * memory and page sizes supported by the IOMMU, arbitrary unmaps may
1248 * or may not have worked. We only guaranteed unmap granularity
1249 * matching the original mapping; even though it was untracked here,
1250 * the original mappings are reflected in IOMMU mappings. This
1251 * resulted in a couple unusual behaviors. First, if a range is not
1252 * able to be unmapped, ex. a set of 4k pages that was mapped as a
1253 * 2M hugepage into the IOMMU, the unmap ioctl returns success but with
1254 * a zero sized unmap. Also, if an unmap request overlaps the first
1255 * address of a hugepage, the IOMMU will unmap the entire hugepage.
1256 * This also returns success and the returned unmap size reflects the
1257 * actual size unmapped.
1259 * We attempt to maintain compatibility with this "v1" interface, but
1260 * we take control out of the hands of the IOMMU. Therefore, an unmap
1261 * request offset from the beginning of the original mapping will
1262 * return success with zero sized unmap. And an unmap request covering
1263 * the first iova of mapping will unmap the entire range.
1265 * The v2 version of this interface intends to be more deterministic.
1266 * Unmap requests must fully cover previous mappings. Multiple
1267 * mappings may still be unmaped by specifying large ranges, but there
1268 * must not be any previous mappings bisected by the range. An error
1269 * will be returned if these conditions are not met. The v2 interface
1270 * will only return success and a size of zero if there were no
1271 * mappings within the range.
1273 if (iommu->v2 && !unmap_all) {
1274 dma = vfio_find_dma(iommu, iova, 1);
1275 if (dma && dma->iova != iova)
1278 dma = vfio_find_dma(iommu, iova + size - 1, 0);
1279 if (dma && dma->iova + dma->size != iova + size)
1284 n = first_n = vfio_find_dma_first_node(iommu, iova, size);
1287 dma = rb_entry(n, struct vfio_dma, node);
1288 if (dma->iova >= iova + size)
1291 if (!iommu->v2 && iova > dma->iova)
1294 * Task with same address space who mapped this iova range is
1295 * allowed to unmap the iova range.
1297 if (dma->task->mm != current->mm)
1300 if (invalidate_vaddr) {
1301 if (dma->vaddr_invalid) {
1302 struct rb_node *last_n = n;
1304 for (n = first_n; n != last_n; n = rb_next(n)) {
1306 struct vfio_dma, node);
1307 dma->vaddr_invalid = false;
1308 iommu->vaddr_invalid_count--;
1314 dma->vaddr_invalid = true;
1315 iommu->vaddr_invalid_count++;
1316 unmapped += dma->size;
1321 if (!RB_EMPTY_ROOT(&dma->pfn_list)) {
1322 struct vfio_iommu_type1_dma_unmap nb_unmap;
1324 if (dma_last == dma) {
1325 BUG_ON(++retries > 10);
1331 nb_unmap.iova = dma->iova;
1332 nb_unmap.size = dma->size;
1335 * Notify anyone (mdev vendor drivers) to invalidate and
1336 * unmap iovas within the range we're about to unmap.
1337 * Vendor drivers MUST unpin pages in response to an
1340 mutex_unlock(&iommu->lock);
1341 blocking_notifier_call_chain(&iommu->notifier,
1342 VFIO_IOMMU_NOTIFY_DMA_UNMAP,
1344 mutex_lock(&iommu->lock);
1348 if (unmap->flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) {
1349 ret = update_user_bitmap(bitmap->data, iommu, dma,
1355 unmapped += dma->size;
1357 vfio_remove_dma(iommu, dma);
1361 mutex_unlock(&iommu->lock);
1363 /* Report how much was unmapped */
1364 unmap->size = unmapped;
1369 static int vfio_iommu_map(struct vfio_iommu *iommu, dma_addr_t iova,
1370 unsigned long pfn, long npage, int prot)
1372 struct vfio_domain *d;
1375 list_for_each_entry(d, &iommu->domain_list, next) {
1376 ret = iommu_map(d->domain, iova, (phys_addr_t)pfn << PAGE_SHIFT,
1377 npage << PAGE_SHIFT, prot | d->prot);
1387 list_for_each_entry_continue_reverse(d, &iommu->domain_list, next) {
1388 iommu_unmap(d->domain, iova, npage << PAGE_SHIFT);
1395 static int vfio_pin_map_dma(struct vfio_iommu *iommu, struct vfio_dma *dma,
1398 dma_addr_t iova = dma->iova;
1399 unsigned long vaddr = dma->vaddr;
1400 size_t size = map_size;
1402 unsigned long pfn, limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1406 /* Pin a contiguous chunk of memory */
1407 npage = vfio_pin_pages_remote(dma, vaddr + dma->size,
1408 size >> PAGE_SHIFT, &pfn, limit);
1416 ret = vfio_iommu_map(iommu, iova + dma->size, pfn, npage,
1419 vfio_unpin_pages_remote(dma, iova + dma->size, pfn,
1424 size -= npage << PAGE_SHIFT;
1425 dma->size += npage << PAGE_SHIFT;
1428 dma->iommu_mapped = true;
1431 vfio_remove_dma(iommu, dma);
1437 * Check dma map request is within a valid iova range
1439 static bool vfio_iommu_iova_dma_valid(struct vfio_iommu *iommu,
1440 dma_addr_t start, dma_addr_t end)
1442 struct list_head *iova = &iommu->iova_list;
1443 struct vfio_iova *node;
1445 list_for_each_entry(node, iova, list) {
1446 if (start >= node->start && end <= node->end)
1451 * Check for list_empty() as well since a container with
1452 * a single mdev device will have an empty list.
1454 return list_empty(iova);
1457 static int vfio_dma_do_map(struct vfio_iommu *iommu,
1458 struct vfio_iommu_type1_dma_map *map)
1460 bool set_vaddr = map->flags & VFIO_DMA_MAP_FLAG_VADDR;
1461 dma_addr_t iova = map->iova;
1462 unsigned long vaddr = map->vaddr;
1463 size_t size = map->size;
1464 int ret = 0, prot = 0;
1466 struct vfio_dma *dma;
1468 /* Verify that none of our __u64 fields overflow */
1469 if (map->size != size || map->vaddr != vaddr || map->iova != iova)
1472 /* READ/WRITE from device perspective */
1473 if (map->flags & VFIO_DMA_MAP_FLAG_WRITE)
1474 prot |= IOMMU_WRITE;
1475 if (map->flags & VFIO_DMA_MAP_FLAG_READ)
1478 if ((prot && set_vaddr) || (!prot && !set_vaddr))
1481 mutex_lock(&iommu->lock);
1483 pgsize = (size_t)1 << __ffs(iommu->pgsize_bitmap);
1485 WARN_ON((pgsize - 1) & PAGE_MASK);
1487 if (!size || (size | iova | vaddr) & (pgsize - 1)) {
1492 /* Don't allow IOVA or virtual address wrap */
1493 if (iova + size - 1 < iova || vaddr + size - 1 < vaddr) {
1498 dma = vfio_find_dma(iommu, iova, size);
1502 } else if (!dma->vaddr_invalid || dma->iova != iova ||
1503 dma->size != size) {
1507 dma->vaddr_invalid = false;
1508 iommu->vaddr_invalid_count--;
1509 wake_up_all(&iommu->vaddr_wait);
1517 if (!iommu->dma_avail) {
1522 if (!vfio_iommu_iova_dma_valid(iommu, iova, iova + size - 1)) {
1527 dma = kzalloc(sizeof(*dma), GFP_KERNEL);
1539 * We need to be able to both add to a task's locked memory and test
1540 * against the locked memory limit and we need to be able to do both
1541 * outside of this call path as pinning can be asynchronous via the
1542 * external interfaces for mdev devices. RLIMIT_MEMLOCK requires a
1543 * task_struct and VM locked pages requires an mm_struct, however
1544 * holding an indefinite mm reference is not recommended, therefore we
1545 * only hold a reference to a task. We could hold a reference to
1546 * current, however QEMU uses this call path through vCPU threads,
1547 * which can be killed resulting in a NULL mm and failure in the unmap
1548 * path when called via a different thread. Avoid this problem by
1549 * using the group_leader as threads within the same group require
1550 * both CLONE_THREAD and CLONE_VM and will therefore use the same
1553 * Previously we also used the task for testing CAP_IPC_LOCK at the
1554 * time of pinning and accounting, however has_capability() makes use
1555 * of real_cred, a copy-on-write field, so we can't guarantee that it
1556 * matches group_leader, or in fact that it might not change by the
1557 * time it's evaluated. If a process were to call MAP_DMA with
1558 * CAP_IPC_LOCK but later drop it, it doesn't make sense that they
1559 * possibly see different results for an iommu_mapped vfio_dma vs
1560 * externally mapped. Therefore track CAP_IPC_LOCK in vfio_dma at the
1561 * time of calling MAP_DMA.
1563 get_task_struct(current->group_leader);
1564 dma->task = current->group_leader;
1565 dma->lock_cap = capable(CAP_IPC_LOCK);
1567 dma->pfn_list = RB_ROOT;
1569 /* Insert zero-sized and grow as we map chunks of it */
1570 vfio_link_dma(iommu, dma);
1572 /* Don't pin and map if container doesn't contain IOMMU capable domain*/
1573 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu))
1576 ret = vfio_pin_map_dma(iommu, dma, size);
1578 if (!ret && iommu->dirty_page_tracking) {
1579 ret = vfio_dma_bitmap_alloc(dma, pgsize);
1581 vfio_remove_dma(iommu, dma);
1585 mutex_unlock(&iommu->lock);
1589 static int vfio_bus_type(struct device *dev, void *data)
1591 struct bus_type **bus = data;
1593 if (*bus && *bus != dev->bus)
1601 static int vfio_iommu_replay(struct vfio_iommu *iommu,
1602 struct vfio_domain *domain)
1604 struct vfio_domain *d = NULL;
1606 unsigned long limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1609 ret = vfio_wait_all_valid(iommu);
1613 /* Arbitrarily pick the first domain in the list for lookups */
1614 if (!list_empty(&iommu->domain_list))
1615 d = list_first_entry(&iommu->domain_list,
1616 struct vfio_domain, next);
1618 n = rb_first(&iommu->dma_list);
1620 for (; n; n = rb_next(n)) {
1621 struct vfio_dma *dma;
1624 dma = rb_entry(n, struct vfio_dma, node);
1627 while (iova < dma->iova + dma->size) {
1631 if (dma->iommu_mapped) {
1635 if (WARN_ON(!d)) { /* mapped w/o a domain?! */
1640 phys = iommu_iova_to_phys(d->domain, iova);
1642 if (WARN_ON(!phys)) {
1650 while (i < dma->iova + dma->size &&
1651 p == iommu_iova_to_phys(d->domain, i)) {
1658 unsigned long vaddr = dma->vaddr +
1660 size_t n = dma->iova + dma->size - iova;
1663 npage = vfio_pin_pages_remote(dma, vaddr,
1672 phys = pfn << PAGE_SHIFT;
1673 size = npage << PAGE_SHIFT;
1676 ret = iommu_map(domain->domain, iova, phys,
1677 size, dma->prot | domain->prot);
1679 if (!dma->iommu_mapped)
1680 vfio_unpin_pages_remote(dma, iova,
1691 /* All dmas are now mapped, defer to second tree walk for unwind */
1692 for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
1693 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1695 dma->iommu_mapped = true;
1701 for (; n; n = rb_prev(n)) {
1702 struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1705 if (dma->iommu_mapped) {
1706 iommu_unmap(domain->domain, dma->iova, dma->size);
1711 while (iova < dma->iova + dma->size) {
1712 phys_addr_t phys, p;
1716 phys = iommu_iova_to_phys(domain->domain, iova);
1725 while (i < dma->iova + dma->size &&
1726 p == iommu_iova_to_phys(domain->domain, i)) {
1732 iommu_unmap(domain->domain, iova, size);
1733 vfio_unpin_pages_remote(dma, iova, phys >> PAGE_SHIFT,
1734 size >> PAGE_SHIFT, true);
1742 * We change our unmap behavior slightly depending on whether the IOMMU
1743 * supports fine-grained superpages. IOMMUs like AMD-Vi will use a superpage
1744 * for practically any contiguous power-of-two mapping we give it. This means
1745 * we don't need to look for contiguous chunks ourselves to make unmapping
1746 * more efficient. On IOMMUs with coarse-grained super pages, like Intel VT-d
1747 * with discrete 2M/1G/512G/1T superpages, identifying contiguous chunks
1748 * significantly boosts non-hugetlbfs mappings and doesn't seem to hurt when
1749 * hugetlbfs is in use.
1751 static void vfio_test_domain_fgsp(struct vfio_domain *domain)
1754 int ret, order = get_order(PAGE_SIZE * 2);
1756 pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order);
1760 ret = iommu_map(domain->domain, 0, page_to_phys(pages), PAGE_SIZE * 2,
1761 IOMMU_READ | IOMMU_WRITE | domain->prot);
1763 size_t unmapped = iommu_unmap(domain->domain, 0, PAGE_SIZE);
1765 if (unmapped == PAGE_SIZE)
1766 iommu_unmap(domain->domain, PAGE_SIZE, PAGE_SIZE);
1768 domain->fgsp = true;
1771 __free_pages(pages, order);
1774 static struct vfio_group *find_iommu_group(struct vfio_domain *domain,
1775 struct iommu_group *iommu_group)
1777 struct vfio_group *g;
1779 list_for_each_entry(g, &domain->group_list, next) {
1780 if (g->iommu_group == iommu_group)
1787 static struct vfio_group *vfio_iommu_find_iommu_group(struct vfio_iommu *iommu,
1788 struct iommu_group *iommu_group)
1790 struct vfio_domain *domain;
1791 struct vfio_group *group = NULL;
1793 list_for_each_entry(domain, &iommu->domain_list, next) {
1794 group = find_iommu_group(domain, iommu_group);
1799 if (iommu->external_domain)
1800 group = find_iommu_group(iommu->external_domain, iommu_group);
1805 static bool vfio_iommu_has_sw_msi(struct list_head *group_resv_regions,
1808 struct iommu_resv_region *region;
1811 list_for_each_entry(region, group_resv_regions, list) {
1813 * The presence of any 'real' MSI regions should take
1814 * precedence over the software-managed one if the
1815 * IOMMU driver happens to advertise both types.
1817 if (region->type == IOMMU_RESV_MSI) {
1822 if (region->type == IOMMU_RESV_SW_MSI) {
1823 *base = region->start;
1831 static struct device *vfio_mdev_get_iommu_device(struct device *dev)
1833 struct device *(*fn)(struct device *dev);
1834 struct device *iommu_device;
1836 fn = symbol_get(mdev_get_iommu_device);
1838 iommu_device = fn(dev);
1839 symbol_put(mdev_get_iommu_device);
1841 return iommu_device;
1847 static int vfio_mdev_attach_domain(struct device *dev, void *data)
1849 struct iommu_domain *domain = data;
1850 struct device *iommu_device;
1852 iommu_device = vfio_mdev_get_iommu_device(dev);
1854 if (iommu_dev_feature_enabled(iommu_device, IOMMU_DEV_FEAT_AUX))
1855 return iommu_aux_attach_device(domain, iommu_device);
1857 return iommu_attach_device(domain, iommu_device);
1863 static int vfio_mdev_detach_domain(struct device *dev, void *data)
1865 struct iommu_domain *domain = data;
1866 struct device *iommu_device;
1868 iommu_device = vfio_mdev_get_iommu_device(dev);
1870 if (iommu_dev_feature_enabled(iommu_device, IOMMU_DEV_FEAT_AUX))
1871 iommu_aux_detach_device(domain, iommu_device);
1873 iommu_detach_device(domain, iommu_device);
1879 static int vfio_iommu_attach_group(struct vfio_domain *domain,
1880 struct vfio_group *group)
1882 if (group->mdev_group)
1883 return iommu_group_for_each_dev(group->iommu_group,
1885 vfio_mdev_attach_domain);
1887 return iommu_attach_group(domain->domain, group->iommu_group);
1890 static void vfio_iommu_detach_group(struct vfio_domain *domain,
1891 struct vfio_group *group)
1893 if (group->mdev_group)
1894 iommu_group_for_each_dev(group->iommu_group, domain->domain,
1895 vfio_mdev_detach_domain);
1897 iommu_detach_group(domain->domain, group->iommu_group);
1900 static bool vfio_bus_is_mdev(struct bus_type *bus)
1902 struct bus_type *mdev_bus;
1905 mdev_bus = symbol_get(mdev_bus_type);
1907 ret = (bus == mdev_bus);
1908 symbol_put(mdev_bus_type);
1914 static int vfio_mdev_iommu_device(struct device *dev, void *data)
1916 struct device **old = data, *new;
1918 new = vfio_mdev_get_iommu_device(dev);
1919 if (!new || (*old && *old != new))
1928 * This is a helper function to insert an address range to iova list.
1929 * The list is initially created with a single entry corresponding to
1930 * the IOMMU domain geometry to which the device group is attached.
1931 * The list aperture gets modified when a new domain is added to the
1932 * container if the new aperture doesn't conflict with the current one
1933 * or with any existing dma mappings. The list is also modified to
1934 * exclude any reserved regions associated with the device group.
1936 static int vfio_iommu_iova_insert(struct list_head *head,
1937 dma_addr_t start, dma_addr_t end)
1939 struct vfio_iova *region;
1941 region = kmalloc(sizeof(*region), GFP_KERNEL);
1945 INIT_LIST_HEAD(®ion->list);
1946 region->start = start;
1949 list_add_tail(®ion->list, head);
1954 * Check the new iommu aperture conflicts with existing aper or with any
1955 * existing dma mappings.
1957 static bool vfio_iommu_aper_conflict(struct vfio_iommu *iommu,
1958 dma_addr_t start, dma_addr_t end)
1960 struct vfio_iova *first, *last;
1961 struct list_head *iova = &iommu->iova_list;
1963 if (list_empty(iova))
1966 /* Disjoint sets, return conflict */
1967 first = list_first_entry(iova, struct vfio_iova, list);
1968 last = list_last_entry(iova, struct vfio_iova, list);
1969 if (start > last->end || end < first->start)
1972 /* Check for any existing dma mappings below the new start */
1973 if (start > first->start) {
1974 if (vfio_find_dma(iommu, first->start, start - first->start))
1978 /* Check for any existing dma mappings beyond the new end */
1979 if (end < last->end) {
1980 if (vfio_find_dma(iommu, end + 1, last->end - end))
1988 * Resize iommu iova aperture window. This is called only if the new
1989 * aperture has no conflict with existing aperture and dma mappings.
1991 static int vfio_iommu_aper_resize(struct list_head *iova,
1992 dma_addr_t start, dma_addr_t end)
1994 struct vfio_iova *node, *next;
1996 if (list_empty(iova))
1997 return vfio_iommu_iova_insert(iova, start, end);
1999 /* Adjust iova list start */
2000 list_for_each_entry_safe(node, next, iova, list) {
2001 if (start < node->start)
2003 if (start >= node->start && start < node->end) {
2004 node->start = start;
2007 /* Delete nodes before new start */
2008 list_del(&node->list);
2012 /* Adjust iova list end */
2013 list_for_each_entry_safe(node, next, iova, list) {
2014 if (end > node->end)
2016 if (end > node->start && end <= node->end) {
2020 /* Delete nodes after new end */
2021 list_del(&node->list);
2029 * Check reserved region conflicts with existing dma mappings
2031 static bool vfio_iommu_resv_conflict(struct vfio_iommu *iommu,
2032 struct list_head *resv_regions)
2034 struct iommu_resv_region *region;
2036 /* Check for conflict with existing dma mappings */
2037 list_for_each_entry(region, resv_regions, list) {
2038 if (region->type == IOMMU_RESV_DIRECT_RELAXABLE)
2041 if (vfio_find_dma(iommu, region->start, region->length))
2049 * Check iova region overlap with reserved regions and
2050 * exclude them from the iommu iova range
2052 static int vfio_iommu_resv_exclude(struct list_head *iova,
2053 struct list_head *resv_regions)
2055 struct iommu_resv_region *resv;
2056 struct vfio_iova *n, *next;
2058 list_for_each_entry(resv, resv_regions, list) {
2059 phys_addr_t start, end;
2061 if (resv->type == IOMMU_RESV_DIRECT_RELAXABLE)
2064 start = resv->start;
2065 end = resv->start + resv->length - 1;
2067 list_for_each_entry_safe(n, next, iova, list) {
2071 if (start > n->end || end < n->start)
2074 * Insert a new node if current node overlaps with the
2075 * reserve region to exlude that from valid iova range.
2076 * Note that, new node is inserted before the current
2077 * node and finally the current node is deleted keeping
2078 * the list updated and sorted.
2080 if (start > n->start)
2081 ret = vfio_iommu_iova_insert(&n->list, n->start,
2083 if (!ret && end < n->end)
2084 ret = vfio_iommu_iova_insert(&n->list, end + 1,
2094 if (list_empty(iova))
2100 static void vfio_iommu_resv_free(struct list_head *resv_regions)
2102 struct iommu_resv_region *n, *next;
2104 list_for_each_entry_safe(n, next, resv_regions, list) {
2110 static void vfio_iommu_iova_free(struct list_head *iova)
2112 struct vfio_iova *n, *next;
2114 list_for_each_entry_safe(n, next, iova, list) {
2120 static int vfio_iommu_iova_get_copy(struct vfio_iommu *iommu,
2121 struct list_head *iova_copy)
2123 struct list_head *iova = &iommu->iova_list;
2124 struct vfio_iova *n;
2127 list_for_each_entry(n, iova, list) {
2128 ret = vfio_iommu_iova_insert(iova_copy, n->start, n->end);
2136 vfio_iommu_iova_free(iova_copy);
2140 static void vfio_iommu_iova_insert_copy(struct vfio_iommu *iommu,
2141 struct list_head *iova_copy)
2143 struct list_head *iova = &iommu->iova_list;
2145 vfio_iommu_iova_free(iova);
2147 list_splice_tail(iova_copy, iova);
2150 static int vfio_iommu_type1_attach_group(void *iommu_data,
2151 struct iommu_group *iommu_group)
2153 struct vfio_iommu *iommu = iommu_data;
2154 struct vfio_group *group;
2155 struct vfio_domain *domain, *d;
2156 struct bus_type *bus = NULL;
2158 bool resv_msi, msi_remap;
2159 phys_addr_t resv_msi_base = 0;
2160 struct iommu_domain_geometry geo;
2161 LIST_HEAD(iova_copy);
2162 LIST_HEAD(group_resv_regions);
2164 mutex_lock(&iommu->lock);
2166 /* Check for duplicates */
2167 if (vfio_iommu_find_iommu_group(iommu, iommu_group)) {
2168 mutex_unlock(&iommu->lock);
2172 group = kzalloc(sizeof(*group), GFP_KERNEL);
2173 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
2174 if (!group || !domain) {
2179 group->iommu_group = iommu_group;
2181 /* Determine bus_type in order to allocate a domain */
2182 ret = iommu_group_for_each_dev(iommu_group, &bus, vfio_bus_type);
2186 if (vfio_bus_is_mdev(bus)) {
2187 struct device *iommu_device = NULL;
2189 group->mdev_group = true;
2191 /* Determine the isolation type */
2192 ret = iommu_group_for_each_dev(iommu_group, &iommu_device,
2193 vfio_mdev_iommu_device);
2194 if (ret || !iommu_device) {
2195 if (!iommu->external_domain) {
2196 INIT_LIST_HEAD(&domain->group_list);
2197 iommu->external_domain = domain;
2198 vfio_update_pgsize_bitmap(iommu);
2203 list_add(&group->next,
2204 &iommu->external_domain->group_list);
2206 * Non-iommu backed group cannot dirty memory directly,
2207 * it can only use interfaces that provide dirty
2209 * The iommu scope can only be promoted with the
2210 * addition of a dirty tracking group.
2212 group->pinned_page_dirty_scope = true;
2213 mutex_unlock(&iommu->lock);
2218 bus = iommu_device->bus;
2221 domain->domain = iommu_domain_alloc(bus);
2222 if (!domain->domain) {
2227 if (iommu->nesting) {
2230 ret = iommu_domain_set_attr(domain->domain, DOMAIN_ATTR_NESTING,
2236 ret = vfio_iommu_attach_group(domain, group);
2240 /* Get aperture info */
2241 iommu_domain_get_attr(domain->domain, DOMAIN_ATTR_GEOMETRY, &geo);
2243 if (vfio_iommu_aper_conflict(iommu, geo.aperture_start,
2244 geo.aperture_end)) {
2249 ret = iommu_get_group_resv_regions(iommu_group, &group_resv_regions);
2253 if (vfio_iommu_resv_conflict(iommu, &group_resv_regions)) {
2259 * We don't want to work on the original iova list as the list
2260 * gets modified and in case of failure we have to retain the
2261 * original list. Get a copy here.
2263 ret = vfio_iommu_iova_get_copy(iommu, &iova_copy);
2267 ret = vfio_iommu_aper_resize(&iova_copy, geo.aperture_start,
2272 ret = vfio_iommu_resv_exclude(&iova_copy, &group_resv_regions);
2276 resv_msi = vfio_iommu_has_sw_msi(&group_resv_regions, &resv_msi_base);
2278 INIT_LIST_HEAD(&domain->group_list);
2279 list_add(&group->next, &domain->group_list);
2281 msi_remap = irq_domain_check_msi_remap() ||
2282 iommu_capable(bus, IOMMU_CAP_INTR_REMAP);
2284 if (!allow_unsafe_interrupts && !msi_remap) {
2285 pr_warn("%s: No interrupt remapping support. Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n",
2291 if (iommu_capable(bus, IOMMU_CAP_CACHE_COHERENCY))
2292 domain->prot |= IOMMU_CACHE;
2295 * Try to match an existing compatible domain. We don't want to
2296 * preclude an IOMMU driver supporting multiple bus_types and being
2297 * able to include different bus_types in the same IOMMU domain, so
2298 * we test whether the domains use the same iommu_ops rather than
2299 * testing if they're on the same bus_type.
2301 list_for_each_entry(d, &iommu->domain_list, next) {
2302 if (d->domain->ops == domain->domain->ops &&
2303 d->prot == domain->prot) {
2304 vfio_iommu_detach_group(domain, group);
2305 if (!vfio_iommu_attach_group(d, group)) {
2306 list_add(&group->next, &d->group_list);
2307 iommu_domain_free(domain->domain);
2312 ret = vfio_iommu_attach_group(domain, group);
2318 vfio_test_domain_fgsp(domain);
2320 /* replay mappings on new domains */
2321 ret = vfio_iommu_replay(iommu, domain);
2326 ret = iommu_get_msi_cookie(domain->domain, resv_msi_base);
2327 if (ret && ret != -ENODEV)
2331 list_add(&domain->next, &iommu->domain_list);
2332 vfio_update_pgsize_bitmap(iommu);
2334 /* Delete the old one and insert new iova list */
2335 vfio_iommu_iova_insert_copy(iommu, &iova_copy);
2338 * An iommu backed group can dirty memory directly and therefore
2339 * demotes the iommu scope until it declares itself dirty tracking
2340 * capable via the page pinning interface.
2342 iommu->num_non_pinned_groups++;
2343 mutex_unlock(&iommu->lock);
2344 vfio_iommu_resv_free(&group_resv_regions);
2349 vfio_iommu_detach_group(domain, group);
2351 iommu_domain_free(domain->domain);
2352 vfio_iommu_iova_free(&iova_copy);
2353 vfio_iommu_resv_free(&group_resv_regions);
2357 mutex_unlock(&iommu->lock);
2361 static void vfio_iommu_unmap_unpin_all(struct vfio_iommu *iommu)
2363 struct rb_node *node;
2365 while ((node = rb_first(&iommu->dma_list)))
2366 vfio_remove_dma(iommu, rb_entry(node, struct vfio_dma, node));
2369 static void vfio_iommu_unmap_unpin_reaccount(struct vfio_iommu *iommu)
2371 struct rb_node *n, *p;
2373 n = rb_first(&iommu->dma_list);
2374 for (; n; n = rb_next(n)) {
2375 struct vfio_dma *dma;
2376 long locked = 0, unlocked = 0;
2378 dma = rb_entry(n, struct vfio_dma, node);
2379 unlocked += vfio_unmap_unpin(iommu, dma, false);
2380 p = rb_first(&dma->pfn_list);
2381 for (; p; p = rb_next(p)) {
2382 struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn,
2385 if (!is_invalid_reserved_pfn(vpfn->pfn))
2388 vfio_lock_acct(dma, locked - unlocked, true);
2393 * Called when a domain is removed in detach. It is possible that
2394 * the removed domain decided the iova aperture window. Modify the
2395 * iova aperture with the smallest window among existing domains.
2397 static void vfio_iommu_aper_expand(struct vfio_iommu *iommu,
2398 struct list_head *iova_copy)
2400 struct vfio_domain *domain;
2401 struct iommu_domain_geometry geo;
2402 struct vfio_iova *node;
2403 dma_addr_t start = 0;
2404 dma_addr_t end = (dma_addr_t)~0;
2406 if (list_empty(iova_copy))
2409 list_for_each_entry(domain, &iommu->domain_list, next) {
2410 iommu_domain_get_attr(domain->domain, DOMAIN_ATTR_GEOMETRY,
2412 if (geo.aperture_start > start)
2413 start = geo.aperture_start;
2414 if (geo.aperture_end < end)
2415 end = geo.aperture_end;
2418 /* Modify aperture limits. The new aper is either same or bigger */
2419 node = list_first_entry(iova_copy, struct vfio_iova, list);
2420 node->start = start;
2421 node = list_last_entry(iova_copy, struct vfio_iova, list);
2426 * Called when a group is detached. The reserved regions for that
2427 * group can be part of valid iova now. But since reserved regions
2428 * may be duplicated among groups, populate the iova valid regions
2431 static int vfio_iommu_resv_refresh(struct vfio_iommu *iommu,
2432 struct list_head *iova_copy)
2434 struct vfio_domain *d;
2435 struct vfio_group *g;
2436 struct vfio_iova *node;
2437 dma_addr_t start, end;
2438 LIST_HEAD(resv_regions);
2441 if (list_empty(iova_copy))
2444 list_for_each_entry(d, &iommu->domain_list, next) {
2445 list_for_each_entry(g, &d->group_list, next) {
2446 ret = iommu_get_group_resv_regions(g->iommu_group,
2453 node = list_first_entry(iova_copy, struct vfio_iova, list);
2454 start = node->start;
2455 node = list_last_entry(iova_copy, struct vfio_iova, list);
2458 /* purge the iova list and create new one */
2459 vfio_iommu_iova_free(iova_copy);
2461 ret = vfio_iommu_aper_resize(iova_copy, start, end);
2465 /* Exclude current reserved regions from iova ranges */
2466 ret = vfio_iommu_resv_exclude(iova_copy, &resv_regions);
2468 vfio_iommu_resv_free(&resv_regions);
2472 static void vfio_iommu_type1_detach_group(void *iommu_data,
2473 struct iommu_group *iommu_group)
2475 struct vfio_iommu *iommu = iommu_data;
2476 struct vfio_domain *domain;
2477 struct vfio_group *group;
2478 bool update_dirty_scope = false;
2479 LIST_HEAD(iova_copy);
2481 mutex_lock(&iommu->lock);
2483 if (iommu->external_domain) {
2484 group = find_iommu_group(iommu->external_domain, iommu_group);
2486 update_dirty_scope = !group->pinned_page_dirty_scope;
2487 list_del(&group->next);
2490 if (list_empty(&iommu->external_domain->group_list)) {
2491 if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu)) {
2492 WARN_ON(iommu->notifier.head);
2493 vfio_iommu_unmap_unpin_all(iommu);
2496 kfree(iommu->external_domain);
2497 iommu->external_domain = NULL;
2499 goto detach_group_done;
2504 * Get a copy of iova list. This will be used to update
2505 * and to replace the current one later. Please note that
2506 * we will leave the original list as it is if update fails.
2508 vfio_iommu_iova_get_copy(iommu, &iova_copy);
2510 list_for_each_entry(domain, &iommu->domain_list, next) {
2511 group = find_iommu_group(domain, iommu_group);
2515 vfio_iommu_detach_group(domain, group);
2516 update_dirty_scope = !group->pinned_page_dirty_scope;
2517 list_del(&group->next);
2520 * Group ownership provides privilege, if the group list is
2521 * empty, the domain goes away. If it's the last domain with
2522 * iommu and external domain doesn't exist, then all the
2523 * mappings go away too. If it's the last domain with iommu and
2524 * external domain exist, update accounting
2526 if (list_empty(&domain->group_list)) {
2527 if (list_is_singular(&iommu->domain_list)) {
2528 if (!iommu->external_domain) {
2529 WARN_ON(iommu->notifier.head);
2530 vfio_iommu_unmap_unpin_all(iommu);
2532 vfio_iommu_unmap_unpin_reaccount(iommu);
2535 iommu_domain_free(domain->domain);
2536 list_del(&domain->next);
2538 vfio_iommu_aper_expand(iommu, &iova_copy);
2539 vfio_update_pgsize_bitmap(iommu);
2544 if (!vfio_iommu_resv_refresh(iommu, &iova_copy))
2545 vfio_iommu_iova_insert_copy(iommu, &iova_copy);
2547 vfio_iommu_iova_free(&iova_copy);
2551 * Removal of a group without dirty tracking may allow the iommu scope
2554 if (update_dirty_scope) {
2555 iommu->num_non_pinned_groups--;
2556 if (iommu->dirty_page_tracking)
2557 vfio_iommu_populate_bitmap_full(iommu);
2559 mutex_unlock(&iommu->lock);
2562 static void *vfio_iommu_type1_open(unsigned long arg)
2564 struct vfio_iommu *iommu;
2566 iommu = kzalloc(sizeof(*iommu), GFP_KERNEL);
2568 return ERR_PTR(-ENOMEM);
2571 case VFIO_TYPE1_IOMMU:
2573 case VFIO_TYPE1_NESTING_IOMMU:
2574 iommu->nesting = true;
2576 case VFIO_TYPE1v2_IOMMU:
2581 return ERR_PTR(-EINVAL);
2584 INIT_LIST_HEAD(&iommu->domain_list);
2585 INIT_LIST_HEAD(&iommu->iova_list);
2586 iommu->dma_list = RB_ROOT;
2587 iommu->dma_avail = dma_entry_limit;
2588 iommu->container_open = true;
2589 mutex_init(&iommu->lock);
2590 BLOCKING_INIT_NOTIFIER_HEAD(&iommu->notifier);
2591 init_waitqueue_head(&iommu->vaddr_wait);
2596 static void vfio_release_domain(struct vfio_domain *domain, bool external)
2598 struct vfio_group *group, *group_tmp;
2600 list_for_each_entry_safe(group, group_tmp,
2601 &domain->group_list, next) {
2603 vfio_iommu_detach_group(domain, group);
2604 list_del(&group->next);
2609 iommu_domain_free(domain->domain);
2612 static void vfio_iommu_type1_release(void *iommu_data)
2614 struct vfio_iommu *iommu = iommu_data;
2615 struct vfio_domain *domain, *domain_tmp;
2617 if (iommu->external_domain) {
2618 vfio_release_domain(iommu->external_domain, true);
2619 kfree(iommu->external_domain);
2622 vfio_iommu_unmap_unpin_all(iommu);
2624 list_for_each_entry_safe(domain, domain_tmp,
2625 &iommu->domain_list, next) {
2626 vfio_release_domain(domain, false);
2627 list_del(&domain->next);
2631 vfio_iommu_iova_free(&iommu->iova_list);
2636 static int vfio_domains_have_iommu_cache(struct vfio_iommu *iommu)
2638 struct vfio_domain *domain;
2641 mutex_lock(&iommu->lock);
2642 list_for_each_entry(domain, &iommu->domain_list, next) {
2643 if (!(domain->prot & IOMMU_CACHE)) {
2648 mutex_unlock(&iommu->lock);
2653 static int vfio_iommu_type1_check_extension(struct vfio_iommu *iommu,
2657 case VFIO_TYPE1_IOMMU:
2658 case VFIO_TYPE1v2_IOMMU:
2659 case VFIO_TYPE1_NESTING_IOMMU:
2660 case VFIO_UNMAP_ALL:
2661 case VFIO_UPDATE_VADDR:
2663 case VFIO_DMA_CC_IOMMU:
2666 return vfio_domains_have_iommu_cache(iommu);
2672 static int vfio_iommu_iova_add_cap(struct vfio_info_cap *caps,
2673 struct vfio_iommu_type1_info_cap_iova_range *cap_iovas,
2676 struct vfio_info_cap_header *header;
2677 struct vfio_iommu_type1_info_cap_iova_range *iova_cap;
2679 header = vfio_info_cap_add(caps, size,
2680 VFIO_IOMMU_TYPE1_INFO_CAP_IOVA_RANGE, 1);
2682 return PTR_ERR(header);
2684 iova_cap = container_of(header,
2685 struct vfio_iommu_type1_info_cap_iova_range,
2687 iova_cap->nr_iovas = cap_iovas->nr_iovas;
2688 memcpy(iova_cap->iova_ranges, cap_iovas->iova_ranges,
2689 cap_iovas->nr_iovas * sizeof(*cap_iovas->iova_ranges));
2693 static int vfio_iommu_iova_build_caps(struct vfio_iommu *iommu,
2694 struct vfio_info_cap *caps)
2696 struct vfio_iommu_type1_info_cap_iova_range *cap_iovas;
2697 struct vfio_iova *iova;
2699 int iovas = 0, i = 0, ret;
2701 list_for_each_entry(iova, &iommu->iova_list, list)
2706 * Return 0 as a container with a single mdev device
2707 * will have an empty list
2712 size = sizeof(*cap_iovas) + (iovas * sizeof(*cap_iovas->iova_ranges));
2714 cap_iovas = kzalloc(size, GFP_KERNEL);
2718 cap_iovas->nr_iovas = iovas;
2720 list_for_each_entry(iova, &iommu->iova_list, list) {
2721 cap_iovas->iova_ranges[i].start = iova->start;
2722 cap_iovas->iova_ranges[i].end = iova->end;
2726 ret = vfio_iommu_iova_add_cap(caps, cap_iovas, size);
2732 static int vfio_iommu_migration_build_caps(struct vfio_iommu *iommu,
2733 struct vfio_info_cap *caps)
2735 struct vfio_iommu_type1_info_cap_migration cap_mig;
2737 cap_mig.header.id = VFIO_IOMMU_TYPE1_INFO_CAP_MIGRATION;
2738 cap_mig.header.version = 1;
2741 /* support minimum pgsize */
2742 cap_mig.pgsize_bitmap = (size_t)1 << __ffs(iommu->pgsize_bitmap);
2743 cap_mig.max_dirty_bitmap_size = DIRTY_BITMAP_SIZE_MAX;
2745 return vfio_info_add_capability(caps, &cap_mig.header, sizeof(cap_mig));
2748 static int vfio_iommu_dma_avail_build_caps(struct vfio_iommu *iommu,
2749 struct vfio_info_cap *caps)
2751 struct vfio_iommu_type1_info_dma_avail cap_dma_avail;
2753 cap_dma_avail.header.id = VFIO_IOMMU_TYPE1_INFO_DMA_AVAIL;
2754 cap_dma_avail.header.version = 1;
2756 cap_dma_avail.avail = iommu->dma_avail;
2758 return vfio_info_add_capability(caps, &cap_dma_avail.header,
2759 sizeof(cap_dma_avail));
2762 static int vfio_iommu_type1_get_info(struct vfio_iommu *iommu,
2765 struct vfio_iommu_type1_info info;
2766 unsigned long minsz;
2767 struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
2768 unsigned long capsz;
2771 minsz = offsetofend(struct vfio_iommu_type1_info, iova_pgsizes);
2773 /* For backward compatibility, cannot require this */
2774 capsz = offsetofend(struct vfio_iommu_type1_info, cap_offset);
2776 if (copy_from_user(&info, (void __user *)arg, minsz))
2779 if (info.argsz < minsz)
2782 if (info.argsz >= capsz) {
2784 info.cap_offset = 0; /* output, no-recopy necessary */
2787 mutex_lock(&iommu->lock);
2788 info.flags = VFIO_IOMMU_INFO_PGSIZES;
2790 info.iova_pgsizes = iommu->pgsize_bitmap;
2792 ret = vfio_iommu_migration_build_caps(iommu, &caps);
2795 ret = vfio_iommu_dma_avail_build_caps(iommu, &caps);
2798 ret = vfio_iommu_iova_build_caps(iommu, &caps);
2800 mutex_unlock(&iommu->lock);
2806 info.flags |= VFIO_IOMMU_INFO_CAPS;
2808 if (info.argsz < sizeof(info) + caps.size) {
2809 info.argsz = sizeof(info) + caps.size;
2811 vfio_info_cap_shift(&caps, sizeof(info));
2812 if (copy_to_user((void __user *)arg +
2813 sizeof(info), caps.buf,
2818 info.cap_offset = sizeof(info);
2824 return copy_to_user((void __user *)arg, &info, minsz) ?
2828 static int vfio_iommu_type1_map_dma(struct vfio_iommu *iommu,
2831 struct vfio_iommu_type1_dma_map map;
2832 unsigned long minsz;
2833 uint32_t mask = VFIO_DMA_MAP_FLAG_READ | VFIO_DMA_MAP_FLAG_WRITE |
2834 VFIO_DMA_MAP_FLAG_VADDR;
2836 minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
2838 if (copy_from_user(&map, (void __user *)arg, minsz))
2841 if (map.argsz < minsz || map.flags & ~mask)
2844 return vfio_dma_do_map(iommu, &map);
2847 static int vfio_iommu_type1_unmap_dma(struct vfio_iommu *iommu,
2850 struct vfio_iommu_type1_dma_unmap unmap;
2851 struct vfio_bitmap bitmap = { 0 };
2852 uint32_t mask = VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP |
2853 VFIO_DMA_UNMAP_FLAG_VADDR |
2854 VFIO_DMA_UNMAP_FLAG_ALL;
2855 unsigned long minsz;
2858 minsz = offsetofend(struct vfio_iommu_type1_dma_unmap, size);
2860 if (copy_from_user(&unmap, (void __user *)arg, minsz))
2863 if (unmap.argsz < minsz || unmap.flags & ~mask)
2866 if ((unmap.flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) &&
2867 (unmap.flags & (VFIO_DMA_UNMAP_FLAG_ALL |
2868 VFIO_DMA_UNMAP_FLAG_VADDR)))
2871 if (unmap.flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) {
2872 unsigned long pgshift;
2874 if (unmap.argsz < (minsz + sizeof(bitmap)))
2877 if (copy_from_user(&bitmap,
2878 (void __user *)(arg + minsz),
2882 if (!access_ok((void __user *)bitmap.data, bitmap.size))
2885 pgshift = __ffs(bitmap.pgsize);
2886 ret = verify_bitmap_size(unmap.size >> pgshift,
2892 ret = vfio_dma_do_unmap(iommu, &unmap, &bitmap);
2896 return copy_to_user((void __user *)arg, &unmap, minsz) ?
2900 static int vfio_iommu_type1_dirty_pages(struct vfio_iommu *iommu,
2903 struct vfio_iommu_type1_dirty_bitmap dirty;
2904 uint32_t mask = VFIO_IOMMU_DIRTY_PAGES_FLAG_START |
2905 VFIO_IOMMU_DIRTY_PAGES_FLAG_STOP |
2906 VFIO_IOMMU_DIRTY_PAGES_FLAG_GET_BITMAP;
2907 unsigned long minsz;
2913 minsz = offsetofend(struct vfio_iommu_type1_dirty_bitmap, flags);
2915 if (copy_from_user(&dirty, (void __user *)arg, minsz))
2918 if (dirty.argsz < minsz || dirty.flags & ~mask)
2921 /* only one flag should be set at a time */
2922 if (__ffs(dirty.flags) != __fls(dirty.flags))
2925 if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_START) {
2928 mutex_lock(&iommu->lock);
2929 pgsize = 1 << __ffs(iommu->pgsize_bitmap);
2930 if (!iommu->dirty_page_tracking) {
2931 ret = vfio_dma_bitmap_alloc_all(iommu, pgsize);
2933 iommu->dirty_page_tracking = true;
2935 mutex_unlock(&iommu->lock);
2937 } else if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_STOP) {
2938 mutex_lock(&iommu->lock);
2939 if (iommu->dirty_page_tracking) {
2940 iommu->dirty_page_tracking = false;
2941 vfio_dma_bitmap_free_all(iommu);
2943 mutex_unlock(&iommu->lock);
2945 } else if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_GET_BITMAP) {
2946 struct vfio_iommu_type1_dirty_bitmap_get range;
2947 unsigned long pgshift;
2948 size_t data_size = dirty.argsz - minsz;
2949 size_t iommu_pgsize;
2951 if (!data_size || data_size < sizeof(range))
2954 if (copy_from_user(&range, (void __user *)(arg + minsz),
2958 if (range.iova + range.size < range.iova)
2960 if (!access_ok((void __user *)range.bitmap.data,
2964 pgshift = __ffs(range.bitmap.pgsize);
2965 ret = verify_bitmap_size(range.size >> pgshift,
2970 mutex_lock(&iommu->lock);
2972 iommu_pgsize = (size_t)1 << __ffs(iommu->pgsize_bitmap);
2974 /* allow only smallest supported pgsize */
2975 if (range.bitmap.pgsize != iommu_pgsize) {
2979 if (range.iova & (iommu_pgsize - 1)) {
2983 if (!range.size || range.size & (iommu_pgsize - 1)) {
2988 if (iommu->dirty_page_tracking)
2989 ret = vfio_iova_dirty_bitmap(range.bitmap.data,
2992 range.bitmap.pgsize);
2996 mutex_unlock(&iommu->lock);
3004 static long vfio_iommu_type1_ioctl(void *iommu_data,
3005 unsigned int cmd, unsigned long arg)
3007 struct vfio_iommu *iommu = iommu_data;
3010 case VFIO_CHECK_EXTENSION:
3011 return vfio_iommu_type1_check_extension(iommu, arg);
3012 case VFIO_IOMMU_GET_INFO:
3013 return vfio_iommu_type1_get_info(iommu, arg);
3014 case VFIO_IOMMU_MAP_DMA:
3015 return vfio_iommu_type1_map_dma(iommu, arg);
3016 case VFIO_IOMMU_UNMAP_DMA:
3017 return vfio_iommu_type1_unmap_dma(iommu, arg);
3018 case VFIO_IOMMU_DIRTY_PAGES:
3019 return vfio_iommu_type1_dirty_pages(iommu, arg);
3025 static int vfio_iommu_type1_register_notifier(void *iommu_data,
3026 unsigned long *events,
3027 struct notifier_block *nb)
3029 struct vfio_iommu *iommu = iommu_data;
3031 /* clear known events */
3032 *events &= ~VFIO_IOMMU_NOTIFY_DMA_UNMAP;
3034 /* refuse to register if still events remaining */
3038 return blocking_notifier_chain_register(&iommu->notifier, nb);
3041 static int vfio_iommu_type1_unregister_notifier(void *iommu_data,
3042 struct notifier_block *nb)
3044 struct vfio_iommu *iommu = iommu_data;
3046 return blocking_notifier_chain_unregister(&iommu->notifier, nb);
3049 static int vfio_iommu_type1_dma_rw_chunk(struct vfio_iommu *iommu,
3050 dma_addr_t user_iova, void *data,
3051 size_t count, bool write,
3054 struct mm_struct *mm;
3055 unsigned long vaddr;
3056 struct vfio_dma *dma;
3057 bool kthread = current->mm == NULL;
3063 ret = vfio_find_dma_valid(iommu, user_iova, 1, &dma);
3067 if ((write && !(dma->prot & IOMMU_WRITE)) ||
3068 !(dma->prot & IOMMU_READ))
3071 mm = get_task_mm(dma->task);
3078 else if (current->mm != mm)
3081 offset = user_iova - dma->iova;
3083 if (count > dma->size - offset)
3084 count = dma->size - offset;
3086 vaddr = dma->vaddr + offset;
3089 *copied = copy_to_user((void __user *)vaddr, data,
3091 if (*copied && iommu->dirty_page_tracking) {
3092 unsigned long pgshift = __ffs(iommu->pgsize_bitmap);
3094 * Bitmap populated with the smallest supported page
3097 bitmap_set(dma->bitmap, offset >> pgshift,
3098 ((offset + *copied - 1) >> pgshift) -
3099 (offset >> pgshift) + 1);
3102 *copied = copy_from_user(data, (void __user *)vaddr,
3105 kthread_unuse_mm(mm);
3108 return *copied ? 0 : -EFAULT;
3111 static int vfio_iommu_type1_dma_rw(void *iommu_data, dma_addr_t user_iova,
3112 void *data, size_t count, bool write)
3114 struct vfio_iommu *iommu = iommu_data;
3118 mutex_lock(&iommu->lock);
3120 ret = vfio_iommu_type1_dma_rw_chunk(iommu, user_iova, data,
3121 count, write, &done);
3130 mutex_unlock(&iommu->lock);
3134 static struct iommu_domain *
3135 vfio_iommu_type1_group_iommu_domain(void *iommu_data,
3136 struct iommu_group *iommu_group)
3138 struct iommu_domain *domain = ERR_PTR(-ENODEV);
3139 struct vfio_iommu *iommu = iommu_data;
3140 struct vfio_domain *d;
3142 if (!iommu || !iommu_group)
3143 return ERR_PTR(-EINVAL);
3145 mutex_lock(&iommu->lock);
3146 list_for_each_entry(d, &iommu->domain_list, next) {
3147 if (find_iommu_group(d, iommu_group)) {
3152 mutex_unlock(&iommu->lock);
3157 static void vfio_iommu_type1_notify(void *iommu_data,
3158 enum vfio_iommu_notify_type event)
3160 struct vfio_iommu *iommu = iommu_data;
3162 if (event != VFIO_IOMMU_CONTAINER_CLOSE)
3164 mutex_lock(&iommu->lock);
3165 iommu->container_open = false;
3166 mutex_unlock(&iommu->lock);
3167 wake_up_all(&iommu->vaddr_wait);
3170 static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = {
3171 .name = "vfio-iommu-type1",
3172 .owner = THIS_MODULE,
3173 .open = vfio_iommu_type1_open,
3174 .release = vfio_iommu_type1_release,
3175 .ioctl = vfio_iommu_type1_ioctl,
3176 .attach_group = vfio_iommu_type1_attach_group,
3177 .detach_group = vfio_iommu_type1_detach_group,
3178 .pin_pages = vfio_iommu_type1_pin_pages,
3179 .unpin_pages = vfio_iommu_type1_unpin_pages,
3180 .register_notifier = vfio_iommu_type1_register_notifier,
3181 .unregister_notifier = vfio_iommu_type1_unregister_notifier,
3182 .dma_rw = vfio_iommu_type1_dma_rw,
3183 .group_iommu_domain = vfio_iommu_type1_group_iommu_domain,
3184 .notify = vfio_iommu_type1_notify,
3187 static int __init vfio_iommu_type1_init(void)
3189 return vfio_register_iommu_driver(&vfio_iommu_driver_ops_type1);
3192 static void __exit vfio_iommu_type1_cleanup(void)
3194 vfio_unregister_iommu_driver(&vfio_iommu_driver_ops_type1);
3197 module_init(vfio_iommu_type1_init);
3198 module_exit(vfio_iommu_type1_cleanup);
3200 MODULE_VERSION(DRIVER_VERSION);
3201 MODULE_LICENSE("GPL v2");
3202 MODULE_AUTHOR(DRIVER_AUTHOR);
3203 MODULE_DESCRIPTION(DRIVER_DESC);